1 //===- InstCombineVectorOps.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 instcombine for ExtractElement, InsertElement and
13 //===----------------------------------------------------------------------===//
15 #include "InstCombine.h"
16 #include "llvm/IR/PatternMatch.h"
18 using namespace PatternMatch;
20 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
21 /// is to leave as a vector operation. isConstant indicates whether we're
22 /// extracting one known element. If false we're extracting a variable index.
23 static bool CheapToScalarize(Value *V, bool isConstant) {
24 if (Constant *C = dyn_cast<Constant>(V)) {
25 if (isConstant) return true;
27 // If all elts are the same, we can extract it and use any of the values.
28 if (Constant *Op0 = C->getAggregateElement(0U)) {
29 for (unsigned i = 1, e = V->getType()->getVectorNumElements(); i != e;
31 if (C->getAggregateElement(i) != Op0)
36 Instruction *I = dyn_cast<Instruction>(V);
39 // Insert element gets simplified to the inserted element or is deleted if
40 // this is constant idx extract element and its a constant idx insertelt.
41 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
42 isa<ConstantInt>(I->getOperand(2)))
44 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
46 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
47 if (BO->hasOneUse() &&
48 (CheapToScalarize(BO->getOperand(0), isConstant) ||
49 CheapToScalarize(BO->getOperand(1), isConstant)))
51 if (CmpInst *CI = dyn_cast<CmpInst>(I))
52 if (CI->hasOneUse() &&
53 (CheapToScalarize(CI->getOperand(0), isConstant) ||
54 CheapToScalarize(CI->getOperand(1), isConstant)))
60 /// FindScalarElement - Given a vector and an element number, see if the scalar
61 /// value is already around as a register, for example if it were inserted then
62 /// extracted from the vector.
63 static Value *FindScalarElement(Value *V, unsigned EltNo) {
64 assert(V->getType()->isVectorTy() && "Not looking at a vector?");
65 VectorType *VTy = cast<VectorType>(V->getType());
66 unsigned Width = VTy->getNumElements();
67 if (EltNo >= Width) // Out of range access.
68 return UndefValue::get(VTy->getElementType());
70 if (Constant *C = dyn_cast<Constant>(V))
71 return C->getAggregateElement(EltNo);
73 if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
74 // If this is an insert to a variable element, we don't know what it is.
75 if (!isa<ConstantInt>(III->getOperand(2)))
77 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
79 // If this is an insert to the element we are looking for, return the
82 return III->getOperand(1);
84 // Otherwise, the insertelement doesn't modify the value, recurse on its
86 return FindScalarElement(III->getOperand(0), EltNo);
89 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
90 unsigned LHSWidth = SVI->getOperand(0)->getType()->getVectorNumElements();
91 int InEl = SVI->getMaskValue(EltNo);
93 return UndefValue::get(VTy->getElementType());
94 if (InEl < (int)LHSWidth)
95 return FindScalarElement(SVI->getOperand(0), InEl);
96 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth);
99 // Extract a value from a vector add operation with a constant zero.
100 Value *Val = 0; Constant *Con = 0;
101 if (match(V, m_Add(m_Value(Val), m_Constant(Con)))) {
102 if (Con->getAggregateElement(EltNo)->isNullValue())
103 return FindScalarElement(Val, EltNo);
106 // Otherwise, we don't know.
110 // If we have a PHI node with a vector type that has only 2 uses: feed
111 // itself and be an operand of extractelement at a constant location,
112 // try to replace the PHI of the vector type with a PHI of a scalar type.
113 Instruction *InstCombiner::scalarizePHI(ExtractElementInst &EI, PHINode *PN) {
114 // Verify that the PHI node has exactly 2 uses. Otherwise return NULL.
115 if (!PN->hasNUses(2))
118 // If so, it's known at this point that one operand is PHI and the other is
119 // an extractelement node. Find the PHI user that is not the extractelement
121 Value::use_iterator iu = PN->use_begin();
122 Instruction *PHIUser = dyn_cast<Instruction>(*iu);
123 if (PHIUser == cast<Instruction>(&EI))
124 PHIUser = cast<Instruction>(*(++iu));
126 // Verify that this PHI user has one use, which is the PHI itself,
127 // and that it is a binary operation which is cheap to scalarize.
128 // otherwise return NULL.
129 if (!PHIUser->hasOneUse() || !(PHIUser->use_back() == PN) ||
130 !(isa<BinaryOperator>(PHIUser)) || !CheapToScalarize(PHIUser, true))
133 // Create a scalar PHI node that will replace the vector PHI node
134 // just before the current PHI node.
135 PHINode *scalarPHI = cast<PHINode>(InsertNewInstWith(
136 PHINode::Create(EI.getType(), PN->getNumIncomingValues(), ""), *PN));
137 // Scalarize each PHI operand.
138 for (unsigned i = 0; i < PN->getNumIncomingValues(); i++) {
139 Value *PHIInVal = PN->getIncomingValue(i);
140 BasicBlock *inBB = PN->getIncomingBlock(i);
141 Value *Elt = EI.getIndexOperand();
142 // If the operand is the PHI induction variable:
143 if (PHIInVal == PHIUser) {
144 // Scalarize the binary operation. Its first operand is the
145 // scalar PHI and the second operand is extracted from the other
147 BinaryOperator *B0 = cast<BinaryOperator>(PHIUser);
148 unsigned opId = (B0->getOperand(0) == PN) ? 1 : 0;
149 Value *Op = InsertNewInstWith(
150 ExtractElementInst::Create(B0->getOperand(opId), Elt,
151 B0->getOperand(opId)->getName() + ".Elt"),
153 Value *newPHIUser = InsertNewInstWith(
154 BinaryOperator::Create(B0->getOpcode(), scalarPHI, Op), *B0);
155 scalarPHI->addIncoming(newPHIUser, inBB);
157 // Scalarize PHI input:
158 Instruction *newEI = ExtractElementInst::Create(PHIInVal, Elt, "");
159 // Insert the new instruction into the predecessor basic block.
160 Instruction *pos = dyn_cast<Instruction>(PHIInVal);
161 BasicBlock::iterator InsertPos;
162 if (pos && !isa<PHINode>(pos)) {
166 InsertPos = inBB->getFirstInsertionPt();
169 InsertNewInstWith(newEI, *InsertPos);
171 scalarPHI->addIncoming(newEI, inBB);
174 return ReplaceInstUsesWith(EI, scalarPHI);
177 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
178 // If vector val is constant with all elements the same, replace EI with
179 // that element. We handle a known element # below.
180 if (Constant *C = dyn_cast<Constant>(EI.getOperand(0)))
181 if (CheapToScalarize(C, false))
182 return ReplaceInstUsesWith(EI, C->getAggregateElement(0U));
184 // If extracting a specified index from the vector, see if we can recursively
185 // find a previously computed scalar that was inserted into the vector.
186 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
187 unsigned IndexVal = IdxC->getZExtValue();
188 unsigned VectorWidth = EI.getVectorOperandType()->getNumElements();
190 // If this is extracting an invalid index, turn this into undef, to avoid
191 // crashing the code below.
192 if (IndexVal >= VectorWidth)
193 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
195 // This instruction only demands the single element from the input vector.
196 // If the input vector has a single use, simplify it based on this use
198 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
199 APInt UndefElts(VectorWidth, 0);
200 APInt DemandedMask(VectorWidth, 0);
201 DemandedMask.setBit(IndexVal);
202 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
203 DemandedMask, UndefElts)) {
209 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
210 return ReplaceInstUsesWith(EI, Elt);
212 // If the this extractelement is directly using a bitcast from a vector of
213 // the same number of elements, see if we can find the source element from
214 // it. In this case, we will end up needing to bitcast the scalars.
215 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
216 if (VectorType *VT = dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
217 if (VT->getNumElements() == VectorWidth)
218 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
219 return new BitCastInst(Elt, EI.getType());
222 // If there's a vector PHI feeding a scalar use through this extractelement
223 // instruction, try to scalarize the PHI.
224 if (PHINode *PN = dyn_cast<PHINode>(EI.getOperand(0))) {
225 Instruction *scalarPHI = scalarizePHI(EI, PN);
231 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
232 // Push extractelement into predecessor operation if legal and
233 // profitable to do so
234 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
235 if (I->hasOneUse() &&
236 CheapToScalarize(BO, isa<ConstantInt>(EI.getOperand(1)))) {
238 Builder->CreateExtractElement(BO->getOperand(0), EI.getOperand(1),
239 EI.getName()+".lhs");
241 Builder->CreateExtractElement(BO->getOperand(1), EI.getOperand(1),
242 EI.getName()+".rhs");
243 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
245 } else if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
246 // Extracting the inserted element?
247 if (IE->getOperand(2) == EI.getOperand(1))
248 return ReplaceInstUsesWith(EI, IE->getOperand(1));
249 // If the inserted and extracted elements are constants, they must not
250 // be the same value, extract from the pre-inserted value instead.
251 if (isa<Constant>(IE->getOperand(2)) && isa<Constant>(EI.getOperand(1))) {
252 Worklist.AddValue(EI.getOperand(0));
253 EI.setOperand(0, IE->getOperand(0));
256 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
257 // If this is extracting an element from a shufflevector, figure out where
258 // it came from and extract from the appropriate input element instead.
259 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
260 int SrcIdx = SVI->getMaskValue(Elt->getZExtValue());
263 SVI->getOperand(0)->getType()->getVectorNumElements();
266 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
267 if (SrcIdx < (int)LHSWidth)
268 Src = SVI->getOperand(0);
271 Src = SVI->getOperand(1);
273 Type *Int32Ty = Type::getInt32Ty(EI.getContext());
274 return ExtractElementInst::Create(Src,
275 ConstantInt::get(Int32Ty,
278 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
279 // Canonicalize extractelement(cast) -> cast(extractelement)
280 // bitcasts can change the number of vector elements and they cost nothing
281 if (CI->hasOneUse() && (CI->getOpcode() != Instruction::BitCast)) {
282 Value *EE = Builder->CreateExtractElement(CI->getOperand(0),
283 EI.getIndexOperand());
284 Worklist.AddValue(EE);
285 return CastInst::Create(CI->getOpcode(), EE, EI.getType());
287 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
288 if (SI->hasOneUse()) {
289 // TODO: For a select on vectors, it might be useful to do this if it
290 // has multiple extractelement uses. For vector select, that seems to
291 // fight the vectorizer.
293 // If we are extracting an element from a vector select or a select on
294 // vectors, a select on the scalars extracted from the vector arguments.
295 Value *TrueVal = SI->getTrueValue();
296 Value *FalseVal = SI->getFalseValue();
298 Value *Cond = SI->getCondition();
299 if (Cond->getType()->isVectorTy()) {
300 Cond = Builder->CreateExtractElement(Cond,
301 EI.getIndexOperand(),
302 Cond->getName() + ".elt");
306 = Builder->CreateExtractElement(TrueVal,
307 EI.getIndexOperand(),
308 TrueVal->getName() + ".elt");
311 = Builder->CreateExtractElement(FalseVal,
312 EI.getIndexOperand(),
313 FalseVal->getName() + ".elt");
314 return SelectInst::Create(Cond,
317 SI->getName() + ".elt");
324 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
325 /// elements from either LHS or RHS, return the shuffle mask and true.
326 /// Otherwise, return false.
327 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
328 SmallVectorImpl<Constant*> &Mask) {
329 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
330 "Invalid CollectSingleShuffleElements");
331 unsigned NumElts = V->getType()->getVectorNumElements();
333 if (isa<UndefValue>(V)) {
334 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
339 for (unsigned i = 0; i != NumElts; ++i)
340 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
345 for (unsigned i = 0; i != NumElts; ++i)
346 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()),
351 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
352 // If this is an insert of an extract from some other vector, include it.
353 Value *VecOp = IEI->getOperand(0);
354 Value *ScalarOp = IEI->getOperand(1);
355 Value *IdxOp = IEI->getOperand(2);
357 if (!isa<ConstantInt>(IdxOp))
359 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
361 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
362 // Okay, we can handle this if the vector we are insertinting into is
364 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
365 // If so, update the mask to reflect the inserted undef.
366 Mask[InsertedIdx] = UndefValue::get(Type::getInt32Ty(V->getContext()));
369 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
370 if (isa<ConstantInt>(EI->getOperand(1)) &&
371 EI->getOperand(0)->getType() == V->getType()) {
372 unsigned ExtractedIdx =
373 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
375 // This must be extracting from either LHS or RHS.
376 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
377 // Okay, we can handle this if the vector we are insertinting into is
379 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
380 // If so, update the mask to reflect the inserted value.
381 if (EI->getOperand(0) == LHS) {
382 Mask[InsertedIdx % NumElts] =
383 ConstantInt::get(Type::getInt32Ty(V->getContext()),
386 assert(EI->getOperand(0) == RHS);
387 Mask[InsertedIdx % NumElts] =
388 ConstantInt::get(Type::getInt32Ty(V->getContext()),
389 ExtractedIdx+NumElts);
397 // TODO: Handle shufflevector here!
402 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
403 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
404 /// that computes V and the LHS value of the shuffle.
405 static Value *CollectShuffleElements(Value *V, SmallVectorImpl<Constant*> &Mask,
407 assert(V->getType()->isVectorTy() &&
408 (RHS == 0 || V->getType() == RHS->getType()) &&
410 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
412 if (isa<UndefValue>(V)) {
413 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
417 if (isa<ConstantAggregateZero>(V)) {
418 Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(V->getContext()),0));
422 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
423 // If this is an insert of an extract from some other vector, include it.
424 Value *VecOp = IEI->getOperand(0);
425 Value *ScalarOp = IEI->getOperand(1);
426 Value *IdxOp = IEI->getOperand(2);
428 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
429 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
430 EI->getOperand(0)->getType() == V->getType()) {
431 unsigned ExtractedIdx =
432 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
433 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
435 // Either the extracted from or inserted into vector must be RHSVec,
436 // otherwise we'd end up with a shuffle of three inputs.
437 if (EI->getOperand(0) == RHS || RHS == 0) {
438 RHS = EI->getOperand(0);
439 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
440 Mask[InsertedIdx % NumElts] =
441 ConstantInt::get(Type::getInt32Ty(V->getContext()),
442 NumElts+ExtractedIdx);
447 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
448 // Update Mask to reflect that `ScalarOp' has been inserted at
449 // position `InsertedIdx' within the vector returned by IEI.
450 Mask[InsertedIdx % NumElts] = Mask[ExtractedIdx];
452 // Everything but the extracted element is replaced with the RHS.
453 for (unsigned i = 0; i != NumElts; ++i) {
454 if (i != InsertedIdx)
455 Mask[i] = ConstantInt::get(Type::getInt32Ty(V->getContext()),
461 // If this insertelement is a chain that comes from exactly these two
462 // vectors, return the vector and the effective shuffle.
463 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
464 return EI->getOperand(0);
468 // TODO: Handle shufflevector here!
470 // Otherwise, can't do anything fancy. Return an identity vector.
471 for (unsigned i = 0; i != NumElts; ++i)
472 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
476 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
477 Value *VecOp = IE.getOperand(0);
478 Value *ScalarOp = IE.getOperand(1);
479 Value *IdxOp = IE.getOperand(2);
481 // Inserting an undef or into an undefined place, remove this.
482 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
483 ReplaceInstUsesWith(IE, VecOp);
485 // If the inserted element was extracted from some other vector, and if the
486 // indexes are constant, try to turn this into a shufflevector operation.
487 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
488 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
489 EI->getOperand(0)->getType() == IE.getType()) {
490 unsigned NumVectorElts = IE.getType()->getNumElements();
491 unsigned ExtractedIdx =
492 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
493 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
495 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
496 return ReplaceInstUsesWith(IE, VecOp);
498 if (InsertedIdx >= NumVectorElts) // Out of range insert.
499 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
501 // If we are extracting a value from a vector, then inserting it right
502 // back into the same place, just use the input vector.
503 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
504 return ReplaceInstUsesWith(IE, VecOp);
506 // If this insertelement isn't used by some other insertelement, turn it
507 // (and any insertelements it points to), into one big shuffle.
508 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
509 SmallVector<Constant*, 16> Mask;
511 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
512 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
513 // We now have a shuffle of LHS, RHS, Mask.
514 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
519 unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements();
520 APInt UndefElts(VWidth, 0);
521 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
522 if (Value *V = SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts)) {
524 return ReplaceInstUsesWith(IE, V);
531 /// Return true if we can evaluate the specified expression tree if the vector
532 /// elements were shuffled in a different order.
533 static bool CanEvaluateShuffled(Value *V, ArrayRef<int> Mask,
534 unsigned Depth = 5) {
535 // We can always reorder the elements of a constant.
536 if (isa<Constant>(V))
539 // We won't reorder vector arguments. No IPO here.
540 Instruction *I = dyn_cast<Instruction>(V);
541 if (!I) return false;
543 // Two users may expect different orders of the elements. Don't try it.
547 if (Depth == 0) return false;
549 switch (I->getOpcode()) {
550 case Instruction::Add:
551 case Instruction::FAdd:
552 case Instruction::Sub:
553 case Instruction::FSub:
554 case Instruction::Mul:
555 case Instruction::FMul:
556 case Instruction::UDiv:
557 case Instruction::SDiv:
558 case Instruction::FDiv:
559 case Instruction::URem:
560 case Instruction::SRem:
561 case Instruction::FRem:
562 case Instruction::Shl:
563 case Instruction::LShr:
564 case Instruction::AShr:
565 case Instruction::And:
566 case Instruction::Or:
567 case Instruction::Xor:
568 case Instruction::ICmp:
569 case Instruction::FCmp:
570 case Instruction::Trunc:
571 case Instruction::ZExt:
572 case Instruction::SExt:
573 case Instruction::FPToUI:
574 case Instruction::FPToSI:
575 case Instruction::UIToFP:
576 case Instruction::SIToFP:
577 case Instruction::FPTrunc:
578 case Instruction::FPExt:
579 case Instruction::GetElementPtr: {
580 for (int i = 0, e = I->getNumOperands(); i != e; ++i) {
581 if (!CanEvaluateShuffled(I->getOperand(i), Mask, Depth-1))
586 case Instruction::InsertElement: {
587 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(2));
588 if (!CI) return false;
589 int ElementNumber = CI->getLimitedValue();
591 // Verify that 'CI' does not occur twice in Mask. A single 'insertelement'
592 // can't put an element into multiple indices.
593 bool SeenOnce = false;
594 for (int i = 0, e = Mask.size(); i != e; ++i) {
595 if (Mask[i] == ElementNumber) {
601 return CanEvaluateShuffled(I->getOperand(0), Mask, Depth-1);
607 /// Rebuild a new instruction just like 'I' but with the new operands given.
608 /// In the event of type mismatch, the type of the operands is correct.
609 static Value *BuildNew(Instruction *I, ArrayRef<Value*> NewOps) {
610 // We don't want to use the IRBuilder here because we want the replacement
611 // instructions to appear next to 'I', not the builder's insertion point.
612 switch (I->getOpcode()) {
613 case Instruction::Add:
614 case Instruction::FAdd:
615 case Instruction::Sub:
616 case Instruction::FSub:
617 case Instruction::Mul:
618 case Instruction::FMul:
619 case Instruction::UDiv:
620 case Instruction::SDiv:
621 case Instruction::FDiv:
622 case Instruction::URem:
623 case Instruction::SRem:
624 case Instruction::FRem:
625 case Instruction::Shl:
626 case Instruction::LShr:
627 case Instruction::AShr:
628 case Instruction::And:
629 case Instruction::Or:
630 case Instruction::Xor: {
631 BinaryOperator *BO = cast<BinaryOperator>(I);
632 assert(NewOps.size() == 2 && "binary operator with #ops != 2");
633 BinaryOperator *New =
634 BinaryOperator::Create(cast<BinaryOperator>(I)->getOpcode(),
635 NewOps[0], NewOps[1], "", BO);
636 if (isa<OverflowingBinaryOperator>(BO)) {
637 New->setHasNoUnsignedWrap(BO->hasNoUnsignedWrap());
638 New->setHasNoSignedWrap(BO->hasNoSignedWrap());
640 if (isa<PossiblyExactOperator>(BO)) {
641 New->setIsExact(BO->isExact());
643 if (isa<FPMathOperator>(BO))
644 New->copyFastMathFlags(I);
647 case Instruction::ICmp:
648 assert(NewOps.size() == 2 && "icmp with #ops != 2");
649 return new ICmpInst(I, cast<ICmpInst>(I)->getPredicate(),
650 NewOps[0], NewOps[1]);
651 case Instruction::FCmp:
652 assert(NewOps.size() == 2 && "fcmp with #ops != 2");
653 return new FCmpInst(I, cast<FCmpInst>(I)->getPredicate(),
654 NewOps[0], NewOps[1]);
655 case Instruction::Trunc:
656 case Instruction::ZExt:
657 case Instruction::SExt:
658 case Instruction::FPToUI:
659 case Instruction::FPToSI:
660 case Instruction::UIToFP:
661 case Instruction::SIToFP:
662 case Instruction::FPTrunc:
663 case Instruction::FPExt: {
664 // It's possible that the mask has a different number of elements from
665 // the original cast. We recompute the destination type to match the mask.
667 VectorType::get(I->getType()->getScalarType(),
668 NewOps[0]->getType()->getVectorNumElements());
669 assert(NewOps.size() == 1 && "cast with #ops != 1");
670 return CastInst::Create(cast<CastInst>(I)->getOpcode(), NewOps[0], DestTy,
673 case Instruction::GetElementPtr: {
674 Value *Ptr = NewOps[0];
675 ArrayRef<Value*> Idx = NewOps.slice(1);
676 GetElementPtrInst *GEP = GetElementPtrInst::Create(Ptr, Idx, "", I);
677 GEP->setIsInBounds(cast<GetElementPtrInst>(I)->isInBounds());
681 llvm_unreachable("failed to rebuild vector instructions");
685 InstCombiner::EvaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask) {
686 // Mask.size() does not need to be equal to the number of vector elements.
688 assert(V->getType()->isVectorTy() && "can't reorder non-vector elements");
689 if (isa<UndefValue>(V)) {
690 return UndefValue::get(VectorType::get(V->getType()->getScalarType(),
693 if (isa<ConstantAggregateZero>(V)) {
694 return ConstantAggregateZero::get(
695 VectorType::get(V->getType()->getScalarType(),
698 if (Constant *C = dyn_cast<Constant>(V)) {
699 SmallVector<Constant *, 16> MaskValues;
700 for (int i = 0, e = Mask.size(); i != e; ++i) {
702 MaskValues.push_back(UndefValue::get(Builder->getInt32Ty()));
704 MaskValues.push_back(Builder->getInt32(Mask[i]));
706 return ConstantExpr::getShuffleVector(C, UndefValue::get(C->getType()),
707 ConstantVector::get(MaskValues));
710 Instruction *I = cast<Instruction>(V);
711 switch (I->getOpcode()) {
712 case Instruction::Add:
713 case Instruction::FAdd:
714 case Instruction::Sub:
715 case Instruction::FSub:
716 case Instruction::Mul:
717 case Instruction::FMul:
718 case Instruction::UDiv:
719 case Instruction::SDiv:
720 case Instruction::FDiv:
721 case Instruction::URem:
722 case Instruction::SRem:
723 case Instruction::FRem:
724 case Instruction::Shl:
725 case Instruction::LShr:
726 case Instruction::AShr:
727 case Instruction::And:
728 case Instruction::Or:
729 case Instruction::Xor:
730 case Instruction::ICmp:
731 case Instruction::FCmp:
732 case Instruction::Trunc:
733 case Instruction::ZExt:
734 case Instruction::SExt:
735 case Instruction::FPToUI:
736 case Instruction::FPToSI:
737 case Instruction::UIToFP:
738 case Instruction::SIToFP:
739 case Instruction::FPTrunc:
740 case Instruction::FPExt:
741 case Instruction::Select:
742 case Instruction::GetElementPtr: {
743 SmallVector<Value*, 8> NewOps;
744 bool NeedsRebuild = (Mask.size() != I->getType()->getVectorNumElements());
745 for (int i = 0, e = I->getNumOperands(); i != e; ++i) {
746 Value *V = EvaluateInDifferentElementOrder(I->getOperand(i), Mask);
748 NeedsRebuild |= (V != I->getOperand(i));
751 return BuildNew(I, NewOps);
755 case Instruction::InsertElement: {
756 int Element = cast<ConstantInt>(I->getOperand(2))->getLimitedValue();
758 // The insertelement was inserting at Element. Figure out which element
759 // that becomes after shuffling. The answer is guaranteed to be unique
760 // by CanEvaluateShuffled.
763 for (int e = Mask.size(); Index != e; ++Index) {
764 if (Mask[Index] == Element) {
770 // If element is not in Mask, no need to handle the operand 1 (element to
771 // be inserted). Just evaluate values in operand 0 according to Mask.
773 return EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
775 Value *V = EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
776 return InsertElementInst::Create(V, I->getOperand(1),
777 Builder->getInt32(Index), "", I);
780 llvm_unreachable("failed to reorder elements of vector instruction!");
783 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
784 Value *LHS = SVI.getOperand(0);
785 Value *RHS = SVI.getOperand(1);
786 SmallVector<int, 16> Mask = SVI.getShuffleMask();
788 bool MadeChange = false;
790 // Undefined shuffle mask -> undefined value.
791 if (isa<UndefValue>(SVI.getOperand(2)))
792 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
794 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
796 APInt UndefElts(VWidth, 0);
797 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
798 if (Value *V = SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
800 return ReplaceInstUsesWith(SVI, V);
801 LHS = SVI.getOperand(0);
802 RHS = SVI.getOperand(1);
806 unsigned LHSWidth = cast<VectorType>(LHS->getType())->getNumElements();
808 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
809 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
810 if (LHS == RHS || isa<UndefValue>(LHS)) {
811 if (isa<UndefValue>(LHS) && LHS == RHS) {
812 // shuffle(undef,undef,mask) -> undef.
813 Value *Result = (VWidth == LHSWidth)
814 ? LHS : UndefValue::get(SVI.getType());
815 return ReplaceInstUsesWith(SVI, Result);
818 // Remap any references to RHS to use LHS.
819 SmallVector<Constant*, 16> Elts;
820 for (unsigned i = 0, e = LHSWidth; i != VWidth; ++i) {
822 Elts.push_back(UndefValue::get(Type::getInt32Ty(SVI.getContext())));
826 if ((Mask[i] >= (int)e && isa<UndefValue>(RHS)) ||
827 (Mask[i] < (int)e && isa<UndefValue>(LHS))) {
828 Mask[i] = -1; // Turn into undef.
829 Elts.push_back(UndefValue::get(Type::getInt32Ty(SVI.getContext())));
831 Mask[i] = Mask[i] % e; // Force to LHS.
832 Elts.push_back(ConstantInt::get(Type::getInt32Ty(SVI.getContext()),
836 SVI.setOperand(0, SVI.getOperand(1));
837 SVI.setOperand(1, UndefValue::get(RHS->getType()));
838 SVI.setOperand(2, ConstantVector::get(Elts));
839 LHS = SVI.getOperand(0);
840 RHS = SVI.getOperand(1);
844 if (VWidth == LHSWidth) {
845 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
846 bool isLHSID = true, isRHSID = true;
848 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
849 if (Mask[i] < 0) continue; // Ignore undef values.
850 // Is this an identity shuffle of the LHS value?
851 isLHSID &= (Mask[i] == (int)i);
853 // Is this an identity shuffle of the RHS value?
854 isRHSID &= (Mask[i]-e == i);
857 // Eliminate identity shuffles.
858 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
859 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
862 if (isa<UndefValue>(RHS) && CanEvaluateShuffled(LHS, Mask)) {
863 Value *V = EvaluateInDifferentElementOrder(LHS, Mask);
864 return ReplaceInstUsesWith(SVI, V);
867 // If the LHS is a shufflevector itself, see if we can combine it with this
868 // one without producing an unusual shuffle.
869 // Cases that might be simplified:
871 // x1=shuffle(v1,v2,mask1)
872 // x=shuffle(x1,undef,mask)
874 // x=shuffle(v1,undef,newMask)
875 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : -1
877 // x1=shuffle(v1,undef,mask1)
878 // x=shuffle(x1,x2,mask)
879 // where v1.size() == mask1.size()
881 // x=shuffle(v1,x2,newMask)
882 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : mask[i]
884 // x2=shuffle(v2,undef,mask2)
885 // x=shuffle(x1,x2,mask)
886 // where v2.size() == mask2.size()
888 // x=shuffle(x1,v2,newMask)
889 // newMask[i] = (mask[i] < x1.size())
890 // ? mask[i] : mask2[mask[i]-x1.size()]+x1.size()
892 // x1=shuffle(v1,undef,mask1)
893 // x2=shuffle(v2,undef,mask2)
894 // x=shuffle(x1,x2,mask)
895 // where v1.size() == v2.size()
897 // x=shuffle(v1,v2,newMask)
898 // newMask[i] = (mask[i] < x1.size())
899 // ? mask1[mask[i]] : mask2[mask[i]-x1.size()]+v1.size()
901 // Here we are really conservative:
902 // we are absolutely afraid of producing a shuffle mask not in the input
903 // program, because the code gen may not be smart enough to turn a merged
904 // shuffle into two specific shuffles: it may produce worse code. As such,
905 // we only merge two shuffles if the result is either a splat or one of the
906 // input shuffle masks. In this case, merging the shuffles just removes
907 // one instruction, which we know is safe. This is good for things like
908 // turning: (splat(splat)) -> splat, or
909 // merge(V[0..n], V[n+1..2n]) -> V[0..2n]
910 ShuffleVectorInst* LHSShuffle = dyn_cast<ShuffleVectorInst>(LHS);
911 ShuffleVectorInst* RHSShuffle = dyn_cast<ShuffleVectorInst>(RHS);
913 if (!isa<UndefValue>(LHSShuffle->getOperand(1)) && !isa<UndefValue>(RHS))
916 if (!isa<UndefValue>(RHSShuffle->getOperand(1)))
918 if (!LHSShuffle && !RHSShuffle)
919 return MadeChange ? &SVI : 0;
921 Value* LHSOp0 = NULL;
922 Value* LHSOp1 = NULL;
923 Value* RHSOp0 = NULL;
924 unsigned LHSOp0Width = 0;
925 unsigned RHSOp0Width = 0;
927 LHSOp0 = LHSShuffle->getOperand(0);
928 LHSOp1 = LHSShuffle->getOperand(1);
929 LHSOp0Width = cast<VectorType>(LHSOp0->getType())->getNumElements();
932 RHSOp0 = RHSShuffle->getOperand(0);
933 RHSOp0Width = cast<VectorType>(RHSOp0->getType())->getNumElements();
939 if (isa<UndefValue>(RHS)) {
944 else if (LHSOp0Width == LHSWidth) {
949 if (RHSShuffle && RHSOp0Width == LHSWidth) {
953 if (LHSOp0 == RHSOp0) {
958 if (newLHS == LHS && newRHS == RHS)
959 return MadeChange ? &SVI : 0;
961 SmallVector<int, 16> LHSMask;
962 SmallVector<int, 16> RHSMask;
964 LHSMask = LHSShuffle->getShuffleMask();
965 if (RHSShuffle && newRHS != RHS)
966 RHSMask = RHSShuffle->getShuffleMask();
968 unsigned newLHSWidth = (newLHS != LHS) ? LHSOp0Width : LHSWidth;
969 SmallVector<int, 16> newMask;
972 // Create a new mask for the new ShuffleVectorInst so that the new
973 // ShuffleVectorInst is equivalent to the original one.
974 for (unsigned i = 0; i < VWidth; ++i) {
977 // This element is an undef value.
979 } else if (Mask[i] < (int)LHSWidth) {
980 // This element is from left hand side vector operand.
982 // If LHS is going to be replaced (case 1, 2, or 4), calculate the
983 // new mask value for the element.
985 eltMask = LHSMask[Mask[i]];
986 // If the value selected is an undef value, explicitly specify it
987 // with a -1 mask value.
988 if (eltMask >= (int)LHSOp0Width && isa<UndefValue>(LHSOp1))
993 // This element is from right hand side vector operand
995 // If the value selected is an undef value, explicitly specify it
996 // with a -1 mask value. (case 1)
997 if (isa<UndefValue>(RHS))
999 // If RHS is going to be replaced (case 3 or 4), calculate the
1000 // new mask value for the element.
1001 else if (newRHS != RHS) {
1002 eltMask = RHSMask[Mask[i]-LHSWidth];
1003 // If the value selected is an undef value, explicitly specify it
1004 // with a -1 mask value.
1005 if (eltMask >= (int)RHSOp0Width) {
1006 assert(isa<UndefValue>(RHSShuffle->getOperand(1))
1007 && "should have been check above");
1011 eltMask = Mask[i]-LHSWidth;
1013 // If LHS's width is changed, shift the mask value accordingly.
1014 // If newRHS == NULL, i.e. LHSOp0 == RHSOp0, we want to remap any
1015 // references from RHSOp0 to LHSOp0, so we don't need to shift the mask.
1016 // If newRHS == newLHS, we want to remap any references from newRHS to
1017 // newLHS so that we can properly identify splats that may occur due to
1018 // obfuscation across the two vectors.
1019 if (eltMask >= 0 && newRHS != NULL && newLHS != newRHS)
1020 eltMask += newLHSWidth;
1023 // Check if this could still be a splat.
1025 if (SplatElt >= 0 && SplatElt != eltMask)
1030 newMask.push_back(eltMask);
1033 // If the result mask is equal to one of the original shuffle masks,
1034 // or is a splat, do the replacement.
1035 if (isSplat || newMask == LHSMask || newMask == RHSMask || newMask == Mask) {
1036 SmallVector<Constant*, 16> Elts;
1037 Type *Int32Ty = Type::getInt32Ty(SVI.getContext());
1038 for (unsigned i = 0, e = newMask.size(); i != e; ++i) {
1039 if (newMask[i] < 0) {
1040 Elts.push_back(UndefValue::get(Int32Ty));
1042 Elts.push_back(ConstantInt::get(Int32Ty, newMask[i]));
1046 newRHS = UndefValue::get(newLHS->getType());
1047 return new ShuffleVectorInst(newLHS, newRHS, ConstantVector::get(Elts));
1050 return MadeChange ? &SVI : 0;