1 //===- InstCombineCalls.cpp -----------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the visitCall and visitInvoke functions.
12 //===----------------------------------------------------------------------===//
14 #include "InstCombine.h"
15 #include "llvm/ADT/Statistic.h"
16 #include "llvm/Analysis/MemoryBuiltins.h"
17 #include "llvm/IR/CallSite.h"
18 #include "llvm/IR/DataLayout.h"
19 #include "llvm/IR/Dominators.h"
20 #include "llvm/IR/PatternMatch.h"
21 #include "llvm/IR/Statepoint.h"
22 #include "llvm/Transforms/Utils/BuildLibCalls.h"
23 #include "llvm/Transforms/Utils/Local.h"
25 using namespace PatternMatch;
27 #define DEBUG_TYPE "instcombine"
29 STATISTIC(NumSimplified, "Number of library calls simplified");
31 /// getPromotedType - Return the specified type promoted as it would be to pass
32 /// though a va_arg area.
33 static Type *getPromotedType(Type *Ty) {
34 if (IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
35 if (ITy->getBitWidth() < 32)
36 return Type::getInt32Ty(Ty->getContext());
41 /// reduceToSingleValueType - Given an aggregate type which ultimately holds a
42 /// single scalar element, like {{{type}}} or [1 x type], return type.
43 static Type *reduceToSingleValueType(Type *T) {
44 while (!T->isSingleValueType()) {
45 if (StructType *STy = dyn_cast<StructType>(T)) {
46 if (STy->getNumElements() == 1)
47 T = STy->getElementType(0);
50 } else if (ArrayType *ATy = dyn_cast<ArrayType>(T)) {
51 if (ATy->getNumElements() == 1)
52 T = ATy->getElementType();
62 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
63 unsigned DstAlign = getKnownAlignment(MI->getArgOperand(0), DL, AT, MI, DT);
64 unsigned SrcAlign = getKnownAlignment(MI->getArgOperand(1), DL, AT, MI, DT);
65 unsigned MinAlign = std::min(DstAlign, SrcAlign);
66 unsigned CopyAlign = MI->getAlignment();
68 if (CopyAlign < MinAlign) {
69 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
74 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
76 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getArgOperand(2));
77 if (!MemOpLength) return nullptr;
79 // Source and destination pointer types are always "i8*" for intrinsic. See
80 // if the size is something we can handle with a single primitive load/store.
81 // A single load+store correctly handles overlapping memory in the memmove
83 uint64_t Size = MemOpLength->getLimitedValue();
84 assert(Size && "0-sized memory transferring should be removed already.");
86 if (Size > 8 || (Size&(Size-1)))
87 return nullptr; // If not 1/2/4/8 bytes, exit.
89 // Use an integer load+store unless we can find something better.
91 cast<PointerType>(MI->getArgOperand(1)->getType())->getAddressSpace();
93 cast<PointerType>(MI->getArgOperand(0)->getType())->getAddressSpace();
95 IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3);
96 Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp);
97 Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp);
99 // Memcpy forces the use of i8* for the source and destination. That means
100 // that if you're using memcpy to move one double around, you'll get a cast
101 // from double* to i8*. We'd much rather use a double load+store rather than
102 // an i64 load+store, here because this improves the odds that the source or
103 // dest address will be promotable. See if we can find a better type than the
105 Value *StrippedDest = MI->getArgOperand(0)->stripPointerCasts();
106 MDNode *CopyMD = nullptr;
107 if (StrippedDest != MI->getArgOperand(0)) {
108 Type *SrcETy = cast<PointerType>(StrippedDest->getType())
110 if (DL && SrcETy->isSized() && DL->getTypeStoreSize(SrcETy) == Size) {
111 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
112 // down through these levels if so.
113 SrcETy = reduceToSingleValueType(SrcETy);
115 if (SrcETy->isSingleValueType()) {
116 NewSrcPtrTy = PointerType::get(SrcETy, SrcAddrSp);
117 NewDstPtrTy = PointerType::get(SrcETy, DstAddrSp);
119 // If the memcpy has metadata describing the members, see if we can
120 // get the TBAA tag describing our copy.
121 if (MDNode *M = MI->getMetadata(LLVMContext::MD_tbaa_struct)) {
122 if (M->getNumOperands() == 3 && M->getOperand(0) &&
123 mdconst::hasa<ConstantInt>(M->getOperand(0)) &&
124 mdconst::extract<ConstantInt>(M->getOperand(0))->isNullValue() &&
126 mdconst::hasa<ConstantInt>(M->getOperand(1)) &&
127 mdconst::extract<ConstantInt>(M->getOperand(1))->getValue() ==
129 M->getOperand(2) && isa<MDNode>(M->getOperand(2)))
130 CopyMD = cast<MDNode>(M->getOperand(2));
136 // If the memcpy/memmove provides better alignment info than we can
138 SrcAlign = std::max(SrcAlign, CopyAlign);
139 DstAlign = std::max(DstAlign, CopyAlign);
141 Value *Src = Builder->CreateBitCast(MI->getArgOperand(1), NewSrcPtrTy);
142 Value *Dest = Builder->CreateBitCast(MI->getArgOperand(0), NewDstPtrTy);
143 LoadInst *L = Builder->CreateLoad(Src, MI->isVolatile());
144 L->setAlignment(SrcAlign);
146 L->setMetadata(LLVMContext::MD_tbaa, CopyMD);
147 StoreInst *S = Builder->CreateStore(L, Dest, MI->isVolatile());
148 S->setAlignment(DstAlign);
150 S->setMetadata(LLVMContext::MD_tbaa, CopyMD);
152 // Set the size of the copy to 0, it will be deleted on the next iteration.
153 MI->setArgOperand(2, Constant::getNullValue(MemOpLength->getType()));
157 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
158 unsigned Alignment = getKnownAlignment(MI->getDest(), DL, AT, MI, DT);
159 if (MI->getAlignment() < Alignment) {
160 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
165 // Extract the length and alignment and fill if they are constant.
166 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
167 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
168 if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8))
170 uint64_t Len = LenC->getLimitedValue();
171 Alignment = MI->getAlignment();
172 assert(Len && "0-sized memory setting should be removed already.");
174 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
175 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
176 Type *ITy = IntegerType::get(MI->getContext(), Len*8); // n=1 -> i8.
178 Value *Dest = MI->getDest();
179 unsigned DstAddrSp = cast<PointerType>(Dest->getType())->getAddressSpace();
180 Type *NewDstPtrTy = PointerType::get(ITy, DstAddrSp);
181 Dest = Builder->CreateBitCast(Dest, NewDstPtrTy);
183 // Alignment 0 is identity for alignment 1 for memset, but not store.
184 if (Alignment == 0) Alignment = 1;
186 // Extract the fill value and store.
187 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
188 StoreInst *S = Builder->CreateStore(ConstantInt::get(ITy, Fill), Dest,
190 S->setAlignment(Alignment);
192 // Set the size of the copy to 0, it will be deleted on the next iteration.
193 MI->setLength(Constant::getNullValue(LenC->getType()));
200 /// visitCallInst - CallInst simplification. This mostly only handles folding
201 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
202 /// the heavy lifting.
204 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
205 if (isFreeCall(&CI, TLI))
206 return visitFree(CI);
208 // If the caller function is nounwind, mark the call as nounwind, even if the
210 if (CI.getParent()->getParent()->doesNotThrow() &&
211 !CI.doesNotThrow()) {
212 CI.setDoesNotThrow();
216 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
217 if (!II) return visitCallSite(&CI);
219 // Intrinsics cannot occur in an invoke, so handle them here instead of in
221 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
222 bool Changed = false;
224 // memmove/cpy/set of zero bytes is a noop.
225 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
226 if (NumBytes->isNullValue())
227 return EraseInstFromFunction(CI);
229 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
230 if (CI->getZExtValue() == 1) {
231 // Replace the instruction with just byte operations. We would
232 // transform other cases to loads/stores, but we don't know if
233 // alignment is sufficient.
237 // No other transformations apply to volatile transfers.
238 if (MI->isVolatile())
241 // If we have a memmove and the source operation is a constant global,
242 // then the source and dest pointers can't alias, so we can change this
243 // into a call to memcpy.
244 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
245 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
246 if (GVSrc->isConstant()) {
247 Module *M = CI.getParent()->getParent()->getParent();
248 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
249 Type *Tys[3] = { CI.getArgOperand(0)->getType(),
250 CI.getArgOperand(1)->getType(),
251 CI.getArgOperand(2)->getType() };
252 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys));
257 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
258 // memmove(x,x,size) -> noop.
259 if (MTI->getSource() == MTI->getDest())
260 return EraseInstFromFunction(CI);
263 // If we can determine a pointer alignment that is bigger than currently
264 // set, update the alignment.
265 if (isa<MemTransferInst>(MI)) {
266 if (Instruction *I = SimplifyMemTransfer(MI))
268 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
269 if (Instruction *I = SimplifyMemSet(MSI))
273 if (Changed) return II;
276 switch (II->getIntrinsicID()) {
278 case Intrinsic::objectsize: {
280 if (getObjectSize(II->getArgOperand(0), Size, DL, TLI))
281 return ReplaceInstUsesWith(CI, ConstantInt::get(CI.getType(), Size));
284 case Intrinsic::bswap: {
285 Value *IIOperand = II->getArgOperand(0);
288 // bswap(bswap(x)) -> x
289 if (match(IIOperand, m_BSwap(m_Value(X))))
290 return ReplaceInstUsesWith(CI, X);
292 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
293 if (match(IIOperand, m_Trunc(m_BSwap(m_Value(X))))) {
294 unsigned C = X->getType()->getPrimitiveSizeInBits() -
295 IIOperand->getType()->getPrimitiveSizeInBits();
296 Value *CV = ConstantInt::get(X->getType(), C);
297 Value *V = Builder->CreateLShr(X, CV);
298 return new TruncInst(V, IIOperand->getType());
303 case Intrinsic::powi:
304 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
307 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
310 return ReplaceInstUsesWith(CI, II->getArgOperand(0));
311 // powi(x, -1) -> 1/x
312 if (Power->isAllOnesValue())
313 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
314 II->getArgOperand(0));
317 case Intrinsic::cttz: {
318 // If all bits below the first known one are known zero,
319 // this value is constant.
320 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
321 // FIXME: Try to simplify vectors of integers.
323 uint32_t BitWidth = IT->getBitWidth();
324 APInt KnownZero(BitWidth, 0);
325 APInt KnownOne(BitWidth, 0);
326 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
327 unsigned TrailingZeros = KnownOne.countTrailingZeros();
328 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros));
329 if ((Mask & KnownZero) == Mask)
330 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
331 APInt(BitWidth, TrailingZeros)));
335 case Intrinsic::ctlz: {
336 // If all bits above the first known one are known zero,
337 // this value is constant.
338 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
339 // FIXME: Try to simplify vectors of integers.
341 uint32_t BitWidth = IT->getBitWidth();
342 APInt KnownZero(BitWidth, 0);
343 APInt KnownOne(BitWidth, 0);
344 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
345 unsigned LeadingZeros = KnownOne.countLeadingZeros();
346 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
347 if ((Mask & KnownZero) == Mask)
348 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
349 APInt(BitWidth, LeadingZeros)));
353 case Intrinsic::uadd_with_overflow: {
354 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
355 IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
356 uint32_t BitWidth = IT->getBitWidth();
357 APInt LHSKnownZero(BitWidth, 0);
358 APInt LHSKnownOne(BitWidth, 0);
359 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, 0, II);
360 bool LHSKnownNegative = LHSKnownOne[BitWidth - 1];
361 bool LHSKnownPositive = LHSKnownZero[BitWidth - 1];
363 if (LHSKnownNegative || LHSKnownPositive) {
364 APInt RHSKnownZero(BitWidth, 0);
365 APInt RHSKnownOne(BitWidth, 0);
366 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, 0, II);
367 bool RHSKnownNegative = RHSKnownOne[BitWidth - 1];
368 bool RHSKnownPositive = RHSKnownZero[BitWidth - 1];
369 if (LHSKnownNegative && RHSKnownNegative) {
370 // The sign bit is set in both cases: this MUST overflow.
371 // Create a simple add instruction, and insert it into the struct.
372 return CreateOverflowTuple(II, Builder->CreateAdd(LHS, RHS), true,
376 if (LHSKnownPositive && RHSKnownPositive) {
377 // The sign bit is clear in both cases: this CANNOT overflow.
378 // Create a simple add instruction, and insert it into the struct.
379 return CreateOverflowTuple(II, Builder->CreateNUWAdd(LHS, RHS), false);
383 // FALL THROUGH uadd into sadd
384 case Intrinsic::sadd_with_overflow:
385 // Canonicalize constants into the RHS.
386 if (isa<Constant>(II->getArgOperand(0)) &&
387 !isa<Constant>(II->getArgOperand(1))) {
388 Value *LHS = II->getArgOperand(0);
389 II->setArgOperand(0, II->getArgOperand(1));
390 II->setArgOperand(1, LHS);
394 // X + undef -> undef
395 if (isa<UndefValue>(II->getArgOperand(1)))
396 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
398 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
399 // X + 0 -> {X, false}
401 return CreateOverflowTuple(II, II->getArgOperand(0), false,
406 // We can strength reduce reduce this signed add into a regular add if we
407 // can prove that it will never overflow.
408 if (II->getIntrinsicID() == Intrinsic::sadd_with_overflow) {
409 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
410 if (WillNotOverflowSignedAdd(LHS, RHS, II)) {
411 return CreateOverflowTuple(II, Builder->CreateNSWAdd(LHS, RHS), false);
416 case Intrinsic::usub_with_overflow:
417 case Intrinsic::ssub_with_overflow: {
418 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
419 // undef - X -> undef
420 // X - undef -> undef
421 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
422 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
424 if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(RHS)) {
425 // X - 0 -> {X, false}
426 if (ConstRHS->isZero()) {
427 return CreateOverflowTuple(II, LHS, false, /*ReUseName*/false);
432 case Intrinsic::umul_with_overflow: {
433 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
434 unsigned BitWidth = cast<IntegerType>(LHS->getType())->getBitWidth();
436 APInt LHSKnownZero(BitWidth, 0);
437 APInt LHSKnownOne(BitWidth, 0);
438 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, 0, II);
439 APInt RHSKnownZero(BitWidth, 0);
440 APInt RHSKnownOne(BitWidth, 0);
441 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, 0, II);
443 // Get the largest possible values for each operand.
444 APInt LHSMax = ~LHSKnownZero;
445 APInt RHSMax = ~RHSKnownZero;
447 // If multiplying the maximum values does not overflow then we can turn
448 // this into a plain NUW mul.
450 LHSMax.umul_ov(RHSMax, Overflow);
452 return CreateOverflowTuple(II, Builder->CreateNUWMul(LHS, RHS), false);
455 case Intrinsic::smul_with_overflow:
456 // Canonicalize constants into the RHS.
457 if (isa<Constant>(II->getArgOperand(0)) &&
458 !isa<Constant>(II->getArgOperand(1))) {
459 Value *LHS = II->getArgOperand(0);
460 II->setArgOperand(0, II->getArgOperand(1));
461 II->setArgOperand(1, LHS);
465 // X * undef -> undef
466 if (isa<UndefValue>(II->getArgOperand(1)))
467 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
469 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
472 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType()));
474 // X * 1 -> {X, false}
475 if (RHSI->equalsInt(1)) {
476 return CreateOverflowTuple(II, II->getArgOperand(0), false,
481 case Intrinsic::minnum:
482 case Intrinsic::maxnum: {
483 Value *Arg0 = II->getArgOperand(0);
484 Value *Arg1 = II->getArgOperand(1);
488 return ReplaceInstUsesWith(CI, Arg0);
490 const ConstantFP *C0 = dyn_cast<ConstantFP>(Arg0);
491 const ConstantFP *C1 = dyn_cast<ConstantFP>(Arg1);
493 // Canonicalize constants into the RHS.
495 II->setArgOperand(0, Arg1);
496 II->setArgOperand(1, Arg0);
501 if (C1 && C1->isNaN())
502 return ReplaceInstUsesWith(CI, Arg0);
504 // This is the value because if undef were NaN, we would return the other
505 // value and cannot return a NaN unless both operands are.
507 // fmin(undef, x) -> x
508 if (isa<UndefValue>(Arg0))
509 return ReplaceInstUsesWith(CI, Arg1);
511 // fmin(x, undef) -> x
512 if (isa<UndefValue>(Arg1))
513 return ReplaceInstUsesWith(CI, Arg0);
517 if (II->getIntrinsicID() == Intrinsic::minnum) {
518 // fmin(x, fmin(x, y)) -> fmin(x, y)
519 // fmin(y, fmin(x, y)) -> fmin(x, y)
520 if (match(Arg1, m_FMin(m_Value(X), m_Value(Y)))) {
521 if (Arg0 == X || Arg0 == Y)
522 return ReplaceInstUsesWith(CI, Arg1);
525 // fmin(fmin(x, y), x) -> fmin(x, y)
526 // fmin(fmin(x, y), y) -> fmin(x, y)
527 if (match(Arg0, m_FMin(m_Value(X), m_Value(Y)))) {
528 if (Arg1 == X || Arg1 == Y)
529 return ReplaceInstUsesWith(CI, Arg0);
532 // TODO: fmin(nnan x, inf) -> x
533 // TODO: fmin(nnan ninf x, flt_max) -> x
534 if (C1 && C1->isInfinity()) {
535 // fmin(x, -inf) -> -inf
536 if (C1->isNegative())
537 return ReplaceInstUsesWith(CI, Arg1);
540 assert(II->getIntrinsicID() == Intrinsic::maxnum);
541 // fmax(x, fmax(x, y)) -> fmax(x, y)
542 // fmax(y, fmax(x, y)) -> fmax(x, y)
543 if (match(Arg1, m_FMax(m_Value(X), m_Value(Y)))) {
544 if (Arg0 == X || Arg0 == Y)
545 return ReplaceInstUsesWith(CI, Arg1);
548 // fmax(fmax(x, y), x) -> fmax(x, y)
549 // fmax(fmax(x, y), y) -> fmax(x, y)
550 if (match(Arg0, m_FMax(m_Value(X), m_Value(Y)))) {
551 if (Arg1 == X || Arg1 == Y)
552 return ReplaceInstUsesWith(CI, Arg0);
555 // TODO: fmax(nnan x, -inf) -> x
556 // TODO: fmax(nnan ninf x, -flt_max) -> x
557 if (C1 && C1->isInfinity()) {
558 // fmax(x, inf) -> inf
559 if (!C1->isNegative())
560 return ReplaceInstUsesWith(CI, Arg1);
565 case Intrinsic::ppc_altivec_lvx:
566 case Intrinsic::ppc_altivec_lvxl:
567 // Turn PPC lvx -> load if the pointer is known aligned.
568 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16,
569 DL, AT, II, DT) >= 16) {
570 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
571 PointerType::getUnqual(II->getType()));
572 return new LoadInst(Ptr);
575 case Intrinsic::ppc_vsx_lxvw4x:
576 case Intrinsic::ppc_vsx_lxvd2x: {
577 // Turn PPC VSX loads into normal loads.
578 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
579 PointerType::getUnqual(II->getType()));
580 return new LoadInst(Ptr, Twine(""), false, 1);
582 case Intrinsic::ppc_altivec_stvx:
583 case Intrinsic::ppc_altivec_stvxl:
584 // Turn stvx -> store if the pointer is known aligned.
585 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16,
586 DL, AT, II, DT) >= 16) {
588 PointerType::getUnqual(II->getArgOperand(0)->getType());
589 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
590 return new StoreInst(II->getArgOperand(0), Ptr);
593 case Intrinsic::ppc_vsx_stxvw4x:
594 case Intrinsic::ppc_vsx_stxvd2x: {
595 // Turn PPC VSX stores into normal stores.
596 Type *OpPtrTy = PointerType::getUnqual(II->getArgOperand(0)->getType());
597 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
598 return new StoreInst(II->getArgOperand(0), Ptr, false, 1);
600 case Intrinsic::x86_sse_storeu_ps:
601 case Intrinsic::x86_sse2_storeu_pd:
602 case Intrinsic::x86_sse2_storeu_dq:
603 // Turn X86 storeu -> store if the pointer is known aligned.
604 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16,
605 DL, AT, II, DT) >= 16) {
607 PointerType::getUnqual(II->getArgOperand(1)->getType());
608 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
609 return new StoreInst(II->getArgOperand(1), Ptr);
613 case Intrinsic::x86_sse_cvtss2si:
614 case Intrinsic::x86_sse_cvtss2si64:
615 case Intrinsic::x86_sse_cvttss2si:
616 case Intrinsic::x86_sse_cvttss2si64:
617 case Intrinsic::x86_sse2_cvtsd2si:
618 case Intrinsic::x86_sse2_cvtsd2si64:
619 case Intrinsic::x86_sse2_cvttsd2si:
620 case Intrinsic::x86_sse2_cvttsd2si64: {
621 // These intrinsics only demand the 0th element of their input vectors. If
622 // we can simplify the input based on that, do so now.
624 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
625 APInt DemandedElts(VWidth, 1);
626 APInt UndefElts(VWidth, 0);
627 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
628 DemandedElts, UndefElts)) {
629 II->setArgOperand(0, V);
635 // Constant fold <A x Bi> << Ci.
636 // FIXME: We don't handle _dq because it's a shift of an i128, but is
637 // represented in the IR as <2 x i64>. A per element shift is wrong.
638 case Intrinsic::x86_sse2_psll_d:
639 case Intrinsic::x86_sse2_psll_q:
640 case Intrinsic::x86_sse2_psll_w:
641 case Intrinsic::x86_sse2_pslli_d:
642 case Intrinsic::x86_sse2_pslli_q:
643 case Intrinsic::x86_sse2_pslli_w:
644 case Intrinsic::x86_avx2_psll_d:
645 case Intrinsic::x86_avx2_psll_q:
646 case Intrinsic::x86_avx2_psll_w:
647 case Intrinsic::x86_avx2_pslli_d:
648 case Intrinsic::x86_avx2_pslli_q:
649 case Intrinsic::x86_avx2_pslli_w:
650 case Intrinsic::x86_sse2_psrl_d:
651 case Intrinsic::x86_sse2_psrl_q:
652 case Intrinsic::x86_sse2_psrl_w:
653 case Intrinsic::x86_sse2_psrli_d:
654 case Intrinsic::x86_sse2_psrli_q:
655 case Intrinsic::x86_sse2_psrli_w:
656 case Intrinsic::x86_avx2_psrl_d:
657 case Intrinsic::x86_avx2_psrl_q:
658 case Intrinsic::x86_avx2_psrl_w:
659 case Intrinsic::x86_avx2_psrli_d:
660 case Intrinsic::x86_avx2_psrli_q:
661 case Intrinsic::x86_avx2_psrli_w: {
662 // Simplify if count is constant. To 0 if >= BitWidth,
663 // otherwise to shl/lshr.
664 auto CDV = dyn_cast<ConstantDataVector>(II->getArgOperand(1));
665 auto CInt = dyn_cast<ConstantInt>(II->getArgOperand(1));
670 Count = cast<ConstantInt>(CDV->getElementAsConstant(0));
674 auto Vec = II->getArgOperand(0);
675 auto VT = cast<VectorType>(Vec->getType());
676 if (Count->getZExtValue() >
677 VT->getElementType()->getPrimitiveSizeInBits() - 1)
678 return ReplaceInstUsesWith(
679 CI, ConstantAggregateZero::get(Vec->getType()));
681 bool isPackedShiftLeft = true;
682 switch (II->getIntrinsicID()) {
684 case Intrinsic::x86_sse2_psrl_d:
685 case Intrinsic::x86_sse2_psrl_q:
686 case Intrinsic::x86_sse2_psrl_w:
687 case Intrinsic::x86_sse2_psrli_d:
688 case Intrinsic::x86_sse2_psrli_q:
689 case Intrinsic::x86_sse2_psrli_w:
690 case Intrinsic::x86_avx2_psrl_d:
691 case Intrinsic::x86_avx2_psrl_q:
692 case Intrinsic::x86_avx2_psrl_w:
693 case Intrinsic::x86_avx2_psrli_d:
694 case Intrinsic::x86_avx2_psrli_q:
695 case Intrinsic::x86_avx2_psrli_w: isPackedShiftLeft = false; break;
698 unsigned VWidth = VT->getNumElements();
699 // Get a constant vector of the same type as the first operand.
700 auto VTCI = ConstantInt::get(VT->getElementType(), Count->getZExtValue());
701 if (isPackedShiftLeft)
702 return BinaryOperator::CreateShl(Vec,
703 Builder->CreateVectorSplat(VWidth, VTCI));
705 return BinaryOperator::CreateLShr(Vec,
706 Builder->CreateVectorSplat(VWidth, VTCI));
709 case Intrinsic::x86_sse41_pmovsxbw:
710 case Intrinsic::x86_sse41_pmovsxwd:
711 case Intrinsic::x86_sse41_pmovsxdq:
712 case Intrinsic::x86_sse41_pmovzxbw:
713 case Intrinsic::x86_sse41_pmovzxwd:
714 case Intrinsic::x86_sse41_pmovzxdq: {
715 // pmov{s|z}x ignores the upper half of their input vectors.
717 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
718 unsigned LowHalfElts = VWidth / 2;
719 APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts));
720 APInt UndefElts(VWidth, 0);
721 if (Value *TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0),
724 II->setArgOperand(0, TmpV);
730 case Intrinsic::x86_sse4a_insertqi: {
731 // insertqi x, y, 64, 0 can just copy y's lower bits and leave the top
733 // TODO: eventually we should lower this intrinsic to IR
734 if (auto CIWidth = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
735 if (auto CIStart = dyn_cast<ConstantInt>(II->getArgOperand(3))) {
736 if (CIWidth->equalsInt(64) && CIStart->isZero()) {
737 Value *Vec = II->getArgOperand(1);
738 Value *Undef = UndefValue::get(Vec->getType());
739 const uint32_t Mask[] = { 0, 2 };
740 return ReplaceInstUsesWith(
742 Builder->CreateShuffleVector(
743 Vec, Undef, ConstantDataVector::get(
744 II->getContext(), makeArrayRef(Mask))));
746 } else if (auto Source =
747 dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
748 if (Source->hasOneUse() &&
749 Source->getArgOperand(1) == II->getArgOperand(1)) {
750 // If the source of the insert has only one use and it's another
751 // insert (and they're both inserting from the same vector), try to
752 // bundle both together.
754 dyn_cast<ConstantInt>(Source->getArgOperand(2));
756 dyn_cast<ConstantInt>(Source->getArgOperand(3));
757 if (CISourceStart && CISourceWidth) {
758 unsigned Start = CIStart->getZExtValue();
759 unsigned Width = CIWidth->getZExtValue();
760 unsigned End = Start + Width;
761 unsigned SourceStart = CISourceStart->getZExtValue();
762 unsigned SourceWidth = CISourceWidth->getZExtValue();
763 unsigned SourceEnd = SourceStart + SourceWidth;
764 unsigned NewStart, NewWidth;
765 bool ShouldReplace = false;
766 if (Start <= SourceStart && SourceStart <= End) {
768 NewWidth = std::max(End, SourceEnd) - NewStart;
769 ShouldReplace = true;
770 } else if (SourceStart <= Start && Start <= SourceEnd) {
771 NewStart = SourceStart;
772 NewWidth = std::max(SourceEnd, End) - NewStart;
773 ShouldReplace = true;
777 Constant *ConstantWidth = ConstantInt::get(
778 II->getArgOperand(2)->getType(), NewWidth, false);
779 Constant *ConstantStart = ConstantInt::get(
780 II->getArgOperand(3)->getType(), NewStart, false);
781 Value *Args[4] = { Source->getArgOperand(0),
782 II->getArgOperand(1), ConstantWidth,
784 Module *M = CI.getParent()->getParent()->getParent();
786 Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi);
787 return ReplaceInstUsesWith(CI, Builder->CreateCall(F, Args));
797 case Intrinsic::x86_sse41_pblendvb:
798 case Intrinsic::x86_sse41_blendvps:
799 case Intrinsic::x86_sse41_blendvpd:
800 case Intrinsic::x86_avx_blendv_ps_256:
801 case Intrinsic::x86_avx_blendv_pd_256:
802 case Intrinsic::x86_avx2_pblendvb: {
803 // Convert blendv* to vector selects if the mask is constant.
804 // This optimization is convoluted because the intrinsic is defined as
805 // getting a vector of floats or doubles for the ps and pd versions.
806 // FIXME: That should be changed.
807 Value *Mask = II->getArgOperand(2);
808 if (auto C = dyn_cast<ConstantDataVector>(Mask)) {
809 auto Tyi1 = Builder->getInt1Ty();
810 auto SelectorType = cast<VectorType>(Mask->getType());
811 auto EltTy = SelectorType->getElementType();
812 unsigned Size = SelectorType->getNumElements();
816 : (EltTy->isDoubleTy() ? 64 : EltTy->getIntegerBitWidth());
817 assert((BitWidth == 64 || BitWidth == 32 || BitWidth == 8) &&
818 "Wrong arguments for variable blend intrinsic");
819 SmallVector<Constant *, 32> Selectors;
820 for (unsigned I = 0; I < Size; ++I) {
821 // The intrinsics only read the top bit
824 Selector = C->getElementAsInteger(I);
826 Selector = C->getElementAsAPFloat(I).bitcastToAPInt().getZExtValue();
827 Selectors.push_back(ConstantInt::get(Tyi1, Selector >> (BitWidth - 1)));
829 auto NewSelector = ConstantVector::get(Selectors);
830 return SelectInst::Create(NewSelector, II->getArgOperand(1),
831 II->getArgOperand(0), "blendv");
837 case Intrinsic::x86_avx_vpermilvar_ps:
838 case Intrinsic::x86_avx_vpermilvar_ps_256:
839 case Intrinsic::x86_avx_vpermilvar_pd:
840 case Intrinsic::x86_avx_vpermilvar_pd_256: {
841 // Convert vpermil* to shufflevector if the mask is constant.
842 Value *V = II->getArgOperand(1);
843 unsigned Size = cast<VectorType>(V->getType())->getNumElements();
844 assert(Size == 8 || Size == 4 || Size == 2);
846 if (auto C = dyn_cast<ConstantDataVector>(V)) {
847 // The intrinsics only read one or two bits, clear the rest.
848 for (unsigned I = 0; I < Size; ++I) {
849 uint32_t Index = C->getElementAsInteger(I) & 0x3;
850 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd ||
851 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256)
855 } else if (isa<ConstantAggregateZero>(V)) {
856 for (unsigned I = 0; I < Size; ++I)
861 // The _256 variants are a bit trickier since the mask bits always index
862 // into the corresponding 128 half. In order to convert to a generic
863 // shuffle, we have to make that explicit.
864 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_ps_256 ||
865 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256) {
866 for (unsigned I = Size / 2; I < Size; ++I)
867 Indexes[I] += Size / 2;
870 ConstantDataVector::get(V->getContext(), makeArrayRef(Indexes, Size));
871 auto V1 = II->getArgOperand(0);
872 auto V2 = UndefValue::get(V1->getType());
873 auto Shuffle = Builder->CreateShuffleVector(V1, V2, NewC);
874 return ReplaceInstUsesWith(CI, Shuffle);
877 case Intrinsic::ppc_altivec_vperm:
878 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
879 // Note that ppc_altivec_vperm has a big-endian bias, so when creating
880 // a vectorshuffle for little endian, we must undo the transformation
881 // performed on vec_perm in altivec.h. That is, we must complement
882 // the permutation mask with respect to 31 and reverse the order of
884 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) {
885 assert(Mask->getType()->getVectorNumElements() == 16 &&
886 "Bad type for intrinsic!");
888 // Check that all of the elements are integer constants or undefs.
889 bool AllEltsOk = true;
890 for (unsigned i = 0; i != 16; ++i) {
891 Constant *Elt = Mask->getAggregateElement(i);
892 if (!Elt || !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) {
899 // Cast the input vectors to byte vectors.
900 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
902 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
904 Value *Result = UndefValue::get(Op0->getType());
906 // Only extract each element once.
907 Value *ExtractedElts[32];
908 memset(ExtractedElts, 0, sizeof(ExtractedElts));
910 for (unsigned i = 0; i != 16; ++i) {
911 if (isa<UndefValue>(Mask->getAggregateElement(i)))
914 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue();
915 Idx &= 31; // Match the hardware behavior.
916 if (DL && DL->isLittleEndian())
919 if (!ExtractedElts[Idx]) {
920 Value *Op0ToUse = (DL && DL->isLittleEndian()) ? Op1 : Op0;
921 Value *Op1ToUse = (DL && DL->isLittleEndian()) ? Op0 : Op1;
923 Builder->CreateExtractElement(Idx < 16 ? Op0ToUse : Op1ToUse,
924 Builder->getInt32(Idx&15));
927 // Insert this value into the result vector.
928 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
929 Builder->getInt32(i));
931 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
936 case Intrinsic::arm_neon_vld1:
937 case Intrinsic::arm_neon_vld2:
938 case Intrinsic::arm_neon_vld3:
939 case Intrinsic::arm_neon_vld4:
940 case Intrinsic::arm_neon_vld2lane:
941 case Intrinsic::arm_neon_vld3lane:
942 case Intrinsic::arm_neon_vld4lane:
943 case Intrinsic::arm_neon_vst1:
944 case Intrinsic::arm_neon_vst2:
945 case Intrinsic::arm_neon_vst3:
946 case Intrinsic::arm_neon_vst4:
947 case Intrinsic::arm_neon_vst2lane:
948 case Intrinsic::arm_neon_vst3lane:
949 case Intrinsic::arm_neon_vst4lane: {
950 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), DL, AT, II, DT);
951 unsigned AlignArg = II->getNumArgOperands() - 1;
952 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
953 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
954 II->setArgOperand(AlignArg,
955 ConstantInt::get(Type::getInt32Ty(II->getContext()),
962 case Intrinsic::arm_neon_vmulls:
963 case Intrinsic::arm_neon_vmullu:
964 case Intrinsic::aarch64_neon_smull:
965 case Intrinsic::aarch64_neon_umull: {
966 Value *Arg0 = II->getArgOperand(0);
967 Value *Arg1 = II->getArgOperand(1);
969 // Handle mul by zero first:
970 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) {
971 return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType()));
974 // Check for constant LHS & RHS - in this case we just simplify.
975 bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu ||
976 II->getIntrinsicID() == Intrinsic::aarch64_neon_umull);
977 VectorType *NewVT = cast<VectorType>(II->getType());
978 if (Constant *CV0 = dyn_cast<Constant>(Arg0)) {
979 if (Constant *CV1 = dyn_cast<Constant>(Arg1)) {
980 CV0 = ConstantExpr::getIntegerCast(CV0, NewVT, /*isSigned=*/!Zext);
981 CV1 = ConstantExpr::getIntegerCast(CV1, NewVT, /*isSigned=*/!Zext);
983 return ReplaceInstUsesWith(CI, ConstantExpr::getMul(CV0, CV1));
986 // Couldn't simplify - canonicalize constant to the RHS.
987 std::swap(Arg0, Arg1);
990 // Handle mul by one:
991 if (Constant *CV1 = dyn_cast<Constant>(Arg1))
992 if (ConstantInt *Splat =
993 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue()))
995 return CastInst::CreateIntegerCast(Arg0, II->getType(),
1001 case Intrinsic::AMDGPU_rcp: {
1002 if (const ConstantFP *C = dyn_cast<ConstantFP>(II->getArgOperand(0))) {
1003 const APFloat &ArgVal = C->getValueAPF();
1004 APFloat Val(ArgVal.getSemantics(), 1.0);
1005 APFloat::opStatus Status = Val.divide(ArgVal,
1006 APFloat::rmNearestTiesToEven);
1007 // Only do this if it was exact and therefore not dependent on the
1009 if (Status == APFloat::opOK)
1010 return ReplaceInstUsesWith(CI, ConstantFP::get(II->getContext(), Val));
1015 case Intrinsic::stackrestore: {
1016 // If the save is right next to the restore, remove the restore. This can
1017 // happen when variable allocas are DCE'd.
1018 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
1019 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
1020 BasicBlock::iterator BI = SS;
1022 return EraseInstFromFunction(CI);
1026 // Scan down this block to see if there is another stack restore in the
1027 // same block without an intervening call/alloca.
1028 BasicBlock::iterator BI = II;
1029 TerminatorInst *TI = II->getParent()->getTerminator();
1030 bool CannotRemove = false;
1031 for (++BI; &*BI != TI; ++BI) {
1032 if (isa<AllocaInst>(BI)) {
1033 CannotRemove = true;
1036 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
1037 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
1038 // If there is a stackrestore below this one, remove this one.
1039 if (II->getIntrinsicID() == Intrinsic::stackrestore)
1040 return EraseInstFromFunction(CI);
1041 // Otherwise, ignore the intrinsic.
1043 // If we found a non-intrinsic call, we can't remove the stack
1045 CannotRemove = true;
1051 // If the stack restore is in a return, resume, or unwind block and if there
1052 // are no allocas or calls between the restore and the return, nuke the
1054 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
1055 return EraseInstFromFunction(CI);
1058 case Intrinsic::assume: {
1059 // Canonicalize assume(a && b) -> assume(a); assume(b);
1060 // Note: New assumption intrinsics created here are registered by
1061 // the InstCombineIRInserter object.
1062 Value *IIOperand = II->getArgOperand(0), *A, *B,
1063 *AssumeIntrinsic = II->getCalledValue();
1064 if (match(IIOperand, m_And(m_Value(A), m_Value(B)))) {
1065 Builder->CreateCall(AssumeIntrinsic, A, II->getName());
1066 Builder->CreateCall(AssumeIntrinsic, B, II->getName());
1067 return EraseInstFromFunction(*II);
1069 // assume(!(a || b)) -> assume(!a); assume(!b);
1070 if (match(IIOperand, m_Not(m_Or(m_Value(A), m_Value(B))))) {
1071 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(A),
1073 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(B),
1075 return EraseInstFromFunction(*II);
1078 // assume( (load addr) != null ) -> add 'nonnull' metadata to load
1079 // (if assume is valid at the load)
1080 if (ICmpInst* ICmp = dyn_cast<ICmpInst>(IIOperand)) {
1081 Value *LHS = ICmp->getOperand(0);
1082 Value *RHS = ICmp->getOperand(1);
1083 if (ICmpInst::ICMP_NE == ICmp->getPredicate() &&
1084 isa<LoadInst>(LHS) &&
1085 isa<Constant>(RHS) &&
1086 RHS->getType()->isPointerTy() &&
1087 cast<Constant>(RHS)->isNullValue()) {
1088 LoadInst* LI = cast<LoadInst>(LHS);
1089 if (isValidAssumeForContext(II, LI, DL, DT)) {
1090 MDNode *MD = MDNode::get(II->getContext(), None);
1091 LI->setMetadata(LLVMContext::MD_nonnull, MD);
1092 return EraseInstFromFunction(*II);
1095 // TODO: apply nonnull return attributes to calls and invokes
1096 // TODO: apply range metadata for range check patterns?
1098 // If there is a dominating assume with the same condition as this one,
1099 // then this one is redundant, and should be removed.
1100 APInt KnownZero(1, 0), KnownOne(1, 0);
1101 computeKnownBits(IIOperand, KnownZero, KnownOne, 0, II);
1102 if (KnownOne.isAllOnesValue())
1103 return EraseInstFromFunction(*II);
1109 return visitCallSite(II);
1112 // InvokeInst simplification
1114 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
1115 return visitCallSite(&II);
1118 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
1119 /// passed through the varargs area, we can eliminate the use of the cast.
1120 static bool isSafeToEliminateVarargsCast(const CallSite CS,
1121 const CastInst * const CI,
1122 const DataLayout * const DL,
1124 if (!CI->isLosslessCast())
1127 // If this is a GC intrinsic, avoid munging types. We need types for
1128 // statepoint reconstruction in SelectionDAG.
1129 // TODO: This is probably something which should be expanded to all
1130 // intrinsics since the entire point of intrinsics is that
1131 // they are understandable by the optimizer.
1132 if (isStatepoint(CS) || isGCRelocate(CS) || isGCResult(CS))
1135 // The size of ByVal or InAlloca arguments is derived from the type, so we
1136 // can't change to a type with a different size. If the size were
1137 // passed explicitly we could avoid this check.
1138 if (!CS.isByValOrInAllocaArgument(ix))
1142 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
1143 Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
1144 if (!SrcTy->isSized() || !DstTy->isSized())
1146 if (!DL || DL->getTypeAllocSize(SrcTy) != DL->getTypeAllocSize(DstTy))
1151 // Try to fold some different type of calls here.
1152 // Currently we're only working with the checking functions, memcpy_chk,
1153 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
1154 // strcat_chk and strncat_chk.
1155 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const DataLayout *DL) {
1156 if (!CI->getCalledFunction()) return nullptr;
1158 if (Value *With = Simplifier->optimizeCall(CI)) {
1160 return CI->use_empty() ? CI : ReplaceInstUsesWith(*CI, With);
1166 static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
1167 // Strip off at most one level of pointer casts, looking for an alloca. This
1168 // is good enough in practice and simpler than handling any number of casts.
1169 Value *Underlying = TrampMem->stripPointerCasts();
1170 if (Underlying != TrampMem &&
1171 (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem))
1173 if (!isa<AllocaInst>(Underlying))
1176 IntrinsicInst *InitTrampoline = nullptr;
1177 for (User *U : TrampMem->users()) {
1178 IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
1181 if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
1183 // More than one init_trampoline writes to this value. Give up.
1185 InitTrampoline = II;
1188 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
1189 // Allow any number of calls to adjust.trampoline.
1194 // No call to init.trampoline found.
1195 if (!InitTrampoline)
1198 // Check that the alloca is being used in the expected way.
1199 if (InitTrampoline->getOperand(0) != TrampMem)
1202 return InitTrampoline;
1205 static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
1207 // Visit all the previous instructions in the basic block, and try to find a
1208 // init.trampoline which has a direct path to the adjust.trampoline.
1209 for (BasicBlock::iterator I = AdjustTramp,
1210 E = AdjustTramp->getParent()->begin(); I != E; ) {
1211 Instruction *Inst = --I;
1212 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1213 if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
1214 II->getOperand(0) == TrampMem)
1216 if (Inst->mayWriteToMemory())
1222 // Given a call to llvm.adjust.trampoline, find and return the corresponding
1223 // call to llvm.init.trampoline if the call to the trampoline can be optimized
1224 // to a direct call to a function. Otherwise return NULL.
1226 static IntrinsicInst *FindInitTrampoline(Value *Callee) {
1227 Callee = Callee->stripPointerCasts();
1228 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
1230 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
1233 Value *TrampMem = AdjustTramp->getOperand(0);
1235 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
1237 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
1242 // visitCallSite - Improvements for call and invoke instructions.
1244 Instruction *InstCombiner::visitCallSite(CallSite CS) {
1245 if (isAllocLikeFn(CS.getInstruction(), TLI))
1246 return visitAllocSite(*CS.getInstruction());
1248 bool Changed = false;
1250 // If the callee is a pointer to a function, attempt to move any casts to the
1251 // arguments of the call/invoke.
1252 Value *Callee = CS.getCalledValue();
1253 if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
1256 if (Function *CalleeF = dyn_cast<Function>(Callee))
1257 // If the call and callee calling conventions don't match, this call must
1258 // be unreachable, as the call is undefined.
1259 if (CalleeF->getCallingConv() != CS.getCallingConv() &&
1260 // Only do this for calls to a function with a body. A prototype may
1261 // not actually end up matching the implementation's calling conv for a
1262 // variety of reasons (e.g. it may be written in assembly).
1263 !CalleeF->isDeclaration()) {
1264 Instruction *OldCall = CS.getInstruction();
1265 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1266 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1268 // If OldCall does not return void then replaceAllUsesWith undef.
1269 // This allows ValueHandlers and custom metadata to adjust itself.
1270 if (!OldCall->getType()->isVoidTy())
1271 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
1272 if (isa<CallInst>(OldCall))
1273 return EraseInstFromFunction(*OldCall);
1275 // We cannot remove an invoke, because it would change the CFG, just
1276 // change the callee to a null pointer.
1277 cast<InvokeInst>(OldCall)->setCalledFunction(
1278 Constant::getNullValue(CalleeF->getType()));
1282 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1283 // If CS does not return void then replaceAllUsesWith undef.
1284 // This allows ValueHandlers and custom metadata to adjust itself.
1285 if (!CS.getInstruction()->getType()->isVoidTy())
1286 ReplaceInstUsesWith(*CS.getInstruction(),
1287 UndefValue::get(CS.getInstruction()->getType()));
1289 if (isa<InvokeInst>(CS.getInstruction())) {
1290 // Can't remove an invoke because we cannot change the CFG.
1294 // This instruction is not reachable, just remove it. We insert a store to
1295 // undef so that we know that this code is not reachable, despite the fact
1296 // that we can't modify the CFG here.
1297 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1298 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1299 CS.getInstruction());
1301 return EraseInstFromFunction(*CS.getInstruction());
1304 if (IntrinsicInst *II = FindInitTrampoline(Callee))
1305 return transformCallThroughTrampoline(CS, II);
1307 PointerType *PTy = cast<PointerType>(Callee->getType());
1308 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1309 if (FTy->isVarArg()) {
1310 int ix = FTy->getNumParams();
1311 // See if we can optimize any arguments passed through the varargs area of
1313 for (CallSite::arg_iterator I = CS.arg_begin() + FTy->getNumParams(),
1314 E = CS.arg_end(); I != E; ++I, ++ix) {
1315 CastInst *CI = dyn_cast<CastInst>(*I);
1316 if (CI && isSafeToEliminateVarargsCast(CS, CI, DL, ix)) {
1317 *I = CI->getOperand(0);
1323 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
1324 // Inline asm calls cannot throw - mark them 'nounwind'.
1325 CS.setDoesNotThrow();
1329 // Try to optimize the call if possible, we require DataLayout for most of
1330 // this. None of these calls are seen as possibly dead so go ahead and
1331 // delete the instruction now.
1332 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
1333 Instruction *I = tryOptimizeCall(CI, DL);
1334 // If we changed something return the result, etc. Otherwise let
1335 // the fallthrough check.
1336 if (I) return EraseInstFromFunction(*I);
1339 return Changed ? CS.getInstruction() : nullptr;
1342 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
1343 // attempt to move the cast to the arguments of the call/invoke.
1345 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
1347 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
1350 Instruction *Caller = CS.getInstruction();
1351 const AttributeSet &CallerPAL = CS.getAttributes();
1353 // Okay, this is a cast from a function to a different type. Unless doing so
1354 // would cause a type conversion of one of our arguments, change this call to
1355 // be a direct call with arguments casted to the appropriate types.
1357 FunctionType *FT = Callee->getFunctionType();
1358 Type *OldRetTy = Caller->getType();
1359 Type *NewRetTy = FT->getReturnType();
1361 // Check to see if we are changing the return type...
1362 if (OldRetTy != NewRetTy) {
1364 if (NewRetTy->isStructTy())
1365 return false; // TODO: Handle multiple return values.
1367 if (!CastInst::isBitCastable(NewRetTy, OldRetTy)) {
1368 if (Callee->isDeclaration())
1369 return false; // Cannot transform this return value.
1371 if (!Caller->use_empty() &&
1372 // void -> non-void is handled specially
1373 !NewRetTy->isVoidTy())
1374 return false; // Cannot transform this return value.
1377 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
1378 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1380 hasAttributes(AttributeFuncs::
1381 typeIncompatible(NewRetTy, AttributeSet::ReturnIndex),
1382 AttributeSet::ReturnIndex))
1383 return false; // Attribute not compatible with transformed value.
1386 // If the callsite is an invoke instruction, and the return value is used by
1387 // a PHI node in a successor, we cannot change the return type of the call
1388 // because there is no place to put the cast instruction (without breaking
1389 // the critical edge). Bail out in this case.
1390 if (!Caller->use_empty())
1391 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
1392 for (User *U : II->users())
1393 if (PHINode *PN = dyn_cast<PHINode>(U))
1394 if (PN->getParent() == II->getNormalDest() ||
1395 PN->getParent() == II->getUnwindDest())
1399 unsigned NumActualArgs = CS.arg_size();
1400 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1402 CallSite::arg_iterator AI = CS.arg_begin();
1403 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1404 Type *ParamTy = FT->getParamType(i);
1405 Type *ActTy = (*AI)->getType();
1407 if (!CastInst::isBitCastable(ActTy, ParamTy))
1408 return false; // Cannot transform this parameter value.
1410 if (AttrBuilder(CallerPAL.getParamAttributes(i + 1), i + 1).
1411 hasAttributes(AttributeFuncs::
1412 typeIncompatible(ParamTy, i + 1), i + 1))
1413 return false; // Attribute not compatible with transformed value.
1415 if (CS.isInAllocaArgument(i))
1416 return false; // Cannot transform to and from inalloca.
1418 // If the parameter is passed as a byval argument, then we have to have a
1419 // sized type and the sized type has to have the same size as the old type.
1420 if (ParamTy != ActTy &&
1421 CallerPAL.getParamAttributes(i + 1).hasAttribute(i + 1,
1422 Attribute::ByVal)) {
1423 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1424 if (!ParamPTy || !ParamPTy->getElementType()->isSized() || !DL)
1427 Type *CurElTy = ActTy->getPointerElementType();
1428 if (DL->getTypeAllocSize(CurElTy) !=
1429 DL->getTypeAllocSize(ParamPTy->getElementType()))
1434 if (Callee->isDeclaration()) {
1435 // Do not delete arguments unless we have a function body.
1436 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1439 // If the callee is just a declaration, don't change the varargsness of the
1440 // call. We don't want to introduce a varargs call where one doesn't
1442 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1443 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1446 // If both the callee and the cast type are varargs, we still have to make
1447 // sure the number of fixed parameters are the same or we have the same
1448 // ABI issues as if we introduce a varargs call.
1449 if (FT->isVarArg() &&
1450 cast<FunctionType>(APTy->getElementType())->isVarArg() &&
1451 FT->getNumParams() !=
1452 cast<FunctionType>(APTy->getElementType())->getNumParams())
1456 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1457 !CallerPAL.isEmpty())
1458 // In this case we have more arguments than the new function type, but we
1459 // won't be dropping them. Check that these extra arguments have attributes
1460 // that are compatible with being a vararg call argument.
1461 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1462 unsigned Index = CallerPAL.getSlotIndex(i - 1);
1463 if (Index <= FT->getNumParams())
1466 // Check if it has an attribute that's incompatible with varargs.
1467 AttributeSet PAttrs = CallerPAL.getSlotAttributes(i - 1);
1468 if (PAttrs.hasAttribute(Index, Attribute::StructRet))
1473 // Okay, we decided that this is a safe thing to do: go ahead and start
1474 // inserting cast instructions as necessary.
1475 std::vector<Value*> Args;
1476 Args.reserve(NumActualArgs);
1477 SmallVector<AttributeSet, 8> attrVec;
1478 attrVec.reserve(NumCommonArgs);
1480 // Get any return attributes.
1481 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1483 // If the return value is not being used, the type may not be compatible
1484 // with the existing attributes. Wipe out any problematic attributes.
1486 removeAttributes(AttributeFuncs::
1487 typeIncompatible(NewRetTy, AttributeSet::ReturnIndex),
1488 AttributeSet::ReturnIndex);
1490 // Add the new return attributes.
1491 if (RAttrs.hasAttributes())
1492 attrVec.push_back(AttributeSet::get(Caller->getContext(),
1493 AttributeSet::ReturnIndex, RAttrs));
1495 AI = CS.arg_begin();
1496 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1497 Type *ParamTy = FT->getParamType(i);
1499 if ((*AI)->getType() == ParamTy) {
1500 Args.push_back(*AI);
1502 Args.push_back(Builder->CreateBitCast(*AI, ParamTy));
1505 // Add any parameter attributes.
1506 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1507 if (PAttrs.hasAttributes())
1508 attrVec.push_back(AttributeSet::get(Caller->getContext(), i + 1,
1512 // If the function takes more arguments than the call was taking, add them
1514 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1515 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1517 // If we are removing arguments to the function, emit an obnoxious warning.
1518 if (FT->getNumParams() < NumActualArgs) {
1519 // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722
1520 if (FT->isVarArg()) {
1521 // Add all of the arguments in their promoted form to the arg list.
1522 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1523 Type *PTy = getPromotedType((*AI)->getType());
1524 if (PTy != (*AI)->getType()) {
1525 // Must promote to pass through va_arg area!
1526 Instruction::CastOps opcode =
1527 CastInst::getCastOpcode(*AI, false, PTy, false);
1528 Args.push_back(Builder->CreateCast(opcode, *AI, PTy));
1530 Args.push_back(*AI);
1533 // Add any parameter attributes.
1534 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1535 if (PAttrs.hasAttributes())
1536 attrVec.push_back(AttributeSet::get(FT->getContext(), i + 1,
1542 AttributeSet FnAttrs = CallerPAL.getFnAttributes();
1543 if (CallerPAL.hasAttributes(AttributeSet::FunctionIndex))
1544 attrVec.push_back(AttributeSet::get(Callee->getContext(), FnAttrs));
1546 if (NewRetTy->isVoidTy())
1547 Caller->setName(""); // Void type should not have a name.
1549 const AttributeSet &NewCallerPAL = AttributeSet::get(Callee->getContext(),
1553 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1554 NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1555 II->getUnwindDest(), Args);
1557 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1558 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1560 CallInst *CI = cast<CallInst>(Caller);
1561 NC = Builder->CreateCall(Callee, Args);
1563 if (CI->isTailCall())
1564 cast<CallInst>(NC)->setTailCall();
1565 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1566 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1569 // Insert a cast of the return type as necessary.
1571 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1572 if (!NV->getType()->isVoidTy()) {
1573 NV = NC = CastInst::Create(CastInst::BitCast, NC, OldRetTy);
1574 NC->setDebugLoc(Caller->getDebugLoc());
1576 // If this is an invoke instruction, we should insert it after the first
1577 // non-phi, instruction in the normal successor block.
1578 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1579 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
1580 InsertNewInstBefore(NC, *I);
1582 // Otherwise, it's a call, just insert cast right after the call.
1583 InsertNewInstBefore(NC, *Caller);
1585 Worklist.AddUsersToWorkList(*Caller);
1587 NV = UndefValue::get(Caller->getType());
1591 if (!Caller->use_empty())
1592 ReplaceInstUsesWith(*Caller, NV);
1593 else if (Caller->hasValueHandle()) {
1594 if (OldRetTy == NV->getType())
1595 ValueHandleBase::ValueIsRAUWd(Caller, NV);
1597 // We cannot call ValueIsRAUWd with a different type, and the
1598 // actual tracked value will disappear.
1599 ValueHandleBase::ValueIsDeleted(Caller);
1602 EraseInstFromFunction(*Caller);
1606 // transformCallThroughTrampoline - Turn a call to a function created by
1607 // init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
1608 // underlying function.
1611 InstCombiner::transformCallThroughTrampoline(CallSite CS,
1612 IntrinsicInst *Tramp) {
1613 Value *Callee = CS.getCalledValue();
1614 PointerType *PTy = cast<PointerType>(Callee->getType());
1615 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1616 const AttributeSet &Attrs = CS.getAttributes();
1618 // If the call already has the 'nest' attribute somewhere then give up -
1619 // otherwise 'nest' would occur twice after splicing in the chain.
1620 if (Attrs.hasAttrSomewhere(Attribute::Nest))
1624 "transformCallThroughTrampoline called with incorrect CallSite.");
1626 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1627 PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1628 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1630 const AttributeSet &NestAttrs = NestF->getAttributes();
1631 if (!NestAttrs.isEmpty()) {
1632 unsigned NestIdx = 1;
1633 Type *NestTy = nullptr;
1634 AttributeSet NestAttr;
1636 // Look for a parameter marked with the 'nest' attribute.
1637 for (FunctionType::param_iterator I = NestFTy->param_begin(),
1638 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1639 if (NestAttrs.hasAttribute(NestIdx, Attribute::Nest)) {
1640 // Record the parameter type and any other attributes.
1642 NestAttr = NestAttrs.getParamAttributes(NestIdx);
1647 Instruction *Caller = CS.getInstruction();
1648 std::vector<Value*> NewArgs;
1649 NewArgs.reserve(CS.arg_size() + 1);
1651 SmallVector<AttributeSet, 8> NewAttrs;
1652 NewAttrs.reserve(Attrs.getNumSlots() + 1);
1654 // Insert the nest argument into the call argument list, which may
1655 // mean appending it. Likewise for attributes.
1657 // Add any result attributes.
1658 if (Attrs.hasAttributes(AttributeSet::ReturnIndex))
1659 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1660 Attrs.getRetAttributes()));
1664 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1666 if (Idx == NestIdx) {
1667 // Add the chain argument and attributes.
1668 Value *NestVal = Tramp->getArgOperand(2);
1669 if (NestVal->getType() != NestTy)
1670 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
1671 NewArgs.push_back(NestVal);
1672 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1679 // Add the original argument and attributes.
1680 NewArgs.push_back(*I);
1681 AttributeSet Attr = Attrs.getParamAttributes(Idx);
1682 if (Attr.hasAttributes(Idx)) {
1683 AttrBuilder B(Attr, Idx);
1684 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1685 Idx + (Idx >= NestIdx), B));
1692 // Add any function attributes.
1693 if (Attrs.hasAttributes(AttributeSet::FunctionIndex))
1694 NewAttrs.push_back(AttributeSet::get(FTy->getContext(),
1695 Attrs.getFnAttributes()));
1697 // The trampoline may have been bitcast to a bogus type (FTy).
1698 // Handle this by synthesizing a new function type, equal to FTy
1699 // with the chain parameter inserted.
1701 std::vector<Type*> NewTypes;
1702 NewTypes.reserve(FTy->getNumParams()+1);
1704 // Insert the chain's type into the list of parameter types, which may
1705 // mean appending it.
1708 FunctionType::param_iterator I = FTy->param_begin(),
1709 E = FTy->param_end();
1713 // Add the chain's type.
1714 NewTypes.push_back(NestTy);
1719 // Add the original type.
1720 NewTypes.push_back(*I);
1726 // Replace the trampoline call with a direct call. Let the generic
1727 // code sort out any function type mismatches.
1728 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1730 Constant *NewCallee =
1731 NestF->getType() == PointerType::getUnqual(NewFTy) ?
1732 NestF : ConstantExpr::getBitCast(NestF,
1733 PointerType::getUnqual(NewFTy));
1734 const AttributeSet &NewPAL =
1735 AttributeSet::get(FTy->getContext(), NewAttrs);
1737 Instruction *NewCaller;
1738 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1739 NewCaller = InvokeInst::Create(NewCallee,
1740 II->getNormalDest(), II->getUnwindDest(),
1742 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1743 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1745 NewCaller = CallInst::Create(NewCallee, NewArgs);
1746 if (cast<CallInst>(Caller)->isTailCall())
1747 cast<CallInst>(NewCaller)->setTailCall();
1748 cast<CallInst>(NewCaller)->
1749 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1750 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1757 // Replace the trampoline call with a direct call. Since there is no 'nest'
1758 // parameter, there is no need to adjust the argument list. Let the generic
1759 // code sort out any function type mismatches.
1760 Constant *NewCallee =
1761 NestF->getType() == PTy ? NestF :
1762 ConstantExpr::getBitCast(NestF, PTy);
1763 CS.setCalledFunction(NewCallee);
1764 return CS.getInstruction();