1 //===- InstCombineCasts.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 visit functions for cast operations.
12 //===----------------------------------------------------------------------===//
14 #include "InstCombineInternal.h"
15 #include "llvm/ADT/SetVector.h"
16 #include "llvm/Analysis/ConstantFolding.h"
17 #include "llvm/IR/DataLayout.h"
18 #include "llvm/IR/PatternMatch.h"
19 #include "llvm/Analysis/TargetLibraryInfo.h"
21 using namespace PatternMatch;
23 #define DEBUG_TYPE "instcombine"
25 /// Analyze 'Val', seeing if it is a simple linear expression.
26 /// If so, decompose it, returning some value X, such that Val is
29 static Value *decomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
31 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
32 Offset = CI->getZExtValue();
34 return ConstantInt::get(Val->getType(), 0);
37 if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
38 // Cannot look past anything that might overflow.
39 OverflowingBinaryOperator *OBI = dyn_cast<OverflowingBinaryOperator>(Val);
40 if (OBI && !OBI->hasNoUnsignedWrap() && !OBI->hasNoSignedWrap()) {
46 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
47 if (I->getOpcode() == Instruction::Shl) {
48 // This is a value scaled by '1 << the shift amt'.
49 Scale = UINT64_C(1) << RHS->getZExtValue();
51 return I->getOperand(0);
54 if (I->getOpcode() == Instruction::Mul) {
55 // This value is scaled by 'RHS'.
56 Scale = RHS->getZExtValue();
58 return I->getOperand(0);
61 if (I->getOpcode() == Instruction::Add) {
62 // We have X+C. Check to see if we really have (X*C2)+C1,
63 // where C1 is divisible by C2.
66 decomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
67 Offset += RHS->getZExtValue();
74 // Otherwise, we can't look past this.
80 /// If we find a cast of an allocation instruction, try to eliminate the cast by
81 /// moving the type information into the alloc.
82 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
84 PointerType *PTy = cast<PointerType>(CI.getType());
86 BuilderTy AllocaBuilder(*Builder);
87 AllocaBuilder.SetInsertPoint(&AI);
89 // Get the type really allocated and the type casted to.
90 Type *AllocElTy = AI.getAllocatedType();
91 Type *CastElTy = PTy->getElementType();
92 if (!AllocElTy->isSized() || !CastElTy->isSized()) return nullptr;
94 unsigned AllocElTyAlign = DL.getABITypeAlignment(AllocElTy);
95 unsigned CastElTyAlign = DL.getABITypeAlignment(CastElTy);
96 if (CastElTyAlign < AllocElTyAlign) return nullptr;
98 // If the allocation has multiple uses, only promote it if we are strictly
99 // increasing the alignment of the resultant allocation. If we keep it the
100 // same, we open the door to infinite loops of various kinds.
101 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return nullptr;
103 uint64_t AllocElTySize = DL.getTypeAllocSize(AllocElTy);
104 uint64_t CastElTySize = DL.getTypeAllocSize(CastElTy);
105 if (CastElTySize == 0 || AllocElTySize == 0) return nullptr;
107 // If the allocation has multiple uses, only promote it if we're not
108 // shrinking the amount of memory being allocated.
109 uint64_t AllocElTyStoreSize = DL.getTypeStoreSize(AllocElTy);
110 uint64_t CastElTyStoreSize = DL.getTypeStoreSize(CastElTy);
111 if (!AI.hasOneUse() && CastElTyStoreSize < AllocElTyStoreSize) return nullptr;
113 // See if we can satisfy the modulus by pulling a scale out of the array
115 unsigned ArraySizeScale;
116 uint64_t ArrayOffset;
117 Value *NumElements = // See if the array size is a decomposable linear expr.
118 decomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
120 // If we can now satisfy the modulus, by using a non-1 scale, we really can
122 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
123 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return nullptr;
125 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
126 Value *Amt = nullptr;
130 Amt = ConstantInt::get(AI.getArraySize()->getType(), Scale);
131 // Insert before the alloca, not before the cast.
132 Amt = AllocaBuilder.CreateMul(Amt, NumElements);
135 if (uint64_t Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
136 Value *Off = ConstantInt::get(AI.getArraySize()->getType(),
138 Amt = AllocaBuilder.CreateAdd(Amt, Off);
141 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt);
142 New->setAlignment(AI.getAlignment());
144 New->setUsedWithInAlloca(AI.isUsedWithInAlloca());
146 // If the allocation has multiple real uses, insert a cast and change all
147 // things that used it to use the new cast. This will also hack on CI, but it
149 if (!AI.hasOneUse()) {
150 // New is the allocation instruction, pointer typed. AI is the original
151 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
152 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast");
153 replaceInstUsesWith(AI, NewCast);
155 return replaceInstUsesWith(CI, New);
158 /// Given an expression that CanEvaluateTruncated or CanEvaluateSExtd returns
159 /// true for, actually insert the code to evaluate the expression.
160 Value *InstCombiner::EvaluateInDifferentType(Value *V, Type *Ty,
162 if (Constant *C = dyn_cast<Constant>(V)) {
163 C = ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
164 // If we got a constantexpr back, try to simplify it with DL info.
165 if (Constant *FoldedC = ConstantFoldConstant(C, DL, &TLI))
170 // Otherwise, it must be an instruction.
171 Instruction *I = cast<Instruction>(V);
172 Instruction *Res = nullptr;
173 unsigned Opc = I->getOpcode();
175 case Instruction::Add:
176 case Instruction::Sub:
177 case Instruction::Mul:
178 case Instruction::And:
179 case Instruction::Or:
180 case Instruction::Xor:
181 case Instruction::AShr:
182 case Instruction::LShr:
183 case Instruction::Shl:
184 case Instruction::UDiv:
185 case Instruction::URem: {
186 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
187 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
188 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
191 case Instruction::Trunc:
192 case Instruction::ZExt:
193 case Instruction::SExt:
194 // If the source type of the cast is the type we're trying for then we can
195 // just return the source. There's no need to insert it because it is not
197 if (I->getOperand(0)->getType() == Ty)
198 return I->getOperand(0);
200 // Otherwise, must be the same type of cast, so just reinsert a new one.
201 // This also handles the case of zext(trunc(x)) -> zext(x).
202 Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty,
203 Opc == Instruction::SExt);
205 case Instruction::Select: {
206 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
207 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
208 Res = SelectInst::Create(I->getOperand(0), True, False);
211 case Instruction::PHI: {
212 PHINode *OPN = cast<PHINode>(I);
213 PHINode *NPN = PHINode::Create(Ty, OPN->getNumIncomingValues());
214 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
216 EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
217 NPN->addIncoming(V, OPN->getIncomingBlock(i));
223 // TODO: Can handle more cases here.
224 llvm_unreachable("Unreachable!");
228 return InsertNewInstWith(Res, *I);
231 Instruction::CastOps InstCombiner::isEliminableCastPair(const CastInst *CI1,
232 const CastInst *CI2) {
233 Type *SrcTy = CI1->getSrcTy();
234 Type *MidTy = CI1->getDestTy();
235 Type *DstTy = CI2->getDestTy();
237 Instruction::CastOps firstOp = Instruction::CastOps(CI1->getOpcode());
238 Instruction::CastOps secondOp = Instruction::CastOps(CI2->getOpcode());
240 SrcTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(SrcTy) : nullptr;
242 MidTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(MidTy) : nullptr;
244 DstTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(DstTy) : nullptr;
245 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
246 DstTy, SrcIntPtrTy, MidIntPtrTy,
249 // We don't want to form an inttoptr or ptrtoint that converts to an integer
250 // type that differs from the pointer size.
251 if ((Res == Instruction::IntToPtr && SrcTy != DstIntPtrTy) ||
252 (Res == Instruction::PtrToInt && DstTy != SrcIntPtrTy))
255 return Instruction::CastOps(Res);
258 /// @brief Implement the transforms common to all CastInst visitors.
259 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
260 Value *Src = CI.getOperand(0);
262 // Try to eliminate a cast of a cast.
263 if (auto *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
264 if (Instruction::CastOps NewOpc = isEliminableCastPair(CSrc, &CI)) {
265 // The first cast (CSrc) is eliminable so we need to fix up or replace
266 // the second cast (CI). CSrc will then have a good chance of being dead.
267 return CastInst::Create(NewOpc, CSrc->getOperand(0), CI.getType());
271 // If we are casting a select, then fold the cast into the select.
272 if (auto *SI = dyn_cast<SelectInst>(Src))
273 if (Instruction *NV = FoldOpIntoSelect(CI, SI))
276 // If we are casting a PHI, then fold the cast into the PHI.
277 if (auto *PN = dyn_cast<PHINode>(Src)) {
278 // Don't do this if it would create a PHI node with an illegal type from a
280 if (!Src->getType()->isIntegerTy() || !CI.getType()->isIntegerTy() ||
281 shouldChangeType(CI.getType(), Src->getType()))
282 if (Instruction *NV = foldOpIntoPhi(CI, PN))
289 /// Return true if we can evaluate the specified expression tree as type Ty
290 /// instead of its larger type, and arrive with the same value.
291 /// This is used by code that tries to eliminate truncates.
293 /// Ty will always be a type smaller than V. We should return true if trunc(V)
294 /// can be computed by computing V in the smaller type. If V is an instruction,
295 /// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only
296 /// makes sense if x and y can be efficiently truncated.
298 /// This function works on both vectors and scalars.
300 static bool canEvaluateTruncated(Value *V, Type *Ty, InstCombiner &IC,
302 // We can always evaluate constants in another type.
303 if (isa<Constant>(V))
306 Instruction *I = dyn_cast<Instruction>(V);
307 if (!I) return false;
309 Type *OrigTy = V->getType();
311 // If this is an extension from the dest type, we can eliminate it, even if it
312 // has multiple uses.
313 if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
314 I->getOperand(0)->getType() == Ty)
317 // We can't extend or shrink something that has multiple uses: doing so would
318 // require duplicating the instruction in general, which isn't profitable.
319 if (!I->hasOneUse()) return false;
321 unsigned Opc = I->getOpcode();
323 case Instruction::Add:
324 case Instruction::Sub:
325 case Instruction::Mul:
326 case Instruction::And:
327 case Instruction::Or:
328 case Instruction::Xor:
329 // These operators can all arbitrarily be extended or truncated.
330 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
331 canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
333 case Instruction::UDiv:
334 case Instruction::URem: {
335 // UDiv and URem can be truncated if all the truncated bits are zero.
336 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
337 uint32_t BitWidth = Ty->getScalarSizeInBits();
338 if (BitWidth < OrigBitWidth) {
339 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
340 if (IC.MaskedValueIsZero(I->getOperand(0), Mask, 0, CxtI) &&
341 IC.MaskedValueIsZero(I->getOperand(1), Mask, 0, CxtI)) {
342 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
343 canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
348 case Instruction::Shl:
349 // If we are truncating the result of this SHL, and if it's a shift of a
350 // constant amount, we can always perform a SHL in a smaller type.
351 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
352 uint32_t BitWidth = Ty->getScalarSizeInBits();
353 if (CI->getLimitedValue(BitWidth) < BitWidth)
354 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI);
357 case Instruction::LShr:
358 // If this is a truncate of a logical shr, we can truncate it to a smaller
359 // lshr iff we know that the bits we would otherwise be shifting in are
361 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
362 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
363 uint32_t BitWidth = Ty->getScalarSizeInBits();
364 if (IC.MaskedValueIsZero(I->getOperand(0),
365 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth), 0, CxtI) &&
366 CI->getLimitedValue(BitWidth) < BitWidth) {
367 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI);
371 case Instruction::Trunc:
372 // trunc(trunc(x)) -> trunc(x)
374 case Instruction::ZExt:
375 case Instruction::SExt:
376 // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest
377 // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest
379 case Instruction::Select: {
380 SelectInst *SI = cast<SelectInst>(I);
381 return canEvaluateTruncated(SI->getTrueValue(), Ty, IC, CxtI) &&
382 canEvaluateTruncated(SI->getFalseValue(), Ty, IC, CxtI);
384 case Instruction::PHI: {
385 // We can change a phi if we can change all operands. Note that we never
386 // get into trouble with cyclic PHIs here because we only consider
387 // instructions with a single use.
388 PHINode *PN = cast<PHINode>(I);
389 for (Value *IncValue : PN->incoming_values())
390 if (!canEvaluateTruncated(IncValue, Ty, IC, CxtI))
395 // TODO: Can handle more cases here.
402 /// Given a vector that is bitcast to an integer, optionally logically
403 /// right-shifted, and truncated, convert it to an extractelement.
404 /// Example (big endian):
405 /// trunc (lshr (bitcast <4 x i32> %X to i128), 32) to i32
407 /// extractelement <4 x i32> %X, 1
408 static Instruction *foldVecTruncToExtElt(TruncInst &Trunc, InstCombiner &IC,
409 const DataLayout &DL) {
410 Value *TruncOp = Trunc.getOperand(0);
411 Type *DestType = Trunc.getType();
412 if (!TruncOp->hasOneUse() || !isa<IntegerType>(DestType))
415 Value *VecInput = nullptr;
416 ConstantInt *ShiftVal = nullptr;
417 if (!match(TruncOp, m_CombineOr(m_BitCast(m_Value(VecInput)),
418 m_LShr(m_BitCast(m_Value(VecInput)),
419 m_ConstantInt(ShiftVal)))) ||
420 !isa<VectorType>(VecInput->getType()))
423 VectorType *VecType = cast<VectorType>(VecInput->getType());
424 unsigned VecWidth = VecType->getPrimitiveSizeInBits();
425 unsigned DestWidth = DestType->getPrimitiveSizeInBits();
426 unsigned ShiftAmount = ShiftVal ? ShiftVal->getZExtValue() : 0;
428 if ((VecWidth % DestWidth != 0) || (ShiftAmount % DestWidth != 0))
431 // If the element type of the vector doesn't match the result type,
432 // bitcast it to a vector type that we can extract from.
433 unsigned NumVecElts = VecWidth / DestWidth;
434 if (VecType->getElementType() != DestType) {
435 VecType = VectorType::get(DestType, NumVecElts);
436 VecInput = IC.Builder->CreateBitCast(VecInput, VecType, "bc");
439 unsigned Elt = ShiftAmount / DestWidth;
440 if (DL.isBigEndian())
441 Elt = NumVecElts - 1 - Elt;
443 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt));
446 /// Try to narrow the width of bitwise logic instructions with constants.
447 Instruction *InstCombiner::shrinkBitwiseLogic(TruncInst &Trunc) {
448 Type *SrcTy = Trunc.getSrcTy();
449 Type *DestTy = Trunc.getType();
450 if (isa<IntegerType>(SrcTy) && !shouldChangeType(SrcTy, DestTy))
453 BinaryOperator *LogicOp;
455 if (!match(Trunc.getOperand(0), m_OneUse(m_BinOp(LogicOp))) ||
456 !LogicOp->isBitwiseLogicOp() ||
457 !match(LogicOp->getOperand(1), m_Constant(C)))
460 // trunc (logic X, C) --> logic (trunc X, C')
461 Constant *NarrowC = ConstantExpr::getTrunc(C, DestTy);
462 Value *NarrowOp0 = Builder->CreateTrunc(LogicOp->getOperand(0), DestTy);
463 return BinaryOperator::Create(LogicOp->getOpcode(), NarrowOp0, NarrowC);
466 /// Try to narrow the width of a splat shuffle. This could be generalized to any
467 /// shuffle with a constant operand, but we limit the transform to avoid
468 /// creating a shuffle type that targets may not be able to lower effectively.
469 static Instruction *shrinkSplatShuffle(TruncInst &Trunc,
470 InstCombiner::BuilderTy &Builder) {
471 auto *Shuf = dyn_cast<ShuffleVectorInst>(Trunc.getOperand(0));
472 if (Shuf && Shuf->hasOneUse() && isa<UndefValue>(Shuf->getOperand(1)) &&
473 Shuf->getMask()->getSplatValue() &&
474 Shuf->getType() == Shuf->getOperand(0)->getType()) {
475 // trunc (shuf X, Undef, SplatMask) --> shuf (trunc X), Undef, SplatMask
476 Constant *NarrowUndef = UndefValue::get(Trunc.getType());
477 Value *NarrowOp = Builder.CreateTrunc(Shuf->getOperand(0), Trunc.getType());
478 return new ShuffleVectorInst(NarrowOp, NarrowUndef, Shuf->getMask());
484 /// Try to narrow the width of an insert element. This could be generalized for
485 /// any vector constant, but we limit the transform to insertion into undef to
486 /// avoid potential backend problems from unsupported insertion widths. This
487 /// could also be extended to handle the case of inserting a scalar constant
488 /// into a vector variable.
489 static Instruction *shrinkInsertElt(CastInst &Trunc,
490 InstCombiner::BuilderTy &Builder) {
491 Instruction::CastOps Opcode = Trunc.getOpcode();
492 assert((Opcode == Instruction::Trunc || Opcode == Instruction::FPTrunc) &&
493 "Unexpected instruction for shrinking");
495 auto *InsElt = dyn_cast<InsertElementInst>(Trunc.getOperand(0));
496 if (!InsElt || !InsElt->hasOneUse())
499 Type *DestTy = Trunc.getType();
500 Type *DestScalarTy = DestTy->getScalarType();
501 Value *VecOp = InsElt->getOperand(0);
502 Value *ScalarOp = InsElt->getOperand(1);
503 Value *Index = InsElt->getOperand(2);
505 if (isa<UndefValue>(VecOp)) {
506 // trunc (inselt undef, X, Index) --> inselt undef, (trunc X), Index
507 // fptrunc (inselt undef, X, Index) --> inselt undef, (fptrunc X), Index
508 UndefValue *NarrowUndef = UndefValue::get(DestTy);
509 Value *NarrowOp = Builder.CreateCast(Opcode, ScalarOp, DestScalarTy);
510 return InsertElementInst::Create(NarrowUndef, NarrowOp, Index);
516 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
517 if (Instruction *Result = commonCastTransforms(CI))
520 // Test if the trunc is the user of a select which is part of a
521 // minimum or maximum operation. If so, don't do any more simplification.
522 // Even simplifying demanded bits can break the canonical form of a
525 if (SelectInst *SI = dyn_cast<SelectInst>(CI.getOperand(0)))
526 if (matchSelectPattern(SI, LHS, RHS).Flavor != SPF_UNKNOWN)
529 // See if we can simplify any instructions used by the input whose sole
530 // purpose is to compute bits we don't care about.
531 if (SimplifyDemandedInstructionBits(CI))
534 Value *Src = CI.getOperand(0);
535 Type *DestTy = CI.getType(), *SrcTy = Src->getType();
537 // Attempt to truncate the entire input expression tree to the destination
538 // type. Only do this if the dest type is a simple type, don't convert the
539 // expression tree to something weird like i93 unless the source is also
541 if ((DestTy->isVectorTy() || shouldChangeType(SrcTy, DestTy)) &&
542 canEvaluateTruncated(Src, DestTy, *this, &CI)) {
544 // If this cast is a truncate, evaluting in a different type always
545 // eliminates the cast, so it is always a win.
546 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
547 " to avoid cast: " << CI << '\n');
548 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
549 assert(Res->getType() == DestTy);
550 return replaceInstUsesWith(CI, Res);
553 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector.
554 if (DestTy->getScalarSizeInBits() == 1) {
555 Constant *One = ConstantInt::get(SrcTy, 1);
556 Src = Builder->CreateAnd(Src, One);
557 Value *Zero = Constant::getNullValue(Src->getType());
558 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
561 // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion.
562 Value *A = nullptr; ConstantInt *Cst = nullptr;
563 if (Src->hasOneUse() &&
564 match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst)))) {
565 // We have three types to worry about here, the type of A, the source of
566 // the truncate (MidSize), and the destination of the truncate. We know that
567 // ASize < MidSize and MidSize > ResultSize, but don't know the relation
568 // between ASize and ResultSize.
569 unsigned ASize = A->getType()->getPrimitiveSizeInBits();
571 // If the shift amount is larger than the size of A, then the result is
572 // known to be zero because all the input bits got shifted out.
573 if (Cst->getZExtValue() >= ASize)
574 return replaceInstUsesWith(CI, Constant::getNullValue(DestTy));
576 // Since we're doing an lshr and a zero extend, and know that the shift
577 // amount is smaller than ASize, it is always safe to do the shift in A's
578 // type, then zero extend or truncate to the result.
579 Value *Shift = Builder->CreateLShr(A, Cst->getZExtValue());
580 Shift->takeName(Src);
581 return CastInst::CreateIntegerCast(Shift, DestTy, false);
584 // Transform trunc(lshr (sext A), Cst) to ashr A, Cst to eliminate type
586 // It works because bits coming from sign extension have the same value as
587 // the sign bit of the original value; performing ashr instead of lshr
588 // generates bits of the same value as the sign bit.
589 if (Src->hasOneUse() &&
590 match(Src, m_LShr(m_SExt(m_Value(A)), m_ConstantInt(Cst))) &&
591 cast<Instruction>(Src)->getOperand(0)->hasOneUse()) {
592 const unsigned ASize = A->getType()->getPrimitiveSizeInBits();
593 // This optimization can be only performed when zero bits generated by
594 // the original lshr aren't pulled into the value after truncation, so we
595 // can only shift by values smaller than the size of destination type (in
597 if (Cst->getValue().ult(ASize)) {
598 Value *Shift = Builder->CreateAShr(A, Cst->getZExtValue());
599 Shift->takeName(Src);
600 return CastInst::CreateIntegerCast(Shift, CI.getType(), true);
604 if (Instruction *I = shrinkBitwiseLogic(CI))
607 if (Instruction *I = shrinkSplatShuffle(CI, *Builder))
610 if (Instruction *I = shrinkInsertElt(CI, *Builder))
613 if (Src->hasOneUse() && isa<IntegerType>(SrcTy) &&
614 shouldChangeType(SrcTy, DestTy)) {
615 // Transform "trunc (shl X, cst)" -> "shl (trunc X), cst" so long as the
616 // dest type is native and cst < dest size.
617 if (match(Src, m_Shl(m_Value(A), m_ConstantInt(Cst))) &&
618 !match(A, m_Shr(m_Value(), m_Constant()))) {
619 // Skip shifts of shift by constants. It undoes a combine in
620 // FoldShiftByConstant and is the extend in reg pattern.
621 const unsigned DestSize = DestTy->getScalarSizeInBits();
622 if (Cst->getValue().ult(DestSize)) {
623 Value *NewTrunc = Builder->CreateTrunc(A, DestTy, A->getName() + ".tr");
625 return BinaryOperator::Create(
626 Instruction::Shl, NewTrunc,
627 ConstantInt::get(DestTy, Cst->getValue().trunc(DestSize)));
632 if (Instruction *I = foldVecTruncToExtElt(CI, *this, DL))
638 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, ZExtInst &CI,
640 // If we are just checking for a icmp eq of a single bit and zext'ing it
641 // to an integer, then shift the bit to the appropriate place and then
642 // cast to integer to avoid the comparison.
643 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
644 const APInt &Op1CV = Op1C->getValue();
646 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
647 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
648 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
649 (ICI->getPredicate() == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
650 if (!DoTransform) return ICI;
652 Value *In = ICI->getOperand(0);
653 Value *Sh = ConstantInt::get(In->getType(),
654 In->getType()->getScalarSizeInBits() - 1);
655 In = Builder->CreateLShr(In, Sh, In->getName() + ".lobit");
656 if (In->getType() != CI.getType())
657 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/);
659 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
660 Constant *One = ConstantInt::get(In->getType(), 1);
661 In = Builder->CreateXor(In, One, In->getName() + ".not");
664 return replaceInstUsesWith(CI, In);
667 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
668 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
669 // zext (X == 1) to i32 --> X iff X has only the low bit set.
670 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
671 // zext (X != 0) to i32 --> X iff X has only the low bit set.
672 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
673 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
674 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
675 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
676 // This only works for EQ and NE
678 // If Op1C some other power of two, convert:
679 uint32_t BitWidth = Op1C->getType()->getBitWidth();
680 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
681 computeKnownBits(ICI->getOperand(0), KnownZero, KnownOne, 0, &CI);
683 APInt KnownZeroMask(~KnownZero);
684 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
685 if (!DoTransform) return ICI;
687 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
688 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
689 // (X&4) == 2 --> false
690 // (X&4) != 2 --> true
691 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()),
693 Res = ConstantExpr::getZExt(Res, CI.getType());
694 return replaceInstUsesWith(CI, Res);
697 uint32_t ShAmt = KnownZeroMask.logBase2();
698 Value *In = ICI->getOperand(0);
700 // Perform a logical shr by shiftamt.
701 // Insert the shift to put the result in the low bit.
702 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(), ShAmt),
703 In->getName() + ".lobit");
706 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
707 Constant *One = ConstantInt::get(In->getType(), 1);
708 In = Builder->CreateXor(In, One);
711 if (CI.getType() == In->getType())
712 return replaceInstUsesWith(CI, In);
714 Value *IntCast = Builder->CreateIntCast(In, CI.getType(), false);
715 return replaceInstUsesWith(CI, IntCast);
720 // icmp ne A, B is equal to xor A, B when A and B only really have one bit.
721 // It is also profitable to transform icmp eq into not(xor(A, B)) because that
722 // may lead to additional simplifications.
723 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) {
724 if (IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) {
725 uint32_t BitWidth = ITy->getBitWidth();
726 Value *LHS = ICI->getOperand(0);
727 Value *RHS = ICI->getOperand(1);
729 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0);
730 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0);
731 computeKnownBits(LHS, KnownZeroLHS, KnownOneLHS, 0, &CI);
732 computeKnownBits(RHS, KnownZeroRHS, KnownOneRHS, 0, &CI);
734 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) {
735 APInt KnownBits = KnownZeroLHS | KnownOneLHS;
736 APInt UnknownBit = ~KnownBits;
737 if (UnknownBit.countPopulation() == 1) {
738 if (!DoTransform) return ICI;
740 Value *Result = Builder->CreateXor(LHS, RHS);
742 // Mask off any bits that are set and won't be shifted away.
743 if (KnownOneLHS.uge(UnknownBit))
744 Result = Builder->CreateAnd(Result,
745 ConstantInt::get(ITy, UnknownBit));
747 // Shift the bit we're testing down to the lsb.
748 Result = Builder->CreateLShr(
749 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros()));
751 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
752 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1));
753 Result->takeName(ICI);
754 return replaceInstUsesWith(CI, Result);
763 /// Determine if the specified value can be computed in the specified wider type
764 /// and produce the same low bits. If not, return false.
766 /// If this function returns true, it can also return a non-zero number of bits
767 /// (in BitsToClear) which indicates that the value it computes is correct for
768 /// the zero extend, but that the additional BitsToClear bits need to be zero'd
769 /// out. For example, to promote something like:
771 /// %B = trunc i64 %A to i32
772 /// %C = lshr i32 %B, 8
773 /// %E = zext i32 %C to i64
775 /// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be
776 /// set to 8 to indicate that the promoted value needs to have bits 24-31
777 /// cleared in addition to bits 32-63. Since an 'and' will be generated to
778 /// clear the top bits anyway, doing this has no extra cost.
780 /// This function works on both vectors and scalars.
781 static bool canEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear,
782 InstCombiner &IC, Instruction *CxtI) {
784 if (isa<Constant>(V))
787 Instruction *I = dyn_cast<Instruction>(V);
788 if (!I) return false;
790 // If the input is a truncate from the destination type, we can trivially
792 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
795 // We can't extend or shrink something that has multiple uses: doing so would
796 // require duplicating the instruction in general, which isn't profitable.
797 if (!I->hasOneUse()) return false;
799 unsigned Opc = I->getOpcode(), Tmp;
801 case Instruction::ZExt: // zext(zext(x)) -> zext(x).
802 case Instruction::SExt: // zext(sext(x)) -> sext(x).
803 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x)
805 case Instruction::And:
806 case Instruction::Or:
807 case Instruction::Xor:
808 case Instruction::Add:
809 case Instruction::Sub:
810 case Instruction::Mul:
811 if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI) ||
812 !canEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI))
814 // These can all be promoted if neither operand has 'bits to clear'.
815 if (BitsToClear == 0 && Tmp == 0)
818 // If the operation is an AND/OR/XOR and the bits to clear are zero in the
819 // other side, BitsToClear is ok.
820 if (Tmp == 0 && I->isBitwiseLogicOp()) {
821 // We use MaskedValueIsZero here for generality, but the case we care
822 // about the most is constant RHS.
823 unsigned VSize = V->getType()->getScalarSizeInBits();
824 if (IC.MaskedValueIsZero(I->getOperand(1),
825 APInt::getHighBitsSet(VSize, BitsToClear),
830 // Otherwise, we don't know how to analyze this BitsToClear case yet.
833 case Instruction::Shl:
834 // We can promote shl(x, cst) if we can promote x. Since shl overwrites the
835 // upper bits we can reduce BitsToClear by the shift amount.
836 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
837 if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI))
839 uint64_t ShiftAmt = Amt->getZExtValue();
840 BitsToClear = ShiftAmt < BitsToClear ? BitsToClear - ShiftAmt : 0;
844 case Instruction::LShr:
845 // We can promote lshr(x, cst) if we can promote x. This requires the
846 // ultimate 'and' to clear out the high zero bits we're clearing out though.
847 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
848 if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI))
850 BitsToClear += Amt->getZExtValue();
851 if (BitsToClear > V->getType()->getScalarSizeInBits())
852 BitsToClear = V->getType()->getScalarSizeInBits();
855 // Cannot promote variable LSHR.
857 case Instruction::Select:
858 if (!canEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI) ||
859 !canEvaluateZExtd(I->getOperand(2), Ty, BitsToClear, IC, CxtI) ||
860 // TODO: If important, we could handle the case when the BitsToClear are
861 // known zero in the disagreeing side.
866 case Instruction::PHI: {
867 // We can change a phi if we can change all operands. Note that we never
868 // get into trouble with cyclic PHIs here because we only consider
869 // instructions with a single use.
870 PHINode *PN = cast<PHINode>(I);
871 if (!canEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear, IC, CxtI))
873 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
874 if (!canEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp, IC, CxtI) ||
875 // TODO: If important, we could handle the case when the BitsToClear
876 // are known zero in the disagreeing input.
882 // TODO: Can handle more cases here.
887 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
888 // If this zero extend is only used by a truncate, let the truncate be
889 // eliminated before we try to optimize this zext.
890 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back()))
893 // If one of the common conversion will work, do it.
894 if (Instruction *Result = commonCastTransforms(CI))
897 Value *Src = CI.getOperand(0);
898 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
900 // Attempt to extend the entire input expression tree to the destination
901 // type. Only do this if the dest type is a simple type, don't convert the
902 // expression tree to something weird like i93 unless the source is also
904 unsigned BitsToClear;
905 if ((DestTy->isVectorTy() || shouldChangeType(SrcTy, DestTy)) &&
906 canEvaluateZExtd(Src, DestTy, BitsToClear, *this, &CI)) {
907 assert(BitsToClear <= SrcTy->getScalarSizeInBits() &&
908 "Can't clear more bits than in SrcTy");
910 // Okay, we can transform this! Insert the new expression now.
911 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
912 " to avoid zero extend: " << CI << '\n');
913 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
914 assert(Res->getType() == DestTy);
916 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear;
917 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
919 // If the high bits are already filled with zeros, just replace this
920 // cast with the result.
921 if (MaskedValueIsZero(Res,
922 APInt::getHighBitsSet(DestBitSize,
923 DestBitSize-SrcBitsKept),
925 return replaceInstUsesWith(CI, Res);
927 // We need to emit an AND to clear the high bits.
928 Constant *C = ConstantInt::get(Res->getType(),
929 APInt::getLowBitsSet(DestBitSize, SrcBitsKept));
930 return BinaryOperator::CreateAnd(Res, C);
933 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
934 // types and if the sizes are just right we can convert this into a logical
935 // 'and' which will be much cheaper than the pair of casts.
936 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
937 // TODO: Subsume this into EvaluateInDifferentType.
939 // Get the sizes of the types involved. We know that the intermediate type
940 // will be smaller than A or C, but don't know the relation between A and C.
941 Value *A = CSrc->getOperand(0);
942 unsigned SrcSize = A->getType()->getScalarSizeInBits();
943 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
944 unsigned DstSize = CI.getType()->getScalarSizeInBits();
945 // If we're actually extending zero bits, then if
946 // SrcSize < DstSize: zext(a & mask)
947 // SrcSize == DstSize: a & mask
948 // SrcSize > DstSize: trunc(a) & mask
949 if (SrcSize < DstSize) {
950 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
951 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
952 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
953 return new ZExtInst(And, CI.getType());
956 if (SrcSize == DstSize) {
957 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
958 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
961 if (SrcSize > DstSize) {
962 Value *Trunc = Builder->CreateTrunc(A, CI.getType());
963 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
964 return BinaryOperator::CreateAnd(Trunc,
965 ConstantInt::get(Trunc->getType(),
970 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
971 return transformZExtICmp(ICI, CI);
973 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
974 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
975 // zext (or icmp, icmp) -> or (zext icmp), (zext icmp) if at least one
976 // of the (zext icmp) can be eliminated. If so, immediately perform the
977 // according elimination.
978 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
979 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
980 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
981 (transformZExtICmp(LHS, CI, false) ||
982 transformZExtICmp(RHS, CI, false))) {
983 // zext (or icmp, icmp) -> or (zext icmp), (zext icmp)
984 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());
985 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());
986 BinaryOperator *Or = BinaryOperator::Create(Instruction::Or, LCast, RCast);
988 // Perform the elimination.
989 if (auto *LZExt = dyn_cast<ZExtInst>(LCast))
990 transformZExtICmp(LHS, *LZExt);
991 if (auto *RZExt = dyn_cast<ZExtInst>(RCast))
992 transformZExtICmp(RHS, *RZExt);
998 // zext(trunc(X) & C) -> (X & zext(C)).
1002 match(SrcI, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Constant(C)))) &&
1003 X->getType() == CI.getType())
1004 return BinaryOperator::CreateAnd(X, ConstantExpr::getZExt(C, CI.getType()));
1006 // zext((trunc(X) & C) ^ C) -> ((X & zext(C)) ^ zext(C)).
1008 if (SrcI && match(SrcI, m_OneUse(m_Xor(m_Value(And), m_Constant(C)))) &&
1009 match(And, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Specific(C)))) &&
1010 X->getType() == CI.getType()) {
1011 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
1012 return BinaryOperator::CreateXor(Builder->CreateAnd(X, ZC), ZC);
1018 /// Transform (sext icmp) to bitwise / integer operations to eliminate the icmp.
1019 Instruction *InstCombiner::transformSExtICmp(ICmpInst *ICI, Instruction &CI) {
1020 Value *Op0 = ICI->getOperand(0), *Op1 = ICI->getOperand(1);
1021 ICmpInst::Predicate Pred = ICI->getPredicate();
1023 // Don't bother if Op1 isn't of vector or integer type.
1024 if (!Op1->getType()->isIntOrIntVectorTy())
1027 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
1028 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if negative
1029 // (x >s -1) ? -1 : 0 -> not (ashr x, 31) -> all ones if positive
1030 if ((Pred == ICmpInst::ICMP_SLT && Op1C->isNullValue()) ||
1031 (Pred == ICmpInst::ICMP_SGT && Op1C->isAllOnesValue())) {
1033 Value *Sh = ConstantInt::get(Op0->getType(),
1034 Op0->getType()->getScalarSizeInBits()-1);
1035 Value *In = Builder->CreateAShr(Op0, Sh, Op0->getName()+".lobit");
1036 if (In->getType() != CI.getType())
1037 In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/);
1039 if (Pred == ICmpInst::ICMP_SGT)
1040 In = Builder->CreateNot(In, In->getName()+".not");
1041 return replaceInstUsesWith(CI, In);
1045 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
1046 // If we know that only one bit of the LHS of the icmp can be set and we
1047 // have an equality comparison with zero or a power of 2, we can transform
1048 // the icmp and sext into bitwise/integer operations.
1049 if (ICI->hasOneUse() &&
1050 ICI->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){
1051 unsigned BitWidth = Op1C->getType()->getBitWidth();
1052 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
1053 computeKnownBits(Op0, KnownZero, KnownOne, 0, &CI);
1055 APInt KnownZeroMask(~KnownZero);
1056 if (KnownZeroMask.isPowerOf2()) {
1057 Value *In = ICI->getOperand(0);
1059 // If the icmp tests for a known zero bit we can constant fold it.
1060 if (!Op1C->isZero() && Op1C->getValue() != KnownZeroMask) {
1061 Value *V = Pred == ICmpInst::ICMP_NE ?
1062 ConstantInt::getAllOnesValue(CI.getType()) :
1063 ConstantInt::getNullValue(CI.getType());
1064 return replaceInstUsesWith(CI, V);
1067 if (!Op1C->isZero() == (Pred == ICmpInst::ICMP_NE)) {
1068 // sext ((x & 2^n) == 0) -> (x >> n) - 1
1069 // sext ((x & 2^n) != 2^n) -> (x >> n) - 1
1070 unsigned ShiftAmt = KnownZeroMask.countTrailingZeros();
1071 // Perform a right shift to place the desired bit in the LSB.
1073 In = Builder->CreateLShr(In,
1074 ConstantInt::get(In->getType(), ShiftAmt));
1076 // At this point "In" is either 1 or 0. Subtract 1 to turn
1077 // {1, 0} -> {0, -1}.
1078 In = Builder->CreateAdd(In,
1079 ConstantInt::getAllOnesValue(In->getType()),
1082 // sext ((x & 2^n) != 0) -> (x << bitwidth-n) a>> bitwidth-1
1083 // sext ((x & 2^n) == 2^n) -> (x << bitwidth-n) a>> bitwidth-1
1084 unsigned ShiftAmt = KnownZeroMask.countLeadingZeros();
1085 // Perform a left shift to place the desired bit in the MSB.
1087 In = Builder->CreateShl(In,
1088 ConstantInt::get(In->getType(), ShiftAmt));
1090 // Distribute the bit over the whole bit width.
1091 In = Builder->CreateAShr(In, ConstantInt::get(In->getType(),
1092 BitWidth - 1), "sext");
1095 if (CI.getType() == In->getType())
1096 return replaceInstUsesWith(CI, In);
1097 return CastInst::CreateIntegerCast(In, CI.getType(), true/*SExt*/);
1105 /// Return true if we can take the specified value and return it as type Ty
1106 /// without inserting any new casts and without changing the value of the common
1107 /// low bits. This is used by code that tries to promote integer operations to
1108 /// a wider types will allow us to eliminate the extension.
1110 /// This function works on both vectors and scalars.
1112 static bool canEvaluateSExtd(Value *V, Type *Ty) {
1113 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() &&
1114 "Can't sign extend type to a smaller type");
1115 // If this is a constant, it can be trivially promoted.
1116 if (isa<Constant>(V))
1119 Instruction *I = dyn_cast<Instruction>(V);
1120 if (!I) return false;
1122 // If this is a truncate from the dest type, we can trivially eliminate it.
1123 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
1126 // We can't extend or shrink something that has multiple uses: doing so would
1127 // require duplicating the instruction in general, which isn't profitable.
1128 if (!I->hasOneUse()) return false;
1130 switch (I->getOpcode()) {
1131 case Instruction::SExt: // sext(sext(x)) -> sext(x)
1132 case Instruction::ZExt: // sext(zext(x)) -> zext(x)
1133 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x)
1135 case Instruction::And:
1136 case Instruction::Or:
1137 case Instruction::Xor:
1138 case Instruction::Add:
1139 case Instruction::Sub:
1140 case Instruction::Mul:
1141 // These operators can all arbitrarily be extended if their inputs can.
1142 return canEvaluateSExtd(I->getOperand(0), Ty) &&
1143 canEvaluateSExtd(I->getOperand(1), Ty);
1145 //case Instruction::Shl: TODO
1146 //case Instruction::LShr: TODO
1148 case Instruction::Select:
1149 return canEvaluateSExtd(I->getOperand(1), Ty) &&
1150 canEvaluateSExtd(I->getOperand(2), Ty);
1152 case Instruction::PHI: {
1153 // We can change a phi if we can change all operands. Note that we never
1154 // get into trouble with cyclic PHIs here because we only consider
1155 // instructions with a single use.
1156 PHINode *PN = cast<PHINode>(I);
1157 for (Value *IncValue : PN->incoming_values())
1158 if (!canEvaluateSExtd(IncValue, Ty)) return false;
1162 // TODO: Can handle more cases here.
1169 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
1170 // If this sign extend is only used by a truncate, let the truncate be
1171 // eliminated before we try to optimize this sext.
1172 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back()))
1175 if (Instruction *I = commonCastTransforms(CI))
1178 Value *Src = CI.getOperand(0);
1179 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
1181 // If we know that the value being extended is positive, we can use a zext
1183 bool KnownZero, KnownOne;
1184 ComputeSignBit(Src, KnownZero, KnownOne, 0, &CI);
1186 Value *ZExt = Builder->CreateZExt(Src, DestTy);
1187 return replaceInstUsesWith(CI, ZExt);
1190 // Attempt to extend the entire input expression tree to the destination
1191 // type. Only do this if the dest type is a simple type, don't convert the
1192 // expression tree to something weird like i93 unless the source is also
1194 if ((DestTy->isVectorTy() || shouldChangeType(SrcTy, DestTy)) &&
1195 canEvaluateSExtd(Src, DestTy)) {
1196 // Okay, we can transform this! Insert the new expression now.
1197 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
1198 " to avoid sign extend: " << CI << '\n');
1199 Value *Res = EvaluateInDifferentType(Src, DestTy, true);
1200 assert(Res->getType() == DestTy);
1202 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1203 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1205 // If the high bits are already filled with sign bit, just replace this
1206 // cast with the result.
1207 if (ComputeNumSignBits(Res, 0, &CI) > DestBitSize - SrcBitSize)
1208 return replaceInstUsesWith(CI, Res);
1210 // We need to emit a shl + ashr to do the sign extend.
1211 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1212 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"),
1216 // If the input is a trunc from the destination type, then turn sext(trunc(x))
1219 if (match(Src, m_OneUse(m_Trunc(m_Value(X)))) && X->getType() == DestTy) {
1220 // sext(trunc(X)) --> ashr(shl(X, C), C)
1221 unsigned SrcBitSize = SrcTy->getScalarSizeInBits();
1222 unsigned DestBitSize = DestTy->getScalarSizeInBits();
1223 Constant *ShAmt = ConstantInt::get(DestTy, DestBitSize - SrcBitSize);
1224 return BinaryOperator::CreateAShr(Builder->CreateShl(X, ShAmt), ShAmt);
1227 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
1228 return transformSExtICmp(ICI, CI);
1230 // If the input is a shl/ashr pair of a same constant, then this is a sign
1231 // extension from a smaller value. If we could trust arbitrary bitwidth
1232 // integers, we could turn this into a truncate to the smaller bit and then
1233 // use a sext for the whole extension. Since we don't, look deeper and check
1234 // for a truncate. If the source and dest are the same type, eliminate the
1235 // trunc and extend and just do shifts. For example, turn:
1236 // %a = trunc i32 %i to i8
1237 // %b = shl i8 %a, 6
1238 // %c = ashr i8 %b, 6
1239 // %d = sext i8 %c to i32
1241 // %a = shl i32 %i, 30
1242 // %d = ashr i32 %a, 30
1244 // TODO: Eventually this could be subsumed by EvaluateInDifferentType.
1245 ConstantInt *BA = nullptr, *CA = nullptr;
1246 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)),
1247 m_ConstantInt(CA))) &&
1248 BA == CA && A->getType() == CI.getType()) {
1249 unsigned MidSize = Src->getType()->getScalarSizeInBits();
1250 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
1251 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
1252 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
1253 A = Builder->CreateShl(A, ShAmtV, CI.getName());
1254 return BinaryOperator::CreateAShr(A, ShAmtV);
1261 /// Return a Constant* for the specified floating-point constant if it fits
1262 /// in the specified FP type without changing its value.
1263 static Constant *fitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
1265 APFloat F = CFP->getValueAPF();
1266 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
1268 return ConstantFP::get(CFP->getContext(), F);
1272 /// Look through floating-point extensions until we get the source value.
1273 static Value *lookThroughFPExtensions(Value *V) {
1274 while (auto *FPExt = dyn_cast<FPExtInst>(V))
1275 V = FPExt->getOperand(0);
1277 // If this value is a constant, return the constant in the smallest FP type
1278 // that can accurately represent it. This allows us to turn
1279 // (float)((double)X+2.0) into x+2.0f.
1280 if (auto *CFP = dyn_cast<ConstantFP>(V)) {
1281 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext()))
1282 return V; // No constant folding of this.
1283 // See if the value can be truncated to half and then reextended.
1284 if (Value *V = fitsInFPType(CFP, APFloat::IEEEhalf()))
1286 // See if the value can be truncated to float and then reextended.
1287 if (Value *V = fitsInFPType(CFP, APFloat::IEEEsingle()))
1289 if (CFP->getType()->isDoubleTy())
1290 return V; // Won't shrink.
1291 if (Value *V = fitsInFPType(CFP, APFloat::IEEEdouble()))
1293 // Don't try to shrink to various long double types.
1299 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
1300 if (Instruction *I = commonCastTransforms(CI))
1302 // If we have fptrunc(OpI (fpextend x), (fpextend y)), we would like to
1303 // simplify this expression to avoid one or more of the trunc/extend
1304 // operations if we can do so without changing the numerical results.
1306 // The exact manner in which the widths of the operands interact to limit
1307 // what we can and cannot do safely varies from operation to operation, and
1308 // is explained below in the various case statements.
1309 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
1310 if (OpI && OpI->hasOneUse()) {
1311 Value *LHSOrig = lookThroughFPExtensions(OpI->getOperand(0));
1312 Value *RHSOrig = lookThroughFPExtensions(OpI->getOperand(1));
1313 unsigned OpWidth = OpI->getType()->getFPMantissaWidth();
1314 unsigned LHSWidth = LHSOrig->getType()->getFPMantissaWidth();
1315 unsigned RHSWidth = RHSOrig->getType()->getFPMantissaWidth();
1316 unsigned SrcWidth = std::max(LHSWidth, RHSWidth);
1317 unsigned DstWidth = CI.getType()->getFPMantissaWidth();
1318 switch (OpI->getOpcode()) {
1320 case Instruction::FAdd:
1321 case Instruction::FSub:
1322 // For addition and subtraction, the infinitely precise result can
1323 // essentially be arbitrarily wide; proving that double rounding
1324 // will not occur because the result of OpI is exact (as we will for
1325 // FMul, for example) is hopeless. However, we *can* nonetheless
1326 // frequently know that double rounding cannot occur (or that it is
1327 // innocuous) by taking advantage of the specific structure of
1328 // infinitely-precise results that admit double rounding.
1330 // Specifically, if OpWidth >= 2*DstWdith+1 and DstWidth is sufficient
1331 // to represent both sources, we can guarantee that the double
1332 // rounding is innocuous (See p50 of Figueroa's 2000 PhD thesis,
1333 // "A Rigorous Framework for Fully Supporting the IEEE Standard ..."
1334 // for proof of this fact).
1336 // Note: Figueroa does not consider the case where DstFormat !=
1337 // SrcFormat. It's possible (likely even!) that this analysis
1338 // could be tightened for those cases, but they are rare (the main
1339 // case of interest here is (float)((double)float + float)).
1340 if (OpWidth >= 2*DstWidth+1 && DstWidth >= SrcWidth) {
1341 if (LHSOrig->getType() != CI.getType())
1342 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1343 if (RHSOrig->getType() != CI.getType())
1344 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1346 BinaryOperator::Create(OpI->getOpcode(), LHSOrig, RHSOrig);
1347 RI->copyFastMathFlags(OpI);
1351 case Instruction::FMul:
1352 // For multiplication, the infinitely precise result has at most
1353 // LHSWidth + RHSWidth significant bits; if OpWidth is sufficient
1354 // that such a value can be exactly represented, then no double
1355 // rounding can possibly occur; we can safely perform the operation
1356 // in the destination format if it can represent both sources.
1357 if (OpWidth >= LHSWidth + RHSWidth && DstWidth >= SrcWidth) {
1358 if (LHSOrig->getType() != CI.getType())
1359 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1360 if (RHSOrig->getType() != CI.getType())
1361 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1363 BinaryOperator::CreateFMul(LHSOrig, RHSOrig);
1364 RI->copyFastMathFlags(OpI);
1368 case Instruction::FDiv:
1369 // For division, we use again use the bound from Figueroa's
1370 // dissertation. I am entirely certain that this bound can be
1371 // tightened in the unbalanced operand case by an analysis based on
1372 // the diophantine rational approximation bound, but the well-known
1373 // condition used here is a good conservative first pass.
1374 // TODO: Tighten bound via rigorous analysis of the unbalanced case.
1375 if (OpWidth >= 2*DstWidth && DstWidth >= SrcWidth) {
1376 if (LHSOrig->getType() != CI.getType())
1377 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1378 if (RHSOrig->getType() != CI.getType())
1379 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1381 BinaryOperator::CreateFDiv(LHSOrig, RHSOrig);
1382 RI->copyFastMathFlags(OpI);
1386 case Instruction::FRem:
1387 // Remainder is straightforward. Remainder is always exact, so the
1388 // type of OpI doesn't enter into things at all. We simply evaluate
1389 // in whichever source type is larger, then convert to the
1390 // destination type.
1391 if (SrcWidth == OpWidth)
1393 if (LHSWidth < SrcWidth)
1394 LHSOrig = Builder->CreateFPExt(LHSOrig, RHSOrig->getType());
1395 else if (RHSWidth <= SrcWidth)
1396 RHSOrig = Builder->CreateFPExt(RHSOrig, LHSOrig->getType());
1397 if (LHSOrig != OpI->getOperand(0) || RHSOrig != OpI->getOperand(1)) {
1398 Value *ExactResult = Builder->CreateFRem(LHSOrig, RHSOrig);
1399 if (Instruction *RI = dyn_cast<Instruction>(ExactResult))
1400 RI->copyFastMathFlags(OpI);
1401 return CastInst::CreateFPCast(ExactResult, CI.getType());
1405 // (fptrunc (fneg x)) -> (fneg (fptrunc x))
1406 if (BinaryOperator::isFNeg(OpI)) {
1407 Value *InnerTrunc = Builder->CreateFPTrunc(OpI->getOperand(1),
1409 Instruction *RI = BinaryOperator::CreateFNeg(InnerTrunc);
1410 RI->copyFastMathFlags(OpI);
1415 // (fptrunc (select cond, R1, Cst)) -->
1416 // (select cond, (fptrunc R1), (fptrunc Cst))
1418 // - but only if this isn't part of a min/max operation, else we'll
1419 // ruin min/max canonical form which is to have the select and
1420 // compare's operands be of the same type with no casts to look through.
1422 SelectInst *SI = dyn_cast<SelectInst>(CI.getOperand(0));
1424 (isa<ConstantFP>(SI->getOperand(1)) ||
1425 isa<ConstantFP>(SI->getOperand(2))) &&
1426 matchSelectPattern(SI, LHS, RHS).Flavor == SPF_UNKNOWN) {
1427 Value *LHSTrunc = Builder->CreateFPTrunc(SI->getOperand(1),
1429 Value *RHSTrunc = Builder->CreateFPTrunc(SI->getOperand(2),
1431 return SelectInst::Create(SI->getOperand(0), LHSTrunc, RHSTrunc);
1434 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI.getOperand(0));
1436 switch (II->getIntrinsicID()) {
1438 case Intrinsic::fabs:
1439 case Intrinsic::ceil:
1440 case Intrinsic::floor:
1441 case Intrinsic::rint:
1442 case Intrinsic::round:
1443 case Intrinsic::nearbyint:
1444 case Intrinsic::trunc: {
1445 Value *Src = II->getArgOperand(0);
1446 if (!Src->hasOneUse())
1449 // Except for fabs, this transformation requires the input of the unary FP
1450 // operation to be itself an fpext from the type to which we're
1452 if (II->getIntrinsicID() != Intrinsic::fabs) {
1453 FPExtInst *FPExtSrc = dyn_cast<FPExtInst>(Src);
1454 if (!FPExtSrc || FPExtSrc->getOperand(0)->getType() != CI.getType())
1458 // Do unary FP operation on smaller type.
1459 // (fptrunc (fabs x)) -> (fabs (fptrunc x))
1460 Value *InnerTrunc = Builder->CreateFPTrunc(Src, CI.getType());
1461 Type *IntrinsicType[] = { CI.getType() };
1462 Function *Overload = Intrinsic::getDeclaration(
1463 CI.getModule(), II->getIntrinsicID(), IntrinsicType);
1465 SmallVector<OperandBundleDef, 1> OpBundles;
1466 II->getOperandBundlesAsDefs(OpBundles);
1468 Value *Args[] = { InnerTrunc };
1469 CallInst *NewCI = CallInst::Create(Overload, Args,
1470 OpBundles, II->getName());
1471 NewCI->copyFastMathFlags(II);
1477 if (Instruction *I = shrinkInsertElt(CI, *Builder))
1483 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
1484 return commonCastTransforms(CI);
1487 // fpto{s/u}i({u/s}itofp(X)) --> X or zext(X) or sext(X) or trunc(X)
1488 // This is safe if the intermediate type has enough bits in its mantissa to
1489 // accurately represent all values of X. For example, this won't work with
1490 // i64 -> float -> i64.
1491 Instruction *InstCombiner::FoldItoFPtoI(Instruction &FI) {
1492 if (!isa<UIToFPInst>(FI.getOperand(0)) && !isa<SIToFPInst>(FI.getOperand(0)))
1494 Instruction *OpI = cast<Instruction>(FI.getOperand(0));
1496 Value *SrcI = OpI->getOperand(0);
1497 Type *FITy = FI.getType();
1498 Type *OpITy = OpI->getType();
1499 Type *SrcTy = SrcI->getType();
1500 bool IsInputSigned = isa<SIToFPInst>(OpI);
1501 bool IsOutputSigned = isa<FPToSIInst>(FI);
1503 // We can safely assume the conversion won't overflow the output range,
1504 // because (for example) (uint8_t)18293.f is undefined behavior.
1506 // Since we can assume the conversion won't overflow, our decision as to
1507 // whether the input will fit in the float should depend on the minimum
1508 // of the input range and output range.
1510 // This means this is also safe for a signed input and unsigned output, since
1511 // a negative input would lead to undefined behavior.
1512 int InputSize = (int)SrcTy->getScalarSizeInBits() - IsInputSigned;
1513 int OutputSize = (int)FITy->getScalarSizeInBits() - IsOutputSigned;
1514 int ActualSize = std::min(InputSize, OutputSize);
1516 if (ActualSize <= OpITy->getFPMantissaWidth()) {
1517 if (FITy->getScalarSizeInBits() > SrcTy->getScalarSizeInBits()) {
1518 if (IsInputSigned && IsOutputSigned)
1519 return new SExtInst(SrcI, FITy);
1520 return new ZExtInst(SrcI, FITy);
1522 if (FITy->getScalarSizeInBits() < SrcTy->getScalarSizeInBits())
1523 return new TruncInst(SrcI, FITy);
1525 return replaceInstUsesWith(FI, SrcI);
1526 return new BitCastInst(SrcI, FITy);
1531 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
1532 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1534 return commonCastTransforms(FI);
1536 if (Instruction *I = FoldItoFPtoI(FI))
1539 return commonCastTransforms(FI);
1542 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
1543 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1545 return commonCastTransforms(FI);
1547 if (Instruction *I = FoldItoFPtoI(FI))
1550 return commonCastTransforms(FI);
1553 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
1554 return commonCastTransforms(CI);
1557 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
1558 return commonCastTransforms(CI);
1561 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
1562 // If the source integer type is not the intptr_t type for this target, do a
1563 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the
1564 // cast to be exposed to other transforms.
1565 unsigned AS = CI.getAddressSpace();
1566 if (CI.getOperand(0)->getType()->getScalarSizeInBits() !=
1567 DL.getPointerSizeInBits(AS)) {
1568 Type *Ty = DL.getIntPtrType(CI.getContext(), AS);
1569 if (CI.getType()->isVectorTy()) // Handle vectors of pointers.
1570 Ty = VectorType::get(Ty, CI.getType()->getVectorNumElements());
1572 Value *P = Builder->CreateZExtOrTrunc(CI.getOperand(0), Ty);
1573 return new IntToPtrInst(P, CI.getType());
1576 if (Instruction *I = commonCastTransforms(CI))
1582 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
1583 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
1584 Value *Src = CI.getOperand(0);
1586 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
1587 // If casting the result of a getelementptr instruction with no offset, turn
1588 // this into a cast of the original pointer!
1589 if (GEP->hasAllZeroIndices() &&
1590 // If CI is an addrspacecast and GEP changes the poiner type, merging
1591 // GEP into CI would undo canonicalizing addrspacecast with different
1592 // pointer types, causing infinite loops.
1593 (!isa<AddrSpaceCastInst>(CI) ||
1594 GEP->getType() == GEP->getPointerOperandType())) {
1595 // Changing the cast operand is usually not a good idea but it is safe
1596 // here because the pointer operand is being replaced with another
1597 // pointer operand so the opcode doesn't need to change.
1599 CI.setOperand(0, GEP->getOperand(0));
1604 return commonCastTransforms(CI);
1607 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
1608 // If the destination integer type is not the intptr_t type for this target,
1609 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast
1610 // to be exposed to other transforms.
1612 Type *Ty = CI.getType();
1613 unsigned AS = CI.getPointerAddressSpace();
1615 if (Ty->getScalarSizeInBits() == DL.getPointerSizeInBits(AS))
1616 return commonPointerCastTransforms(CI);
1618 Type *PtrTy = DL.getIntPtrType(CI.getContext(), AS);
1619 if (Ty->isVectorTy()) // Handle vectors of pointers.
1620 PtrTy = VectorType::get(PtrTy, Ty->getVectorNumElements());
1622 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), PtrTy);
1623 return CastInst::CreateIntegerCast(P, Ty, /*isSigned=*/false);
1626 /// This input value (which is known to have vector type) is being zero extended
1627 /// or truncated to the specified vector type.
1628 /// Try to replace it with a shuffle (and vector/vector bitcast) if possible.
1630 /// The source and destination vector types may have different element types.
1631 static Instruction *optimizeVectorResize(Value *InVal, VectorType *DestTy,
1633 // We can only do this optimization if the output is a multiple of the input
1634 // element size, or the input is a multiple of the output element size.
1635 // Convert the input type to have the same element type as the output.
1636 VectorType *SrcTy = cast<VectorType>(InVal->getType());
1638 if (SrcTy->getElementType() != DestTy->getElementType()) {
1639 // The input types don't need to be identical, but for now they must be the
1640 // same size. There is no specific reason we couldn't handle things like
1641 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten
1643 if (SrcTy->getElementType()->getPrimitiveSizeInBits() !=
1644 DestTy->getElementType()->getPrimitiveSizeInBits())
1647 SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements());
1648 InVal = IC.Builder->CreateBitCast(InVal, SrcTy);
1651 // Now that the element types match, get the shuffle mask and RHS of the
1652 // shuffle to use, which depends on whether we're increasing or decreasing the
1653 // size of the input.
1654 SmallVector<uint32_t, 16> ShuffleMask;
1657 if (SrcTy->getNumElements() > DestTy->getNumElements()) {
1658 // If we're shrinking the number of elements, just shuffle in the low
1659 // elements from the input and use undef as the second shuffle input.
1660 V2 = UndefValue::get(SrcTy);
1661 for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i)
1662 ShuffleMask.push_back(i);
1665 // If we're increasing the number of elements, shuffle in all of the
1666 // elements from InVal and fill the rest of the result elements with zeros
1667 // from a constant zero.
1668 V2 = Constant::getNullValue(SrcTy);
1669 unsigned SrcElts = SrcTy->getNumElements();
1670 for (unsigned i = 0, e = SrcElts; i != e; ++i)
1671 ShuffleMask.push_back(i);
1673 // The excess elements reference the first element of the zero input.
1674 for (unsigned i = 0, e = DestTy->getNumElements()-SrcElts; i != e; ++i)
1675 ShuffleMask.push_back(SrcElts);
1678 return new ShuffleVectorInst(InVal, V2,
1679 ConstantDataVector::get(V2->getContext(),
1683 static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) {
1684 return Value % Ty->getPrimitiveSizeInBits() == 0;
1687 static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) {
1688 return Value / Ty->getPrimitiveSizeInBits();
1691 /// V is a value which is inserted into a vector of VecEltTy.
1692 /// Look through the value to see if we can decompose it into
1693 /// insertions into the vector. See the example in the comment for
1694 /// OptimizeIntegerToVectorInsertions for the pattern this handles.
1695 /// The type of V is always a non-zero multiple of VecEltTy's size.
1696 /// Shift is the number of bits between the lsb of V and the lsb of
1699 /// This returns false if the pattern can't be matched or true if it can,
1700 /// filling in Elements with the elements found here.
1701 static bool collectInsertionElements(Value *V, unsigned Shift,
1702 SmallVectorImpl<Value *> &Elements,
1703 Type *VecEltTy, bool isBigEndian) {
1704 assert(isMultipleOfTypeSize(Shift, VecEltTy) &&
1705 "Shift should be a multiple of the element type size");
1707 // Undef values never contribute useful bits to the result.
1708 if (isa<UndefValue>(V)) return true;
1710 // If we got down to a value of the right type, we win, try inserting into the
1712 if (V->getType() == VecEltTy) {
1713 // Inserting null doesn't actually insert any elements.
1714 if (Constant *C = dyn_cast<Constant>(V))
1715 if (C->isNullValue())
1718 unsigned ElementIndex = getTypeSizeIndex(Shift, VecEltTy);
1720 ElementIndex = Elements.size() - ElementIndex - 1;
1722 // Fail if multiple elements are inserted into this slot.
1723 if (Elements[ElementIndex])
1726 Elements[ElementIndex] = V;
1730 if (Constant *C = dyn_cast<Constant>(V)) {
1731 // Figure out the # elements this provides, and bitcast it or slice it up
1733 unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(),
1735 // If the constant is the size of a vector element, we just need to bitcast
1736 // it to the right type so it gets properly inserted.
1738 return collectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy),
1739 Shift, Elements, VecEltTy, isBigEndian);
1741 // Okay, this is a constant that covers multiple elements. Slice it up into
1742 // pieces and insert each element-sized piece into the vector.
1743 if (!isa<IntegerType>(C->getType()))
1744 C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(),
1745 C->getType()->getPrimitiveSizeInBits()));
1746 unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits();
1747 Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize);
1749 for (unsigned i = 0; i != NumElts; ++i) {
1750 unsigned ShiftI = Shift+i*ElementSize;
1751 Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(),
1753 Piece = ConstantExpr::getTrunc(Piece, ElementIntTy);
1754 if (!collectInsertionElements(Piece, ShiftI, Elements, VecEltTy,
1761 if (!V->hasOneUse()) return false;
1763 Instruction *I = dyn_cast<Instruction>(V);
1764 if (!I) return false;
1765 switch (I->getOpcode()) {
1766 default: return false; // Unhandled case.
1767 case Instruction::BitCast:
1768 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
1770 case Instruction::ZExt:
1771 if (!isMultipleOfTypeSize(
1772 I->getOperand(0)->getType()->getPrimitiveSizeInBits(),
1775 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
1777 case Instruction::Or:
1778 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
1780 collectInsertionElements(I->getOperand(1), Shift, Elements, VecEltTy,
1782 case Instruction::Shl: {
1783 // Must be shifting by a constant that is a multiple of the element size.
1784 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1));
1785 if (!CI) return false;
1786 Shift += CI->getZExtValue();
1787 if (!isMultipleOfTypeSize(Shift, VecEltTy)) return false;
1788 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
1796 /// If the input is an 'or' instruction, we may be doing shifts and ors to
1797 /// assemble the elements of the vector manually.
1798 /// Try to rip the code out and replace it with insertelements. This is to
1799 /// optimize code like this:
1801 /// %tmp37 = bitcast float %inc to i32
1802 /// %tmp38 = zext i32 %tmp37 to i64
1803 /// %tmp31 = bitcast float %inc5 to i32
1804 /// %tmp32 = zext i32 %tmp31 to i64
1805 /// %tmp33 = shl i64 %tmp32, 32
1806 /// %ins35 = or i64 %tmp33, %tmp38
1807 /// %tmp43 = bitcast i64 %ins35 to <2 x float>
1809 /// Into two insertelements that do "buildvector{%inc, %inc5}".
1810 static Value *optimizeIntegerToVectorInsertions(BitCastInst &CI,
1812 VectorType *DestVecTy = cast<VectorType>(CI.getType());
1813 Value *IntInput = CI.getOperand(0);
1815 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements());
1816 if (!collectInsertionElements(IntInput, 0, Elements,
1817 DestVecTy->getElementType(),
1818 IC.getDataLayout().isBigEndian()))
1821 // If we succeeded, we know that all of the element are specified by Elements
1822 // or are zero if Elements has a null entry. Recast this as a set of
1824 Value *Result = Constant::getNullValue(CI.getType());
1825 for (unsigned i = 0, e = Elements.size(); i != e; ++i) {
1826 if (!Elements[i]) continue; // Unset element.
1828 Result = IC.Builder->CreateInsertElement(Result, Elements[i],
1829 IC.Builder->getInt32(i));
1835 /// Canonicalize scalar bitcasts of extracted elements into a bitcast of the
1836 /// vector followed by extract element. The backend tends to handle bitcasts of
1837 /// vectors better than bitcasts of scalars because vector registers are
1838 /// usually not type-specific like scalar integer or scalar floating-point.
1839 static Instruction *canonicalizeBitCastExtElt(BitCastInst &BitCast,
1841 const DataLayout &DL) {
1842 // TODO: Create and use a pattern matcher for ExtractElementInst.
1843 auto *ExtElt = dyn_cast<ExtractElementInst>(BitCast.getOperand(0));
1844 if (!ExtElt || !ExtElt->hasOneUse())
1847 // The bitcast must be to a vectorizable type, otherwise we can't make a new
1848 // type to extract from.
1849 Type *DestType = BitCast.getType();
1850 if (!VectorType::isValidElementType(DestType))
1853 unsigned NumElts = ExtElt->getVectorOperandType()->getNumElements();
1854 auto *NewVecType = VectorType::get(DestType, NumElts);
1855 auto *NewBC = IC.Builder->CreateBitCast(ExtElt->getVectorOperand(),
1857 return ExtractElementInst::Create(NewBC, ExtElt->getIndexOperand());
1860 /// Change the type of a bitwise logic operation if we can eliminate a bitcast.
1861 static Instruction *foldBitCastBitwiseLogic(BitCastInst &BitCast,
1862 InstCombiner::BuilderTy &Builder) {
1863 Type *DestTy = BitCast.getType();
1865 if (!DestTy->getScalarType()->isIntegerTy() ||
1866 !match(BitCast.getOperand(0), m_OneUse(m_BinOp(BO))) ||
1867 !BO->isBitwiseLogicOp())
1870 // FIXME: This transform is restricted to vector types to avoid backend
1871 // problems caused by creating potentially illegal operations. If a fix-up is
1872 // added to handle that situation, we can remove this check.
1873 if (!DestTy->isVectorTy() || !BO->getType()->isVectorTy())
1877 if (match(BO->getOperand(0), m_OneUse(m_BitCast(m_Value(X)))) &&
1878 X->getType() == DestTy && !isa<Constant>(X)) {
1879 // bitcast(logic(bitcast(X), Y)) --> logic'(X, bitcast(Y))
1880 Value *CastedOp1 = Builder.CreateBitCast(BO->getOperand(1), DestTy);
1881 return BinaryOperator::Create(BO->getOpcode(), X, CastedOp1);
1884 if (match(BO->getOperand(1), m_OneUse(m_BitCast(m_Value(X)))) &&
1885 X->getType() == DestTy && !isa<Constant>(X)) {
1886 // bitcast(logic(Y, bitcast(X))) --> logic'(bitcast(Y), X)
1887 Value *CastedOp0 = Builder.CreateBitCast(BO->getOperand(0), DestTy);
1888 return BinaryOperator::Create(BO->getOpcode(), CastedOp0, X);
1894 /// Change the type of a select if we can eliminate a bitcast.
1895 static Instruction *foldBitCastSelect(BitCastInst &BitCast,
1896 InstCombiner::BuilderTy &Builder) {
1897 Value *Cond, *TVal, *FVal;
1898 if (!match(BitCast.getOperand(0),
1899 m_OneUse(m_Select(m_Value(Cond), m_Value(TVal), m_Value(FVal)))))
1902 // A vector select must maintain the same number of elements in its operands.
1903 Type *CondTy = Cond->getType();
1904 Type *DestTy = BitCast.getType();
1905 if (CondTy->isVectorTy()) {
1906 if (!DestTy->isVectorTy())
1908 if (DestTy->getVectorNumElements() != CondTy->getVectorNumElements())
1912 // FIXME: This transform is restricted from changing the select between
1913 // scalars and vectors to avoid backend problems caused by creating
1914 // potentially illegal operations. If a fix-up is added to handle that
1915 // situation, we can remove this check.
1916 if (DestTy->isVectorTy() != TVal->getType()->isVectorTy())
1919 auto *Sel = cast<Instruction>(BitCast.getOperand(0));
1921 if (match(TVal, m_OneUse(m_BitCast(m_Value(X)))) && X->getType() == DestTy &&
1922 !isa<Constant>(X)) {
1923 // bitcast(select(Cond, bitcast(X), Y)) --> select'(Cond, X, bitcast(Y))
1924 Value *CastedVal = Builder.CreateBitCast(FVal, DestTy);
1925 return SelectInst::Create(Cond, X, CastedVal, "", nullptr, Sel);
1928 if (match(FVal, m_OneUse(m_BitCast(m_Value(X)))) && X->getType() == DestTy &&
1929 !isa<Constant>(X)) {
1930 // bitcast(select(Cond, Y, bitcast(X))) --> select'(Cond, bitcast(Y), X)
1931 Value *CastedVal = Builder.CreateBitCast(TVal, DestTy);
1932 return SelectInst::Create(Cond, CastedVal, X, "", nullptr, Sel);
1938 /// Check if all users of CI are StoreInsts.
1939 static bool hasStoreUsersOnly(CastInst &CI) {
1940 for (User *U : CI.users()) {
1941 if (!isa<StoreInst>(U))
1947 /// This function handles following case
1953 /// All the related PHI nodes can be replaced by new PHI nodes with type A.
1954 /// The uses of \p CI can be changed to the new PHI node corresponding to \p PN.
1955 Instruction *InstCombiner::optimizeBitCastFromPhi(CastInst &CI, PHINode *PN) {
1956 // BitCast used by Store can be handled in InstCombineLoadStoreAlloca.cpp.
1957 if (hasStoreUsersOnly(CI))
1960 Value *Src = CI.getOperand(0);
1961 Type *SrcTy = Src->getType(); // Type B
1962 Type *DestTy = CI.getType(); // Type A
1964 SmallVector<PHINode *, 4> PhiWorklist;
1965 SmallSetVector<PHINode *, 4> OldPhiNodes;
1967 // Find all of the A->B casts and PHI nodes.
1968 // We need to inpect all related PHI nodes, but PHIs can be cyclic, so
1969 // OldPhiNodes is used to track all known PHI nodes, before adding a new
1970 // PHI to PhiWorklist, it is checked against and added to OldPhiNodes first.
1971 PhiWorklist.push_back(PN);
1972 OldPhiNodes.insert(PN);
1973 while (!PhiWorklist.empty()) {
1974 auto *OldPN = PhiWorklist.pop_back_val();
1975 for (Value *IncValue : OldPN->incoming_values()) {
1976 if (isa<Constant>(IncValue))
1979 if (auto *LI = dyn_cast<LoadInst>(IncValue)) {
1980 // If there is a sequence of one or more load instructions, each loaded
1981 // value is used as address of later load instruction, bitcast is
1982 // necessary to change the value type, don't optimize it. For
1983 // simplicity we give up if the load address comes from another load.
1984 Value *Addr = LI->getOperand(0);
1985 if (Addr == &CI || isa<LoadInst>(Addr))
1987 if (LI->hasOneUse() && LI->isSimple())
1989 // If a LoadInst has more than one use, changing the type of loaded
1990 // value may create another bitcast.
1994 if (auto *PNode = dyn_cast<PHINode>(IncValue)) {
1995 if (OldPhiNodes.insert(PNode))
1996 PhiWorklist.push_back(PNode);
2000 auto *BCI = dyn_cast<BitCastInst>(IncValue);
2001 // We can't handle other instructions.
2005 // Verify it's a A->B cast.
2006 Type *TyA = BCI->getOperand(0)->getType();
2007 Type *TyB = BCI->getType();
2008 if (TyA != DestTy || TyB != SrcTy)
2013 // For each old PHI node, create a corresponding new PHI node with a type A.
2014 SmallDenseMap<PHINode *, PHINode *> NewPNodes;
2015 for (auto *OldPN : OldPhiNodes) {
2016 Builder->SetInsertPoint(OldPN);
2017 PHINode *NewPN = Builder->CreatePHI(DestTy, OldPN->getNumOperands());
2018 NewPNodes[OldPN] = NewPN;
2021 // Fill in the operands of new PHI nodes.
2022 for (auto *OldPN : OldPhiNodes) {
2023 PHINode *NewPN = NewPNodes[OldPN];
2024 for (unsigned j = 0, e = OldPN->getNumOperands(); j != e; ++j) {
2025 Value *V = OldPN->getOperand(j);
2026 Value *NewV = nullptr;
2027 if (auto *C = dyn_cast<Constant>(V)) {
2028 NewV = ConstantExpr::getBitCast(C, DestTy);
2029 } else if (auto *LI = dyn_cast<LoadInst>(V)) {
2030 Builder->SetInsertPoint(LI->getNextNode());
2031 NewV = Builder->CreateBitCast(LI, DestTy);
2033 } else if (auto *BCI = dyn_cast<BitCastInst>(V)) {
2034 NewV = BCI->getOperand(0);
2035 } else if (auto *PrevPN = dyn_cast<PHINode>(V)) {
2036 NewV = NewPNodes[PrevPN];
2039 NewPN->addIncoming(NewV, OldPN->getIncomingBlock(j));
2043 // If there is a store with type B, change it to type A.
2044 for (User *U : PN->users()) {
2045 auto *SI = dyn_cast<StoreInst>(U);
2046 if (SI && SI->isSimple() && SI->getOperand(0) == PN) {
2047 Builder->SetInsertPoint(SI);
2049 cast<BitCastInst>(Builder->CreateBitCast(NewPNodes[PN], SrcTy));
2050 SI->setOperand(0, NewBC);
2052 assert(hasStoreUsersOnly(*NewBC));
2056 return replaceInstUsesWith(CI, NewPNodes[PN]);
2059 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
2060 // If the operands are integer typed then apply the integer transforms,
2061 // otherwise just apply the common ones.
2062 Value *Src = CI.getOperand(0);
2063 Type *SrcTy = Src->getType();
2064 Type *DestTy = CI.getType();
2066 // Get rid of casts from one type to the same type. These are useless and can
2067 // be replaced by the operand.
2068 if (DestTy == Src->getType())
2069 return replaceInstUsesWith(CI, Src);
2071 if (PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
2072 PointerType *SrcPTy = cast<PointerType>(SrcTy);
2073 Type *DstElTy = DstPTy->getElementType();
2074 Type *SrcElTy = SrcPTy->getElementType();
2076 // If we are casting a alloca to a pointer to a type of the same
2077 // size, rewrite the allocation instruction to allocate the "right" type.
2078 // There is no need to modify malloc calls because it is their bitcast that
2079 // needs to be cleaned up.
2080 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src))
2081 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
2084 // When the type pointed to is not sized the cast cannot be
2085 // turned into a gep.
2087 cast<PointerType>(Src->getType()->getScalarType())->getElementType();
2088 if (!PointeeType->isSized())
2091 // If the source and destination are pointers, and this cast is equivalent
2092 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
2093 // This can enhance SROA and other transforms that want type-safe pointers.
2094 unsigned NumZeros = 0;
2095 while (SrcElTy != DstElTy &&
2096 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() &&
2097 SrcElTy->getNumContainedTypes() /* not "{}" */) {
2098 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(0U);
2102 // If we found a path from the src to dest, create the getelementptr now.
2103 if (SrcElTy == DstElTy) {
2104 SmallVector<Value *, 8> Idxs(NumZeros + 1, Builder->getInt32(0));
2105 return GetElementPtrInst::CreateInBounds(Src, Idxs);
2109 if (VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
2110 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) {
2111 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
2112 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
2113 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
2114 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
2117 if (isa<IntegerType>(SrcTy)) {
2118 // If this is a cast from an integer to vector, check to see if the input
2119 // is a trunc or zext of a bitcast from vector. If so, we can replace all
2120 // the casts with a shuffle and (potentially) a bitcast.
2121 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) {
2122 CastInst *SrcCast = cast<CastInst>(Src);
2123 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0)))
2124 if (isa<VectorType>(BCIn->getOperand(0)->getType()))
2125 if (Instruction *I = optimizeVectorResize(BCIn->getOperand(0),
2126 cast<VectorType>(DestTy), *this))
2130 // If the input is an 'or' instruction, we may be doing shifts and ors to
2131 // assemble the elements of the vector manually. Try to rip the code out
2132 // and replace it with insertelements.
2133 if (Value *V = optimizeIntegerToVectorInsertions(CI, *this))
2134 return replaceInstUsesWith(CI, V);
2138 if (VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
2139 if (SrcVTy->getNumElements() == 1) {
2140 // If our destination is not a vector, then make this a straight
2141 // scalar-scalar cast.
2142 if (!DestTy->isVectorTy()) {
2144 Builder->CreateExtractElement(Src,
2145 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
2146 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
2149 // Otherwise, see if our source is an insert. If so, then use the scalar
2150 // component directly.
2151 if (InsertElementInst *IEI =
2152 dyn_cast<InsertElementInst>(CI.getOperand(0)))
2153 return CastInst::Create(Instruction::BitCast, IEI->getOperand(1),
2158 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
2159 // Okay, we have (bitcast (shuffle ..)). Check to see if this is
2160 // a bitcast to a vector with the same # elts.
2161 if (SVI->hasOneUse() && DestTy->isVectorTy() &&
2162 DestTy->getVectorNumElements() == SVI->getType()->getNumElements() &&
2163 SVI->getType()->getNumElements() ==
2164 SVI->getOperand(0)->getType()->getVectorNumElements()) {
2166 // If either of the operands is a cast from CI.getType(), then
2167 // evaluating the shuffle in the casted destination's type will allow
2168 // us to eliminate at least one cast.
2169 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) &&
2170 Tmp->getOperand(0)->getType() == DestTy) ||
2171 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) &&
2172 Tmp->getOperand(0)->getType() == DestTy)) {
2173 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy);
2174 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy);
2175 // Return a new shuffle vector. Use the same element ID's, as we
2176 // know the vector types match #elts.
2177 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
2182 // Handle the A->B->A cast, and there is an intervening PHI node.
2183 if (PHINode *PN = dyn_cast<PHINode>(Src))
2184 if (Instruction *I = optimizeBitCastFromPhi(CI, PN))
2187 if (Instruction *I = canonicalizeBitCastExtElt(CI, *this, DL))
2190 if (Instruction *I = foldBitCastBitwiseLogic(CI, *Builder))
2193 if (Instruction *I = foldBitCastSelect(CI, *Builder))
2196 if (SrcTy->isPointerTy())
2197 return commonPointerCastTransforms(CI);
2198 return commonCastTransforms(CI);
2201 Instruction *InstCombiner::visitAddrSpaceCast(AddrSpaceCastInst &CI) {
2202 // If the destination pointer element type is not the same as the source's
2203 // first do a bitcast to the destination type, and then the addrspacecast.
2204 // This allows the cast to be exposed to other transforms.
2205 Value *Src = CI.getOperand(0);
2206 PointerType *SrcTy = cast<PointerType>(Src->getType()->getScalarType());
2207 PointerType *DestTy = cast<PointerType>(CI.getType()->getScalarType());
2209 Type *DestElemTy = DestTy->getElementType();
2210 if (SrcTy->getElementType() != DestElemTy) {
2211 Type *MidTy = PointerType::get(DestElemTy, SrcTy->getAddressSpace());
2212 if (VectorType *VT = dyn_cast<VectorType>(CI.getType())) {
2213 // Handle vectors of pointers.
2214 MidTy = VectorType::get(MidTy, VT->getNumElements());
2217 Value *NewBitCast = Builder->CreateBitCast(Src, MidTy);
2218 return new AddrSpaceCastInst(NewBitCast, CI.getType());
2221 return commonPointerCastTransforms(CI);