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 const DataLayout &DL) {
411 Value *TruncOp = Trunc.getOperand(0);
412 Type *DestType = Trunc.getType();
413 if (!TruncOp->hasOneUse() || !isa<IntegerType>(DestType))
416 Value *VecInput = nullptr;
417 ConstantInt *ShiftVal = nullptr;
418 if (!match(TruncOp, m_CombineOr(m_BitCast(m_Value(VecInput)),
419 m_LShr(m_BitCast(m_Value(VecInput)),
420 m_ConstantInt(ShiftVal)))) ||
421 !isa<VectorType>(VecInput->getType()))
424 VectorType *VecType = cast<VectorType>(VecInput->getType());
425 unsigned VecWidth = VecType->getPrimitiveSizeInBits();
426 unsigned DestWidth = DestType->getPrimitiveSizeInBits();
427 unsigned ShiftAmount = ShiftVal ? ShiftVal->getZExtValue() : 0;
429 if ((VecWidth % DestWidth != 0) || (ShiftAmount % DestWidth != 0))
432 // If the element type of the vector doesn't match the result type,
433 // bitcast it to a vector type that we can extract from.
434 unsigned NumVecElts = VecWidth / DestWidth;
435 if (VecType->getElementType() != DestType) {
436 VecType = VectorType::get(DestType, NumVecElts);
437 VecInput = IC.Builder->CreateBitCast(VecInput, VecType, "bc");
440 unsigned Elt = ShiftAmount / DestWidth;
441 if (DL.isBigEndian())
442 Elt = NumVecElts - 1 - Elt;
444 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt));
447 /// Try to narrow the width of bitwise logic instructions with constants.
448 Instruction *InstCombiner::shrinkBitwiseLogic(TruncInst &Trunc) {
449 Type *SrcTy = Trunc.getSrcTy();
450 Type *DestTy = Trunc.getType();
451 if (isa<IntegerType>(SrcTy) && !shouldChangeType(SrcTy, DestTy))
454 BinaryOperator *LogicOp;
456 if (!match(Trunc.getOperand(0), m_OneUse(m_BinOp(LogicOp))) ||
457 !LogicOp->isBitwiseLogicOp() ||
458 !match(LogicOp->getOperand(1), m_Constant(C)))
461 // trunc (logic X, C) --> logic (trunc X, C')
462 Constant *NarrowC = ConstantExpr::getTrunc(C, DestTy);
463 Value *NarrowOp0 = Builder->CreateTrunc(LogicOp->getOperand(0), DestTy);
464 return BinaryOperator::Create(LogicOp->getOpcode(), NarrowOp0, NarrowC);
467 /// Try to narrow the width of a splat shuffle. This could be generalized to any
468 /// shuffle with a constant operand, but we limit the transform to avoid
469 /// creating a shuffle type that targets may not be able to lower effectively.
470 static Instruction *shrinkSplatShuffle(TruncInst &Trunc,
471 InstCombiner::BuilderTy &Builder) {
472 auto *Shuf = dyn_cast<ShuffleVectorInst>(Trunc.getOperand(0));
473 if (Shuf && Shuf->hasOneUse() && isa<UndefValue>(Shuf->getOperand(1)) &&
474 Shuf->getMask()->getSplatValue() &&
475 Shuf->getType() == Shuf->getOperand(0)->getType()) {
476 // trunc (shuf X, Undef, SplatMask) --> shuf (trunc X), Undef, SplatMask
477 Constant *NarrowUndef = UndefValue::get(Trunc.getType());
478 Value *NarrowOp = Builder.CreateTrunc(Shuf->getOperand(0), Trunc.getType());
479 return new ShuffleVectorInst(NarrowOp, NarrowUndef, Shuf->getMask());
485 /// Try to narrow the width of an insert element. This could be generalized for
486 /// any vector constant, but we limit the transform to insertion into undef to
487 /// avoid potential backend problems from unsupported insertion widths. This
488 /// could also be extended to handle the case of inserting a scalar constant
489 /// into a vector variable.
490 static Instruction *shrinkInsertElt(CastInst &Trunc,
491 InstCombiner::BuilderTy &Builder) {
492 Instruction::CastOps Opcode = Trunc.getOpcode();
493 assert((Opcode == Instruction::Trunc || Opcode == Instruction::FPTrunc) &&
494 "Unexpected instruction for shrinking");
496 auto *InsElt = dyn_cast<InsertElementInst>(Trunc.getOperand(0));
497 if (!InsElt || !InsElt->hasOneUse())
500 Type *DestTy = Trunc.getType();
501 Type *DestScalarTy = DestTy->getScalarType();
502 Value *VecOp = InsElt->getOperand(0);
503 Value *ScalarOp = InsElt->getOperand(1);
504 Value *Index = InsElt->getOperand(2);
506 if (isa<UndefValue>(VecOp)) {
507 // trunc (inselt undef, X, Index) --> inselt undef, (trunc X), Index
508 // fptrunc (inselt undef, X, Index) --> inselt undef, (fptrunc X), Index
509 UndefValue *NarrowUndef = UndefValue::get(DestTy);
510 Value *NarrowOp = Builder.CreateCast(Opcode, ScalarOp, DestScalarTy);
511 return InsertElementInst::Create(NarrowUndef, NarrowOp, Index);
517 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
518 if (Instruction *Result = commonCastTransforms(CI))
521 // Test if the trunc is the user of a select which is part of a
522 // minimum or maximum operation. If so, don't do any more simplification.
523 // Even simplifying demanded bits can break the canonical form of a
526 if (SelectInst *SI = dyn_cast<SelectInst>(CI.getOperand(0)))
527 if (matchSelectPattern(SI, LHS, RHS).Flavor != SPF_UNKNOWN)
530 // See if we can simplify any instructions used by the input whose sole
531 // purpose is to compute bits we don't care about.
532 if (SimplifyDemandedInstructionBits(CI))
535 Value *Src = CI.getOperand(0);
536 Type *DestTy = CI.getType(), *SrcTy = Src->getType();
538 // Attempt to truncate the entire input expression tree to the destination
539 // type. Only do this if the dest type is a simple type, don't convert the
540 // expression tree to something weird like i93 unless the source is also
542 if ((DestTy->isVectorTy() || shouldChangeType(SrcTy, DestTy)) &&
543 canEvaluateTruncated(Src, DestTy, *this, &CI)) {
545 // If this cast is a truncate, evaluting in a different type always
546 // eliminates the cast, so it is always a win.
547 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
548 " to avoid cast: " << CI << '\n');
549 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
550 assert(Res->getType() == DestTy);
551 return replaceInstUsesWith(CI, Res);
554 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector.
555 if (DestTy->getScalarSizeInBits() == 1) {
556 Constant *One = ConstantInt::get(SrcTy, 1);
557 Src = Builder->CreateAnd(Src, One);
558 Value *Zero = Constant::getNullValue(Src->getType());
559 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
562 // FIXME: Maybe combine the next two transforms to handle the no cast case
563 // more efficiently. Support vector types. Cleanup code by using m_OneUse.
565 // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion.
566 Value *A = nullptr; ConstantInt *Cst = nullptr;
567 if (Src->hasOneUse() &&
568 match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst)))) {
569 // We have three types to worry about here, the type of A, the source of
570 // the truncate (MidSize), and the destination of the truncate. We know that
571 // ASize < MidSize and MidSize > ResultSize, but don't know the relation
572 // between ASize and ResultSize.
573 unsigned ASize = A->getType()->getPrimitiveSizeInBits();
575 // If the shift amount is larger than the size of A, then the result is
576 // known to be zero because all the input bits got shifted out.
577 if (Cst->getZExtValue() >= ASize)
578 return replaceInstUsesWith(CI, Constant::getNullValue(DestTy));
580 // Since we're doing an lshr and a zero extend, and know that the shift
581 // amount is smaller than ASize, it is always safe to do the shift in A's
582 // type, then zero extend or truncate to the result.
583 Value *Shift = Builder->CreateLShr(A, Cst->getZExtValue());
584 Shift->takeName(Src);
585 return CastInst::CreateIntegerCast(Shift, DestTy, false);
588 // FIXME: We should canonicalize to zext/trunc and remove this transform.
589 // Transform trunc(lshr (sext A), Cst) to ashr A, Cst to eliminate type
591 // It works because bits coming from sign extension have the same value as
592 // the sign bit of the original value; performing ashr instead of lshr
593 // generates bits of the same value as the sign bit.
594 if (Src->hasOneUse() &&
595 match(Src, m_LShr(m_SExt(m_Value(A)), m_ConstantInt(Cst)))) {
596 Value *SExt = cast<Instruction>(Src)->getOperand(0);
597 const unsigned SExtSize = SExt->getType()->getPrimitiveSizeInBits();
598 const unsigned ASize = A->getType()->getPrimitiveSizeInBits();
599 const unsigned CISize = CI.getType()->getPrimitiveSizeInBits();
600 const unsigned MaxAmt = SExtSize - std::max(CISize, ASize);
601 unsigned ShiftAmt = Cst->getZExtValue();
603 // This optimization can be only performed when zero bits generated by
604 // the original lshr aren't pulled into the value after truncation, so we
605 // can only shift by values no larger than the number of extension bits.
606 // FIXME: Instead of bailing when the shift is too large, use and to clear
608 if (ShiftAmt <= MaxAmt) {
610 return BinaryOperator::CreateAShr(A, ConstantInt::get(CI.getType(),
611 std::min(ShiftAmt, ASize - 1)));
612 if (SExt->hasOneUse()) {
613 Value *Shift = Builder->CreateAShr(A, std::min(ShiftAmt, ASize-1));
614 Shift->takeName(Src);
615 return CastInst::CreateIntegerCast(Shift, CI.getType(), true);
620 if (Instruction *I = shrinkBitwiseLogic(CI))
623 if (Instruction *I = shrinkSplatShuffle(CI, *Builder))
626 if (Instruction *I = shrinkInsertElt(CI, *Builder))
629 if (Src->hasOneUse() && isa<IntegerType>(SrcTy) &&
630 shouldChangeType(SrcTy, DestTy)) {
631 // Transform "trunc (shl X, cst)" -> "shl (trunc X), cst" so long as the
632 // dest type is native and cst < dest size.
633 if (match(Src, m_Shl(m_Value(A), m_ConstantInt(Cst))) &&
634 !match(A, m_Shr(m_Value(), m_Constant()))) {
635 // Skip shifts of shift by constants. It undoes a combine in
636 // FoldShiftByConstant and is the extend in reg pattern.
637 const unsigned DestSize = DestTy->getScalarSizeInBits();
638 if (Cst->getValue().ult(DestSize)) {
639 Value *NewTrunc = Builder->CreateTrunc(A, DestTy, A->getName() + ".tr");
641 return BinaryOperator::Create(
642 Instruction::Shl, NewTrunc,
643 ConstantInt::get(DestTy, Cst->getValue().trunc(DestSize)));
648 if (Instruction *I = foldVecTruncToExtElt(CI, *this, DL))
654 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, ZExtInst &CI,
656 // If we are just checking for a icmp eq of a single bit and zext'ing it
657 // to an integer, then shift the bit to the appropriate place and then
658 // cast to integer to avoid the comparison.
659 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
660 const APInt &Op1CV = Op1C->getValue();
662 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
663 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
664 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
665 (ICI->getPredicate() == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
666 if (!DoTransform) return ICI;
668 Value *In = ICI->getOperand(0);
669 Value *Sh = ConstantInt::get(In->getType(),
670 In->getType()->getScalarSizeInBits() - 1);
671 In = Builder->CreateLShr(In, Sh, In->getName() + ".lobit");
672 if (In->getType() != CI.getType())
673 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/);
675 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
676 Constant *One = ConstantInt::get(In->getType(), 1);
677 In = Builder->CreateXor(In, One, In->getName() + ".not");
680 return replaceInstUsesWith(CI, In);
683 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
684 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
685 // zext (X == 1) to i32 --> X iff X has only the low bit set.
686 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
687 // zext (X != 0) to i32 --> X iff X has only the low bit set.
688 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
689 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
690 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
691 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
692 // This only works for EQ and NE
694 // If Op1C some other power of two, convert:
695 KnownBits Known(Op1C->getType()->getBitWidth());
696 computeKnownBits(ICI->getOperand(0), Known, 0, &CI);
698 APInt KnownZeroMask(~Known.Zero);
699 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
700 if (!DoTransform) return ICI;
702 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
703 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
704 // (X&4) == 2 --> false
705 // (X&4) != 2 --> true
706 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()),
708 Res = ConstantExpr::getZExt(Res, CI.getType());
709 return replaceInstUsesWith(CI, Res);
712 uint32_t ShAmt = KnownZeroMask.logBase2();
713 Value *In = ICI->getOperand(0);
715 // Perform a logical shr by shiftamt.
716 // Insert the shift to put the result in the low bit.
717 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(), ShAmt),
718 In->getName() + ".lobit");
721 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
722 Constant *One = ConstantInt::get(In->getType(), 1);
723 In = Builder->CreateXor(In, One);
726 if (CI.getType() == In->getType())
727 return replaceInstUsesWith(CI, In);
729 Value *IntCast = Builder->CreateIntCast(In, CI.getType(), false);
730 return replaceInstUsesWith(CI, IntCast);
735 // icmp ne A, B is equal to xor A, B when A and B only really have one bit.
736 // It is also profitable to transform icmp eq into not(xor(A, B)) because that
737 // may lead to additional simplifications.
738 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) {
739 if (IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) {
740 uint32_t BitWidth = ITy->getBitWidth();
741 Value *LHS = ICI->getOperand(0);
742 Value *RHS = ICI->getOperand(1);
744 KnownBits KnownLHS(BitWidth);
745 KnownBits KnownRHS(BitWidth);
746 computeKnownBits(LHS, KnownLHS, 0, &CI);
747 computeKnownBits(RHS, KnownRHS, 0, &CI);
749 if (KnownLHS.Zero == KnownRHS.Zero && KnownLHS.One == KnownRHS.One) {
750 APInt KnownBits = KnownLHS.Zero | KnownLHS.One;
751 APInt UnknownBit = ~KnownBits;
752 if (UnknownBit.countPopulation() == 1) {
753 if (!DoTransform) return ICI;
755 Value *Result = Builder->CreateXor(LHS, RHS);
757 // Mask off any bits that are set and won't be shifted away.
758 if (KnownLHS.One.uge(UnknownBit))
759 Result = Builder->CreateAnd(Result,
760 ConstantInt::get(ITy, UnknownBit));
762 // Shift the bit we're testing down to the lsb.
763 Result = Builder->CreateLShr(
764 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros()));
766 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
767 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1));
768 Result->takeName(ICI);
769 return replaceInstUsesWith(CI, Result);
778 /// Determine if the specified value can be computed in the specified wider type
779 /// and produce the same low bits. If not, return false.
781 /// If this function returns true, it can also return a non-zero number of bits
782 /// (in BitsToClear) which indicates that the value it computes is correct for
783 /// the zero extend, but that the additional BitsToClear bits need to be zero'd
784 /// out. For example, to promote something like:
786 /// %B = trunc i64 %A to i32
787 /// %C = lshr i32 %B, 8
788 /// %E = zext i32 %C to i64
790 /// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be
791 /// set to 8 to indicate that the promoted value needs to have bits 24-31
792 /// cleared in addition to bits 32-63. Since an 'and' will be generated to
793 /// clear the top bits anyway, doing this has no extra cost.
795 /// This function works on both vectors and scalars.
796 static bool canEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear,
797 InstCombiner &IC, Instruction *CxtI) {
799 if (isa<Constant>(V))
802 Instruction *I = dyn_cast<Instruction>(V);
803 if (!I) return false;
805 // If the input is a truncate from the destination type, we can trivially
807 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
810 // We can't extend or shrink something that has multiple uses: doing so would
811 // require duplicating the instruction in general, which isn't profitable.
812 if (!I->hasOneUse()) return false;
814 unsigned Opc = I->getOpcode(), Tmp;
816 case Instruction::ZExt: // zext(zext(x)) -> zext(x).
817 case Instruction::SExt: // zext(sext(x)) -> sext(x).
818 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x)
820 case Instruction::And:
821 case Instruction::Or:
822 case Instruction::Xor:
823 case Instruction::Add:
824 case Instruction::Sub:
825 case Instruction::Mul:
826 if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI) ||
827 !canEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI))
829 // These can all be promoted if neither operand has 'bits to clear'.
830 if (BitsToClear == 0 && Tmp == 0)
833 // If the operation is an AND/OR/XOR and the bits to clear are zero in the
834 // other side, BitsToClear is ok.
835 if (Tmp == 0 && I->isBitwiseLogicOp()) {
836 // We use MaskedValueIsZero here for generality, but the case we care
837 // about the most is constant RHS.
838 unsigned VSize = V->getType()->getScalarSizeInBits();
839 if (IC.MaskedValueIsZero(I->getOperand(1),
840 APInt::getHighBitsSet(VSize, BitsToClear),
845 // Otherwise, we don't know how to analyze this BitsToClear case yet.
848 case Instruction::Shl:
849 // We can promote shl(x, cst) if we can promote x. Since shl overwrites the
850 // upper bits we can reduce BitsToClear by the shift amount.
851 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
852 if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI))
854 uint64_t ShiftAmt = Amt->getZExtValue();
855 BitsToClear = ShiftAmt < BitsToClear ? BitsToClear - ShiftAmt : 0;
859 case Instruction::LShr:
860 // We can promote lshr(x, cst) if we can promote x. This requires the
861 // ultimate 'and' to clear out the high zero bits we're clearing out though.
862 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
863 if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI))
865 BitsToClear += Amt->getZExtValue();
866 if (BitsToClear > V->getType()->getScalarSizeInBits())
867 BitsToClear = V->getType()->getScalarSizeInBits();
870 // Cannot promote variable LSHR.
872 case Instruction::Select:
873 if (!canEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI) ||
874 !canEvaluateZExtd(I->getOperand(2), Ty, BitsToClear, IC, CxtI) ||
875 // TODO: If important, we could handle the case when the BitsToClear are
876 // known zero in the disagreeing side.
881 case Instruction::PHI: {
882 // We can change a phi if we can change all operands. Note that we never
883 // get into trouble with cyclic PHIs here because we only consider
884 // instructions with a single use.
885 PHINode *PN = cast<PHINode>(I);
886 if (!canEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear, IC, CxtI))
888 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
889 if (!canEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp, IC, CxtI) ||
890 // TODO: If important, we could handle the case when the BitsToClear
891 // are known zero in the disagreeing input.
897 // TODO: Can handle more cases here.
902 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
903 // If this zero extend is only used by a truncate, let the truncate be
904 // eliminated before we try to optimize this zext.
905 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back()))
908 // If one of the common conversion will work, do it.
909 if (Instruction *Result = commonCastTransforms(CI))
912 Value *Src = CI.getOperand(0);
913 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
915 // Attempt to extend the entire input expression tree to the destination
916 // type. Only do this if the dest type is a simple type, don't convert the
917 // expression tree to something weird like i93 unless the source is also
919 unsigned BitsToClear;
920 if ((DestTy->isVectorTy() || shouldChangeType(SrcTy, DestTy)) &&
921 canEvaluateZExtd(Src, DestTy, BitsToClear, *this, &CI)) {
922 assert(BitsToClear <= SrcTy->getScalarSizeInBits() &&
923 "Can't clear more bits than in SrcTy");
925 // Okay, we can transform this! Insert the new expression now.
926 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
927 " to avoid zero extend: " << CI << '\n');
928 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
929 assert(Res->getType() == DestTy);
931 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear;
932 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
934 // If the high bits are already filled with zeros, just replace this
935 // cast with the result.
936 if (MaskedValueIsZero(Res,
937 APInt::getHighBitsSet(DestBitSize,
938 DestBitSize-SrcBitsKept),
940 return replaceInstUsesWith(CI, Res);
942 // We need to emit an AND to clear the high bits.
943 Constant *C = ConstantInt::get(Res->getType(),
944 APInt::getLowBitsSet(DestBitSize, SrcBitsKept));
945 return BinaryOperator::CreateAnd(Res, C);
948 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
949 // types and if the sizes are just right we can convert this into a logical
950 // 'and' which will be much cheaper than the pair of casts.
951 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
952 // TODO: Subsume this into EvaluateInDifferentType.
954 // Get the sizes of the types involved. We know that the intermediate type
955 // will be smaller than A or C, but don't know the relation between A and C.
956 Value *A = CSrc->getOperand(0);
957 unsigned SrcSize = A->getType()->getScalarSizeInBits();
958 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
959 unsigned DstSize = CI.getType()->getScalarSizeInBits();
960 // If we're actually extending zero bits, then if
961 // SrcSize < DstSize: zext(a & mask)
962 // SrcSize == DstSize: a & mask
963 // SrcSize > DstSize: trunc(a) & mask
964 if (SrcSize < DstSize) {
965 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
966 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
967 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
968 return new ZExtInst(And, CI.getType());
971 if (SrcSize == DstSize) {
972 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
973 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
976 if (SrcSize > DstSize) {
977 Value *Trunc = Builder->CreateTrunc(A, CI.getType());
978 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
979 return BinaryOperator::CreateAnd(Trunc,
980 ConstantInt::get(Trunc->getType(),
985 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
986 return transformZExtICmp(ICI, CI);
988 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
989 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
990 // zext (or icmp, icmp) -> or (zext icmp), (zext icmp) if at least one
991 // of the (zext icmp) can be eliminated. If so, immediately perform the
992 // according elimination.
993 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
994 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
995 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
996 (transformZExtICmp(LHS, CI, false) ||
997 transformZExtICmp(RHS, CI, false))) {
998 // zext (or icmp, icmp) -> or (zext icmp), (zext icmp)
999 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());
1000 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());
1001 BinaryOperator *Or = BinaryOperator::Create(Instruction::Or, LCast, RCast);
1003 // Perform the elimination.
1004 if (auto *LZExt = dyn_cast<ZExtInst>(LCast))
1005 transformZExtICmp(LHS, *LZExt);
1006 if (auto *RZExt = dyn_cast<ZExtInst>(RCast))
1007 transformZExtICmp(RHS, *RZExt);
1013 // zext(trunc(X) & C) -> (X & zext(C)).
1017 match(SrcI, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Constant(C)))) &&
1018 X->getType() == CI.getType())
1019 return BinaryOperator::CreateAnd(X, ConstantExpr::getZExt(C, CI.getType()));
1021 // zext((trunc(X) & C) ^ C) -> ((X & zext(C)) ^ zext(C)).
1023 if (SrcI && match(SrcI, m_OneUse(m_Xor(m_Value(And), m_Constant(C)))) &&
1024 match(And, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Specific(C)))) &&
1025 X->getType() == CI.getType()) {
1026 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
1027 return BinaryOperator::CreateXor(Builder->CreateAnd(X, ZC), ZC);
1033 /// Transform (sext icmp) to bitwise / integer operations to eliminate the icmp.
1034 Instruction *InstCombiner::transformSExtICmp(ICmpInst *ICI, Instruction &CI) {
1035 Value *Op0 = ICI->getOperand(0), *Op1 = ICI->getOperand(1);
1036 ICmpInst::Predicate Pred = ICI->getPredicate();
1038 // Don't bother if Op1 isn't of vector or integer type.
1039 if (!Op1->getType()->isIntOrIntVectorTy())
1042 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
1043 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if negative
1044 // (x >s -1) ? -1 : 0 -> not (ashr x, 31) -> all ones if positive
1045 if ((Pred == ICmpInst::ICMP_SLT && Op1C->isNullValue()) ||
1046 (Pred == ICmpInst::ICMP_SGT && Op1C->isAllOnesValue())) {
1048 Value *Sh = ConstantInt::get(Op0->getType(),
1049 Op0->getType()->getScalarSizeInBits()-1);
1050 Value *In = Builder->CreateAShr(Op0, Sh, Op0->getName()+".lobit");
1051 if (In->getType() != CI.getType())
1052 In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/);
1054 if (Pred == ICmpInst::ICMP_SGT)
1055 In = Builder->CreateNot(In, In->getName()+".not");
1056 return replaceInstUsesWith(CI, In);
1060 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
1061 // If we know that only one bit of the LHS of the icmp can be set and we
1062 // have an equality comparison with zero or a power of 2, we can transform
1063 // the icmp and sext into bitwise/integer operations.
1064 if (ICI->hasOneUse() &&
1065 ICI->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){
1066 unsigned BitWidth = Op1C->getType()->getBitWidth();
1067 KnownBits Known(BitWidth);
1068 computeKnownBits(Op0, Known, 0, &CI);
1070 APInt KnownZeroMask(~Known.Zero);
1071 if (KnownZeroMask.isPowerOf2()) {
1072 Value *In = ICI->getOperand(0);
1074 // If the icmp tests for a known zero bit we can constant fold it.
1075 if (!Op1C->isZero() && Op1C->getValue() != KnownZeroMask) {
1076 Value *V = Pred == ICmpInst::ICMP_NE ?
1077 ConstantInt::getAllOnesValue(CI.getType()) :
1078 ConstantInt::getNullValue(CI.getType());
1079 return replaceInstUsesWith(CI, V);
1082 if (!Op1C->isZero() == (Pred == ICmpInst::ICMP_NE)) {
1083 // sext ((x & 2^n) == 0) -> (x >> n) - 1
1084 // sext ((x & 2^n) != 2^n) -> (x >> n) - 1
1085 unsigned ShiftAmt = KnownZeroMask.countTrailingZeros();
1086 // Perform a right shift to place the desired bit in the LSB.
1088 In = Builder->CreateLShr(In,
1089 ConstantInt::get(In->getType(), ShiftAmt));
1091 // At this point "In" is either 1 or 0. Subtract 1 to turn
1092 // {1, 0} -> {0, -1}.
1093 In = Builder->CreateAdd(In,
1094 ConstantInt::getAllOnesValue(In->getType()),
1097 // sext ((x & 2^n) != 0) -> (x << bitwidth-n) a>> bitwidth-1
1098 // sext ((x & 2^n) == 2^n) -> (x << bitwidth-n) a>> bitwidth-1
1099 unsigned ShiftAmt = KnownZeroMask.countLeadingZeros();
1100 // Perform a left shift to place the desired bit in the MSB.
1102 In = Builder->CreateShl(In,
1103 ConstantInt::get(In->getType(), ShiftAmt));
1105 // Distribute the bit over the whole bit width.
1106 In = Builder->CreateAShr(In, ConstantInt::get(In->getType(),
1107 BitWidth - 1), "sext");
1110 if (CI.getType() == In->getType())
1111 return replaceInstUsesWith(CI, In);
1112 return CastInst::CreateIntegerCast(In, CI.getType(), true/*SExt*/);
1120 /// Return true if we can take the specified value and return it as type Ty
1121 /// without inserting any new casts and without changing the value of the common
1122 /// low bits. This is used by code that tries to promote integer operations to
1123 /// a wider types will allow us to eliminate the extension.
1125 /// This function works on both vectors and scalars.
1127 static bool canEvaluateSExtd(Value *V, Type *Ty) {
1128 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() &&
1129 "Can't sign extend type to a smaller type");
1130 // If this is a constant, it can be trivially promoted.
1131 if (isa<Constant>(V))
1134 Instruction *I = dyn_cast<Instruction>(V);
1135 if (!I) return false;
1137 // If this is a truncate from the dest type, we can trivially eliminate it.
1138 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
1141 // We can't extend or shrink something that has multiple uses: doing so would
1142 // require duplicating the instruction in general, which isn't profitable.
1143 if (!I->hasOneUse()) return false;
1145 switch (I->getOpcode()) {
1146 case Instruction::SExt: // sext(sext(x)) -> sext(x)
1147 case Instruction::ZExt: // sext(zext(x)) -> zext(x)
1148 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x)
1150 case Instruction::And:
1151 case Instruction::Or:
1152 case Instruction::Xor:
1153 case Instruction::Add:
1154 case Instruction::Sub:
1155 case Instruction::Mul:
1156 // These operators can all arbitrarily be extended if their inputs can.
1157 return canEvaluateSExtd(I->getOperand(0), Ty) &&
1158 canEvaluateSExtd(I->getOperand(1), Ty);
1160 //case Instruction::Shl: TODO
1161 //case Instruction::LShr: TODO
1163 case Instruction::Select:
1164 return canEvaluateSExtd(I->getOperand(1), Ty) &&
1165 canEvaluateSExtd(I->getOperand(2), Ty);
1167 case Instruction::PHI: {
1168 // We can change a phi if we can change all operands. Note that we never
1169 // get into trouble with cyclic PHIs here because we only consider
1170 // instructions with a single use.
1171 PHINode *PN = cast<PHINode>(I);
1172 for (Value *IncValue : PN->incoming_values())
1173 if (!canEvaluateSExtd(IncValue, Ty)) return false;
1177 // TODO: Can handle more cases here.
1184 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
1185 // If this sign extend is only used by a truncate, let the truncate be
1186 // eliminated before we try to optimize this sext.
1187 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back()))
1190 if (Instruction *I = commonCastTransforms(CI))
1193 Value *Src = CI.getOperand(0);
1194 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
1196 // If we know that the value being extended is positive, we can use a zext
1198 KnownBits Known = computeKnownBits(Src, 0, &CI);
1199 if (Known.isNonNegative()) {
1200 Value *ZExt = Builder->CreateZExt(Src, DestTy);
1201 return replaceInstUsesWith(CI, ZExt);
1204 // Attempt to extend the entire input expression tree to the destination
1205 // type. Only do this if the dest type is a simple type, don't convert the
1206 // expression tree to something weird like i93 unless the source is also
1208 if ((DestTy->isVectorTy() || shouldChangeType(SrcTy, DestTy)) &&
1209 canEvaluateSExtd(Src, DestTy)) {
1210 // Okay, we can transform this! Insert the new expression now.
1211 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
1212 " to avoid sign extend: " << CI << '\n');
1213 Value *Res = EvaluateInDifferentType(Src, DestTy, true);
1214 assert(Res->getType() == DestTy);
1216 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1217 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1219 // If the high bits are already filled with sign bit, just replace this
1220 // cast with the result.
1221 if (ComputeNumSignBits(Res, 0, &CI) > DestBitSize - SrcBitSize)
1222 return replaceInstUsesWith(CI, Res);
1224 // We need to emit a shl + ashr to do the sign extend.
1225 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1226 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"),
1230 // If the input is a trunc from the destination type, then turn sext(trunc(x))
1233 if (match(Src, m_OneUse(m_Trunc(m_Value(X)))) && X->getType() == DestTy) {
1234 // sext(trunc(X)) --> ashr(shl(X, C), C)
1235 unsigned SrcBitSize = SrcTy->getScalarSizeInBits();
1236 unsigned DestBitSize = DestTy->getScalarSizeInBits();
1237 Constant *ShAmt = ConstantInt::get(DestTy, DestBitSize - SrcBitSize);
1238 return BinaryOperator::CreateAShr(Builder->CreateShl(X, ShAmt), ShAmt);
1241 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
1242 return transformSExtICmp(ICI, CI);
1244 // If the input is a shl/ashr pair of a same constant, then this is a sign
1245 // extension from a smaller value. If we could trust arbitrary bitwidth
1246 // integers, we could turn this into a truncate to the smaller bit and then
1247 // use a sext for the whole extension. Since we don't, look deeper and check
1248 // for a truncate. If the source and dest are the same type, eliminate the
1249 // trunc and extend and just do shifts. For example, turn:
1250 // %a = trunc i32 %i to i8
1251 // %b = shl i8 %a, 6
1252 // %c = ashr i8 %b, 6
1253 // %d = sext i8 %c to i32
1255 // %a = shl i32 %i, 30
1256 // %d = ashr i32 %a, 30
1258 // TODO: Eventually this could be subsumed by EvaluateInDifferentType.
1259 ConstantInt *BA = nullptr, *CA = nullptr;
1260 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)),
1261 m_ConstantInt(CA))) &&
1262 BA == CA && A->getType() == CI.getType()) {
1263 unsigned MidSize = Src->getType()->getScalarSizeInBits();
1264 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
1265 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
1266 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
1267 A = Builder->CreateShl(A, ShAmtV, CI.getName());
1268 return BinaryOperator::CreateAShr(A, ShAmtV);
1275 /// Return a Constant* for the specified floating-point constant if it fits
1276 /// in the specified FP type without changing its value.
1277 static Constant *fitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
1279 APFloat F = CFP->getValueAPF();
1280 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
1282 return ConstantFP::get(CFP->getContext(), F);
1286 /// Look through floating-point extensions until we get the source value.
1287 static Value *lookThroughFPExtensions(Value *V) {
1288 while (auto *FPExt = dyn_cast<FPExtInst>(V))
1289 V = FPExt->getOperand(0);
1291 // If this value is a constant, return the constant in the smallest FP type
1292 // that can accurately represent it. This allows us to turn
1293 // (float)((double)X+2.0) into x+2.0f.
1294 if (auto *CFP = dyn_cast<ConstantFP>(V)) {
1295 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext()))
1296 return V; // No constant folding of this.
1297 // See if the value can be truncated to half and then reextended.
1298 if (Value *V = fitsInFPType(CFP, APFloat::IEEEhalf()))
1300 // See if the value can be truncated to float and then reextended.
1301 if (Value *V = fitsInFPType(CFP, APFloat::IEEEsingle()))
1303 if (CFP->getType()->isDoubleTy())
1304 return V; // Won't shrink.
1305 if (Value *V = fitsInFPType(CFP, APFloat::IEEEdouble()))
1307 // Don't try to shrink to various long double types.
1313 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
1314 if (Instruction *I = commonCastTransforms(CI))
1316 // If we have fptrunc(OpI (fpextend x), (fpextend y)), we would like to
1317 // simplify this expression to avoid one or more of the trunc/extend
1318 // operations if we can do so without changing the numerical results.
1320 // The exact manner in which the widths of the operands interact to limit
1321 // what we can and cannot do safely varies from operation to operation, and
1322 // is explained below in the various case statements.
1323 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
1324 if (OpI && OpI->hasOneUse()) {
1325 Value *LHSOrig = lookThroughFPExtensions(OpI->getOperand(0));
1326 Value *RHSOrig = lookThroughFPExtensions(OpI->getOperand(1));
1327 unsigned OpWidth = OpI->getType()->getFPMantissaWidth();
1328 unsigned LHSWidth = LHSOrig->getType()->getFPMantissaWidth();
1329 unsigned RHSWidth = RHSOrig->getType()->getFPMantissaWidth();
1330 unsigned SrcWidth = std::max(LHSWidth, RHSWidth);
1331 unsigned DstWidth = CI.getType()->getFPMantissaWidth();
1332 switch (OpI->getOpcode()) {
1334 case Instruction::FAdd:
1335 case Instruction::FSub:
1336 // For addition and subtraction, the infinitely precise result can
1337 // essentially be arbitrarily wide; proving that double rounding
1338 // will not occur because the result of OpI is exact (as we will for
1339 // FMul, for example) is hopeless. However, we *can* nonetheless
1340 // frequently know that double rounding cannot occur (or that it is
1341 // innocuous) by taking advantage of the specific structure of
1342 // infinitely-precise results that admit double rounding.
1344 // Specifically, if OpWidth >= 2*DstWdith+1 and DstWidth is sufficient
1345 // to represent both sources, we can guarantee that the double
1346 // rounding is innocuous (See p50 of Figueroa's 2000 PhD thesis,
1347 // "A Rigorous Framework for Fully Supporting the IEEE Standard ..."
1348 // for proof of this fact).
1350 // Note: Figueroa does not consider the case where DstFormat !=
1351 // SrcFormat. It's possible (likely even!) that this analysis
1352 // could be tightened for those cases, but they are rare (the main
1353 // case of interest here is (float)((double)float + float)).
1354 if (OpWidth >= 2*DstWidth+1 && DstWidth >= SrcWidth) {
1355 if (LHSOrig->getType() != CI.getType())
1356 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1357 if (RHSOrig->getType() != CI.getType())
1358 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1360 BinaryOperator::Create(OpI->getOpcode(), LHSOrig, RHSOrig);
1361 RI->copyFastMathFlags(OpI);
1365 case Instruction::FMul:
1366 // For multiplication, the infinitely precise result has at most
1367 // LHSWidth + RHSWidth significant bits; if OpWidth is sufficient
1368 // that such a value can be exactly represented, then no double
1369 // rounding can possibly occur; we can safely perform the operation
1370 // in the destination format if it can represent both sources.
1371 if (OpWidth >= LHSWidth + RHSWidth && DstWidth >= SrcWidth) {
1372 if (LHSOrig->getType() != CI.getType())
1373 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1374 if (RHSOrig->getType() != CI.getType())
1375 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1377 BinaryOperator::CreateFMul(LHSOrig, RHSOrig);
1378 RI->copyFastMathFlags(OpI);
1382 case Instruction::FDiv:
1383 // For division, we use again use the bound from Figueroa's
1384 // dissertation. I am entirely certain that this bound can be
1385 // tightened in the unbalanced operand case by an analysis based on
1386 // the diophantine rational approximation bound, but the well-known
1387 // condition used here is a good conservative first pass.
1388 // TODO: Tighten bound via rigorous analysis of the unbalanced case.
1389 if (OpWidth >= 2*DstWidth && DstWidth >= SrcWidth) {
1390 if (LHSOrig->getType() != CI.getType())
1391 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1392 if (RHSOrig->getType() != CI.getType())
1393 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1395 BinaryOperator::CreateFDiv(LHSOrig, RHSOrig);
1396 RI->copyFastMathFlags(OpI);
1400 case Instruction::FRem:
1401 // Remainder is straightforward. Remainder is always exact, so the
1402 // type of OpI doesn't enter into things at all. We simply evaluate
1403 // in whichever source type is larger, then convert to the
1404 // destination type.
1405 if (SrcWidth == OpWidth)
1407 if (LHSWidth < SrcWidth)
1408 LHSOrig = Builder->CreateFPExt(LHSOrig, RHSOrig->getType());
1409 else if (RHSWidth <= SrcWidth)
1410 RHSOrig = Builder->CreateFPExt(RHSOrig, LHSOrig->getType());
1411 if (LHSOrig != OpI->getOperand(0) || RHSOrig != OpI->getOperand(1)) {
1412 Value *ExactResult = Builder->CreateFRem(LHSOrig, RHSOrig);
1413 if (Instruction *RI = dyn_cast<Instruction>(ExactResult))
1414 RI->copyFastMathFlags(OpI);
1415 return CastInst::CreateFPCast(ExactResult, CI.getType());
1419 // (fptrunc (fneg x)) -> (fneg (fptrunc x))
1420 if (BinaryOperator::isFNeg(OpI)) {
1421 Value *InnerTrunc = Builder->CreateFPTrunc(OpI->getOperand(1),
1423 Instruction *RI = BinaryOperator::CreateFNeg(InnerTrunc);
1424 RI->copyFastMathFlags(OpI);
1429 // (fptrunc (select cond, R1, Cst)) -->
1430 // (select cond, (fptrunc R1), (fptrunc Cst))
1432 // - but only if this isn't part of a min/max operation, else we'll
1433 // ruin min/max canonical form which is to have the select and
1434 // compare's operands be of the same type with no casts to look through.
1436 SelectInst *SI = dyn_cast<SelectInst>(CI.getOperand(0));
1438 (isa<ConstantFP>(SI->getOperand(1)) ||
1439 isa<ConstantFP>(SI->getOperand(2))) &&
1440 matchSelectPattern(SI, LHS, RHS).Flavor == SPF_UNKNOWN) {
1441 Value *LHSTrunc = Builder->CreateFPTrunc(SI->getOperand(1),
1443 Value *RHSTrunc = Builder->CreateFPTrunc(SI->getOperand(2),
1445 return SelectInst::Create(SI->getOperand(0), LHSTrunc, RHSTrunc);
1448 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI.getOperand(0));
1450 switch (II->getIntrinsicID()) {
1452 case Intrinsic::fabs:
1453 case Intrinsic::ceil:
1454 case Intrinsic::floor:
1455 case Intrinsic::rint:
1456 case Intrinsic::round:
1457 case Intrinsic::nearbyint:
1458 case Intrinsic::trunc: {
1459 Value *Src = II->getArgOperand(0);
1460 if (!Src->hasOneUse())
1463 // Except for fabs, this transformation requires the input of the unary FP
1464 // operation to be itself an fpext from the type to which we're
1466 if (II->getIntrinsicID() != Intrinsic::fabs) {
1467 FPExtInst *FPExtSrc = dyn_cast<FPExtInst>(Src);
1468 if (!FPExtSrc || FPExtSrc->getOperand(0)->getType() != CI.getType())
1472 // Do unary FP operation on smaller type.
1473 // (fptrunc (fabs x)) -> (fabs (fptrunc x))
1474 Value *InnerTrunc = Builder->CreateFPTrunc(Src, CI.getType());
1475 Type *IntrinsicType[] = { CI.getType() };
1476 Function *Overload = Intrinsic::getDeclaration(
1477 CI.getModule(), II->getIntrinsicID(), IntrinsicType);
1479 SmallVector<OperandBundleDef, 1> OpBundles;
1480 II->getOperandBundlesAsDefs(OpBundles);
1482 Value *Args[] = { InnerTrunc };
1483 CallInst *NewCI = CallInst::Create(Overload, Args,
1484 OpBundles, II->getName());
1485 NewCI->copyFastMathFlags(II);
1491 if (Instruction *I = shrinkInsertElt(CI, *Builder))
1497 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
1498 return commonCastTransforms(CI);
1501 // fpto{s/u}i({u/s}itofp(X)) --> X or zext(X) or sext(X) or trunc(X)
1502 // This is safe if the intermediate type has enough bits in its mantissa to
1503 // accurately represent all values of X. For example, this won't work with
1504 // i64 -> float -> i64.
1505 Instruction *InstCombiner::FoldItoFPtoI(Instruction &FI) {
1506 if (!isa<UIToFPInst>(FI.getOperand(0)) && !isa<SIToFPInst>(FI.getOperand(0)))
1508 Instruction *OpI = cast<Instruction>(FI.getOperand(0));
1510 Value *SrcI = OpI->getOperand(0);
1511 Type *FITy = FI.getType();
1512 Type *OpITy = OpI->getType();
1513 Type *SrcTy = SrcI->getType();
1514 bool IsInputSigned = isa<SIToFPInst>(OpI);
1515 bool IsOutputSigned = isa<FPToSIInst>(FI);
1517 // We can safely assume the conversion won't overflow the output range,
1518 // because (for example) (uint8_t)18293.f is undefined behavior.
1520 // Since we can assume the conversion won't overflow, our decision as to
1521 // whether the input will fit in the float should depend on the minimum
1522 // of the input range and output range.
1524 // This means this is also safe for a signed input and unsigned output, since
1525 // a negative input would lead to undefined behavior.
1526 int InputSize = (int)SrcTy->getScalarSizeInBits() - IsInputSigned;
1527 int OutputSize = (int)FITy->getScalarSizeInBits() - IsOutputSigned;
1528 int ActualSize = std::min(InputSize, OutputSize);
1530 if (ActualSize <= OpITy->getFPMantissaWidth()) {
1531 if (FITy->getScalarSizeInBits() > SrcTy->getScalarSizeInBits()) {
1532 if (IsInputSigned && IsOutputSigned)
1533 return new SExtInst(SrcI, FITy);
1534 return new ZExtInst(SrcI, FITy);
1536 if (FITy->getScalarSizeInBits() < SrcTy->getScalarSizeInBits())
1537 return new TruncInst(SrcI, FITy);
1539 return replaceInstUsesWith(FI, SrcI);
1540 return new BitCastInst(SrcI, FITy);
1545 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
1546 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1548 return commonCastTransforms(FI);
1550 if (Instruction *I = FoldItoFPtoI(FI))
1553 return commonCastTransforms(FI);
1556 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
1557 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1559 return commonCastTransforms(FI);
1561 if (Instruction *I = FoldItoFPtoI(FI))
1564 return commonCastTransforms(FI);
1567 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
1568 return commonCastTransforms(CI);
1571 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
1572 return commonCastTransforms(CI);
1575 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
1576 // If the source integer type is not the intptr_t type for this target, do a
1577 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the
1578 // cast to be exposed to other transforms.
1579 unsigned AS = CI.getAddressSpace();
1580 if (CI.getOperand(0)->getType()->getScalarSizeInBits() !=
1581 DL.getPointerSizeInBits(AS)) {
1582 Type *Ty = DL.getIntPtrType(CI.getContext(), AS);
1583 if (CI.getType()->isVectorTy()) // Handle vectors of pointers.
1584 Ty = VectorType::get(Ty, CI.getType()->getVectorNumElements());
1586 Value *P = Builder->CreateZExtOrTrunc(CI.getOperand(0), Ty);
1587 return new IntToPtrInst(P, CI.getType());
1590 if (Instruction *I = commonCastTransforms(CI))
1596 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
1597 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
1598 Value *Src = CI.getOperand(0);
1600 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
1601 // If casting the result of a getelementptr instruction with no offset, turn
1602 // this into a cast of the original pointer!
1603 if (GEP->hasAllZeroIndices() &&
1604 // If CI is an addrspacecast and GEP changes the poiner type, merging
1605 // GEP into CI would undo canonicalizing addrspacecast with different
1606 // pointer types, causing infinite loops.
1607 (!isa<AddrSpaceCastInst>(CI) ||
1608 GEP->getType() == GEP->getPointerOperandType())) {
1609 // Changing the cast operand is usually not a good idea but it is safe
1610 // here because the pointer operand is being replaced with another
1611 // pointer operand so the opcode doesn't need to change.
1613 CI.setOperand(0, GEP->getOperand(0));
1618 return commonCastTransforms(CI);
1621 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
1622 // If the destination integer type is not the intptr_t type for this target,
1623 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast
1624 // to be exposed to other transforms.
1626 Type *Ty = CI.getType();
1627 unsigned AS = CI.getPointerAddressSpace();
1629 if (Ty->getScalarSizeInBits() == DL.getPointerSizeInBits(AS))
1630 return commonPointerCastTransforms(CI);
1632 Type *PtrTy = DL.getIntPtrType(CI.getContext(), AS);
1633 if (Ty->isVectorTy()) // Handle vectors of pointers.
1634 PtrTy = VectorType::get(PtrTy, Ty->getVectorNumElements());
1636 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), PtrTy);
1637 return CastInst::CreateIntegerCast(P, Ty, /*isSigned=*/false);
1640 /// This input value (which is known to have vector type) is being zero extended
1641 /// or truncated to the specified vector type.
1642 /// Try to replace it with a shuffle (and vector/vector bitcast) if possible.
1644 /// The source and destination vector types may have different element types.
1645 static Instruction *optimizeVectorResize(Value *InVal, VectorType *DestTy,
1647 // We can only do this optimization if the output is a multiple of the input
1648 // element size, or the input is a multiple of the output element size.
1649 // Convert the input type to have the same element type as the output.
1650 VectorType *SrcTy = cast<VectorType>(InVal->getType());
1652 if (SrcTy->getElementType() != DestTy->getElementType()) {
1653 // The input types don't need to be identical, but for now they must be the
1654 // same size. There is no specific reason we couldn't handle things like
1655 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten
1657 if (SrcTy->getElementType()->getPrimitiveSizeInBits() !=
1658 DestTy->getElementType()->getPrimitiveSizeInBits())
1661 SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements());
1662 InVal = IC.Builder->CreateBitCast(InVal, SrcTy);
1665 // Now that the element types match, get the shuffle mask and RHS of the
1666 // shuffle to use, which depends on whether we're increasing or decreasing the
1667 // size of the input.
1668 SmallVector<uint32_t, 16> ShuffleMask;
1671 if (SrcTy->getNumElements() > DestTy->getNumElements()) {
1672 // If we're shrinking the number of elements, just shuffle in the low
1673 // elements from the input and use undef as the second shuffle input.
1674 V2 = UndefValue::get(SrcTy);
1675 for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i)
1676 ShuffleMask.push_back(i);
1679 // If we're increasing the number of elements, shuffle in all of the
1680 // elements from InVal and fill the rest of the result elements with zeros
1681 // from a constant zero.
1682 V2 = Constant::getNullValue(SrcTy);
1683 unsigned SrcElts = SrcTy->getNumElements();
1684 for (unsigned i = 0, e = SrcElts; i != e; ++i)
1685 ShuffleMask.push_back(i);
1687 // The excess elements reference the first element of the zero input.
1688 for (unsigned i = 0, e = DestTy->getNumElements()-SrcElts; i != e; ++i)
1689 ShuffleMask.push_back(SrcElts);
1692 return new ShuffleVectorInst(InVal, V2,
1693 ConstantDataVector::get(V2->getContext(),
1697 static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) {
1698 return Value % Ty->getPrimitiveSizeInBits() == 0;
1701 static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) {
1702 return Value / Ty->getPrimitiveSizeInBits();
1705 /// V is a value which is inserted into a vector of VecEltTy.
1706 /// Look through the value to see if we can decompose it into
1707 /// insertions into the vector. See the example in the comment for
1708 /// OptimizeIntegerToVectorInsertions for the pattern this handles.
1709 /// The type of V is always a non-zero multiple of VecEltTy's size.
1710 /// Shift is the number of bits between the lsb of V and the lsb of
1713 /// This returns false if the pattern can't be matched or true if it can,
1714 /// filling in Elements with the elements found here.
1715 static bool collectInsertionElements(Value *V, unsigned Shift,
1716 SmallVectorImpl<Value *> &Elements,
1717 Type *VecEltTy, bool isBigEndian) {
1718 assert(isMultipleOfTypeSize(Shift, VecEltTy) &&
1719 "Shift should be a multiple of the element type size");
1721 // Undef values never contribute useful bits to the result.
1722 if (isa<UndefValue>(V)) return true;
1724 // If we got down to a value of the right type, we win, try inserting into the
1726 if (V->getType() == VecEltTy) {
1727 // Inserting null doesn't actually insert any elements.
1728 if (Constant *C = dyn_cast<Constant>(V))
1729 if (C->isNullValue())
1732 unsigned ElementIndex = getTypeSizeIndex(Shift, VecEltTy);
1734 ElementIndex = Elements.size() - ElementIndex - 1;
1736 // Fail if multiple elements are inserted into this slot.
1737 if (Elements[ElementIndex])
1740 Elements[ElementIndex] = V;
1744 if (Constant *C = dyn_cast<Constant>(V)) {
1745 // Figure out the # elements this provides, and bitcast it or slice it up
1747 unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(),
1749 // If the constant is the size of a vector element, we just need to bitcast
1750 // it to the right type so it gets properly inserted.
1752 return collectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy),
1753 Shift, Elements, VecEltTy, isBigEndian);
1755 // Okay, this is a constant that covers multiple elements. Slice it up into
1756 // pieces and insert each element-sized piece into the vector.
1757 if (!isa<IntegerType>(C->getType()))
1758 C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(),
1759 C->getType()->getPrimitiveSizeInBits()));
1760 unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits();
1761 Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize);
1763 for (unsigned i = 0; i != NumElts; ++i) {
1764 unsigned ShiftI = Shift+i*ElementSize;
1765 Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(),
1767 Piece = ConstantExpr::getTrunc(Piece, ElementIntTy);
1768 if (!collectInsertionElements(Piece, ShiftI, Elements, VecEltTy,
1775 if (!V->hasOneUse()) return false;
1777 Instruction *I = dyn_cast<Instruction>(V);
1778 if (!I) return false;
1779 switch (I->getOpcode()) {
1780 default: return false; // Unhandled case.
1781 case Instruction::BitCast:
1782 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
1784 case Instruction::ZExt:
1785 if (!isMultipleOfTypeSize(
1786 I->getOperand(0)->getType()->getPrimitiveSizeInBits(),
1789 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
1791 case Instruction::Or:
1792 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
1794 collectInsertionElements(I->getOperand(1), Shift, Elements, VecEltTy,
1796 case Instruction::Shl: {
1797 // Must be shifting by a constant that is a multiple of the element size.
1798 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1));
1799 if (!CI) return false;
1800 Shift += CI->getZExtValue();
1801 if (!isMultipleOfTypeSize(Shift, VecEltTy)) return false;
1802 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
1810 /// If the input is an 'or' instruction, we may be doing shifts and ors to
1811 /// assemble the elements of the vector manually.
1812 /// Try to rip the code out and replace it with insertelements. This is to
1813 /// optimize code like this:
1815 /// %tmp37 = bitcast float %inc to i32
1816 /// %tmp38 = zext i32 %tmp37 to i64
1817 /// %tmp31 = bitcast float %inc5 to i32
1818 /// %tmp32 = zext i32 %tmp31 to i64
1819 /// %tmp33 = shl i64 %tmp32, 32
1820 /// %ins35 = or i64 %tmp33, %tmp38
1821 /// %tmp43 = bitcast i64 %ins35 to <2 x float>
1823 /// Into two insertelements that do "buildvector{%inc, %inc5}".
1824 static Value *optimizeIntegerToVectorInsertions(BitCastInst &CI,
1826 VectorType *DestVecTy = cast<VectorType>(CI.getType());
1827 Value *IntInput = CI.getOperand(0);
1829 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements());
1830 if (!collectInsertionElements(IntInput, 0, Elements,
1831 DestVecTy->getElementType(),
1832 IC.getDataLayout().isBigEndian()))
1835 // If we succeeded, we know that all of the element are specified by Elements
1836 // or are zero if Elements has a null entry. Recast this as a set of
1838 Value *Result = Constant::getNullValue(CI.getType());
1839 for (unsigned i = 0, e = Elements.size(); i != e; ++i) {
1840 if (!Elements[i]) continue; // Unset element.
1842 Result = IC.Builder->CreateInsertElement(Result, Elements[i],
1843 IC.Builder->getInt32(i));
1849 /// Canonicalize scalar bitcasts of extracted elements into a bitcast of the
1850 /// vector followed by extract element. The backend tends to handle bitcasts of
1851 /// vectors better than bitcasts of scalars because vector registers are
1852 /// usually not type-specific like scalar integer or scalar floating-point.
1853 static Instruction *canonicalizeBitCastExtElt(BitCastInst &BitCast,
1855 const DataLayout &DL) {
1856 // TODO: Create and use a pattern matcher for ExtractElementInst.
1857 auto *ExtElt = dyn_cast<ExtractElementInst>(BitCast.getOperand(0));
1858 if (!ExtElt || !ExtElt->hasOneUse())
1861 // The bitcast must be to a vectorizable type, otherwise we can't make a new
1862 // type to extract from.
1863 Type *DestType = BitCast.getType();
1864 if (!VectorType::isValidElementType(DestType))
1867 unsigned NumElts = ExtElt->getVectorOperandType()->getNumElements();
1868 auto *NewVecType = VectorType::get(DestType, NumElts);
1869 auto *NewBC = IC.Builder->CreateBitCast(ExtElt->getVectorOperand(),
1871 return ExtractElementInst::Create(NewBC, ExtElt->getIndexOperand());
1874 /// Change the type of a bitwise logic operation if we can eliminate a bitcast.
1875 static Instruction *foldBitCastBitwiseLogic(BitCastInst &BitCast,
1876 InstCombiner::BuilderTy &Builder) {
1877 Type *DestTy = BitCast.getType();
1879 if (!DestTy->getScalarType()->isIntegerTy() ||
1880 !match(BitCast.getOperand(0), m_OneUse(m_BinOp(BO))) ||
1881 !BO->isBitwiseLogicOp())
1884 // FIXME: This transform is restricted to vector types to avoid backend
1885 // problems caused by creating potentially illegal operations. If a fix-up is
1886 // added to handle that situation, we can remove this check.
1887 if (!DestTy->isVectorTy() || !BO->getType()->isVectorTy())
1891 if (match(BO->getOperand(0), m_OneUse(m_BitCast(m_Value(X)))) &&
1892 X->getType() == DestTy && !isa<Constant>(X)) {
1893 // bitcast(logic(bitcast(X), Y)) --> logic'(X, bitcast(Y))
1894 Value *CastedOp1 = Builder.CreateBitCast(BO->getOperand(1), DestTy);
1895 return BinaryOperator::Create(BO->getOpcode(), X, CastedOp1);
1898 if (match(BO->getOperand(1), m_OneUse(m_BitCast(m_Value(X)))) &&
1899 X->getType() == DestTy && !isa<Constant>(X)) {
1900 // bitcast(logic(Y, bitcast(X))) --> logic'(bitcast(Y), X)
1901 Value *CastedOp0 = Builder.CreateBitCast(BO->getOperand(0), DestTy);
1902 return BinaryOperator::Create(BO->getOpcode(), CastedOp0, X);
1908 /// Change the type of a select if we can eliminate a bitcast.
1909 static Instruction *foldBitCastSelect(BitCastInst &BitCast,
1910 InstCombiner::BuilderTy &Builder) {
1911 Value *Cond, *TVal, *FVal;
1912 if (!match(BitCast.getOperand(0),
1913 m_OneUse(m_Select(m_Value(Cond), m_Value(TVal), m_Value(FVal)))))
1916 // A vector select must maintain the same number of elements in its operands.
1917 Type *CondTy = Cond->getType();
1918 Type *DestTy = BitCast.getType();
1919 if (CondTy->isVectorTy()) {
1920 if (!DestTy->isVectorTy())
1922 if (DestTy->getVectorNumElements() != CondTy->getVectorNumElements())
1926 // FIXME: This transform is restricted from changing the select between
1927 // scalars and vectors to avoid backend problems caused by creating
1928 // potentially illegal operations. If a fix-up is added to handle that
1929 // situation, we can remove this check.
1930 if (DestTy->isVectorTy() != TVal->getType()->isVectorTy())
1933 auto *Sel = cast<Instruction>(BitCast.getOperand(0));
1935 if (match(TVal, m_OneUse(m_BitCast(m_Value(X)))) && X->getType() == DestTy &&
1936 !isa<Constant>(X)) {
1937 // bitcast(select(Cond, bitcast(X), Y)) --> select'(Cond, X, bitcast(Y))
1938 Value *CastedVal = Builder.CreateBitCast(FVal, DestTy);
1939 return SelectInst::Create(Cond, X, CastedVal, "", nullptr, Sel);
1942 if (match(FVal, m_OneUse(m_BitCast(m_Value(X)))) && X->getType() == DestTy &&
1943 !isa<Constant>(X)) {
1944 // bitcast(select(Cond, Y, bitcast(X))) --> select'(Cond, bitcast(Y), X)
1945 Value *CastedVal = Builder.CreateBitCast(TVal, DestTy);
1946 return SelectInst::Create(Cond, CastedVal, X, "", nullptr, Sel);
1952 /// Check if all users of CI are StoreInsts.
1953 static bool hasStoreUsersOnly(CastInst &CI) {
1954 for (User *U : CI.users()) {
1955 if (!isa<StoreInst>(U))
1961 /// This function handles following case
1967 /// All the related PHI nodes can be replaced by new PHI nodes with type A.
1968 /// The uses of \p CI can be changed to the new PHI node corresponding to \p PN.
1969 Instruction *InstCombiner::optimizeBitCastFromPhi(CastInst &CI, PHINode *PN) {
1970 // BitCast used by Store can be handled in InstCombineLoadStoreAlloca.cpp.
1971 if (hasStoreUsersOnly(CI))
1974 Value *Src = CI.getOperand(0);
1975 Type *SrcTy = Src->getType(); // Type B
1976 Type *DestTy = CI.getType(); // Type A
1978 SmallVector<PHINode *, 4> PhiWorklist;
1979 SmallSetVector<PHINode *, 4> OldPhiNodes;
1981 // Find all of the A->B casts and PHI nodes.
1982 // We need to inpect all related PHI nodes, but PHIs can be cyclic, so
1983 // OldPhiNodes is used to track all known PHI nodes, before adding a new
1984 // PHI to PhiWorklist, it is checked against and added to OldPhiNodes first.
1985 PhiWorklist.push_back(PN);
1986 OldPhiNodes.insert(PN);
1987 while (!PhiWorklist.empty()) {
1988 auto *OldPN = PhiWorklist.pop_back_val();
1989 for (Value *IncValue : OldPN->incoming_values()) {
1990 if (isa<Constant>(IncValue))
1993 if (auto *LI = dyn_cast<LoadInst>(IncValue)) {
1994 // If there is a sequence of one or more load instructions, each loaded
1995 // value is used as address of later load instruction, bitcast is
1996 // necessary to change the value type, don't optimize it. For
1997 // simplicity we give up if the load address comes from another load.
1998 Value *Addr = LI->getOperand(0);
1999 if (Addr == &CI || isa<LoadInst>(Addr))
2001 if (LI->hasOneUse() && LI->isSimple())
2003 // If a LoadInst has more than one use, changing the type of loaded
2004 // value may create another bitcast.
2008 if (auto *PNode = dyn_cast<PHINode>(IncValue)) {
2009 if (OldPhiNodes.insert(PNode))
2010 PhiWorklist.push_back(PNode);
2014 auto *BCI = dyn_cast<BitCastInst>(IncValue);
2015 // We can't handle other instructions.
2019 // Verify it's a A->B cast.
2020 Type *TyA = BCI->getOperand(0)->getType();
2021 Type *TyB = BCI->getType();
2022 if (TyA != DestTy || TyB != SrcTy)
2027 // For each old PHI node, create a corresponding new PHI node with a type A.
2028 SmallDenseMap<PHINode *, PHINode *> NewPNodes;
2029 for (auto *OldPN : OldPhiNodes) {
2030 Builder->SetInsertPoint(OldPN);
2031 PHINode *NewPN = Builder->CreatePHI(DestTy, OldPN->getNumOperands());
2032 NewPNodes[OldPN] = NewPN;
2035 // Fill in the operands of new PHI nodes.
2036 for (auto *OldPN : OldPhiNodes) {
2037 PHINode *NewPN = NewPNodes[OldPN];
2038 for (unsigned j = 0, e = OldPN->getNumOperands(); j != e; ++j) {
2039 Value *V = OldPN->getOperand(j);
2040 Value *NewV = nullptr;
2041 if (auto *C = dyn_cast<Constant>(V)) {
2042 NewV = ConstantExpr::getBitCast(C, DestTy);
2043 } else if (auto *LI = dyn_cast<LoadInst>(V)) {
2044 Builder->SetInsertPoint(LI->getNextNode());
2045 NewV = Builder->CreateBitCast(LI, DestTy);
2047 } else if (auto *BCI = dyn_cast<BitCastInst>(V)) {
2048 NewV = BCI->getOperand(0);
2049 } else if (auto *PrevPN = dyn_cast<PHINode>(V)) {
2050 NewV = NewPNodes[PrevPN];
2053 NewPN->addIncoming(NewV, OldPN->getIncomingBlock(j));
2057 // If there is a store with type B, change it to type A.
2058 for (User *U : PN->users()) {
2059 auto *SI = dyn_cast<StoreInst>(U);
2060 if (SI && SI->isSimple() && SI->getOperand(0) == PN) {
2061 Builder->SetInsertPoint(SI);
2063 cast<BitCastInst>(Builder->CreateBitCast(NewPNodes[PN], SrcTy));
2064 SI->setOperand(0, NewBC);
2066 assert(hasStoreUsersOnly(*NewBC));
2070 return replaceInstUsesWith(CI, NewPNodes[PN]);
2073 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
2074 // If the operands are integer typed then apply the integer transforms,
2075 // otherwise just apply the common ones.
2076 Value *Src = CI.getOperand(0);
2077 Type *SrcTy = Src->getType();
2078 Type *DestTy = CI.getType();
2080 // Get rid of casts from one type to the same type. These are useless and can
2081 // be replaced by the operand.
2082 if (DestTy == Src->getType())
2083 return replaceInstUsesWith(CI, Src);
2085 if (PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
2086 PointerType *SrcPTy = cast<PointerType>(SrcTy);
2087 Type *DstElTy = DstPTy->getElementType();
2088 Type *SrcElTy = SrcPTy->getElementType();
2090 // If we are casting a alloca to a pointer to a type of the same
2091 // size, rewrite the allocation instruction to allocate the "right" type.
2092 // There is no need to modify malloc calls because it is their bitcast that
2093 // needs to be cleaned up.
2094 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src))
2095 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
2098 // When the type pointed to is not sized the cast cannot be
2099 // turned into a gep.
2101 cast<PointerType>(Src->getType()->getScalarType())->getElementType();
2102 if (!PointeeType->isSized())
2105 // If the source and destination are pointers, and this cast is equivalent
2106 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
2107 // This can enhance SROA and other transforms that want type-safe pointers.
2108 unsigned NumZeros = 0;
2109 while (SrcElTy != DstElTy &&
2110 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() &&
2111 SrcElTy->getNumContainedTypes() /* not "{}" */) {
2112 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(0U);
2116 // If we found a path from the src to dest, create the getelementptr now.
2117 if (SrcElTy == DstElTy) {
2118 SmallVector<Value *, 8> Idxs(NumZeros + 1, Builder->getInt32(0));
2119 return GetElementPtrInst::CreateInBounds(Src, Idxs);
2123 if (VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
2124 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) {
2125 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
2126 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
2127 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
2128 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
2131 if (isa<IntegerType>(SrcTy)) {
2132 // If this is a cast from an integer to vector, check to see if the input
2133 // is a trunc or zext of a bitcast from vector. If so, we can replace all
2134 // the casts with a shuffle and (potentially) a bitcast.
2135 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) {
2136 CastInst *SrcCast = cast<CastInst>(Src);
2137 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0)))
2138 if (isa<VectorType>(BCIn->getOperand(0)->getType()))
2139 if (Instruction *I = optimizeVectorResize(BCIn->getOperand(0),
2140 cast<VectorType>(DestTy), *this))
2144 // If the input is an 'or' instruction, we may be doing shifts and ors to
2145 // assemble the elements of the vector manually. Try to rip the code out
2146 // and replace it with insertelements.
2147 if (Value *V = optimizeIntegerToVectorInsertions(CI, *this))
2148 return replaceInstUsesWith(CI, V);
2152 if (VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
2153 if (SrcVTy->getNumElements() == 1) {
2154 // If our destination is not a vector, then make this a straight
2155 // scalar-scalar cast.
2156 if (!DestTy->isVectorTy()) {
2158 Builder->CreateExtractElement(Src,
2159 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
2160 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
2163 // Otherwise, see if our source is an insert. If so, then use the scalar
2164 // component directly.
2165 if (InsertElementInst *IEI =
2166 dyn_cast<InsertElementInst>(CI.getOperand(0)))
2167 return CastInst::Create(Instruction::BitCast, IEI->getOperand(1),
2172 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
2173 // Okay, we have (bitcast (shuffle ..)). Check to see if this is
2174 // a bitcast to a vector with the same # elts.
2175 if (SVI->hasOneUse() && DestTy->isVectorTy() &&
2176 DestTy->getVectorNumElements() == SVI->getType()->getNumElements() &&
2177 SVI->getType()->getNumElements() ==
2178 SVI->getOperand(0)->getType()->getVectorNumElements()) {
2180 // If either of the operands is a cast from CI.getType(), then
2181 // evaluating the shuffle in the casted destination's type will allow
2182 // us to eliminate at least one cast.
2183 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) &&
2184 Tmp->getOperand(0)->getType() == DestTy) ||
2185 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) &&
2186 Tmp->getOperand(0)->getType() == DestTy)) {
2187 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy);
2188 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy);
2189 // Return a new shuffle vector. Use the same element ID's, as we
2190 // know the vector types match #elts.
2191 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
2196 // Handle the A->B->A cast, and there is an intervening PHI node.
2197 if (PHINode *PN = dyn_cast<PHINode>(Src))
2198 if (Instruction *I = optimizeBitCastFromPhi(CI, PN))
2201 if (Instruction *I = canonicalizeBitCastExtElt(CI, *this, DL))
2204 if (Instruction *I = foldBitCastBitwiseLogic(CI, *Builder))
2207 if (Instruction *I = foldBitCastSelect(CI, *Builder))
2210 if (SrcTy->isPointerTy())
2211 return commonPointerCastTransforms(CI);
2212 return commonCastTransforms(CI);
2215 Instruction *InstCombiner::visitAddrSpaceCast(AddrSpaceCastInst &CI) {
2216 // If the destination pointer element type is not the same as the source's
2217 // first do a bitcast to the destination type, and then the addrspacecast.
2218 // This allows the cast to be exposed to other transforms.
2219 Value *Src = CI.getOperand(0);
2220 PointerType *SrcTy = cast<PointerType>(Src->getType()->getScalarType());
2221 PointerType *DestTy = cast<PointerType>(CI.getType()->getScalarType());
2223 Type *DestElemTy = DestTy->getElementType();
2224 if (SrcTy->getElementType() != DestElemTy) {
2225 Type *MidTy = PointerType::get(DestElemTy, SrcTy->getAddressSpace());
2226 if (VectorType *VT = dyn_cast<VectorType>(CI.getType())) {
2227 // Handle vectors of pointers.
2228 MidTy = VectorType::get(MidTy, VT->getNumElements());
2231 Value *NewBitCast = Builder->CreateBitCast(Src, MidTy);
2232 return new AddrSpaceCastInst(NewBitCast, CI.getType());
2235 return commonPointerCastTransforms(CI);