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/Analysis/TargetLibraryInfo.h"
18 #include "llvm/IR/DataLayout.h"
19 #include "llvm/IR/PatternMatch.h"
20 #include "llvm/Support/KnownBits.h"
22 using namespace PatternMatch;
24 #define DEBUG_TYPE "instcombine"
26 /// Analyze 'Val', seeing if it is a simple linear expression.
27 /// If so, decompose it, returning some value X, such that Val is
30 static Value *decomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
32 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
33 Offset = CI->getZExtValue();
35 return ConstantInt::get(Val->getType(), 0);
38 if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
39 // Cannot look past anything that might overflow.
40 OverflowingBinaryOperator *OBI = dyn_cast<OverflowingBinaryOperator>(Val);
41 if (OBI && !OBI->hasNoUnsignedWrap() && !OBI->hasNoSignedWrap()) {
47 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
48 if (I->getOpcode() == Instruction::Shl) {
49 // This is a value scaled by '1 << the shift amt'.
50 Scale = UINT64_C(1) << RHS->getZExtValue();
52 return I->getOperand(0);
55 if (I->getOpcode() == Instruction::Mul) {
56 // This value is scaled by 'RHS'.
57 Scale = RHS->getZExtValue();
59 return I->getOperand(0);
62 if (I->getOpcode() == Instruction::Add) {
63 // We have X+C. Check to see if we really have (X*C2)+C1,
64 // where C1 is divisible by C2.
67 decomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
68 Offset += RHS->getZExtValue();
75 // Otherwise, we can't look past this.
81 /// If we find a cast of an allocation instruction, try to eliminate the cast by
82 /// moving the type information into the alloc.
83 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
85 PointerType *PTy = cast<PointerType>(CI.getType());
87 BuilderTy AllocaBuilder(Builder);
88 AllocaBuilder.SetInsertPoint(&AI);
90 // Get the type really allocated and the type casted to.
91 Type *AllocElTy = AI.getAllocatedType();
92 Type *CastElTy = PTy->getElementType();
93 if (!AllocElTy->isSized() || !CastElTy->isSized()) return nullptr;
95 unsigned AllocElTyAlign = DL.getABITypeAlignment(AllocElTy);
96 unsigned CastElTyAlign = DL.getABITypeAlignment(CastElTy);
97 if (CastElTyAlign < AllocElTyAlign) return nullptr;
99 // If the allocation has multiple uses, only promote it if we are strictly
100 // increasing the alignment of the resultant allocation. If we keep it the
101 // same, we open the door to infinite loops of various kinds.
102 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return nullptr;
104 uint64_t AllocElTySize = DL.getTypeAllocSize(AllocElTy);
105 uint64_t CastElTySize = DL.getTypeAllocSize(CastElTy);
106 if (CastElTySize == 0 || AllocElTySize == 0) return nullptr;
108 // If the allocation has multiple uses, only promote it if we're not
109 // shrinking the amount of memory being allocated.
110 uint64_t AllocElTyStoreSize = DL.getTypeStoreSize(AllocElTy);
111 uint64_t CastElTyStoreSize = DL.getTypeStoreSize(CastElTy);
112 if (!AI.hasOneUse() && CastElTyStoreSize < AllocElTyStoreSize) return nullptr;
114 // See if we can satisfy the modulus by pulling a scale out of the array
116 unsigned ArraySizeScale;
117 uint64_t ArrayOffset;
118 Value *NumElements = // See if the array size is a decomposable linear expr.
119 decomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
121 // If we can now satisfy the modulus, by using a non-1 scale, we really can
123 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
124 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return nullptr;
126 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
127 Value *Amt = nullptr;
131 Amt = ConstantInt::get(AI.getArraySize()->getType(), Scale);
132 // Insert before the alloca, not before the cast.
133 Amt = AllocaBuilder.CreateMul(Amt, NumElements);
136 if (uint64_t Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
137 Value *Off = ConstantInt::get(AI.getArraySize()->getType(),
139 Amt = AllocaBuilder.CreateAdd(Amt, Off);
142 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt);
143 New->setAlignment(AI.getAlignment());
145 New->setUsedWithInAlloca(AI.isUsedWithInAlloca());
147 // If the allocation has multiple real uses, insert a cast and change all
148 // things that used it to use the new cast. This will also hack on CI, but it
150 if (!AI.hasOneUse()) {
151 // New is the allocation instruction, pointer typed. AI is the original
152 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
153 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast");
154 replaceInstUsesWith(AI, NewCast);
156 return replaceInstUsesWith(CI, New);
159 /// Given an expression that CanEvaluateTruncated or CanEvaluateSExtd returns
160 /// true for, actually insert the code to evaluate the expression.
161 Value *InstCombiner::EvaluateInDifferentType(Value *V, Type *Ty,
163 if (Constant *C = dyn_cast<Constant>(V)) {
164 C = ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
165 // If we got a constantexpr back, try to simplify it with DL info.
166 if (Constant *FoldedC = ConstantFoldConstant(C, DL, &TLI))
171 // Otherwise, it must be an instruction.
172 Instruction *I = cast<Instruction>(V);
173 Instruction *Res = nullptr;
174 unsigned Opc = I->getOpcode();
176 case Instruction::Add:
177 case Instruction::Sub:
178 case Instruction::Mul:
179 case Instruction::And:
180 case Instruction::Or:
181 case Instruction::Xor:
182 case Instruction::AShr:
183 case Instruction::LShr:
184 case Instruction::Shl:
185 case Instruction::UDiv:
186 case Instruction::URem: {
187 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
188 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
189 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
192 case Instruction::Trunc:
193 case Instruction::ZExt:
194 case Instruction::SExt:
195 // If the source type of the cast is the type we're trying for then we can
196 // just return the source. There's no need to insert it because it is not
198 if (I->getOperand(0)->getType() == Ty)
199 return I->getOperand(0);
201 // Otherwise, must be the same type of cast, so just reinsert a new one.
202 // This also handles the case of zext(trunc(x)) -> zext(x).
203 Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty,
204 Opc == Instruction::SExt);
206 case Instruction::Select: {
207 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
208 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
209 Res = SelectInst::Create(I->getOperand(0), True, False);
212 case Instruction::PHI: {
213 PHINode *OPN = cast<PHINode>(I);
214 PHINode *NPN = PHINode::Create(Ty, OPN->getNumIncomingValues());
215 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
217 EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
218 NPN->addIncoming(V, OPN->getIncomingBlock(i));
224 // TODO: Can handle more cases here.
225 llvm_unreachable("Unreachable!");
229 return InsertNewInstWith(Res, *I);
232 Instruction::CastOps InstCombiner::isEliminableCastPair(const CastInst *CI1,
233 const CastInst *CI2) {
234 Type *SrcTy = CI1->getSrcTy();
235 Type *MidTy = CI1->getDestTy();
236 Type *DstTy = CI2->getDestTy();
238 Instruction::CastOps firstOp = Instruction::CastOps(CI1->getOpcode());
239 Instruction::CastOps secondOp = Instruction::CastOps(CI2->getOpcode());
241 SrcTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(SrcTy) : nullptr;
243 MidTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(MidTy) : nullptr;
245 DstTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(DstTy) : nullptr;
246 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
247 DstTy, SrcIntPtrTy, MidIntPtrTy,
250 // We don't want to form an inttoptr or ptrtoint that converts to an integer
251 // type that differs from the pointer size.
252 if ((Res == Instruction::IntToPtr && SrcTy != DstIntPtrTy) ||
253 (Res == Instruction::PtrToInt && DstTy != SrcIntPtrTy))
256 return Instruction::CastOps(Res);
259 /// @brief Implement the transforms common to all CastInst visitors.
260 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
261 Value *Src = CI.getOperand(0);
263 // Try to eliminate a cast of a cast.
264 if (auto *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
265 if (Instruction::CastOps NewOpc = isEliminableCastPair(CSrc, &CI)) {
266 // The first cast (CSrc) is eliminable so we need to fix up or replace
267 // the second cast (CI). CSrc will then have a good chance of being dead.
268 return CastInst::Create(NewOpc, CSrc->getOperand(0), CI.getType());
272 // If we are casting a select, then fold the cast into the select.
273 if (auto *SI = dyn_cast<SelectInst>(Src))
274 if (Instruction *NV = FoldOpIntoSelect(CI, SI))
277 // If we are casting a PHI, then fold the cast into the PHI.
278 if (auto *PN = dyn_cast<PHINode>(Src)) {
279 // Don't do this if it would create a PHI node with an illegal type from a
281 if (!Src->getType()->isIntegerTy() || !CI.getType()->isIntegerTy() ||
282 shouldChangeType(CI.getType(), Src->getType()))
283 if (Instruction *NV = foldOpIntoPhi(CI, PN))
290 /// Return true if we can evaluate the specified expression tree as type Ty
291 /// instead of its larger type, and arrive with the same value.
292 /// This is used by code that tries to eliminate truncates.
294 /// Ty will always be a type smaller than V. We should return true if trunc(V)
295 /// can be computed by computing V in the smaller type. If V is an instruction,
296 /// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only
297 /// makes sense if x and y can be efficiently truncated.
299 /// This function works on both vectors and scalars.
301 static bool canEvaluateTruncated(Value *V, Type *Ty, InstCombiner &IC,
303 // We can always evaluate constants in another type.
304 if (isa<Constant>(V))
307 Instruction *I = dyn_cast<Instruction>(V);
308 if (!I) return false;
310 Type *OrigTy = V->getType();
312 // If this is an extension from the dest type, we can eliminate it, even if it
313 // has multiple uses.
314 if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
315 I->getOperand(0)->getType() == Ty)
318 // We can't extend or shrink something that has multiple uses: doing so would
319 // require duplicating the instruction in general, which isn't profitable.
320 if (!I->hasOneUse()) return false;
322 unsigned Opc = I->getOpcode();
324 case Instruction::Add:
325 case Instruction::Sub:
326 case Instruction::Mul:
327 case Instruction::And:
328 case Instruction::Or:
329 case Instruction::Xor:
330 // These operators can all arbitrarily be extended or truncated.
331 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
332 canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
334 case Instruction::UDiv:
335 case Instruction::URem: {
336 // UDiv and URem can be truncated if all the truncated bits are zero.
337 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
338 uint32_t BitWidth = Ty->getScalarSizeInBits();
339 if (BitWidth < OrigBitWidth) {
340 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
341 if (IC.MaskedValueIsZero(I->getOperand(0), Mask, 0, CxtI) &&
342 IC.MaskedValueIsZero(I->getOperand(1), Mask, 0, CxtI)) {
343 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
344 canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
349 case Instruction::Shl:
350 // If we are truncating the result of this SHL, and if it's a shift of a
351 // constant amount, we can always perform a SHL in a smaller type.
352 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
353 uint32_t BitWidth = Ty->getScalarSizeInBits();
354 if (CI->getLimitedValue(BitWidth) < BitWidth)
355 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI);
358 case Instruction::LShr:
359 // If this is a truncate of a logical shr, we can truncate it to a smaller
360 // lshr iff we know that the bits we would otherwise be shifting in are
362 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
363 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
364 uint32_t BitWidth = Ty->getScalarSizeInBits();
365 if (IC.MaskedValueIsZero(I->getOperand(0),
366 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth), 0, CxtI) &&
367 CI->getLimitedValue(BitWidth) < BitWidth) {
368 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI);
372 case Instruction::Trunc:
373 // trunc(trunc(x)) -> trunc(x)
375 case Instruction::ZExt:
376 case Instruction::SExt:
377 // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest
378 // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest
380 case Instruction::Select: {
381 SelectInst *SI = cast<SelectInst>(I);
382 return canEvaluateTruncated(SI->getTrueValue(), Ty, IC, CxtI) &&
383 canEvaluateTruncated(SI->getFalseValue(), Ty, IC, CxtI);
385 case Instruction::PHI: {
386 // We can change a phi if we can change all operands. Note that we never
387 // get into trouble with cyclic PHIs here because we only consider
388 // instructions with a single use.
389 PHINode *PN = cast<PHINode>(I);
390 for (Value *IncValue : PN->incoming_values())
391 if (!canEvaluateTruncated(IncValue, Ty, IC, CxtI))
396 // TODO: Can handle more cases here.
403 /// Given a vector that is bitcast to an integer, optionally logically
404 /// right-shifted, and truncated, convert it to an extractelement.
405 /// Example (big endian):
406 /// trunc (lshr (bitcast <4 x i32> %X to i128), 32) to i32
408 /// extractelement <4 x i32> %X, 1
409 static Instruction *foldVecTruncToExtElt(TruncInst &Trunc, InstCombiner &IC) {
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 (IC.getDataLayout().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 // FIXME: Maybe combine the next two transforms to handle the no cast case
562 // more efficiently. Support vector types. Cleanup code by using m_OneUse.
564 // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion.
565 Value *A = nullptr; ConstantInt *Cst = nullptr;
566 if (Src->hasOneUse() &&
567 match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst)))) {
568 // We have three types to worry about here, the type of A, the source of
569 // the truncate (MidSize), and the destination of the truncate. We know that
570 // ASize < MidSize and MidSize > ResultSize, but don't know the relation
571 // between ASize and ResultSize.
572 unsigned ASize = A->getType()->getPrimitiveSizeInBits();
574 // If the shift amount is larger than the size of A, then the result is
575 // known to be zero because all the input bits got shifted out.
576 if (Cst->getZExtValue() >= ASize)
577 return replaceInstUsesWith(CI, Constant::getNullValue(DestTy));
579 // Since we're doing an lshr and a zero extend, and know that the shift
580 // amount is smaller than ASize, it is always safe to do the shift in A's
581 // type, then zero extend or truncate to the result.
582 Value *Shift = Builder.CreateLShr(A, Cst->getZExtValue());
583 Shift->takeName(Src);
584 return CastInst::CreateIntegerCast(Shift, DestTy, false);
587 // FIXME: We should canonicalize to zext/trunc and remove this transform.
588 // Transform trunc(lshr (sext A), Cst) to ashr A, Cst to eliminate type
590 // It works because bits coming from sign extension have the same value as
591 // the sign bit of the original value; performing ashr instead of lshr
592 // generates bits of the same value as the sign bit.
593 if (Src->hasOneUse() &&
594 match(Src, m_LShr(m_SExt(m_Value(A)), m_ConstantInt(Cst)))) {
595 Value *SExt = cast<Instruction>(Src)->getOperand(0);
596 const unsigned SExtSize = SExt->getType()->getPrimitiveSizeInBits();
597 const unsigned ASize = A->getType()->getPrimitiveSizeInBits();
598 const unsigned CISize = CI.getType()->getPrimitiveSizeInBits();
599 const unsigned MaxAmt = SExtSize - std::max(CISize, ASize);
600 unsigned ShiftAmt = Cst->getZExtValue();
602 // This optimization can be only performed when zero bits generated by
603 // the original lshr aren't pulled into the value after truncation, so we
604 // can only shift by values no larger than the number of extension bits.
605 // FIXME: Instead of bailing when the shift is too large, use and to clear
607 if (ShiftAmt <= MaxAmt) {
609 return BinaryOperator::CreateAShr(A, ConstantInt::get(CI.getType(),
610 std::min(ShiftAmt, ASize - 1)));
611 if (SExt->hasOneUse()) {
612 Value *Shift = Builder.CreateAShr(A, std::min(ShiftAmt, ASize - 1));
613 Shift->takeName(Src);
614 return CastInst::CreateIntegerCast(Shift, CI.getType(), true);
619 if (Instruction *I = shrinkBitwiseLogic(CI))
622 if (Instruction *I = shrinkSplatShuffle(CI, Builder))
625 if (Instruction *I = shrinkInsertElt(CI, Builder))
628 if (Src->hasOneUse() && isa<IntegerType>(SrcTy) &&
629 shouldChangeType(SrcTy, DestTy)) {
630 // Transform "trunc (shl X, cst)" -> "shl (trunc X), cst" so long as the
631 // dest type is native and cst < dest size.
632 if (match(Src, m_Shl(m_Value(A), m_ConstantInt(Cst))) &&
633 !match(A, m_Shr(m_Value(), m_Constant()))) {
634 // Skip shifts of shift by constants. It undoes a combine in
635 // FoldShiftByConstant and is the extend in reg pattern.
636 const unsigned DestSize = DestTy->getScalarSizeInBits();
637 if (Cst->getValue().ult(DestSize)) {
638 Value *NewTrunc = Builder.CreateTrunc(A, DestTy, A->getName() + ".tr");
640 return BinaryOperator::Create(
641 Instruction::Shl, NewTrunc,
642 ConstantInt::get(DestTy, Cst->getValue().trunc(DestSize)));
647 if (Instruction *I = foldVecTruncToExtElt(CI, *this))
653 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, ZExtInst &CI,
655 // If we are just checking for a icmp eq of a single bit and zext'ing it
656 // to an integer, then shift the bit to the appropriate place and then
657 // cast to integer to avoid the comparison.
658 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
659 const APInt &Op1CV = Op1C->getValue();
661 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
662 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
663 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV.isNullValue()) ||
664 (ICI->getPredicate() == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
665 if (!DoTransform) return ICI;
667 Value *In = ICI->getOperand(0);
668 Value *Sh = ConstantInt::get(In->getType(),
669 In->getType()->getScalarSizeInBits() - 1);
670 In = Builder.CreateLShr(In, Sh, In->getName() + ".lobit");
671 if (In->getType() != CI.getType())
672 In = Builder.CreateIntCast(In, CI.getType(), false /*ZExt*/);
674 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
675 Constant *One = ConstantInt::get(In->getType(), 1);
676 In = Builder.CreateXor(In, One, In->getName() + ".not");
679 return replaceInstUsesWith(CI, In);
682 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
683 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
684 // zext (X == 1) to i32 --> X iff X has only the low bit set.
685 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
686 // zext (X != 0) to i32 --> X iff X has only the low bit set.
687 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
688 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
689 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
690 if ((Op1CV.isNullValue() || Op1CV.isPowerOf2()) &&
691 // This only works for EQ and NE
693 // If Op1C some other power of two, convert:
694 KnownBits Known = computeKnownBits(ICI->getOperand(0), 0, &CI);
696 APInt KnownZeroMask(~Known.Zero);
697 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
698 if (!DoTransform) return ICI;
700 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
701 if (!Op1CV.isNullValue() && (Op1CV != KnownZeroMask)) {
702 // (X&4) == 2 --> false
703 // (X&4) != 2 --> true
704 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()),
706 Res = ConstantExpr::getZExt(Res, CI.getType());
707 return replaceInstUsesWith(CI, Res);
710 uint32_t ShAmt = KnownZeroMask.logBase2();
711 Value *In = ICI->getOperand(0);
713 // Perform a logical shr by shiftamt.
714 // Insert the shift to put the result in the low bit.
715 In = Builder.CreateLShr(In, ConstantInt::get(In->getType(), ShAmt),
716 In->getName() + ".lobit");
719 if (!Op1CV.isNullValue() == isNE) { // Toggle the low bit.
720 Constant *One = ConstantInt::get(In->getType(), 1);
721 In = Builder.CreateXor(In, One);
724 if (CI.getType() == In->getType())
725 return replaceInstUsesWith(CI, In);
727 Value *IntCast = Builder.CreateIntCast(In, CI.getType(), false);
728 return replaceInstUsesWith(CI, IntCast);
733 // icmp ne A, B is equal to xor A, B when A and B only really have one bit.
734 // It is also profitable to transform icmp eq into not(xor(A, B)) because that
735 // may lead to additional simplifications.
736 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) {
737 if (IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) {
738 Value *LHS = ICI->getOperand(0);
739 Value *RHS = ICI->getOperand(1);
741 KnownBits KnownLHS = computeKnownBits(LHS, 0, &CI);
742 KnownBits KnownRHS = computeKnownBits(RHS, 0, &CI);
744 if (KnownLHS.Zero == KnownRHS.Zero && KnownLHS.One == KnownRHS.One) {
745 APInt KnownBits = KnownLHS.Zero | KnownLHS.One;
746 APInt UnknownBit = ~KnownBits;
747 if (UnknownBit.countPopulation() == 1) {
748 if (!DoTransform) return ICI;
750 Value *Result = Builder.CreateXor(LHS, RHS);
752 // Mask off any bits that are set and won't be shifted away.
753 if (KnownLHS.One.uge(UnknownBit))
754 Result = Builder.CreateAnd(Result,
755 ConstantInt::get(ITy, UnknownBit));
757 // Shift the bit we're testing down to the lsb.
758 Result = Builder.CreateLShr(
759 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros()));
761 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
762 Result = Builder.CreateXor(Result, ConstantInt::get(ITy, 1));
763 Result->takeName(ICI);
764 return replaceInstUsesWith(CI, Result);
773 /// Determine if the specified value can be computed in the specified wider type
774 /// and produce the same low bits. If not, return false.
776 /// If this function returns true, it can also return a non-zero number of bits
777 /// (in BitsToClear) which indicates that the value it computes is correct for
778 /// the zero extend, but that the additional BitsToClear bits need to be zero'd
779 /// out. For example, to promote something like:
781 /// %B = trunc i64 %A to i32
782 /// %C = lshr i32 %B, 8
783 /// %E = zext i32 %C to i64
785 /// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be
786 /// set to 8 to indicate that the promoted value needs to have bits 24-31
787 /// cleared in addition to bits 32-63. Since an 'and' will be generated to
788 /// clear the top bits anyway, doing this has no extra cost.
790 /// This function works on both vectors and scalars.
791 static bool canEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear,
792 InstCombiner &IC, Instruction *CxtI) {
794 if (isa<Constant>(V))
797 Instruction *I = dyn_cast<Instruction>(V);
798 if (!I) return false;
800 // If the input is a truncate from the destination type, we can trivially
802 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
805 // We can't extend or shrink something that has multiple uses: doing so would
806 // require duplicating the instruction in general, which isn't profitable.
807 if (!I->hasOneUse()) return false;
809 unsigned Opc = I->getOpcode(), Tmp;
811 case Instruction::ZExt: // zext(zext(x)) -> zext(x).
812 case Instruction::SExt: // zext(sext(x)) -> sext(x).
813 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x)
815 case Instruction::And:
816 case Instruction::Or:
817 case Instruction::Xor:
818 case Instruction::Add:
819 case Instruction::Sub:
820 case Instruction::Mul:
821 if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI) ||
822 !canEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI))
824 // These can all be promoted if neither operand has 'bits to clear'.
825 if (BitsToClear == 0 && Tmp == 0)
828 // If the operation is an AND/OR/XOR and the bits to clear are zero in the
829 // other side, BitsToClear is ok.
830 if (Tmp == 0 && I->isBitwiseLogicOp()) {
831 // We use MaskedValueIsZero here for generality, but the case we care
832 // about the most is constant RHS.
833 unsigned VSize = V->getType()->getScalarSizeInBits();
834 if (IC.MaskedValueIsZero(I->getOperand(1),
835 APInt::getHighBitsSet(VSize, BitsToClear),
840 // Otherwise, we don't know how to analyze this BitsToClear case yet.
843 case Instruction::Shl:
844 // We can promote shl(x, cst) if we can promote x. Since shl overwrites the
845 // upper bits we can reduce BitsToClear by the shift amount.
846 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
847 if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI))
849 uint64_t ShiftAmt = Amt->getZExtValue();
850 BitsToClear = ShiftAmt < BitsToClear ? BitsToClear - ShiftAmt : 0;
854 case Instruction::LShr:
855 // We can promote lshr(x, cst) if we can promote x. This requires the
856 // ultimate 'and' to clear out the high zero bits we're clearing out though.
857 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
858 if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI))
860 BitsToClear += Amt->getZExtValue();
861 if (BitsToClear > V->getType()->getScalarSizeInBits())
862 BitsToClear = V->getType()->getScalarSizeInBits();
865 // Cannot promote variable LSHR.
867 case Instruction::Select:
868 if (!canEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI) ||
869 !canEvaluateZExtd(I->getOperand(2), Ty, BitsToClear, IC, CxtI) ||
870 // TODO: If important, we could handle the case when the BitsToClear are
871 // known zero in the disagreeing side.
876 case Instruction::PHI: {
877 // We can change a phi if we can change all operands. Note that we never
878 // get into trouble with cyclic PHIs here because we only consider
879 // instructions with a single use.
880 PHINode *PN = cast<PHINode>(I);
881 if (!canEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear, IC, CxtI))
883 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
884 if (!canEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp, IC, CxtI) ||
885 // TODO: If important, we could handle the case when the BitsToClear
886 // are known zero in the disagreeing input.
892 // TODO: Can handle more cases here.
897 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
898 // If this zero extend is only used by a truncate, let the truncate be
899 // eliminated before we try to optimize this zext.
900 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back()))
903 // If one of the common conversion will work, do it.
904 if (Instruction *Result = commonCastTransforms(CI))
907 Value *Src = CI.getOperand(0);
908 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
910 // Attempt to extend the entire input expression tree to the destination
911 // type. Only do this if the dest type is a simple type, don't convert the
912 // expression tree to something weird like i93 unless the source is also
914 unsigned BitsToClear;
915 if ((DestTy->isVectorTy() || shouldChangeType(SrcTy, DestTy)) &&
916 canEvaluateZExtd(Src, DestTy, BitsToClear, *this, &CI)) {
917 assert(BitsToClear <= SrcTy->getScalarSizeInBits() &&
918 "Can't clear more bits than in SrcTy");
920 // Okay, we can transform this! Insert the new expression now.
921 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
922 " to avoid zero extend: " << CI << '\n');
923 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
924 assert(Res->getType() == DestTy);
926 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear;
927 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
929 // If the high bits are already filled with zeros, just replace this
930 // cast with the result.
931 if (MaskedValueIsZero(Res,
932 APInt::getHighBitsSet(DestBitSize,
933 DestBitSize-SrcBitsKept),
935 return replaceInstUsesWith(CI, Res);
937 // We need to emit an AND to clear the high bits.
938 Constant *C = ConstantInt::get(Res->getType(),
939 APInt::getLowBitsSet(DestBitSize, SrcBitsKept));
940 return BinaryOperator::CreateAnd(Res, C);
943 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
944 // types and if the sizes are just right we can convert this into a logical
945 // 'and' which will be much cheaper than the pair of casts.
946 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
947 // TODO: Subsume this into EvaluateInDifferentType.
949 // Get the sizes of the types involved. We know that the intermediate type
950 // will be smaller than A or C, but don't know the relation between A and C.
951 Value *A = CSrc->getOperand(0);
952 unsigned SrcSize = A->getType()->getScalarSizeInBits();
953 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
954 unsigned DstSize = CI.getType()->getScalarSizeInBits();
955 // If we're actually extending zero bits, then if
956 // SrcSize < DstSize: zext(a & mask)
957 // SrcSize == DstSize: a & mask
958 // SrcSize > DstSize: trunc(a) & mask
959 if (SrcSize < DstSize) {
960 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
961 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
962 Value *And = Builder.CreateAnd(A, AndConst, CSrc->getName() + ".mask");
963 return new ZExtInst(And, CI.getType());
966 if (SrcSize == DstSize) {
967 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
968 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
971 if (SrcSize > DstSize) {
972 Value *Trunc = Builder.CreateTrunc(A, CI.getType());
973 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
974 return BinaryOperator::CreateAnd(Trunc,
975 ConstantInt::get(Trunc->getType(),
980 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
981 return transformZExtICmp(ICI, CI);
983 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
984 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
985 // zext (or icmp, icmp) -> or (zext icmp), (zext icmp) if at least one
986 // of the (zext icmp) can be eliminated. If so, immediately perform the
987 // according elimination.
988 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
989 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
990 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
991 (transformZExtICmp(LHS, CI, false) ||
992 transformZExtICmp(RHS, CI, false))) {
993 // zext (or icmp, icmp) -> or (zext icmp), (zext icmp)
994 Value *LCast = Builder.CreateZExt(LHS, CI.getType(), LHS->getName());
995 Value *RCast = Builder.CreateZExt(RHS, CI.getType(), RHS->getName());
996 BinaryOperator *Or = BinaryOperator::Create(Instruction::Or, LCast, RCast);
998 // Perform the elimination.
999 if (auto *LZExt = dyn_cast<ZExtInst>(LCast))
1000 transformZExtICmp(LHS, *LZExt);
1001 if (auto *RZExt = dyn_cast<ZExtInst>(RCast))
1002 transformZExtICmp(RHS, *RZExt);
1008 // zext(trunc(X) & C) -> (X & zext(C)).
1012 match(SrcI, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Constant(C)))) &&
1013 X->getType() == CI.getType())
1014 return BinaryOperator::CreateAnd(X, ConstantExpr::getZExt(C, CI.getType()));
1016 // zext((trunc(X) & C) ^ C) -> ((X & zext(C)) ^ zext(C)).
1018 if (SrcI && match(SrcI, m_OneUse(m_Xor(m_Value(And), m_Constant(C)))) &&
1019 match(And, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Specific(C)))) &&
1020 X->getType() == CI.getType()) {
1021 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
1022 return BinaryOperator::CreateXor(Builder.CreateAnd(X, ZC), ZC);
1028 /// Transform (sext icmp) to bitwise / integer operations to eliminate the icmp.
1029 Instruction *InstCombiner::transformSExtICmp(ICmpInst *ICI, Instruction &CI) {
1030 Value *Op0 = ICI->getOperand(0), *Op1 = ICI->getOperand(1);
1031 ICmpInst::Predicate Pred = ICI->getPredicate();
1033 // Don't bother if Op1 isn't of vector or integer type.
1034 if (!Op1->getType()->isIntOrIntVectorTy())
1037 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
1038 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if negative
1039 // (x >s -1) ? -1 : 0 -> not (ashr x, 31) -> all ones if positive
1040 if ((Pred == ICmpInst::ICMP_SLT && Op1C->isNullValue()) ||
1041 (Pred == ICmpInst::ICMP_SGT && Op1C->isAllOnesValue())) {
1043 Value *Sh = ConstantInt::get(Op0->getType(),
1044 Op0->getType()->getScalarSizeInBits()-1);
1045 Value *In = Builder.CreateAShr(Op0, Sh, Op0->getName() + ".lobit");
1046 if (In->getType() != CI.getType())
1047 In = Builder.CreateIntCast(In, CI.getType(), true /*SExt*/);
1049 if (Pred == ICmpInst::ICMP_SGT)
1050 In = Builder.CreateNot(In, In->getName() + ".not");
1051 return replaceInstUsesWith(CI, In);
1055 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
1056 // If we know that only one bit of the LHS of the icmp can be set and we
1057 // have an equality comparison with zero or a power of 2, we can transform
1058 // the icmp and sext into bitwise/integer operations.
1059 if (ICI->hasOneUse() &&
1060 ICI->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){
1061 KnownBits Known = computeKnownBits(Op0, 0, &CI);
1063 APInt KnownZeroMask(~Known.Zero);
1064 if (KnownZeroMask.isPowerOf2()) {
1065 Value *In = ICI->getOperand(0);
1067 // If the icmp tests for a known zero bit we can constant fold it.
1068 if (!Op1C->isZero() && Op1C->getValue() != KnownZeroMask) {
1069 Value *V = Pred == ICmpInst::ICMP_NE ?
1070 ConstantInt::getAllOnesValue(CI.getType()) :
1071 ConstantInt::getNullValue(CI.getType());
1072 return replaceInstUsesWith(CI, V);
1075 if (!Op1C->isZero() == (Pred == ICmpInst::ICMP_NE)) {
1076 // sext ((x & 2^n) == 0) -> (x >> n) - 1
1077 // sext ((x & 2^n) != 2^n) -> (x >> n) - 1
1078 unsigned ShiftAmt = KnownZeroMask.countTrailingZeros();
1079 // Perform a right shift to place the desired bit in the LSB.
1081 In = Builder.CreateLShr(In,
1082 ConstantInt::get(In->getType(), ShiftAmt));
1084 // At this point "In" is either 1 or 0. Subtract 1 to turn
1085 // {1, 0} -> {0, -1}.
1086 In = Builder.CreateAdd(In,
1087 ConstantInt::getAllOnesValue(In->getType()),
1090 // sext ((x & 2^n) != 0) -> (x << bitwidth-n) a>> bitwidth-1
1091 // sext ((x & 2^n) == 2^n) -> (x << bitwidth-n) a>> bitwidth-1
1092 unsigned ShiftAmt = KnownZeroMask.countLeadingZeros();
1093 // Perform a left shift to place the desired bit in the MSB.
1095 In = Builder.CreateShl(In,
1096 ConstantInt::get(In->getType(), ShiftAmt));
1098 // Distribute the bit over the whole bit width.
1099 In = Builder.CreateAShr(In, ConstantInt::get(In->getType(),
1100 KnownZeroMask.getBitWidth() - 1), "sext");
1103 if (CI.getType() == In->getType())
1104 return replaceInstUsesWith(CI, In);
1105 return CastInst::CreateIntegerCast(In, CI.getType(), true/*SExt*/);
1113 /// Return true if we can take the specified value and return it as type Ty
1114 /// without inserting any new casts and without changing the value of the common
1115 /// low bits. This is used by code that tries to promote integer operations to
1116 /// a wider types will allow us to eliminate the extension.
1118 /// This function works on both vectors and scalars.
1120 static bool canEvaluateSExtd(Value *V, Type *Ty) {
1121 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() &&
1122 "Can't sign extend type to a smaller type");
1123 // If this is a constant, it can be trivially promoted.
1124 if (isa<Constant>(V))
1127 Instruction *I = dyn_cast<Instruction>(V);
1128 if (!I) return false;
1130 // If this is a truncate from the dest type, we can trivially eliminate it.
1131 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
1134 // We can't extend or shrink something that has multiple uses: doing so would
1135 // require duplicating the instruction in general, which isn't profitable.
1136 if (!I->hasOneUse()) return false;
1138 switch (I->getOpcode()) {
1139 case Instruction::SExt: // sext(sext(x)) -> sext(x)
1140 case Instruction::ZExt: // sext(zext(x)) -> zext(x)
1141 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x)
1143 case Instruction::And:
1144 case Instruction::Or:
1145 case Instruction::Xor:
1146 case Instruction::Add:
1147 case Instruction::Sub:
1148 case Instruction::Mul:
1149 // These operators can all arbitrarily be extended if their inputs can.
1150 return canEvaluateSExtd(I->getOperand(0), Ty) &&
1151 canEvaluateSExtd(I->getOperand(1), Ty);
1153 //case Instruction::Shl: TODO
1154 //case Instruction::LShr: TODO
1156 case Instruction::Select:
1157 return canEvaluateSExtd(I->getOperand(1), Ty) &&
1158 canEvaluateSExtd(I->getOperand(2), Ty);
1160 case Instruction::PHI: {
1161 // We can change a phi if we can change all operands. Note that we never
1162 // get into trouble with cyclic PHIs here because we only consider
1163 // instructions with a single use.
1164 PHINode *PN = cast<PHINode>(I);
1165 for (Value *IncValue : PN->incoming_values())
1166 if (!canEvaluateSExtd(IncValue, Ty)) return false;
1170 // TODO: Can handle more cases here.
1177 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
1178 // If this sign extend is only used by a truncate, let the truncate be
1179 // eliminated before we try to optimize this sext.
1180 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back()))
1183 if (Instruction *I = commonCastTransforms(CI))
1186 Value *Src = CI.getOperand(0);
1187 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
1189 // If we know that the value being extended is positive, we can use a zext
1191 KnownBits Known = computeKnownBits(Src, 0, &CI);
1192 if (Known.isNonNegative()) {
1193 Value *ZExt = Builder.CreateZExt(Src, DestTy);
1194 return replaceInstUsesWith(CI, ZExt);
1197 // Attempt to extend the entire input expression tree to the destination
1198 // type. Only do this if the dest type is a simple type, don't convert the
1199 // expression tree to something weird like i93 unless the source is also
1201 if ((DestTy->isVectorTy() || shouldChangeType(SrcTy, DestTy)) &&
1202 canEvaluateSExtd(Src, DestTy)) {
1203 // Okay, we can transform this! Insert the new expression now.
1204 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
1205 " to avoid sign extend: " << CI << '\n');
1206 Value *Res = EvaluateInDifferentType(Src, DestTy, true);
1207 assert(Res->getType() == DestTy);
1209 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1210 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1212 // If the high bits are already filled with sign bit, just replace this
1213 // cast with the result.
1214 if (ComputeNumSignBits(Res, 0, &CI) > DestBitSize - SrcBitSize)
1215 return replaceInstUsesWith(CI, Res);
1217 // We need to emit a shl + ashr to do the sign extend.
1218 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1219 return BinaryOperator::CreateAShr(Builder.CreateShl(Res, ShAmt, "sext"),
1223 // If the input is a trunc from the destination type, then turn sext(trunc(x))
1226 if (match(Src, m_OneUse(m_Trunc(m_Value(X)))) && X->getType() == DestTy) {
1227 // sext(trunc(X)) --> ashr(shl(X, C), C)
1228 unsigned SrcBitSize = SrcTy->getScalarSizeInBits();
1229 unsigned DestBitSize = DestTy->getScalarSizeInBits();
1230 Constant *ShAmt = ConstantInt::get(DestTy, DestBitSize - SrcBitSize);
1231 return BinaryOperator::CreateAShr(Builder.CreateShl(X, ShAmt), ShAmt);
1234 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
1235 return transformSExtICmp(ICI, CI);
1237 // If the input is a shl/ashr pair of a same constant, then this is a sign
1238 // extension from a smaller value. If we could trust arbitrary bitwidth
1239 // integers, we could turn this into a truncate to the smaller bit and then
1240 // use a sext for the whole extension. Since we don't, look deeper and check
1241 // for a truncate. If the source and dest are the same type, eliminate the
1242 // trunc and extend and just do shifts. For example, turn:
1243 // %a = trunc i32 %i to i8
1244 // %b = shl i8 %a, 6
1245 // %c = ashr i8 %b, 6
1246 // %d = sext i8 %c to i32
1248 // %a = shl i32 %i, 30
1249 // %d = ashr i32 %a, 30
1251 // TODO: Eventually this could be subsumed by EvaluateInDifferentType.
1252 ConstantInt *BA = nullptr, *CA = nullptr;
1253 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)),
1254 m_ConstantInt(CA))) &&
1255 BA == CA && A->getType() == CI.getType()) {
1256 unsigned MidSize = Src->getType()->getScalarSizeInBits();
1257 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
1258 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
1259 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
1260 A = Builder.CreateShl(A, ShAmtV, CI.getName());
1261 return BinaryOperator::CreateAShr(A, ShAmtV);
1268 /// Return a Constant* for the specified floating-point constant if it fits
1269 /// in the specified FP type without changing its value.
1270 static Constant *fitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
1272 APFloat F = CFP->getValueAPF();
1273 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
1275 return ConstantFP::get(CFP->getContext(), F);
1279 /// Look through floating-point extensions until we get the source value.
1280 static Value *lookThroughFPExtensions(Value *V) {
1281 while (auto *FPExt = dyn_cast<FPExtInst>(V))
1282 V = FPExt->getOperand(0);
1284 // If this value is a constant, return the constant in the smallest FP type
1285 // that can accurately represent it. This allows us to turn
1286 // (float)((double)X+2.0) into x+2.0f.
1287 if (auto *CFP = dyn_cast<ConstantFP>(V)) {
1288 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext()))
1289 return V; // No constant folding of this.
1290 // See if the value can be truncated to half and then reextended.
1291 if (Value *V = fitsInFPType(CFP, APFloat::IEEEhalf()))
1293 // See if the value can be truncated to float and then reextended.
1294 if (Value *V = fitsInFPType(CFP, APFloat::IEEEsingle()))
1296 if (CFP->getType()->isDoubleTy())
1297 return V; // Won't shrink.
1298 if (Value *V = fitsInFPType(CFP, APFloat::IEEEdouble()))
1300 // Don't try to shrink to various long double types.
1306 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
1307 if (Instruction *I = commonCastTransforms(CI))
1309 // If we have fptrunc(OpI (fpextend x), (fpextend y)), we would like to
1310 // simplify this expression to avoid one or more of the trunc/extend
1311 // operations if we can do so without changing the numerical results.
1313 // The exact manner in which the widths of the operands interact to limit
1314 // what we can and cannot do safely varies from operation to operation, and
1315 // is explained below in the various case statements.
1316 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
1317 if (OpI && OpI->hasOneUse()) {
1318 Value *LHSOrig = lookThroughFPExtensions(OpI->getOperand(0));
1319 Value *RHSOrig = lookThroughFPExtensions(OpI->getOperand(1));
1320 unsigned OpWidth = OpI->getType()->getFPMantissaWidth();
1321 unsigned LHSWidth = LHSOrig->getType()->getFPMantissaWidth();
1322 unsigned RHSWidth = RHSOrig->getType()->getFPMantissaWidth();
1323 unsigned SrcWidth = std::max(LHSWidth, RHSWidth);
1324 unsigned DstWidth = CI.getType()->getFPMantissaWidth();
1325 switch (OpI->getOpcode()) {
1327 case Instruction::FAdd:
1328 case Instruction::FSub:
1329 // For addition and subtraction, the infinitely precise result can
1330 // essentially be arbitrarily wide; proving that double rounding
1331 // will not occur because the result of OpI is exact (as we will for
1332 // FMul, for example) is hopeless. However, we *can* nonetheless
1333 // frequently know that double rounding cannot occur (or that it is
1334 // innocuous) by taking advantage of the specific structure of
1335 // infinitely-precise results that admit double rounding.
1337 // Specifically, if OpWidth >= 2*DstWdith+1 and DstWidth is sufficient
1338 // to represent both sources, we can guarantee that the double
1339 // rounding is innocuous (See p50 of Figueroa's 2000 PhD thesis,
1340 // "A Rigorous Framework for Fully Supporting the IEEE Standard ..."
1341 // for proof of this fact).
1343 // Note: Figueroa does not consider the case where DstFormat !=
1344 // SrcFormat. It's possible (likely even!) that this analysis
1345 // could be tightened for those cases, but they are rare (the main
1346 // case of interest here is (float)((double)float + float)).
1347 if (OpWidth >= 2*DstWidth+1 && DstWidth >= SrcWidth) {
1348 if (LHSOrig->getType() != CI.getType())
1349 LHSOrig = Builder.CreateFPExt(LHSOrig, CI.getType());
1350 if (RHSOrig->getType() != CI.getType())
1351 RHSOrig = Builder.CreateFPExt(RHSOrig, CI.getType());
1353 BinaryOperator::Create(OpI->getOpcode(), LHSOrig, RHSOrig);
1354 RI->copyFastMathFlags(OpI);
1358 case Instruction::FMul:
1359 // For multiplication, the infinitely precise result has at most
1360 // LHSWidth + RHSWidth significant bits; if OpWidth is sufficient
1361 // that such a value can be exactly represented, then no double
1362 // rounding can possibly occur; we can safely perform the operation
1363 // in the destination format if it can represent both sources.
1364 if (OpWidth >= LHSWidth + RHSWidth && DstWidth >= SrcWidth) {
1365 if (LHSOrig->getType() != CI.getType())
1366 LHSOrig = Builder.CreateFPExt(LHSOrig, CI.getType());
1367 if (RHSOrig->getType() != CI.getType())
1368 RHSOrig = Builder.CreateFPExt(RHSOrig, CI.getType());
1370 BinaryOperator::CreateFMul(LHSOrig, RHSOrig);
1371 RI->copyFastMathFlags(OpI);
1375 case Instruction::FDiv:
1376 // For division, we use again use the bound from Figueroa's
1377 // dissertation. I am entirely certain that this bound can be
1378 // tightened in the unbalanced operand case by an analysis based on
1379 // the diophantine rational approximation bound, but the well-known
1380 // condition used here is a good conservative first pass.
1381 // TODO: Tighten bound via rigorous analysis of the unbalanced case.
1382 if (OpWidth >= 2*DstWidth && DstWidth >= SrcWidth) {
1383 if (LHSOrig->getType() != CI.getType())
1384 LHSOrig = Builder.CreateFPExt(LHSOrig, CI.getType());
1385 if (RHSOrig->getType() != CI.getType())
1386 RHSOrig = Builder.CreateFPExt(RHSOrig, CI.getType());
1388 BinaryOperator::CreateFDiv(LHSOrig, RHSOrig);
1389 RI->copyFastMathFlags(OpI);
1393 case Instruction::FRem:
1394 // Remainder is straightforward. Remainder is always exact, so the
1395 // type of OpI doesn't enter into things at all. We simply evaluate
1396 // in whichever source type is larger, then convert to the
1397 // destination type.
1398 if (SrcWidth == OpWidth)
1400 if (LHSWidth < SrcWidth)
1401 LHSOrig = Builder.CreateFPExt(LHSOrig, RHSOrig->getType());
1402 else if (RHSWidth <= SrcWidth)
1403 RHSOrig = Builder.CreateFPExt(RHSOrig, LHSOrig->getType());
1404 if (LHSOrig != OpI->getOperand(0) || RHSOrig != OpI->getOperand(1)) {
1405 Value *ExactResult = Builder.CreateFRem(LHSOrig, RHSOrig);
1406 if (Instruction *RI = dyn_cast<Instruction>(ExactResult))
1407 RI->copyFastMathFlags(OpI);
1408 return CastInst::CreateFPCast(ExactResult, CI.getType());
1412 // (fptrunc (fneg x)) -> (fneg (fptrunc x))
1413 if (BinaryOperator::isFNeg(OpI)) {
1414 Value *InnerTrunc = Builder.CreateFPTrunc(OpI->getOperand(1),
1416 Instruction *RI = BinaryOperator::CreateFNeg(InnerTrunc);
1417 RI->copyFastMathFlags(OpI);
1422 // (fptrunc (select cond, R1, Cst)) -->
1423 // (select cond, (fptrunc R1), (fptrunc Cst))
1425 // - but only if this isn't part of a min/max operation, else we'll
1426 // ruin min/max canonical form which is to have the select and
1427 // compare's operands be of the same type with no casts to look through.
1429 SelectInst *SI = dyn_cast<SelectInst>(CI.getOperand(0));
1431 (isa<ConstantFP>(SI->getOperand(1)) ||
1432 isa<ConstantFP>(SI->getOperand(2))) &&
1433 matchSelectPattern(SI, LHS, RHS).Flavor == SPF_UNKNOWN) {
1434 Value *LHSTrunc = Builder.CreateFPTrunc(SI->getOperand(1), CI.getType());
1435 Value *RHSTrunc = Builder.CreateFPTrunc(SI->getOperand(2), CI.getType());
1436 return SelectInst::Create(SI->getOperand(0), LHSTrunc, RHSTrunc);
1439 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI.getOperand(0));
1441 switch (II->getIntrinsicID()) {
1443 case Intrinsic::fabs:
1444 case Intrinsic::ceil:
1445 case Intrinsic::floor:
1446 case Intrinsic::rint:
1447 case Intrinsic::round:
1448 case Intrinsic::nearbyint:
1449 case Intrinsic::trunc: {
1450 Value *Src = II->getArgOperand(0);
1451 if (!Src->hasOneUse())
1454 // Except for fabs, this transformation requires the input of the unary FP
1455 // operation to be itself an fpext from the type to which we're
1457 if (II->getIntrinsicID() != Intrinsic::fabs) {
1458 FPExtInst *FPExtSrc = dyn_cast<FPExtInst>(Src);
1459 if (!FPExtSrc || FPExtSrc->getOperand(0)->getType() != CI.getType())
1463 // Do unary FP operation on smaller type.
1464 // (fptrunc (fabs x)) -> (fabs (fptrunc x))
1465 Value *InnerTrunc = Builder.CreateFPTrunc(Src, CI.getType());
1466 Type *IntrinsicType[] = { CI.getType() };
1467 Function *Overload = Intrinsic::getDeclaration(
1468 CI.getModule(), II->getIntrinsicID(), IntrinsicType);
1470 SmallVector<OperandBundleDef, 1> OpBundles;
1471 II->getOperandBundlesAsDefs(OpBundles);
1473 Value *Args[] = { InnerTrunc };
1474 CallInst *NewCI = CallInst::Create(Overload, Args,
1475 OpBundles, II->getName());
1476 NewCI->copyFastMathFlags(II);
1482 if (Instruction *I = shrinkInsertElt(CI, Builder))
1488 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
1489 return commonCastTransforms(CI);
1492 // fpto{s/u}i({u/s}itofp(X)) --> X or zext(X) or sext(X) or trunc(X)
1493 // This is safe if the intermediate type has enough bits in its mantissa to
1494 // accurately represent all values of X. For example, this won't work with
1495 // i64 -> float -> i64.
1496 Instruction *InstCombiner::FoldItoFPtoI(Instruction &FI) {
1497 if (!isa<UIToFPInst>(FI.getOperand(0)) && !isa<SIToFPInst>(FI.getOperand(0)))
1499 Instruction *OpI = cast<Instruction>(FI.getOperand(0));
1501 Value *SrcI = OpI->getOperand(0);
1502 Type *FITy = FI.getType();
1503 Type *OpITy = OpI->getType();
1504 Type *SrcTy = SrcI->getType();
1505 bool IsInputSigned = isa<SIToFPInst>(OpI);
1506 bool IsOutputSigned = isa<FPToSIInst>(FI);
1508 // We can safely assume the conversion won't overflow the output range,
1509 // because (for example) (uint8_t)18293.f is undefined behavior.
1511 // Since we can assume the conversion won't overflow, our decision as to
1512 // whether the input will fit in the float should depend on the minimum
1513 // of the input range and output range.
1515 // This means this is also safe for a signed input and unsigned output, since
1516 // a negative input would lead to undefined behavior.
1517 int InputSize = (int)SrcTy->getScalarSizeInBits() - IsInputSigned;
1518 int OutputSize = (int)FITy->getScalarSizeInBits() - IsOutputSigned;
1519 int ActualSize = std::min(InputSize, OutputSize);
1521 if (ActualSize <= OpITy->getFPMantissaWidth()) {
1522 if (FITy->getScalarSizeInBits() > SrcTy->getScalarSizeInBits()) {
1523 if (IsInputSigned && IsOutputSigned)
1524 return new SExtInst(SrcI, FITy);
1525 return new ZExtInst(SrcI, FITy);
1527 if (FITy->getScalarSizeInBits() < SrcTy->getScalarSizeInBits())
1528 return new TruncInst(SrcI, FITy);
1530 return replaceInstUsesWith(FI, SrcI);
1531 return new BitCastInst(SrcI, FITy);
1536 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
1537 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1539 return commonCastTransforms(FI);
1541 if (Instruction *I = FoldItoFPtoI(FI))
1544 return commonCastTransforms(FI);
1547 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
1548 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1550 return commonCastTransforms(FI);
1552 if (Instruction *I = FoldItoFPtoI(FI))
1555 return commonCastTransforms(FI);
1558 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
1559 return commonCastTransforms(CI);
1562 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
1563 return commonCastTransforms(CI);
1566 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
1567 // If the source integer type is not the intptr_t type for this target, do a
1568 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the
1569 // cast to be exposed to other transforms.
1570 unsigned AS = CI.getAddressSpace();
1571 if (CI.getOperand(0)->getType()->getScalarSizeInBits() !=
1572 DL.getPointerSizeInBits(AS)) {
1573 Type *Ty = DL.getIntPtrType(CI.getContext(), AS);
1574 if (CI.getType()->isVectorTy()) // Handle vectors of pointers.
1575 Ty = VectorType::get(Ty, CI.getType()->getVectorNumElements());
1577 Value *P = Builder.CreateZExtOrTrunc(CI.getOperand(0), Ty);
1578 return new IntToPtrInst(P, CI.getType());
1581 if (Instruction *I = commonCastTransforms(CI))
1587 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
1588 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
1589 Value *Src = CI.getOperand(0);
1591 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
1592 // If casting the result of a getelementptr instruction with no offset, turn
1593 // this into a cast of the original pointer!
1594 if (GEP->hasAllZeroIndices() &&
1595 // If CI is an addrspacecast and GEP changes the poiner type, merging
1596 // GEP into CI would undo canonicalizing addrspacecast with different
1597 // pointer types, causing infinite loops.
1598 (!isa<AddrSpaceCastInst>(CI) ||
1599 GEP->getType() == GEP->getPointerOperandType())) {
1600 // Changing the cast operand is usually not a good idea but it is safe
1601 // here because the pointer operand is being replaced with another
1602 // pointer operand so the opcode doesn't need to change.
1604 CI.setOperand(0, GEP->getOperand(0));
1609 return commonCastTransforms(CI);
1612 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
1613 // If the destination integer type is not the intptr_t type for this target,
1614 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast
1615 // to be exposed to other transforms.
1617 Type *Ty = CI.getType();
1618 unsigned AS = CI.getPointerAddressSpace();
1620 if (Ty->getScalarSizeInBits() == DL.getPointerSizeInBits(AS))
1621 return commonPointerCastTransforms(CI);
1623 Type *PtrTy = DL.getIntPtrType(CI.getContext(), AS);
1624 if (Ty->isVectorTy()) // Handle vectors of pointers.
1625 PtrTy = VectorType::get(PtrTy, Ty->getVectorNumElements());
1627 Value *P = Builder.CreatePtrToInt(CI.getOperand(0), PtrTy);
1628 return CastInst::CreateIntegerCast(P, Ty, /*isSigned=*/false);
1631 /// This input value (which is known to have vector type) is being zero extended
1632 /// or truncated to the specified vector type.
1633 /// Try to replace it with a shuffle (and vector/vector bitcast) if possible.
1635 /// The source and destination vector types may have different element types.
1636 static Instruction *optimizeVectorResize(Value *InVal, VectorType *DestTy,
1638 // We can only do this optimization if the output is a multiple of the input
1639 // element size, or the input is a multiple of the output element size.
1640 // Convert the input type to have the same element type as the output.
1641 VectorType *SrcTy = cast<VectorType>(InVal->getType());
1643 if (SrcTy->getElementType() != DestTy->getElementType()) {
1644 // The input types don't need to be identical, but for now they must be the
1645 // same size. There is no specific reason we couldn't handle things like
1646 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten
1648 if (SrcTy->getElementType()->getPrimitiveSizeInBits() !=
1649 DestTy->getElementType()->getPrimitiveSizeInBits())
1652 SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements());
1653 InVal = IC.Builder.CreateBitCast(InVal, SrcTy);
1656 // Now that the element types match, get the shuffle mask and RHS of the
1657 // shuffle to use, which depends on whether we're increasing or decreasing the
1658 // size of the input.
1659 SmallVector<uint32_t, 16> ShuffleMask;
1662 if (SrcTy->getNumElements() > DestTy->getNumElements()) {
1663 // If we're shrinking the number of elements, just shuffle in the low
1664 // elements from the input and use undef as the second shuffle input.
1665 V2 = UndefValue::get(SrcTy);
1666 for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i)
1667 ShuffleMask.push_back(i);
1670 // If we're increasing the number of elements, shuffle in all of the
1671 // elements from InVal and fill the rest of the result elements with zeros
1672 // from a constant zero.
1673 V2 = Constant::getNullValue(SrcTy);
1674 unsigned SrcElts = SrcTy->getNumElements();
1675 for (unsigned i = 0, e = SrcElts; i != e; ++i)
1676 ShuffleMask.push_back(i);
1678 // The excess elements reference the first element of the zero input.
1679 for (unsigned i = 0, e = DestTy->getNumElements()-SrcElts; i != e; ++i)
1680 ShuffleMask.push_back(SrcElts);
1683 return new ShuffleVectorInst(InVal, V2,
1684 ConstantDataVector::get(V2->getContext(),
1688 static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) {
1689 return Value % Ty->getPrimitiveSizeInBits() == 0;
1692 static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) {
1693 return Value / Ty->getPrimitiveSizeInBits();
1696 /// V is a value which is inserted into a vector of VecEltTy.
1697 /// Look through the value to see if we can decompose it into
1698 /// insertions into the vector. See the example in the comment for
1699 /// OptimizeIntegerToVectorInsertions for the pattern this handles.
1700 /// The type of V is always a non-zero multiple of VecEltTy's size.
1701 /// Shift is the number of bits between the lsb of V and the lsb of
1704 /// This returns false if the pattern can't be matched or true if it can,
1705 /// filling in Elements with the elements found here.
1706 static bool collectInsertionElements(Value *V, unsigned Shift,
1707 SmallVectorImpl<Value *> &Elements,
1708 Type *VecEltTy, bool isBigEndian) {
1709 assert(isMultipleOfTypeSize(Shift, VecEltTy) &&
1710 "Shift should be a multiple of the element type size");
1712 // Undef values never contribute useful bits to the result.
1713 if (isa<UndefValue>(V)) return true;
1715 // If we got down to a value of the right type, we win, try inserting into the
1717 if (V->getType() == VecEltTy) {
1718 // Inserting null doesn't actually insert any elements.
1719 if (Constant *C = dyn_cast<Constant>(V))
1720 if (C->isNullValue())
1723 unsigned ElementIndex = getTypeSizeIndex(Shift, VecEltTy);
1725 ElementIndex = Elements.size() - ElementIndex - 1;
1727 // Fail if multiple elements are inserted into this slot.
1728 if (Elements[ElementIndex])
1731 Elements[ElementIndex] = V;
1735 if (Constant *C = dyn_cast<Constant>(V)) {
1736 // Figure out the # elements this provides, and bitcast it or slice it up
1738 unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(),
1740 // If the constant is the size of a vector element, we just need to bitcast
1741 // it to the right type so it gets properly inserted.
1743 return collectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy),
1744 Shift, Elements, VecEltTy, isBigEndian);
1746 // Okay, this is a constant that covers multiple elements. Slice it up into
1747 // pieces and insert each element-sized piece into the vector.
1748 if (!isa<IntegerType>(C->getType()))
1749 C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(),
1750 C->getType()->getPrimitiveSizeInBits()));
1751 unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits();
1752 Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize);
1754 for (unsigned i = 0; i != NumElts; ++i) {
1755 unsigned ShiftI = Shift+i*ElementSize;
1756 Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(),
1758 Piece = ConstantExpr::getTrunc(Piece, ElementIntTy);
1759 if (!collectInsertionElements(Piece, ShiftI, Elements, VecEltTy,
1766 if (!V->hasOneUse()) return false;
1768 Instruction *I = dyn_cast<Instruction>(V);
1769 if (!I) return false;
1770 switch (I->getOpcode()) {
1771 default: return false; // Unhandled case.
1772 case Instruction::BitCast:
1773 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
1775 case Instruction::ZExt:
1776 if (!isMultipleOfTypeSize(
1777 I->getOperand(0)->getType()->getPrimitiveSizeInBits(),
1780 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
1782 case Instruction::Or:
1783 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
1785 collectInsertionElements(I->getOperand(1), Shift, Elements, VecEltTy,
1787 case Instruction::Shl: {
1788 // Must be shifting by a constant that is a multiple of the element size.
1789 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1));
1790 if (!CI) return false;
1791 Shift += CI->getZExtValue();
1792 if (!isMultipleOfTypeSize(Shift, VecEltTy)) return false;
1793 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
1801 /// If the input is an 'or' instruction, we may be doing shifts and ors to
1802 /// assemble the elements of the vector manually.
1803 /// Try to rip the code out and replace it with insertelements. This is to
1804 /// optimize code like this:
1806 /// %tmp37 = bitcast float %inc to i32
1807 /// %tmp38 = zext i32 %tmp37 to i64
1808 /// %tmp31 = bitcast float %inc5 to i32
1809 /// %tmp32 = zext i32 %tmp31 to i64
1810 /// %tmp33 = shl i64 %tmp32, 32
1811 /// %ins35 = or i64 %tmp33, %tmp38
1812 /// %tmp43 = bitcast i64 %ins35 to <2 x float>
1814 /// Into two insertelements that do "buildvector{%inc, %inc5}".
1815 static Value *optimizeIntegerToVectorInsertions(BitCastInst &CI,
1817 VectorType *DestVecTy = cast<VectorType>(CI.getType());
1818 Value *IntInput = CI.getOperand(0);
1820 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements());
1821 if (!collectInsertionElements(IntInput, 0, Elements,
1822 DestVecTy->getElementType(),
1823 IC.getDataLayout().isBigEndian()))
1826 // If we succeeded, we know that all of the element are specified by Elements
1827 // or are zero if Elements has a null entry. Recast this as a set of
1829 Value *Result = Constant::getNullValue(CI.getType());
1830 for (unsigned i = 0, e = Elements.size(); i != e; ++i) {
1831 if (!Elements[i]) continue; // Unset element.
1833 Result = IC.Builder.CreateInsertElement(Result, Elements[i],
1834 IC.Builder.getInt32(i));
1840 /// Canonicalize scalar bitcasts of extracted elements into a bitcast of the
1841 /// vector followed by extract element. The backend tends to handle bitcasts of
1842 /// vectors better than bitcasts of scalars because vector registers are
1843 /// usually not type-specific like scalar integer or scalar floating-point.
1844 static Instruction *canonicalizeBitCastExtElt(BitCastInst &BitCast,
1846 // TODO: Create and use a pattern matcher for ExtractElementInst.
1847 auto *ExtElt = dyn_cast<ExtractElementInst>(BitCast.getOperand(0));
1848 if (!ExtElt || !ExtElt->hasOneUse())
1851 // The bitcast must be to a vectorizable type, otherwise we can't make a new
1852 // type to extract from.
1853 Type *DestType = BitCast.getType();
1854 if (!VectorType::isValidElementType(DestType))
1857 unsigned NumElts = ExtElt->getVectorOperandType()->getNumElements();
1858 auto *NewVecType = VectorType::get(DestType, NumElts);
1859 auto *NewBC = IC.Builder.CreateBitCast(ExtElt->getVectorOperand(),
1861 return ExtractElementInst::Create(NewBC, ExtElt->getIndexOperand());
1864 /// Change the type of a bitwise logic operation if we can eliminate a bitcast.
1865 static Instruction *foldBitCastBitwiseLogic(BitCastInst &BitCast,
1866 InstCombiner::BuilderTy &Builder) {
1867 Type *DestTy = BitCast.getType();
1869 if (!DestTy->isIntOrIntVectorTy() ||
1870 !match(BitCast.getOperand(0), m_OneUse(m_BinOp(BO))) ||
1871 !BO->isBitwiseLogicOp())
1874 // FIXME: This transform is restricted to vector types to avoid backend
1875 // problems caused by creating potentially illegal operations. If a fix-up is
1876 // added to handle that situation, we can remove this check.
1877 if (!DestTy->isVectorTy() || !BO->getType()->isVectorTy())
1881 if (match(BO->getOperand(0), m_OneUse(m_BitCast(m_Value(X)))) &&
1882 X->getType() == DestTy && !isa<Constant>(X)) {
1883 // bitcast(logic(bitcast(X), Y)) --> logic'(X, bitcast(Y))
1884 Value *CastedOp1 = Builder.CreateBitCast(BO->getOperand(1), DestTy);
1885 return BinaryOperator::Create(BO->getOpcode(), X, CastedOp1);
1888 if (match(BO->getOperand(1), m_OneUse(m_BitCast(m_Value(X)))) &&
1889 X->getType() == DestTy && !isa<Constant>(X)) {
1890 // bitcast(logic(Y, bitcast(X))) --> logic'(bitcast(Y), X)
1891 Value *CastedOp0 = Builder.CreateBitCast(BO->getOperand(0), DestTy);
1892 return BinaryOperator::Create(BO->getOpcode(), CastedOp0, X);
1895 // Canonicalize vector bitcasts to come before vector bitwise logic with a
1896 // constant. This eases recognition of special constants for later ops.
1898 // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
1900 if (match(BO->getOperand(1), m_Constant(C))) {
1901 // bitcast (logic X, C) --> logic (bitcast X, C')
1902 Value *CastedOp0 = Builder.CreateBitCast(BO->getOperand(0), DestTy);
1903 Value *CastedC = ConstantExpr::getBitCast(C, DestTy);
1904 return BinaryOperator::Create(BO->getOpcode(), CastedOp0, CastedC);
1910 /// Change the type of a select if we can eliminate a bitcast.
1911 static Instruction *foldBitCastSelect(BitCastInst &BitCast,
1912 InstCombiner::BuilderTy &Builder) {
1913 Value *Cond, *TVal, *FVal;
1914 if (!match(BitCast.getOperand(0),
1915 m_OneUse(m_Select(m_Value(Cond), m_Value(TVal), m_Value(FVal)))))
1918 // A vector select must maintain the same number of elements in its operands.
1919 Type *CondTy = Cond->getType();
1920 Type *DestTy = BitCast.getType();
1921 if (CondTy->isVectorTy()) {
1922 if (!DestTy->isVectorTy())
1924 if (DestTy->getVectorNumElements() != CondTy->getVectorNumElements())
1928 // FIXME: This transform is restricted from changing the select between
1929 // scalars and vectors to avoid backend problems caused by creating
1930 // potentially illegal operations. If a fix-up is added to handle that
1931 // situation, we can remove this check.
1932 if (DestTy->isVectorTy() != TVal->getType()->isVectorTy())
1935 auto *Sel = cast<Instruction>(BitCast.getOperand(0));
1937 if (match(TVal, m_OneUse(m_BitCast(m_Value(X)))) && X->getType() == DestTy &&
1938 !isa<Constant>(X)) {
1939 // bitcast(select(Cond, bitcast(X), Y)) --> select'(Cond, X, bitcast(Y))
1940 Value *CastedVal = Builder.CreateBitCast(FVal, DestTy);
1941 return SelectInst::Create(Cond, X, CastedVal, "", nullptr, Sel);
1944 if (match(FVal, m_OneUse(m_BitCast(m_Value(X)))) && X->getType() == DestTy &&
1945 !isa<Constant>(X)) {
1946 // bitcast(select(Cond, Y, bitcast(X))) --> select'(Cond, bitcast(Y), X)
1947 Value *CastedVal = Builder.CreateBitCast(TVal, DestTy);
1948 return SelectInst::Create(Cond, CastedVal, X, "", nullptr, Sel);
1954 /// Check if all users of CI are StoreInsts.
1955 static bool hasStoreUsersOnly(CastInst &CI) {
1956 for (User *U : CI.users()) {
1957 if (!isa<StoreInst>(U))
1963 /// This function handles following case
1969 /// All the related PHI nodes can be replaced by new PHI nodes with type A.
1970 /// The uses of \p CI can be changed to the new PHI node corresponding to \p PN.
1971 Instruction *InstCombiner::optimizeBitCastFromPhi(CastInst &CI, PHINode *PN) {
1972 // BitCast used by Store can be handled in InstCombineLoadStoreAlloca.cpp.
1973 if (hasStoreUsersOnly(CI))
1976 Value *Src = CI.getOperand(0);
1977 Type *SrcTy = Src->getType(); // Type B
1978 Type *DestTy = CI.getType(); // Type A
1980 SmallVector<PHINode *, 4> PhiWorklist;
1981 SmallSetVector<PHINode *, 4> OldPhiNodes;
1983 // Find all of the A->B casts and PHI nodes.
1984 // We need to inpect all related PHI nodes, but PHIs can be cyclic, so
1985 // OldPhiNodes is used to track all known PHI nodes, before adding a new
1986 // PHI to PhiWorklist, it is checked against and added to OldPhiNodes first.
1987 PhiWorklist.push_back(PN);
1988 OldPhiNodes.insert(PN);
1989 while (!PhiWorklist.empty()) {
1990 auto *OldPN = PhiWorklist.pop_back_val();
1991 for (Value *IncValue : OldPN->incoming_values()) {
1992 if (isa<Constant>(IncValue))
1995 if (auto *LI = dyn_cast<LoadInst>(IncValue)) {
1996 // If there is a sequence of one or more load instructions, each loaded
1997 // value is used as address of later load instruction, bitcast is
1998 // necessary to change the value type, don't optimize it. For
1999 // simplicity we give up if the load address comes from another load.
2000 Value *Addr = LI->getOperand(0);
2001 if (Addr == &CI || isa<LoadInst>(Addr))
2003 if (LI->hasOneUse() && LI->isSimple())
2005 // If a LoadInst has more than one use, changing the type of loaded
2006 // value may create another bitcast.
2010 if (auto *PNode = dyn_cast<PHINode>(IncValue)) {
2011 if (OldPhiNodes.insert(PNode))
2012 PhiWorklist.push_back(PNode);
2016 auto *BCI = dyn_cast<BitCastInst>(IncValue);
2017 // We can't handle other instructions.
2021 // Verify it's a A->B cast.
2022 Type *TyA = BCI->getOperand(0)->getType();
2023 Type *TyB = BCI->getType();
2024 if (TyA != DestTy || TyB != SrcTy)
2029 // For each old PHI node, create a corresponding new PHI node with a type A.
2030 SmallDenseMap<PHINode *, PHINode *> NewPNodes;
2031 for (auto *OldPN : OldPhiNodes) {
2032 Builder.SetInsertPoint(OldPN);
2033 PHINode *NewPN = Builder.CreatePHI(DestTy, OldPN->getNumOperands());
2034 NewPNodes[OldPN] = NewPN;
2037 // Fill in the operands of new PHI nodes.
2038 for (auto *OldPN : OldPhiNodes) {
2039 PHINode *NewPN = NewPNodes[OldPN];
2040 for (unsigned j = 0, e = OldPN->getNumOperands(); j != e; ++j) {
2041 Value *V = OldPN->getOperand(j);
2042 Value *NewV = nullptr;
2043 if (auto *C = dyn_cast<Constant>(V)) {
2044 NewV = ConstantExpr::getBitCast(C, DestTy);
2045 } else if (auto *LI = dyn_cast<LoadInst>(V)) {
2046 Builder.SetInsertPoint(LI->getNextNode());
2047 NewV = Builder.CreateBitCast(LI, DestTy);
2049 } else if (auto *BCI = dyn_cast<BitCastInst>(V)) {
2050 NewV = BCI->getOperand(0);
2051 } else if (auto *PrevPN = dyn_cast<PHINode>(V)) {
2052 NewV = NewPNodes[PrevPN];
2055 NewPN->addIncoming(NewV, OldPN->getIncomingBlock(j));
2059 // If there is a store with type B, change it to type A.
2060 for (User *U : PN->users()) {
2061 auto *SI = dyn_cast<StoreInst>(U);
2062 if (SI && SI->isSimple() && SI->getOperand(0) == PN) {
2063 Builder.SetInsertPoint(SI);
2065 cast<BitCastInst>(Builder.CreateBitCast(NewPNodes[PN], SrcTy));
2066 SI->setOperand(0, NewBC);
2068 assert(hasStoreUsersOnly(*NewBC));
2072 return replaceInstUsesWith(CI, NewPNodes[PN]);
2075 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
2076 // If the operands are integer typed then apply the integer transforms,
2077 // otherwise just apply the common ones.
2078 Value *Src = CI.getOperand(0);
2079 Type *SrcTy = Src->getType();
2080 Type *DestTy = CI.getType();
2082 // Get rid of casts from one type to the same type. These are useless and can
2083 // be replaced by the operand.
2084 if (DestTy == Src->getType())
2085 return replaceInstUsesWith(CI, Src);
2087 if (PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
2088 PointerType *SrcPTy = cast<PointerType>(SrcTy);
2089 Type *DstElTy = DstPTy->getElementType();
2090 Type *SrcElTy = SrcPTy->getElementType();
2092 // If we are casting a alloca to a pointer to a type of the same
2093 // size, rewrite the allocation instruction to allocate the "right" type.
2094 // There is no need to modify malloc calls because it is their bitcast that
2095 // needs to be cleaned up.
2096 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src))
2097 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
2100 // When the type pointed to is not sized the cast cannot be
2101 // turned into a gep.
2103 cast<PointerType>(Src->getType()->getScalarType())->getElementType();
2104 if (!PointeeType->isSized())
2107 // If the source and destination are pointers, and this cast is equivalent
2108 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
2109 // This can enhance SROA and other transforms that want type-safe pointers.
2110 unsigned NumZeros = 0;
2111 while (SrcElTy != DstElTy &&
2112 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() &&
2113 SrcElTy->getNumContainedTypes() /* not "{}" */) {
2114 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(0U);
2118 // If we found a path from the src to dest, create the getelementptr now.
2119 if (SrcElTy == DstElTy) {
2120 SmallVector<Value *, 8> Idxs(NumZeros + 1, Builder.getInt32(0));
2121 return GetElementPtrInst::CreateInBounds(Src, Idxs);
2125 if (VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
2126 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) {
2127 Value *Elem = Builder.CreateBitCast(Src, DestVTy->getElementType());
2128 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
2129 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
2130 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
2133 if (isa<IntegerType>(SrcTy)) {
2134 // If this is a cast from an integer to vector, check to see if the input
2135 // is a trunc or zext of a bitcast from vector. If so, we can replace all
2136 // the casts with a shuffle and (potentially) a bitcast.
2137 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) {
2138 CastInst *SrcCast = cast<CastInst>(Src);
2139 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0)))
2140 if (isa<VectorType>(BCIn->getOperand(0)->getType()))
2141 if (Instruction *I = optimizeVectorResize(BCIn->getOperand(0),
2142 cast<VectorType>(DestTy), *this))
2146 // If the input is an 'or' instruction, we may be doing shifts and ors to
2147 // assemble the elements of the vector manually. Try to rip the code out
2148 // and replace it with insertelements.
2149 if (Value *V = optimizeIntegerToVectorInsertions(CI, *this))
2150 return replaceInstUsesWith(CI, V);
2154 if (VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
2155 if (SrcVTy->getNumElements() == 1) {
2156 // If our destination is not a vector, then make this a straight
2157 // scalar-scalar cast.
2158 if (!DestTy->isVectorTy()) {
2160 Builder.CreateExtractElement(Src,
2161 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
2162 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
2165 // Otherwise, see if our source is an insert. If so, then use the scalar
2166 // component directly.
2167 if (InsertElementInst *IEI =
2168 dyn_cast<InsertElementInst>(CI.getOperand(0)))
2169 return CastInst::Create(Instruction::BitCast, IEI->getOperand(1),
2174 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
2175 // Okay, we have (bitcast (shuffle ..)). Check to see if this is
2176 // a bitcast to a vector with the same # elts.
2177 if (SVI->hasOneUse() && DestTy->isVectorTy() &&
2178 DestTy->getVectorNumElements() == SVI->getType()->getNumElements() &&
2179 SVI->getType()->getNumElements() ==
2180 SVI->getOperand(0)->getType()->getVectorNumElements()) {
2182 // If either of the operands is a cast from CI.getType(), then
2183 // evaluating the shuffle in the casted destination's type will allow
2184 // us to eliminate at least one cast.
2185 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) &&
2186 Tmp->getOperand(0)->getType() == DestTy) ||
2187 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) &&
2188 Tmp->getOperand(0)->getType() == DestTy)) {
2189 Value *LHS = Builder.CreateBitCast(SVI->getOperand(0), DestTy);
2190 Value *RHS = Builder.CreateBitCast(SVI->getOperand(1), DestTy);
2191 // Return a new shuffle vector. Use the same element ID's, as we
2192 // know the vector types match #elts.
2193 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
2198 // Handle the A->B->A cast, and there is an intervening PHI node.
2199 if (PHINode *PN = dyn_cast<PHINode>(Src))
2200 if (Instruction *I = optimizeBitCastFromPhi(CI, PN))
2203 if (Instruction *I = canonicalizeBitCastExtElt(CI, *this))
2206 if (Instruction *I = foldBitCastBitwiseLogic(CI, Builder))
2209 if (Instruction *I = foldBitCastSelect(CI, Builder))
2212 if (SrcTy->isPointerTy())
2213 return commonPointerCastTransforms(CI);
2214 return commonCastTransforms(CI);
2217 Instruction *InstCombiner::visitAddrSpaceCast(AddrSpaceCastInst &CI) {
2218 // If the destination pointer element type is not the same as the source's
2219 // first do a bitcast to the destination type, and then the addrspacecast.
2220 // This allows the cast to be exposed to other transforms.
2221 Value *Src = CI.getOperand(0);
2222 PointerType *SrcTy = cast<PointerType>(Src->getType()->getScalarType());
2223 PointerType *DestTy = cast<PointerType>(CI.getType()->getScalarType());
2225 Type *DestElemTy = DestTy->getElementType();
2226 if (SrcTy->getElementType() != DestElemTy) {
2227 Type *MidTy = PointerType::get(DestElemTy, SrcTy->getAddressSpace());
2228 if (VectorType *VT = dyn_cast<VectorType>(CI.getType())) {
2229 // Handle vectors of pointers.
2230 MidTy = VectorType::get(MidTy, VT->getNumElements());
2233 Value *NewBitCast = Builder.CreateBitCast(Src, MidTy);
2234 return new AddrSpaceCastInst(NewBitCast, CI.getType());
2237 return commonPointerCastTransforms(CI);