1 //===- InstCombineCalls.cpp -----------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the visitCall and visitInvoke functions.
12 //===----------------------------------------------------------------------===//
14 #include "InstCombine.h"
15 #include "llvm/ADT/Statistic.h"
16 #include "llvm/Analysis/MemoryBuiltins.h"
17 #include "llvm/IR/CallSite.h"
18 #include "llvm/IR/DataLayout.h"
19 #include "llvm/IR/PatternMatch.h"
20 #include "llvm/Transforms/Utils/BuildLibCalls.h"
21 #include "llvm/Transforms/Utils/Local.h"
23 using namespace PatternMatch;
25 #define DEBUG_TYPE "instcombine"
27 STATISTIC(NumSimplified, "Number of library calls simplified");
29 /// getPromotedType - Return the specified type promoted as it would be to pass
30 /// though a va_arg area.
31 static Type *getPromotedType(Type *Ty) {
32 if (IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
33 if (ITy->getBitWidth() < 32)
34 return Type::getInt32Ty(Ty->getContext());
39 /// reduceToSingleValueType - Given an aggregate type which ultimately holds a
40 /// single scalar element, like {{{type}}} or [1 x type], return type.
41 static Type *reduceToSingleValueType(Type *T) {
42 while (!T->isSingleValueType()) {
43 if (StructType *STy = dyn_cast<StructType>(T)) {
44 if (STy->getNumElements() == 1)
45 T = STy->getElementType(0);
48 } else if (ArrayType *ATy = dyn_cast<ArrayType>(T)) {
49 if (ATy->getNumElements() == 1)
50 T = ATy->getElementType();
60 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
61 unsigned DstAlign = getKnownAlignment(MI->getArgOperand(0), DL);
62 unsigned SrcAlign = getKnownAlignment(MI->getArgOperand(1), DL);
63 unsigned MinAlign = std::min(DstAlign, SrcAlign);
64 unsigned CopyAlign = MI->getAlignment();
66 if (CopyAlign < MinAlign) {
67 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
72 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
74 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getArgOperand(2));
75 if (!MemOpLength) return nullptr;
77 // Source and destination pointer types are always "i8*" for intrinsic. See
78 // if the size is something we can handle with a single primitive load/store.
79 // A single load+store correctly handles overlapping memory in the memmove
81 uint64_t Size = MemOpLength->getLimitedValue();
82 assert(Size && "0-sized memory transferring should be removed already.");
84 if (Size > 8 || (Size&(Size-1)))
85 return nullptr; // If not 1/2/4/8 bytes, exit.
87 // Use an integer load+store unless we can find something better.
89 cast<PointerType>(MI->getArgOperand(1)->getType())->getAddressSpace();
91 cast<PointerType>(MI->getArgOperand(0)->getType())->getAddressSpace();
93 IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3);
94 Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp);
95 Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp);
97 // Memcpy forces the use of i8* for the source and destination. That means
98 // that if you're using memcpy to move one double around, you'll get a cast
99 // from double* to i8*. We'd much rather use a double load+store rather than
100 // an i64 load+store, here because this improves the odds that the source or
101 // dest address will be promotable. See if we can find a better type than the
103 Value *StrippedDest = MI->getArgOperand(0)->stripPointerCasts();
104 MDNode *CopyMD = nullptr;
105 if (StrippedDest != MI->getArgOperand(0)) {
106 Type *SrcETy = cast<PointerType>(StrippedDest->getType())
108 if (DL && SrcETy->isSized() && DL->getTypeStoreSize(SrcETy) == Size) {
109 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
110 // down through these levels if so.
111 SrcETy = reduceToSingleValueType(SrcETy);
113 if (SrcETy->isSingleValueType()) {
114 NewSrcPtrTy = PointerType::get(SrcETy, SrcAddrSp);
115 NewDstPtrTy = PointerType::get(SrcETy, DstAddrSp);
117 // If the memcpy has metadata describing the members, see if we can
118 // get the TBAA tag describing our copy.
119 if (MDNode *M = MI->getMetadata(LLVMContext::MD_tbaa_struct)) {
120 if (M->getNumOperands() == 3 &&
122 isa<ConstantInt>(M->getOperand(0)) &&
123 cast<ConstantInt>(M->getOperand(0))->isNullValue() &&
125 isa<ConstantInt>(M->getOperand(1)) &&
126 cast<ConstantInt>(M->getOperand(1))->getValue() == Size &&
128 isa<MDNode>(M->getOperand(2)))
129 CopyMD = cast<MDNode>(M->getOperand(2));
135 // If the memcpy/memmove provides better alignment info than we can
137 SrcAlign = std::max(SrcAlign, CopyAlign);
138 DstAlign = std::max(DstAlign, CopyAlign);
140 Value *Src = Builder->CreateBitCast(MI->getArgOperand(1), NewSrcPtrTy);
141 Value *Dest = Builder->CreateBitCast(MI->getArgOperand(0), NewDstPtrTy);
142 LoadInst *L = Builder->CreateLoad(Src, MI->isVolatile());
143 L->setAlignment(SrcAlign);
145 L->setMetadata(LLVMContext::MD_tbaa, CopyMD);
146 StoreInst *S = Builder->CreateStore(L, Dest, MI->isVolatile());
147 S->setAlignment(DstAlign);
149 S->setMetadata(LLVMContext::MD_tbaa, CopyMD);
151 // Set the size of the copy to 0, it will be deleted on the next iteration.
152 MI->setArgOperand(2, Constant::getNullValue(MemOpLength->getType()));
156 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
157 unsigned Alignment = getKnownAlignment(MI->getDest(), DL);
158 if (MI->getAlignment() < Alignment) {
159 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
164 // Extract the length and alignment and fill if they are constant.
165 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
166 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
167 if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8))
169 uint64_t Len = LenC->getLimitedValue();
170 Alignment = MI->getAlignment();
171 assert(Len && "0-sized memory setting should be removed already.");
173 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
174 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
175 Type *ITy = IntegerType::get(MI->getContext(), Len*8); // n=1 -> i8.
177 Value *Dest = MI->getDest();
178 unsigned DstAddrSp = cast<PointerType>(Dest->getType())->getAddressSpace();
179 Type *NewDstPtrTy = PointerType::get(ITy, DstAddrSp);
180 Dest = Builder->CreateBitCast(Dest, NewDstPtrTy);
182 // Alignment 0 is identity for alignment 1 for memset, but not store.
183 if (Alignment == 0) Alignment = 1;
185 // Extract the fill value and store.
186 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
187 StoreInst *S = Builder->CreateStore(ConstantInt::get(ITy, Fill), Dest,
189 S->setAlignment(Alignment);
191 // Set the size of the copy to 0, it will be deleted on the next iteration.
192 MI->setLength(Constant::getNullValue(LenC->getType()));
199 /// visitCallInst - CallInst simplification. This mostly only handles folding
200 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
201 /// the heavy lifting.
203 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
204 if (isFreeCall(&CI, TLI))
205 return visitFree(CI);
207 // If the caller function is nounwind, mark the call as nounwind, even if the
209 if (CI.getParent()->getParent()->doesNotThrow() &&
210 !CI.doesNotThrow()) {
211 CI.setDoesNotThrow();
215 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
216 if (!II) return visitCallSite(&CI);
218 // Intrinsics cannot occur in an invoke, so handle them here instead of in
220 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
221 bool Changed = false;
223 // memmove/cpy/set of zero bytes is a noop.
224 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
225 if (NumBytes->isNullValue())
226 return EraseInstFromFunction(CI);
228 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
229 if (CI->getZExtValue() == 1) {
230 // Replace the instruction with just byte operations. We would
231 // transform other cases to loads/stores, but we don't know if
232 // alignment is sufficient.
236 // No other transformations apply to volatile transfers.
237 if (MI->isVolatile())
240 // If we have a memmove and the source operation is a constant global,
241 // then the source and dest pointers can't alias, so we can change this
242 // into a call to memcpy.
243 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
244 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
245 if (GVSrc->isConstant()) {
246 Module *M = CI.getParent()->getParent()->getParent();
247 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
248 Type *Tys[3] = { CI.getArgOperand(0)->getType(),
249 CI.getArgOperand(1)->getType(),
250 CI.getArgOperand(2)->getType() };
251 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys));
256 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
257 // memmove(x,x,size) -> noop.
258 if (MTI->getSource() == MTI->getDest())
259 return EraseInstFromFunction(CI);
262 // If we can determine a pointer alignment that is bigger than currently
263 // set, update the alignment.
264 if (isa<MemTransferInst>(MI)) {
265 if (Instruction *I = SimplifyMemTransfer(MI))
267 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
268 if (Instruction *I = SimplifyMemSet(MSI))
272 if (Changed) return II;
275 switch (II->getIntrinsicID()) {
277 case Intrinsic::objectsize: {
279 if (getObjectSize(II->getArgOperand(0), Size, DL, TLI))
280 return ReplaceInstUsesWith(CI, ConstantInt::get(CI.getType(), Size));
283 case Intrinsic::bswap: {
284 Value *IIOperand = II->getArgOperand(0);
287 // bswap(bswap(x)) -> x
288 if (match(IIOperand, m_BSwap(m_Value(X))))
289 return ReplaceInstUsesWith(CI, X);
291 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
292 if (match(IIOperand, m_Trunc(m_BSwap(m_Value(X))))) {
293 unsigned C = X->getType()->getPrimitiveSizeInBits() -
294 IIOperand->getType()->getPrimitiveSizeInBits();
295 Value *CV = ConstantInt::get(X->getType(), C);
296 Value *V = Builder->CreateLShr(X, CV);
297 return new TruncInst(V, IIOperand->getType());
302 case Intrinsic::powi:
303 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
306 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
309 return ReplaceInstUsesWith(CI, II->getArgOperand(0));
310 // powi(x, -1) -> 1/x
311 if (Power->isAllOnesValue())
312 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
313 II->getArgOperand(0));
316 case Intrinsic::cttz: {
317 // If all bits below the first known one are known zero,
318 // this value is constant.
319 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
320 // FIXME: Try to simplify vectors of integers.
322 uint32_t BitWidth = IT->getBitWidth();
323 APInt KnownZero(BitWidth, 0);
324 APInt KnownOne(BitWidth, 0);
325 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne);
326 unsigned TrailingZeros = KnownOne.countTrailingZeros();
327 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros));
328 if ((Mask & KnownZero) == Mask)
329 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
330 APInt(BitWidth, TrailingZeros)));
334 case Intrinsic::ctlz: {
335 // If all bits above the first known one are known zero,
336 // this value is constant.
337 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
338 // FIXME: Try to simplify vectors of integers.
340 uint32_t BitWidth = IT->getBitWidth();
341 APInt KnownZero(BitWidth, 0);
342 APInt KnownOne(BitWidth, 0);
343 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne);
344 unsigned LeadingZeros = KnownOne.countLeadingZeros();
345 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
346 if ((Mask & KnownZero) == Mask)
347 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
348 APInt(BitWidth, LeadingZeros)));
352 case Intrinsic::uadd_with_overflow: {
353 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
354 IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
355 uint32_t BitWidth = IT->getBitWidth();
356 APInt LHSKnownZero(BitWidth, 0);
357 APInt LHSKnownOne(BitWidth, 0);
358 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne);
359 bool LHSKnownNegative = LHSKnownOne[BitWidth - 1];
360 bool LHSKnownPositive = LHSKnownZero[BitWidth - 1];
362 if (LHSKnownNegative || LHSKnownPositive) {
363 APInt RHSKnownZero(BitWidth, 0);
364 APInt RHSKnownOne(BitWidth, 0);
365 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne);
366 bool RHSKnownNegative = RHSKnownOne[BitWidth - 1];
367 bool RHSKnownPositive = RHSKnownZero[BitWidth - 1];
368 if (LHSKnownNegative && RHSKnownNegative) {
369 // The sign bit is set in both cases: this MUST overflow.
370 // Create a simple add instruction, and insert it into the struct.
371 Value *Add = Builder->CreateAdd(LHS, RHS);
374 UndefValue::get(LHS->getType()),
375 ConstantInt::getTrue(II->getContext())
377 StructType *ST = cast<StructType>(II->getType());
378 Constant *Struct = ConstantStruct::get(ST, V);
379 return InsertValueInst::Create(Struct, Add, 0);
382 if (LHSKnownPositive && RHSKnownPositive) {
383 // The sign bit is clear in both cases: this CANNOT overflow.
384 // Create a simple add instruction, and insert it into the struct.
385 Value *Add = Builder->CreateNUWAdd(LHS, RHS);
388 UndefValue::get(LHS->getType()),
389 ConstantInt::getFalse(II->getContext())
391 StructType *ST = cast<StructType>(II->getType());
392 Constant *Struct = ConstantStruct::get(ST, V);
393 return InsertValueInst::Create(Struct, Add, 0);
397 // FALL THROUGH uadd into sadd
398 case Intrinsic::sadd_with_overflow:
399 // Canonicalize constants into the RHS.
400 if (isa<Constant>(II->getArgOperand(0)) &&
401 !isa<Constant>(II->getArgOperand(1))) {
402 Value *LHS = II->getArgOperand(0);
403 II->setArgOperand(0, II->getArgOperand(1));
404 II->setArgOperand(1, LHS);
408 // X + undef -> undef
409 if (isa<UndefValue>(II->getArgOperand(1)))
410 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
412 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
413 // X + 0 -> {X, false}
416 UndefValue::get(II->getArgOperand(0)->getType()),
417 ConstantInt::getFalse(II->getContext())
420 ConstantStruct::get(cast<StructType>(II->getType()), V);
421 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
425 case Intrinsic::usub_with_overflow:
426 case Intrinsic::ssub_with_overflow:
427 // undef - X -> undef
428 // X - undef -> undef
429 if (isa<UndefValue>(II->getArgOperand(0)) ||
430 isa<UndefValue>(II->getArgOperand(1)))
431 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
433 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
434 // X - 0 -> {X, false}
437 UndefValue::get(II->getArgOperand(0)->getType()),
438 ConstantInt::getFalse(II->getContext())
441 ConstantStruct::get(cast<StructType>(II->getType()), V);
442 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
446 case Intrinsic::umul_with_overflow: {
447 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
448 unsigned BitWidth = cast<IntegerType>(LHS->getType())->getBitWidth();
450 APInt LHSKnownZero(BitWidth, 0);
451 APInt LHSKnownOne(BitWidth, 0);
452 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne);
453 APInt RHSKnownZero(BitWidth, 0);
454 APInt RHSKnownOne(BitWidth, 0);
455 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne);
457 // Get the largest possible values for each operand.
458 APInt LHSMax = ~LHSKnownZero;
459 APInt RHSMax = ~RHSKnownZero;
461 // If multiplying the maximum values does not overflow then we can turn
462 // this into a plain NUW mul.
464 LHSMax.umul_ov(RHSMax, Overflow);
466 Value *Mul = Builder->CreateNUWMul(LHS, RHS, "umul_with_overflow");
468 UndefValue::get(LHS->getType()),
471 Constant *Struct = ConstantStruct::get(cast<StructType>(II->getType()),V);
472 return InsertValueInst::Create(Struct, Mul, 0);
475 case Intrinsic::smul_with_overflow:
476 // Canonicalize constants into the RHS.
477 if (isa<Constant>(II->getArgOperand(0)) &&
478 !isa<Constant>(II->getArgOperand(1))) {
479 Value *LHS = II->getArgOperand(0);
480 II->setArgOperand(0, II->getArgOperand(1));
481 II->setArgOperand(1, LHS);
485 // X * undef -> undef
486 if (isa<UndefValue>(II->getArgOperand(1)))
487 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
489 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
492 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType()));
494 // X * 1 -> {X, false}
495 if (RHSI->equalsInt(1)) {
497 UndefValue::get(II->getArgOperand(0)->getType()),
498 ConstantInt::getFalse(II->getContext())
501 ConstantStruct::get(cast<StructType>(II->getType()), V);
502 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
506 case Intrinsic::ppc_altivec_lvx:
507 case Intrinsic::ppc_altivec_lvxl:
508 // Turn PPC lvx -> load if the pointer is known aligned.
509 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL) >= 16) {
510 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
511 PointerType::getUnqual(II->getType()));
512 return new LoadInst(Ptr);
515 case Intrinsic::ppc_altivec_stvx:
516 case Intrinsic::ppc_altivec_stvxl:
517 // Turn stvx -> store if the pointer is known aligned.
518 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL) >= 16) {
520 PointerType::getUnqual(II->getArgOperand(0)->getType());
521 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
522 return new StoreInst(II->getArgOperand(0), Ptr);
525 case Intrinsic::x86_sse_storeu_ps:
526 case Intrinsic::x86_sse2_storeu_pd:
527 case Intrinsic::x86_sse2_storeu_dq:
528 // Turn X86 storeu -> store if the pointer is known aligned.
529 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL) >= 16) {
531 PointerType::getUnqual(II->getArgOperand(1)->getType());
532 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
533 return new StoreInst(II->getArgOperand(1), Ptr);
537 case Intrinsic::x86_sse_cvtss2si:
538 case Intrinsic::x86_sse_cvtss2si64:
539 case Intrinsic::x86_sse_cvttss2si:
540 case Intrinsic::x86_sse_cvttss2si64:
541 case Intrinsic::x86_sse2_cvtsd2si:
542 case Intrinsic::x86_sse2_cvtsd2si64:
543 case Intrinsic::x86_sse2_cvttsd2si:
544 case Intrinsic::x86_sse2_cvttsd2si64: {
545 // These intrinsics only demand the 0th element of their input vectors. If
546 // we can simplify the input based on that, do so now.
548 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
549 APInt DemandedElts(VWidth, 1);
550 APInt UndefElts(VWidth, 0);
551 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
552 DemandedElts, UndefElts)) {
553 II->setArgOperand(0, V);
559 // Constant fold <A x Bi> << Ci.
560 // FIXME: We don't handle _dq because it's a shift of an i128, but is
561 // represented in the IR as <2 x i64>. A per element shift is wrong.
562 case Intrinsic::x86_sse2_psll_d:
563 case Intrinsic::x86_sse2_psll_q:
564 case Intrinsic::x86_sse2_psll_w:
565 case Intrinsic::x86_sse2_pslli_d:
566 case Intrinsic::x86_sse2_pslli_q:
567 case Intrinsic::x86_sse2_pslli_w:
568 case Intrinsic::x86_avx2_psll_d:
569 case Intrinsic::x86_avx2_psll_q:
570 case Intrinsic::x86_avx2_psll_w:
571 case Intrinsic::x86_avx2_pslli_d:
572 case Intrinsic::x86_avx2_pslli_q:
573 case Intrinsic::x86_avx2_pslli_w:
574 case Intrinsic::x86_sse2_psrl_d:
575 case Intrinsic::x86_sse2_psrl_q:
576 case Intrinsic::x86_sse2_psrl_w:
577 case Intrinsic::x86_sse2_psrli_d:
578 case Intrinsic::x86_sse2_psrli_q:
579 case Intrinsic::x86_sse2_psrli_w:
580 case Intrinsic::x86_avx2_psrl_d:
581 case Intrinsic::x86_avx2_psrl_q:
582 case Intrinsic::x86_avx2_psrl_w:
583 case Intrinsic::x86_avx2_psrli_d:
584 case Intrinsic::x86_avx2_psrli_q:
585 case Intrinsic::x86_avx2_psrli_w: {
586 // Simplify if count is constant. To 0 if >= BitWidth,
587 // otherwise to shl/lshr.
588 auto CDV = dyn_cast<ConstantDataVector>(II->getArgOperand(1));
589 auto CInt = dyn_cast<ConstantInt>(II->getArgOperand(1));
594 Count = cast<ConstantInt>(CDV->getElementAsConstant(0));
598 auto Vec = II->getArgOperand(0);
599 auto VT = cast<VectorType>(Vec->getType());
600 if (Count->getZExtValue() >
601 VT->getElementType()->getPrimitiveSizeInBits() - 1)
602 return ReplaceInstUsesWith(
603 CI, ConstantAggregateZero::get(Vec->getType()));
605 bool isPackedShiftLeft = true;
606 switch (II->getIntrinsicID()) {
608 case Intrinsic::x86_sse2_psrl_d:
609 case Intrinsic::x86_sse2_psrl_q:
610 case Intrinsic::x86_sse2_psrl_w:
611 case Intrinsic::x86_sse2_psrli_d:
612 case Intrinsic::x86_sse2_psrli_q:
613 case Intrinsic::x86_sse2_psrli_w:
614 case Intrinsic::x86_avx2_psrl_d:
615 case Intrinsic::x86_avx2_psrl_q:
616 case Intrinsic::x86_avx2_psrl_w:
617 case Intrinsic::x86_avx2_psrli_d:
618 case Intrinsic::x86_avx2_psrli_q:
619 case Intrinsic::x86_avx2_psrli_w: isPackedShiftLeft = false; break;
622 unsigned VWidth = VT->getNumElements();
623 // Get a constant vector of the same type as the first operand.
624 auto VTCI = ConstantInt::get(VT->getElementType(), Count->getZExtValue());
625 if (isPackedShiftLeft)
626 return BinaryOperator::CreateShl(Vec,
627 Builder->CreateVectorSplat(VWidth, VTCI));
629 return BinaryOperator::CreateLShr(Vec,
630 Builder->CreateVectorSplat(VWidth, VTCI));
633 case Intrinsic::x86_sse41_pmovsxbw:
634 case Intrinsic::x86_sse41_pmovsxwd:
635 case Intrinsic::x86_sse41_pmovsxdq:
636 case Intrinsic::x86_sse41_pmovzxbw:
637 case Intrinsic::x86_sse41_pmovzxwd:
638 case Intrinsic::x86_sse41_pmovzxdq: {
639 // pmov{s|z}x ignores the upper half of their input vectors.
641 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
642 unsigned LowHalfElts = VWidth / 2;
643 APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts));
644 APInt UndefElts(VWidth, 0);
645 if (Value *TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0),
648 II->setArgOperand(0, TmpV);
654 case Intrinsic::x86_sse4a_insertqi: {
655 // insertqi x, y, 64, 0 can just copy y's lower bits and leave the top
657 // TODO: eventually we should lower this intrinsic to IR
658 if (auto CIWidth = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
659 if (auto CIStart = dyn_cast<ConstantInt>(II->getArgOperand(3))) {
660 if (CIWidth->equalsInt(64) && CIStart->isZero()) {
661 Value *Vec = II->getArgOperand(1);
662 Value *Undef = UndefValue::get(Vec->getType());
663 const uint32_t Mask[] = { 0, 2 };
664 return ReplaceInstUsesWith(
666 Builder->CreateShuffleVector(
667 Vec, Undef, ConstantDataVector::get(
668 II->getContext(), ArrayRef<uint32_t>(Mask))));
670 } else if (auto Source =
671 dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
672 if (Source->hasOneUse() &&
673 Source->getArgOperand(1) == II->getArgOperand(1)) {
674 // If the source of the insert has only one use and it's another
675 // insert (and they're both inserting from the same vector), try to
676 // bundle both together.
678 dyn_cast<ConstantInt>(Source->getArgOperand(2));
680 dyn_cast<ConstantInt>(Source->getArgOperand(3));
681 if (CISourceStart && CISourceWidth) {
682 unsigned Start = CIStart->getZExtValue();
683 unsigned Width = CIWidth->getZExtValue();
684 unsigned End = Start + Width;
685 unsigned SourceStart = CISourceStart->getZExtValue();
686 unsigned SourceWidth = CISourceWidth->getZExtValue();
687 unsigned SourceEnd = SourceStart + SourceWidth;
688 unsigned NewStart, NewWidth;
689 bool ShouldReplace = false;
690 if (Start <= SourceStart && SourceStart <= End) {
692 NewWidth = std::max(End, SourceEnd) - NewStart;
693 ShouldReplace = true;
694 } else if (SourceStart <= Start && Start <= SourceEnd) {
695 NewStart = SourceStart;
696 NewWidth = std::max(SourceEnd, End) - NewStart;
697 ShouldReplace = true;
701 Constant *ConstantWidth = ConstantInt::get(
702 II->getArgOperand(2)->getType(), NewWidth, false);
703 Constant *ConstantStart = ConstantInt::get(
704 II->getArgOperand(3)->getType(), NewStart, false);
705 Value *Args[4] = { Source->getArgOperand(0),
706 II->getArgOperand(1), ConstantWidth,
708 Module *M = CI.getParent()->getParent()->getParent();
710 Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi);
711 return ReplaceInstUsesWith(CI, Builder->CreateCall(F, Args));
721 case Intrinsic::x86_sse41_pblendvb:
722 case Intrinsic::x86_sse41_blendvps:
723 case Intrinsic::x86_sse41_blendvpd:
724 case Intrinsic::x86_avx_blendv_ps_256:
725 case Intrinsic::x86_avx_blendv_pd_256:
726 case Intrinsic::x86_avx2_pblendvb: {
727 // Convert blendv* to vector selects if the mask is constant.
728 // This optimization is convoluted because the intrinsic is defined as
729 // getting a vector of floats or doubles for the ps and pd versions.
730 // FIXME: That should be changed.
731 Value *Mask = II->getArgOperand(2);
732 if (auto C = dyn_cast<ConstantDataVector>(Mask)) {
733 auto Tyi1 = Builder->getInt1Ty();
734 auto SelectorType = cast<VectorType>(Mask->getType());
735 auto EltTy = SelectorType->getElementType();
736 unsigned Size = SelectorType->getNumElements();
737 unsigned BitWidth = EltTy->isFloatTy() ? 32 : (EltTy->isDoubleTy() ? 64 : EltTy->getIntegerBitWidth());
738 assert(BitWidth == 64 || BitWidth == 32 || BitWidth == 8 && "Wrong arguments for variable blend intrinsic");
739 SmallVector<Constant*, 32> Selectors;
740 for (unsigned I = 0; I < Size; ++I) {
741 // The intrinsics only read the top bit
744 Selector = C->getElementAsInteger(I);
746 Selector = C->getElementAsAPFloat(I).bitcastToAPInt().getZExtValue();
747 Selectors.push_back(ConstantInt::get(Tyi1, Selector >> (BitWidth - 1)));
749 auto NewSelector = ConstantVector::get(Selectors);
750 return SelectInst::Create(NewSelector, II->getArgOperand(0), II->getArgOperand(1), "blendv");
756 case Intrinsic::x86_avx_vpermilvar_ps:
757 case Intrinsic::x86_avx_vpermilvar_ps_256:
758 case Intrinsic::x86_avx_vpermilvar_pd:
759 case Intrinsic::x86_avx_vpermilvar_pd_256: {
760 // Convert vpermil* to shufflevector if the mask is constant.
761 Value *V = II->getArgOperand(1);
762 unsigned Size = cast<VectorType>(V->getType())->getNumElements();
763 assert(Size == 8 || Size == 4 || Size == 2);
765 if (auto C = dyn_cast<ConstantDataVector>(V)) {
766 // The intrinsics only read one or two bits, clear the rest.
767 for (unsigned I = 0; I < Size; ++I) {
768 uint32_t Index = C->getElementAsInteger(I) & 0x3;
769 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd ||
770 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256)
774 } else if (isa<ConstantAggregateZero>(V)) {
775 for (unsigned I = 0; I < Size; ++I)
780 // The _256 variants are a bit trickier since the mask bits always index
781 // into the corresponding 128 half. In order to convert to a generic
782 // shuffle, we have to make that explicit.
783 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_ps_256 ||
784 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256) {
785 for (unsigned I = Size / 2; I < Size; ++I)
786 Indexes[I] += Size / 2;
789 ConstantDataVector::get(V->getContext(), makeArrayRef(Indexes, Size));
790 auto V1 = II->getArgOperand(0);
791 auto V2 = UndefValue::get(V1->getType());
792 auto Shuffle = Builder->CreateShuffleVector(V1, V2, NewC);
793 return ReplaceInstUsesWith(CI, Shuffle);
796 case Intrinsic::ppc_altivec_vperm:
797 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
798 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) {
799 assert(Mask->getType()->getVectorNumElements() == 16 &&
800 "Bad type for intrinsic!");
802 // Check that all of the elements are integer constants or undefs.
803 bool AllEltsOk = true;
804 for (unsigned i = 0; i != 16; ++i) {
805 Constant *Elt = Mask->getAggregateElement(i);
806 if (!Elt || !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) {
813 // Cast the input vectors to byte vectors.
814 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
816 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
818 Value *Result = UndefValue::get(Op0->getType());
820 // Only extract each element once.
821 Value *ExtractedElts[32];
822 memset(ExtractedElts, 0, sizeof(ExtractedElts));
824 for (unsigned i = 0; i != 16; ++i) {
825 if (isa<UndefValue>(Mask->getAggregateElement(i)))
828 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue();
829 Idx &= 31; // Match the hardware behavior.
831 if (!ExtractedElts[Idx]) {
833 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1,
834 Builder->getInt32(Idx&15));
837 // Insert this value into the result vector.
838 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
839 Builder->getInt32(i));
841 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
846 case Intrinsic::arm_neon_vld1:
847 case Intrinsic::arm_neon_vld2:
848 case Intrinsic::arm_neon_vld3:
849 case Intrinsic::arm_neon_vld4:
850 case Intrinsic::arm_neon_vld2lane:
851 case Intrinsic::arm_neon_vld3lane:
852 case Intrinsic::arm_neon_vld4lane:
853 case Intrinsic::arm_neon_vst1:
854 case Intrinsic::arm_neon_vst2:
855 case Intrinsic::arm_neon_vst3:
856 case Intrinsic::arm_neon_vst4:
857 case Intrinsic::arm_neon_vst2lane:
858 case Intrinsic::arm_neon_vst3lane:
859 case Intrinsic::arm_neon_vst4lane: {
860 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), DL);
861 unsigned AlignArg = II->getNumArgOperands() - 1;
862 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
863 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
864 II->setArgOperand(AlignArg,
865 ConstantInt::get(Type::getInt32Ty(II->getContext()),
872 case Intrinsic::arm_neon_vmulls:
873 case Intrinsic::arm_neon_vmullu:
874 case Intrinsic::aarch64_neon_smull:
875 case Intrinsic::aarch64_neon_umull: {
876 Value *Arg0 = II->getArgOperand(0);
877 Value *Arg1 = II->getArgOperand(1);
879 // Handle mul by zero first:
880 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) {
881 return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType()));
884 // Check for constant LHS & RHS - in this case we just simplify.
885 bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu ||
886 II->getIntrinsicID() == Intrinsic::aarch64_neon_umull);
887 VectorType *NewVT = cast<VectorType>(II->getType());
888 if (Constant *CV0 = dyn_cast<Constant>(Arg0)) {
889 if (Constant *CV1 = dyn_cast<Constant>(Arg1)) {
890 CV0 = ConstantExpr::getIntegerCast(CV0, NewVT, /*isSigned=*/!Zext);
891 CV1 = ConstantExpr::getIntegerCast(CV1, NewVT, /*isSigned=*/!Zext);
893 return ReplaceInstUsesWith(CI, ConstantExpr::getMul(CV0, CV1));
896 // Couldn't simplify - canonicalize constant to the RHS.
897 std::swap(Arg0, Arg1);
900 // Handle mul by one:
901 if (Constant *CV1 = dyn_cast<Constant>(Arg1))
902 if (ConstantInt *Splat =
903 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue()))
905 return CastInst::CreateIntegerCast(Arg0, II->getType(),
911 case Intrinsic::stackrestore: {
912 // If the save is right next to the restore, remove the restore. This can
913 // happen when variable allocas are DCE'd.
914 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
915 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
916 BasicBlock::iterator BI = SS;
918 return EraseInstFromFunction(CI);
922 // Scan down this block to see if there is another stack restore in the
923 // same block without an intervening call/alloca.
924 BasicBlock::iterator BI = II;
925 TerminatorInst *TI = II->getParent()->getTerminator();
926 bool CannotRemove = false;
927 for (++BI; &*BI != TI; ++BI) {
928 if (isa<AllocaInst>(BI)) {
932 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
933 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
934 // If there is a stackrestore below this one, remove this one.
935 if (II->getIntrinsicID() == Intrinsic::stackrestore)
936 return EraseInstFromFunction(CI);
937 // Otherwise, ignore the intrinsic.
939 // If we found a non-intrinsic call, we can't remove the stack
947 // If the stack restore is in a return, resume, or unwind block and if there
948 // are no allocas or calls between the restore and the return, nuke the
950 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
951 return EraseInstFromFunction(CI);
956 return visitCallSite(II);
959 // InvokeInst simplification
961 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
962 return visitCallSite(&II);
965 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
966 /// passed through the varargs area, we can eliminate the use of the cast.
967 static bool isSafeToEliminateVarargsCast(const CallSite CS,
968 const CastInst * const CI,
969 const DataLayout * const DL,
971 if (!CI->isLosslessCast())
974 // The size of ByVal or InAlloca arguments is derived from the type, so we
975 // can't change to a type with a different size. If the size were
976 // passed explicitly we could avoid this check.
977 if (!CS.isByValOrInAllocaArgument(ix))
981 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
982 Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
983 if (!SrcTy->isSized() || !DstTy->isSized())
985 if (!DL || DL->getTypeAllocSize(SrcTy) != DL->getTypeAllocSize(DstTy))
990 // Try to fold some different type of calls here.
991 // Currently we're only working with the checking functions, memcpy_chk,
992 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
993 // strcat_chk and strncat_chk.
994 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const DataLayout *DL) {
995 if (!CI->getCalledFunction()) return nullptr;
997 if (Value *With = Simplifier->optimizeCall(CI)) {
999 return CI->use_empty() ? CI : ReplaceInstUsesWith(*CI, With);
1005 static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
1006 // Strip off at most one level of pointer casts, looking for an alloca. This
1007 // is good enough in practice and simpler than handling any number of casts.
1008 Value *Underlying = TrampMem->stripPointerCasts();
1009 if (Underlying != TrampMem &&
1010 (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem))
1012 if (!isa<AllocaInst>(Underlying))
1015 IntrinsicInst *InitTrampoline = nullptr;
1016 for (User *U : TrampMem->users()) {
1017 IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
1020 if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
1022 // More than one init_trampoline writes to this value. Give up.
1024 InitTrampoline = II;
1027 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
1028 // Allow any number of calls to adjust.trampoline.
1033 // No call to init.trampoline found.
1034 if (!InitTrampoline)
1037 // Check that the alloca is being used in the expected way.
1038 if (InitTrampoline->getOperand(0) != TrampMem)
1041 return InitTrampoline;
1044 static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
1046 // Visit all the previous instructions in the basic block, and try to find a
1047 // init.trampoline which has a direct path to the adjust.trampoline.
1048 for (BasicBlock::iterator I = AdjustTramp,
1049 E = AdjustTramp->getParent()->begin(); I != E; ) {
1050 Instruction *Inst = --I;
1051 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1052 if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
1053 II->getOperand(0) == TrampMem)
1055 if (Inst->mayWriteToMemory())
1061 // Given a call to llvm.adjust.trampoline, find and return the corresponding
1062 // call to llvm.init.trampoline if the call to the trampoline can be optimized
1063 // to a direct call to a function. Otherwise return NULL.
1065 static IntrinsicInst *FindInitTrampoline(Value *Callee) {
1066 Callee = Callee->stripPointerCasts();
1067 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
1069 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
1072 Value *TrampMem = AdjustTramp->getOperand(0);
1074 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
1076 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
1081 // visitCallSite - Improvements for call and invoke instructions.
1083 Instruction *InstCombiner::visitCallSite(CallSite CS) {
1084 if (isAllocLikeFn(CS.getInstruction(), TLI))
1085 return visitAllocSite(*CS.getInstruction());
1087 bool Changed = false;
1089 // If the callee is a pointer to a function, attempt to move any casts to the
1090 // arguments of the call/invoke.
1091 Value *Callee = CS.getCalledValue();
1092 if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
1095 if (Function *CalleeF = dyn_cast<Function>(Callee))
1096 // If the call and callee calling conventions don't match, this call must
1097 // be unreachable, as the call is undefined.
1098 if (CalleeF->getCallingConv() != CS.getCallingConv() &&
1099 // Only do this for calls to a function with a body. A prototype may
1100 // not actually end up matching the implementation's calling conv for a
1101 // variety of reasons (e.g. it may be written in assembly).
1102 !CalleeF->isDeclaration()) {
1103 Instruction *OldCall = CS.getInstruction();
1104 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1105 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1107 // If OldCall does not return void then replaceAllUsesWith undef.
1108 // This allows ValueHandlers and custom metadata to adjust itself.
1109 if (!OldCall->getType()->isVoidTy())
1110 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
1111 if (isa<CallInst>(OldCall))
1112 return EraseInstFromFunction(*OldCall);
1114 // We cannot remove an invoke, because it would change the CFG, just
1115 // change the callee to a null pointer.
1116 cast<InvokeInst>(OldCall)->setCalledFunction(
1117 Constant::getNullValue(CalleeF->getType()));
1121 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1122 // If CS does not return void then replaceAllUsesWith undef.
1123 // This allows ValueHandlers and custom metadata to adjust itself.
1124 if (!CS.getInstruction()->getType()->isVoidTy())
1125 ReplaceInstUsesWith(*CS.getInstruction(),
1126 UndefValue::get(CS.getInstruction()->getType()));
1128 if (isa<InvokeInst>(CS.getInstruction())) {
1129 // Can't remove an invoke because we cannot change the CFG.
1133 // This instruction is not reachable, just remove it. We insert a store to
1134 // undef so that we know that this code is not reachable, despite the fact
1135 // that we can't modify the CFG here.
1136 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1137 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1138 CS.getInstruction());
1140 return EraseInstFromFunction(*CS.getInstruction());
1143 if (IntrinsicInst *II = FindInitTrampoline(Callee))
1144 return transformCallThroughTrampoline(CS, II);
1146 PointerType *PTy = cast<PointerType>(Callee->getType());
1147 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1148 if (FTy->isVarArg()) {
1149 int ix = FTy->getNumParams();
1150 // See if we can optimize any arguments passed through the varargs area of
1152 for (CallSite::arg_iterator I = CS.arg_begin() + FTy->getNumParams(),
1153 E = CS.arg_end(); I != E; ++I, ++ix) {
1154 CastInst *CI = dyn_cast<CastInst>(*I);
1155 if (CI && isSafeToEliminateVarargsCast(CS, CI, DL, ix)) {
1156 *I = CI->getOperand(0);
1162 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
1163 // Inline asm calls cannot throw - mark them 'nounwind'.
1164 CS.setDoesNotThrow();
1168 // Try to optimize the call if possible, we require DataLayout for most of
1169 // this. None of these calls are seen as possibly dead so go ahead and
1170 // delete the instruction now.
1171 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
1172 Instruction *I = tryOptimizeCall(CI, DL);
1173 // If we changed something return the result, etc. Otherwise let
1174 // the fallthrough check.
1175 if (I) return EraseInstFromFunction(*I);
1178 return Changed ? CS.getInstruction() : nullptr;
1181 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
1182 // attempt to move the cast to the arguments of the call/invoke.
1184 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
1186 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
1189 Instruction *Caller = CS.getInstruction();
1190 const AttributeSet &CallerPAL = CS.getAttributes();
1192 // Okay, this is a cast from a function to a different type. Unless doing so
1193 // would cause a type conversion of one of our arguments, change this call to
1194 // be a direct call with arguments casted to the appropriate types.
1196 FunctionType *FT = Callee->getFunctionType();
1197 Type *OldRetTy = Caller->getType();
1198 Type *NewRetTy = FT->getReturnType();
1200 // Check to see if we are changing the return type...
1201 if (OldRetTy != NewRetTy) {
1203 if (NewRetTy->isStructTy())
1204 return false; // TODO: Handle multiple return values.
1206 if (!CastInst::isBitCastable(NewRetTy, OldRetTy)) {
1207 if (Callee->isDeclaration())
1208 return false; // Cannot transform this return value.
1210 if (!Caller->use_empty() &&
1211 // void -> non-void is handled specially
1212 !NewRetTy->isVoidTy())
1213 return false; // Cannot transform this return value.
1216 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
1217 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1219 hasAttributes(AttributeFuncs::
1220 typeIncompatible(NewRetTy, AttributeSet::ReturnIndex),
1221 AttributeSet::ReturnIndex))
1222 return false; // Attribute not compatible with transformed value.
1225 // If the callsite is an invoke instruction, and the return value is used by
1226 // a PHI node in a successor, we cannot change the return type of the call
1227 // because there is no place to put the cast instruction (without breaking
1228 // the critical edge). Bail out in this case.
1229 if (!Caller->use_empty())
1230 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
1231 for (User *U : II->users())
1232 if (PHINode *PN = dyn_cast<PHINode>(U))
1233 if (PN->getParent() == II->getNormalDest() ||
1234 PN->getParent() == II->getUnwindDest())
1238 unsigned NumActualArgs = CS.arg_size();
1239 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1241 CallSite::arg_iterator AI = CS.arg_begin();
1242 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1243 Type *ParamTy = FT->getParamType(i);
1244 Type *ActTy = (*AI)->getType();
1246 if (!CastInst::isBitCastable(ActTy, ParamTy))
1247 return false; // Cannot transform this parameter value.
1249 if (AttrBuilder(CallerPAL.getParamAttributes(i + 1), i + 1).
1250 hasAttributes(AttributeFuncs::
1251 typeIncompatible(ParamTy, i + 1), i + 1))
1252 return false; // Attribute not compatible with transformed value.
1254 if (CS.isInAllocaArgument(i))
1255 return false; // Cannot transform to and from inalloca.
1257 // If the parameter is passed as a byval argument, then we have to have a
1258 // sized type and the sized type has to have the same size as the old type.
1259 if (ParamTy != ActTy &&
1260 CallerPAL.getParamAttributes(i + 1).hasAttribute(i + 1,
1261 Attribute::ByVal)) {
1262 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1263 if (!ParamPTy || !ParamPTy->getElementType()->isSized() || !DL)
1266 Type *CurElTy = ActTy->getPointerElementType();
1267 if (DL->getTypeAllocSize(CurElTy) !=
1268 DL->getTypeAllocSize(ParamPTy->getElementType()))
1273 if (Callee->isDeclaration()) {
1274 // Do not delete arguments unless we have a function body.
1275 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1278 // If the callee is just a declaration, don't change the varargsness of the
1279 // call. We don't want to introduce a varargs call where one doesn't
1281 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1282 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1285 // If both the callee and the cast type are varargs, we still have to make
1286 // sure the number of fixed parameters are the same or we have the same
1287 // ABI issues as if we introduce a varargs call.
1288 if (FT->isVarArg() &&
1289 cast<FunctionType>(APTy->getElementType())->isVarArg() &&
1290 FT->getNumParams() !=
1291 cast<FunctionType>(APTy->getElementType())->getNumParams())
1295 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1296 !CallerPAL.isEmpty())
1297 // In this case we have more arguments than the new function type, but we
1298 // won't be dropping them. Check that these extra arguments have attributes
1299 // that are compatible with being a vararg call argument.
1300 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1301 unsigned Index = CallerPAL.getSlotIndex(i - 1);
1302 if (Index <= FT->getNumParams())
1305 // Check if it has an attribute that's incompatible with varargs.
1306 AttributeSet PAttrs = CallerPAL.getSlotAttributes(i - 1);
1307 if (PAttrs.hasAttribute(Index, Attribute::StructRet))
1312 // Okay, we decided that this is a safe thing to do: go ahead and start
1313 // inserting cast instructions as necessary.
1314 std::vector<Value*> Args;
1315 Args.reserve(NumActualArgs);
1316 SmallVector<AttributeSet, 8> attrVec;
1317 attrVec.reserve(NumCommonArgs);
1319 // Get any return attributes.
1320 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1322 // If the return value is not being used, the type may not be compatible
1323 // with the existing attributes. Wipe out any problematic attributes.
1325 removeAttributes(AttributeFuncs::
1326 typeIncompatible(NewRetTy, AttributeSet::ReturnIndex),
1327 AttributeSet::ReturnIndex);
1329 // Add the new return attributes.
1330 if (RAttrs.hasAttributes())
1331 attrVec.push_back(AttributeSet::get(Caller->getContext(),
1332 AttributeSet::ReturnIndex, RAttrs));
1334 AI = CS.arg_begin();
1335 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1336 Type *ParamTy = FT->getParamType(i);
1338 if ((*AI)->getType() == ParamTy) {
1339 Args.push_back(*AI);
1341 Args.push_back(Builder->CreateBitCast(*AI, ParamTy));
1344 // Add any parameter attributes.
1345 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1346 if (PAttrs.hasAttributes())
1347 attrVec.push_back(AttributeSet::get(Caller->getContext(), i + 1,
1351 // If the function takes more arguments than the call was taking, add them
1353 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1354 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1356 // If we are removing arguments to the function, emit an obnoxious warning.
1357 if (FT->getNumParams() < NumActualArgs) {
1358 // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722
1359 if (FT->isVarArg()) {
1360 // Add all of the arguments in their promoted form to the arg list.
1361 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1362 Type *PTy = getPromotedType((*AI)->getType());
1363 if (PTy != (*AI)->getType()) {
1364 // Must promote to pass through va_arg area!
1365 Instruction::CastOps opcode =
1366 CastInst::getCastOpcode(*AI, false, PTy, false);
1367 Args.push_back(Builder->CreateCast(opcode, *AI, PTy));
1369 Args.push_back(*AI);
1372 // Add any parameter attributes.
1373 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1374 if (PAttrs.hasAttributes())
1375 attrVec.push_back(AttributeSet::get(FT->getContext(), i + 1,
1381 AttributeSet FnAttrs = CallerPAL.getFnAttributes();
1382 if (CallerPAL.hasAttributes(AttributeSet::FunctionIndex))
1383 attrVec.push_back(AttributeSet::get(Callee->getContext(), FnAttrs));
1385 if (NewRetTy->isVoidTy())
1386 Caller->setName(""); // Void type should not have a name.
1388 const AttributeSet &NewCallerPAL = AttributeSet::get(Callee->getContext(),
1392 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1393 NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1394 II->getUnwindDest(), Args);
1396 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1397 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1399 CallInst *CI = cast<CallInst>(Caller);
1400 NC = Builder->CreateCall(Callee, Args);
1402 if (CI->isTailCall())
1403 cast<CallInst>(NC)->setTailCall();
1404 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1405 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1408 // Insert a cast of the return type as necessary.
1410 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1411 if (!NV->getType()->isVoidTy()) {
1412 NV = NC = CastInst::Create(CastInst::BitCast, NC, OldRetTy);
1413 NC->setDebugLoc(Caller->getDebugLoc());
1415 // If this is an invoke instruction, we should insert it after the first
1416 // non-phi, instruction in the normal successor block.
1417 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1418 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
1419 InsertNewInstBefore(NC, *I);
1421 // Otherwise, it's a call, just insert cast right after the call.
1422 InsertNewInstBefore(NC, *Caller);
1424 Worklist.AddUsersToWorkList(*Caller);
1426 NV = UndefValue::get(Caller->getType());
1430 if (!Caller->use_empty())
1431 ReplaceInstUsesWith(*Caller, NV);
1432 else if (Caller->hasValueHandle())
1433 ValueHandleBase::ValueIsRAUWd(Caller, NV);
1435 EraseInstFromFunction(*Caller);
1439 // transformCallThroughTrampoline - Turn a call to a function created by
1440 // init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
1441 // underlying function.
1444 InstCombiner::transformCallThroughTrampoline(CallSite CS,
1445 IntrinsicInst *Tramp) {
1446 Value *Callee = CS.getCalledValue();
1447 PointerType *PTy = cast<PointerType>(Callee->getType());
1448 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1449 const AttributeSet &Attrs = CS.getAttributes();
1451 // If the call already has the 'nest' attribute somewhere then give up -
1452 // otherwise 'nest' would occur twice after splicing in the chain.
1453 if (Attrs.hasAttrSomewhere(Attribute::Nest))
1457 "transformCallThroughTrampoline called with incorrect CallSite.");
1459 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1460 PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1461 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1463 const AttributeSet &NestAttrs = NestF->getAttributes();
1464 if (!NestAttrs.isEmpty()) {
1465 unsigned NestIdx = 1;
1466 Type *NestTy = nullptr;
1467 AttributeSet NestAttr;
1469 // Look for a parameter marked with the 'nest' attribute.
1470 for (FunctionType::param_iterator I = NestFTy->param_begin(),
1471 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1472 if (NestAttrs.hasAttribute(NestIdx, Attribute::Nest)) {
1473 // Record the parameter type and any other attributes.
1475 NestAttr = NestAttrs.getParamAttributes(NestIdx);
1480 Instruction *Caller = CS.getInstruction();
1481 std::vector<Value*> NewArgs;
1482 NewArgs.reserve(CS.arg_size() + 1);
1484 SmallVector<AttributeSet, 8> NewAttrs;
1485 NewAttrs.reserve(Attrs.getNumSlots() + 1);
1487 // Insert the nest argument into the call argument list, which may
1488 // mean appending it. Likewise for attributes.
1490 // Add any result attributes.
1491 if (Attrs.hasAttributes(AttributeSet::ReturnIndex))
1492 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1493 Attrs.getRetAttributes()));
1497 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1499 if (Idx == NestIdx) {
1500 // Add the chain argument and attributes.
1501 Value *NestVal = Tramp->getArgOperand(2);
1502 if (NestVal->getType() != NestTy)
1503 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
1504 NewArgs.push_back(NestVal);
1505 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1512 // Add the original argument and attributes.
1513 NewArgs.push_back(*I);
1514 AttributeSet Attr = Attrs.getParamAttributes(Idx);
1515 if (Attr.hasAttributes(Idx)) {
1516 AttrBuilder B(Attr, Idx);
1517 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1518 Idx + (Idx >= NestIdx), B));
1525 // Add any function attributes.
1526 if (Attrs.hasAttributes(AttributeSet::FunctionIndex))
1527 NewAttrs.push_back(AttributeSet::get(FTy->getContext(),
1528 Attrs.getFnAttributes()));
1530 // The trampoline may have been bitcast to a bogus type (FTy).
1531 // Handle this by synthesizing a new function type, equal to FTy
1532 // with the chain parameter inserted.
1534 std::vector<Type*> NewTypes;
1535 NewTypes.reserve(FTy->getNumParams()+1);
1537 // Insert the chain's type into the list of parameter types, which may
1538 // mean appending it.
1541 FunctionType::param_iterator I = FTy->param_begin(),
1542 E = FTy->param_end();
1546 // Add the chain's type.
1547 NewTypes.push_back(NestTy);
1552 // Add the original type.
1553 NewTypes.push_back(*I);
1559 // Replace the trampoline call with a direct call. Let the generic
1560 // code sort out any function type mismatches.
1561 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1563 Constant *NewCallee =
1564 NestF->getType() == PointerType::getUnqual(NewFTy) ?
1565 NestF : ConstantExpr::getBitCast(NestF,
1566 PointerType::getUnqual(NewFTy));
1567 const AttributeSet &NewPAL =
1568 AttributeSet::get(FTy->getContext(), NewAttrs);
1570 Instruction *NewCaller;
1571 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1572 NewCaller = InvokeInst::Create(NewCallee,
1573 II->getNormalDest(), II->getUnwindDest(),
1575 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1576 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1578 NewCaller = CallInst::Create(NewCallee, NewArgs);
1579 if (cast<CallInst>(Caller)->isTailCall())
1580 cast<CallInst>(NewCaller)->setTailCall();
1581 cast<CallInst>(NewCaller)->
1582 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1583 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1590 // Replace the trampoline call with a direct call. Since there is no 'nest'
1591 // parameter, there is no need to adjust the argument list. Let the generic
1592 // code sort out any function type mismatches.
1593 Constant *NewCallee =
1594 NestF->getType() == PTy ? NestF :
1595 ConstantExpr::getBitCast(NestF, PTy);
1596 CS.setCalledFunction(NewCallee);
1597 return CS.getInstruction();