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/Support/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 Constant *Op0 = C->getAggregateElement(0U);
29 for (unsigned i = 1, e = V->getType()->getVectorNumElements(); i != e; ++i)
30 if (C->getAggregateElement(i) != Op0)
34 Instruction *I = dyn_cast<Instruction>(V);
37 // Insert element gets simplified to the inserted element or is deleted if
38 // this is constant idx extract element and its a constant idx insertelt.
39 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
40 isa<ConstantInt>(I->getOperand(2)))
42 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
44 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
45 if (BO->hasOneUse() &&
46 (CheapToScalarize(BO->getOperand(0), isConstant) ||
47 CheapToScalarize(BO->getOperand(1), isConstant)))
49 if (CmpInst *CI = dyn_cast<CmpInst>(I))
50 if (CI->hasOneUse() &&
51 (CheapToScalarize(CI->getOperand(0), isConstant) ||
52 CheapToScalarize(CI->getOperand(1), isConstant)))
58 /// FindScalarElement - Given a vector and an element number, see if the scalar
59 /// value is already around as a register, for example if it were inserted then
60 /// extracted from the vector.
61 static Value *FindScalarElement(Value *V, unsigned EltNo) {
62 assert(V->getType()->isVectorTy() && "Not looking at a vector?");
63 VectorType *VTy = cast<VectorType>(V->getType());
64 unsigned Width = VTy->getNumElements();
65 if (EltNo >= Width) // Out of range access.
66 return UndefValue::get(VTy->getElementType());
68 if (Constant *C = dyn_cast<Constant>(V))
69 return C->getAggregateElement(EltNo);
71 if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
72 // If this is an insert to a variable element, we don't know what it is.
73 if (!isa<ConstantInt>(III->getOperand(2)))
75 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
77 // If this is an insert to the element we are looking for, return the
80 return III->getOperand(1);
82 // Otherwise, the insertelement doesn't modify the value, recurse on its
84 return FindScalarElement(III->getOperand(0), EltNo);
87 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
88 unsigned LHSWidth = SVI->getOperand(0)->getType()->getVectorNumElements();
89 int InEl = SVI->getMaskValue(EltNo);
91 return UndefValue::get(VTy->getElementType());
92 if (InEl < (int)LHSWidth)
93 return FindScalarElement(SVI->getOperand(0), InEl);
94 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth);
97 // Extract a value from a vector add operation with a constant zero.
98 Value *Val = 0; Constant *Con = 0;
99 if (match(V, m_Add(m_Value(Val), m_Constant(Con)))) {
100 if (Con->getAggregateElement(EltNo)->isNullValue())
101 return FindScalarElement(Val, EltNo);
104 // Otherwise, we don't know.
108 // If we have a PHI node with a vector type that has only 2 uses: feed
109 // itself and be an operand of extractelement at a constant location,
110 // try to replace the PHI of the vector type with a PHI of a scalar type.
111 Instruction *InstCombiner::scalarizePHI(ExtractElementInst &EI, PHINode *PN) {
112 // Verify that the PHI node has exactly 2 uses. Otherwise return NULL.
113 if (!PN->hasNUses(2))
116 // If so, it's known at this point that one operand is PHI and the other is
117 // an extractelement node. Find the PHI user that is not the extractelement
119 Value::use_iterator iu = PN->use_begin();
120 Instruction *PHIUser = dyn_cast<Instruction>(*iu);
121 if (PHIUser == cast<Instruction>(&EI))
122 PHIUser = cast<Instruction>(*(++iu));
124 // Verify that this PHI user has one use, which is the PHI itself,
125 // and that it is a binary operation which is cheap to scalarize.
126 // otherwise return NULL.
127 if (!PHIUser->hasOneUse() || !(PHIUser->use_back() == PN) ||
128 !(isa<BinaryOperator>(PHIUser)) || !CheapToScalarize(PHIUser, true))
131 // Create a scalar PHI node that will replace the vector PHI node
132 // just before the current PHI node.
133 PHINode *scalarPHI = cast<PHINode>(InsertNewInstWith(
134 PHINode::Create(EI.getType(), PN->getNumIncomingValues(), ""), *PN));
135 // Scalarize each PHI operand.
136 for (unsigned i = 0; i < PN->getNumIncomingValues(); i++) {
137 Value *PHIInVal = PN->getIncomingValue(i);
138 BasicBlock *inBB = PN->getIncomingBlock(i);
139 Value *Elt = EI.getIndexOperand();
140 // If the operand is the PHI induction variable:
141 if (PHIInVal == PHIUser) {
142 // Scalarize the binary operation. Its first operand is the
143 // scalar PHI and the second operand is extracted from the other
145 BinaryOperator *B0 = cast<BinaryOperator>(PHIUser);
146 unsigned opId = (B0->getOperand(0) == PN) ? 1 : 0;
147 Value *Op = InsertNewInstWith(
148 ExtractElementInst::Create(B0->getOperand(opId), Elt,
149 B0->getOperand(opId)->getName() + ".Elt"),
151 Value *newPHIUser = InsertNewInstWith(
152 BinaryOperator::Create(B0->getOpcode(), scalarPHI, Op), *B0);
153 scalarPHI->addIncoming(newPHIUser, inBB);
155 // Scalarize PHI input:
156 Instruction *newEI = ExtractElementInst::Create(PHIInVal, Elt, "");
157 // Insert the new instruction into the predecessor basic block.
158 Instruction *pos = dyn_cast<Instruction>(PHIInVal);
159 BasicBlock::iterator InsertPos;
160 if (pos && !isa<PHINode>(pos)) {
164 InsertPos = inBB->getFirstInsertionPt();
167 InsertNewInstWith(newEI, *InsertPos);
169 scalarPHI->addIncoming(newEI, inBB);
172 return ReplaceInstUsesWith(EI, scalarPHI);
175 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
176 // If vector val is constant with all elements the same, replace EI with
177 // that element. We handle a known element # below.
178 if (Constant *C = dyn_cast<Constant>(EI.getOperand(0)))
179 if (CheapToScalarize(C, false))
180 return ReplaceInstUsesWith(EI, C->getAggregateElement(0U));
182 // If extracting a specified index from the vector, see if we can recursively
183 // find a previously computed scalar that was inserted into the vector.
184 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
185 unsigned IndexVal = IdxC->getZExtValue();
186 unsigned VectorWidth = EI.getVectorOperandType()->getNumElements();
188 // If this is extracting an invalid index, turn this into undef, to avoid
189 // crashing the code below.
190 if (IndexVal >= VectorWidth)
191 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
193 // This instruction only demands the single element from the input vector.
194 // If the input vector has a single use, simplify it based on this use
196 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
197 APInt UndefElts(VectorWidth, 0);
198 APInt DemandedMask(VectorWidth, 0);
199 DemandedMask.setBit(IndexVal);
200 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
201 DemandedMask, UndefElts)) {
207 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
208 return ReplaceInstUsesWith(EI, Elt);
210 // If the this extractelement is directly using a bitcast from a vector of
211 // the same number of elements, see if we can find the source element from
212 // it. In this case, we will end up needing to bitcast the scalars.
213 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
214 if (VectorType *VT = dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
215 if (VT->getNumElements() == VectorWidth)
216 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
217 return new BitCastInst(Elt, EI.getType());
220 // If there's a vector PHI feeding a scalar use through this extractelement
221 // instruction, try to scalarize the PHI.
222 if (PHINode *PN = dyn_cast<PHINode>(EI.getOperand(0))) {
223 Instruction *scalarPHI = scalarizePHI(EI, PN);
229 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
230 // Push extractelement into predecessor operation if legal and
231 // profitable to do so
232 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
233 if (I->hasOneUse() &&
234 CheapToScalarize(BO, isa<ConstantInt>(EI.getOperand(1)))) {
236 Builder->CreateExtractElement(BO->getOperand(0), EI.getOperand(1),
237 EI.getName()+".lhs");
239 Builder->CreateExtractElement(BO->getOperand(1), EI.getOperand(1),
240 EI.getName()+".rhs");
241 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
243 } else if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
244 // Extracting the inserted element?
245 if (IE->getOperand(2) == EI.getOperand(1))
246 return ReplaceInstUsesWith(EI, IE->getOperand(1));
247 // If the inserted and extracted elements are constants, they must not
248 // be the same value, extract from the pre-inserted value instead.
249 if (isa<Constant>(IE->getOperand(2)) && isa<Constant>(EI.getOperand(1))) {
250 Worklist.AddValue(EI.getOperand(0));
251 EI.setOperand(0, IE->getOperand(0));
254 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
255 // If this is extracting an element from a shufflevector, figure out where
256 // it came from and extract from the appropriate input element instead.
257 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
258 int SrcIdx = SVI->getMaskValue(Elt->getZExtValue());
261 SVI->getOperand(0)->getType()->getVectorNumElements();
264 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
265 if (SrcIdx < (int)LHSWidth)
266 Src = SVI->getOperand(0);
269 Src = SVI->getOperand(1);
271 Type *Int32Ty = Type::getInt32Ty(EI.getContext());
272 return ExtractElementInst::Create(Src,
273 ConstantInt::get(Int32Ty,
276 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
277 // Canonicalize extractelement(cast) -> cast(extractelement)
278 // bitcasts can change the number of vector elements and they cost nothing
279 if (CI->hasOneUse() && (CI->getOpcode() != Instruction::BitCast)) {
280 Value *EE = Builder->CreateExtractElement(CI->getOperand(0),
281 EI.getIndexOperand());
282 Worklist.AddValue(EE);
283 return CastInst::Create(CI->getOpcode(), EE, EI.getType());
285 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
286 if (SI->hasOneUse()) {
287 // TODO: For a select on vectors, it might be useful to do this if it
288 // has multiple extractelement uses. For vector select, that seems to
289 // fight the vectorizer.
291 // If we are extracting an element from a vector select or a select on
292 // vectors, a select on the scalars extracted from the vector arguments.
293 Value *TrueVal = SI->getTrueValue();
294 Value *FalseVal = SI->getFalseValue();
296 Value *Cond = SI->getCondition();
297 if (Cond->getType()->isVectorTy()) {
298 Cond = Builder->CreateExtractElement(Cond,
299 EI.getIndexOperand(),
300 Cond->getName() + ".elt");
304 = Builder->CreateExtractElement(TrueVal,
305 EI.getIndexOperand(),
306 TrueVal->getName() + ".elt");
309 = Builder->CreateExtractElement(FalseVal,
310 EI.getIndexOperand(),
311 FalseVal->getName() + ".elt");
312 return SelectInst::Create(Cond,
315 SI->getName() + ".elt");
322 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
323 /// elements from either LHS or RHS, return the shuffle mask and true.
324 /// Otherwise, return false.
325 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
326 SmallVectorImpl<Constant*> &Mask) {
327 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
328 "Invalid CollectSingleShuffleElements");
329 unsigned NumElts = V->getType()->getVectorNumElements();
331 if (isa<UndefValue>(V)) {
332 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
337 for (unsigned i = 0; i != NumElts; ++i)
338 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
343 for (unsigned i = 0; i != NumElts; ++i)
344 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()),
349 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
350 // If this is an insert of an extract from some other vector, include it.
351 Value *VecOp = IEI->getOperand(0);
352 Value *ScalarOp = IEI->getOperand(1);
353 Value *IdxOp = IEI->getOperand(2);
355 if (!isa<ConstantInt>(IdxOp))
357 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
359 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
360 // Okay, we can handle this if the vector we are insertinting into is
362 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
363 // If so, update the mask to reflect the inserted undef.
364 Mask[InsertedIdx] = UndefValue::get(Type::getInt32Ty(V->getContext()));
367 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
368 if (isa<ConstantInt>(EI->getOperand(1)) &&
369 EI->getOperand(0)->getType() == V->getType()) {
370 unsigned ExtractedIdx =
371 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
373 // This must be extracting from either LHS or RHS.
374 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
375 // Okay, we can handle this if the vector we are insertinting into is
377 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
378 // If so, update the mask to reflect the inserted value.
379 if (EI->getOperand(0) == LHS) {
380 Mask[InsertedIdx % NumElts] =
381 ConstantInt::get(Type::getInt32Ty(V->getContext()),
384 assert(EI->getOperand(0) == RHS);
385 Mask[InsertedIdx % NumElts] =
386 ConstantInt::get(Type::getInt32Ty(V->getContext()),
387 ExtractedIdx+NumElts);
395 // TODO: Handle shufflevector here!
400 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
401 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
402 /// that computes V and the LHS value of the shuffle.
403 static Value *CollectShuffleElements(Value *V, SmallVectorImpl<Constant*> &Mask,
405 assert(V->getType()->isVectorTy() &&
406 (RHS == 0 || V->getType() == RHS->getType()) &&
408 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
410 if (isa<UndefValue>(V)) {
411 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
415 if (isa<ConstantAggregateZero>(V)) {
416 Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(V->getContext()),0));
420 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
421 // If this is an insert of an extract from some other vector, include it.
422 Value *VecOp = IEI->getOperand(0);
423 Value *ScalarOp = IEI->getOperand(1);
424 Value *IdxOp = IEI->getOperand(2);
426 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
427 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
428 EI->getOperand(0)->getType() == V->getType()) {
429 unsigned ExtractedIdx =
430 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
431 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
433 // Either the extracted from or inserted into vector must be RHSVec,
434 // otherwise we'd end up with a shuffle of three inputs.
435 if (EI->getOperand(0) == RHS || RHS == 0) {
436 RHS = EI->getOperand(0);
437 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
438 Mask[InsertedIdx % NumElts] =
439 ConstantInt::get(Type::getInt32Ty(V->getContext()),
440 NumElts+ExtractedIdx);
445 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
446 // Update Mask to reflect that `ScalarOp' has been inserted at
447 // position `InsertedIdx' within the vector returned by IEI.
448 Mask[InsertedIdx % NumElts] = Mask[ExtractedIdx];
450 // Everything but the extracted element is replaced with the RHS.
451 for (unsigned i = 0; i != NumElts; ++i) {
452 if (i != InsertedIdx)
453 Mask[i] = ConstantInt::get(Type::getInt32Ty(V->getContext()),
459 // If this insertelement is a chain that comes from exactly these two
460 // vectors, return the vector and the effective shuffle.
461 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
462 return EI->getOperand(0);
466 // TODO: Handle shufflevector here!
468 // Otherwise, can't do anything fancy. Return an identity vector.
469 for (unsigned i = 0; i != NumElts; ++i)
470 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
474 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
475 Value *VecOp = IE.getOperand(0);
476 Value *ScalarOp = IE.getOperand(1);
477 Value *IdxOp = IE.getOperand(2);
479 // Inserting an undef or into an undefined place, remove this.
480 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
481 ReplaceInstUsesWith(IE, VecOp);
483 // If the inserted element was extracted from some other vector, and if the
484 // indexes are constant, try to turn this into a shufflevector operation.
485 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
486 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
487 EI->getOperand(0)->getType() == IE.getType()) {
488 unsigned NumVectorElts = IE.getType()->getNumElements();
489 unsigned ExtractedIdx =
490 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
491 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
493 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
494 return ReplaceInstUsesWith(IE, VecOp);
496 if (InsertedIdx >= NumVectorElts) // Out of range insert.
497 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
499 // If we are extracting a value from a vector, then inserting it right
500 // back into the same place, just use the input vector.
501 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
502 return ReplaceInstUsesWith(IE, VecOp);
504 // If this insertelement isn't used by some other insertelement, turn it
505 // (and any insertelements it points to), into one big shuffle.
506 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
507 SmallVector<Constant*, 16> Mask;
509 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
510 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
511 // We now have a shuffle of LHS, RHS, Mask.
512 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
517 unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements();
518 APInt UndefElts(VWidth, 0);
519 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
520 if (Value *V = SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts)) {
522 return ReplaceInstUsesWith(IE, V);
529 /// Return true if we can evaluate the specified expression tree if the vector
530 /// elements were shuffled in a different order.
531 static bool CanEvaluateShuffled(Value *V, ArrayRef<int> Mask,
532 unsigned Depth = 5) {
533 // We can always reorder the elements of a constant.
534 if (isa<Constant>(V))
537 // We won't reorder vector arguments. No IPO here.
538 Instruction *I = dyn_cast<Instruction>(V);
539 if (!I) return false;
541 // Two users may expect different orders of the elements. Don't try it.
545 if (Depth == 0) return false;
547 switch (I->getOpcode()) {
548 case Instruction::Add:
549 case Instruction::FAdd:
550 case Instruction::Sub:
551 case Instruction::FSub:
552 case Instruction::Mul:
553 case Instruction::FMul:
554 case Instruction::UDiv:
555 case Instruction::SDiv:
556 case Instruction::FDiv:
557 case Instruction::URem:
558 case Instruction::SRem:
559 case Instruction::FRem:
560 case Instruction::Shl:
561 case Instruction::LShr:
562 case Instruction::AShr:
563 case Instruction::And:
564 case Instruction::Or:
565 case Instruction::Xor:
566 case Instruction::ICmp:
567 case Instruction::FCmp:
568 case Instruction::Trunc:
569 case Instruction::ZExt:
570 case Instruction::SExt:
571 case Instruction::FPToUI:
572 case Instruction::FPToSI:
573 case Instruction::UIToFP:
574 case Instruction::SIToFP:
575 case Instruction::FPTrunc:
576 case Instruction::FPExt:
577 case Instruction::GetElementPtr: {
578 for (int i = 0, e = I->getNumOperands(); i != e; ++i) {
579 if (!CanEvaluateShuffled(I->getOperand(i), Mask, Depth-1))
584 case Instruction::InsertElement: {
585 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(2));
586 if (!CI) return false;
587 int ElementNumber = CI->getLimitedValue();
589 // Verify that 'CI' does not occur twice in Mask. A single 'insertelement'
590 // can't put an element into multiple indices.
591 bool SeenOnce = false;
592 for (int i = 0, e = Mask.size(); i != e; ++i) {
593 if (Mask[i] == ElementNumber) {
599 return CanEvaluateShuffled(I->getOperand(0), Mask, Depth-1);
605 /// Rebuild a new instruction just like 'I' but with the new operands given.
606 /// In the event of type mismatch, the type of the operands is correct.
607 static Value *BuildNew(Instruction *I, ArrayRef<Value*> NewOps) {
608 // We don't want to use the IRBuilder here because we want the replacement
609 // instructions to appear next to 'I', not the builder's insertion point.
610 switch (I->getOpcode()) {
611 case Instruction::Add:
612 case Instruction::FAdd:
613 case Instruction::Sub:
614 case Instruction::FSub:
615 case Instruction::Mul:
616 case Instruction::FMul:
617 case Instruction::UDiv:
618 case Instruction::SDiv:
619 case Instruction::FDiv:
620 case Instruction::URem:
621 case Instruction::SRem:
622 case Instruction::FRem:
623 case Instruction::Shl:
624 case Instruction::LShr:
625 case Instruction::AShr:
626 case Instruction::And:
627 case Instruction::Or:
628 case Instruction::Xor: {
629 BinaryOperator *BO = cast<BinaryOperator>(I);
630 assert(NewOps.size() == 2 && "binary operator with #ops != 2");
631 BinaryOperator *New =
632 BinaryOperator::Create(cast<BinaryOperator>(I)->getOpcode(),
633 NewOps[0], NewOps[1], "", BO);
634 if (isa<OverflowingBinaryOperator>(BO)) {
635 New->setHasNoUnsignedWrap(BO->hasNoUnsignedWrap());
636 New->setHasNoSignedWrap(BO->hasNoSignedWrap());
638 if (isa<PossiblyExactOperator>(BO)) {
639 New->setIsExact(BO->isExact());
643 case Instruction::ICmp:
644 assert(NewOps.size() == 2 && "icmp with #ops != 2");
645 return new ICmpInst(I, cast<ICmpInst>(I)->getPredicate(),
646 NewOps[0], NewOps[1]);
647 case Instruction::FCmp:
648 assert(NewOps.size() == 2 && "fcmp with #ops != 2");
649 return new FCmpInst(I, cast<FCmpInst>(I)->getPredicate(),
650 NewOps[0], NewOps[1]);
651 case Instruction::Trunc:
652 case Instruction::ZExt:
653 case Instruction::SExt:
654 case Instruction::FPToUI:
655 case Instruction::FPToSI:
656 case Instruction::UIToFP:
657 case Instruction::SIToFP:
658 case Instruction::FPTrunc:
659 case Instruction::FPExt: {
660 // It's possible that the mask has a different number of elements from
661 // the original cast. We recompute the destination type to match the mask.
663 VectorType::get(I->getType()->getScalarType(),
664 NewOps[0]->getType()->getVectorNumElements());
665 assert(NewOps.size() == 1 && "cast with #ops != 1");
666 return CastInst::Create(cast<CastInst>(I)->getOpcode(), NewOps[0], DestTy,
669 case Instruction::GetElementPtr: {
670 Value *Ptr = NewOps[0];
671 ArrayRef<Value*> Idx = NewOps.slice(1);
672 GetElementPtrInst *GEP = GetElementPtrInst::Create(Ptr, Idx, "", I);
673 GEP->setIsInBounds(cast<GetElementPtrInst>(I)->isInBounds());
677 llvm_unreachable("failed to rebuild vector instructions");
681 InstCombiner::EvaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask) {
682 // Mask.size() does not need to be equal to the number of vector elements.
684 assert(V->getType()->isVectorTy() && "can't reorder non-vector elements");
685 if (isa<UndefValue>(V)) {
686 return UndefValue::get(VectorType::get(V->getType()->getScalarType(),
689 if (isa<ConstantAggregateZero>(V)) {
690 return ConstantAggregateZero::get(
691 VectorType::get(V->getType()->getScalarType(),
694 if (Constant *C = dyn_cast<Constant>(V)) {
695 SmallVector<Constant *, 16> MaskValues;
696 for (int i = 0, e = Mask.size(); i != e; ++i) {
698 MaskValues.push_back(UndefValue::get(Builder->getInt32Ty()));
700 MaskValues.push_back(Builder->getInt32(Mask[i]));
702 return ConstantExpr::getShuffleVector(C, UndefValue::get(C->getType()),
703 ConstantVector::get(MaskValues));
706 Instruction *I = cast<Instruction>(V);
707 switch (I->getOpcode()) {
708 case Instruction::Add:
709 case Instruction::FAdd:
710 case Instruction::Sub:
711 case Instruction::FSub:
712 case Instruction::Mul:
713 case Instruction::FMul:
714 case Instruction::UDiv:
715 case Instruction::SDiv:
716 case Instruction::FDiv:
717 case Instruction::URem:
718 case Instruction::SRem:
719 case Instruction::FRem:
720 case Instruction::Shl:
721 case Instruction::LShr:
722 case Instruction::AShr:
723 case Instruction::And:
724 case Instruction::Or:
725 case Instruction::Xor:
726 case Instruction::ICmp:
727 case Instruction::FCmp:
728 case Instruction::Trunc:
729 case Instruction::ZExt:
730 case Instruction::SExt:
731 case Instruction::FPToUI:
732 case Instruction::FPToSI:
733 case Instruction::UIToFP:
734 case Instruction::SIToFP:
735 case Instruction::FPTrunc:
736 case Instruction::FPExt:
737 case Instruction::Select:
738 case Instruction::GetElementPtr: {
739 SmallVector<Value*, 8> NewOps;
740 bool NeedsRebuild = (Mask.size() != I->getType()->getVectorNumElements());
741 for (int i = 0, e = I->getNumOperands(); i != e; ++i) {
742 Value *V = EvaluateInDifferentElementOrder(I->getOperand(i), Mask);
744 NeedsRebuild |= (V != I->getOperand(i));
747 return BuildNew(I, NewOps);
751 case Instruction::InsertElement: {
752 int Element = cast<ConstantInt>(I->getOperand(2))->getLimitedValue();
754 // The insertelement was inserting at Element. Figure out which element
755 // that becomes after shuffling. The answer is guaranteed to be unique
756 // by CanEvaluateShuffled.
759 for (int e = Mask.size(); Index != e; ++Index) {
760 if (Mask[Index] == Element) {
766 // If element is not in Mask, no need to handle the operand 1 (element to
767 // be inserted). Just evaluate values in operand 0 according to Mask.
769 return EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
771 Value *V = EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
772 return InsertElementInst::Create(V, I->getOperand(1),
773 Builder->getInt32(Index), "", I);
776 llvm_unreachable("failed to reorder elements of vector instruction!");
779 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
780 Value *LHS = SVI.getOperand(0);
781 Value *RHS = SVI.getOperand(1);
782 SmallVector<int, 16> Mask = SVI.getShuffleMask();
784 bool MadeChange = false;
786 // Undefined shuffle mask -> undefined value.
787 if (isa<UndefValue>(SVI.getOperand(2)))
788 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
790 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
792 APInt UndefElts(VWidth, 0);
793 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
794 if (Value *V = SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
796 return ReplaceInstUsesWith(SVI, V);
797 LHS = SVI.getOperand(0);
798 RHS = SVI.getOperand(1);
802 unsigned LHSWidth = cast<VectorType>(LHS->getType())->getNumElements();
804 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
805 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
806 if (LHS == RHS || isa<UndefValue>(LHS)) {
807 if (isa<UndefValue>(LHS) && LHS == RHS) {
808 // shuffle(undef,undef,mask) -> undef.
809 Value *Result = (VWidth == LHSWidth)
810 ? LHS : UndefValue::get(SVI.getType());
811 return ReplaceInstUsesWith(SVI, Result);
814 // Remap any references to RHS to use LHS.
815 SmallVector<Constant*, 16> Elts;
816 for (unsigned i = 0, e = LHSWidth; i != VWidth; ++i) {
818 Elts.push_back(UndefValue::get(Type::getInt32Ty(SVI.getContext())));
822 if ((Mask[i] >= (int)e && isa<UndefValue>(RHS)) ||
823 (Mask[i] < (int)e && isa<UndefValue>(LHS))) {
824 Mask[i] = -1; // Turn into undef.
825 Elts.push_back(UndefValue::get(Type::getInt32Ty(SVI.getContext())));
827 Mask[i] = Mask[i] % e; // Force to LHS.
828 Elts.push_back(ConstantInt::get(Type::getInt32Ty(SVI.getContext()),
832 SVI.setOperand(0, SVI.getOperand(1));
833 SVI.setOperand(1, UndefValue::get(RHS->getType()));
834 SVI.setOperand(2, ConstantVector::get(Elts));
835 LHS = SVI.getOperand(0);
836 RHS = SVI.getOperand(1);
840 if (VWidth == LHSWidth) {
841 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
842 bool isLHSID = true, isRHSID = true;
844 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
845 if (Mask[i] < 0) continue; // Ignore undef values.
846 // Is this an identity shuffle of the LHS value?
847 isLHSID &= (Mask[i] == (int)i);
849 // Is this an identity shuffle of the RHS value?
850 isRHSID &= (Mask[i]-e == i);
853 // Eliminate identity shuffles.
854 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
855 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
858 if (isa<UndefValue>(RHS) && CanEvaluateShuffled(LHS, Mask)) {
859 Value *V = EvaluateInDifferentElementOrder(LHS, Mask);
860 return ReplaceInstUsesWith(SVI, V);
863 // If the LHS is a shufflevector itself, see if we can combine it with this
864 // one without producing an unusual shuffle.
865 // Cases that might be simplified:
867 // x1=shuffle(v1,v2,mask1)
868 // x=shuffle(x1,undef,mask)
870 // x=shuffle(v1,undef,newMask)
871 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : -1
873 // x1=shuffle(v1,undef,mask1)
874 // x=shuffle(x1,x2,mask)
875 // where v1.size() == mask1.size()
877 // x=shuffle(v1,x2,newMask)
878 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : mask[i]
880 // x2=shuffle(v2,undef,mask2)
881 // x=shuffle(x1,x2,mask)
882 // where v2.size() == mask2.size()
884 // x=shuffle(x1,v2,newMask)
885 // newMask[i] = (mask[i] < x1.size())
886 // ? mask[i] : mask2[mask[i]-x1.size()]+x1.size()
888 // x1=shuffle(v1,undef,mask1)
889 // x2=shuffle(v2,undef,mask2)
890 // x=shuffle(x1,x2,mask)
891 // where v1.size() == v2.size()
893 // x=shuffle(v1,v2,newMask)
894 // newMask[i] = (mask[i] < x1.size())
895 // ? mask1[mask[i]] : mask2[mask[i]-x1.size()]+v1.size()
897 // Here we are really conservative:
898 // we are absolutely afraid of producing a shuffle mask not in the input
899 // program, because the code gen may not be smart enough to turn a merged
900 // shuffle into two specific shuffles: it may produce worse code. As such,
901 // we only merge two shuffles if the result is either a splat or one of the
902 // input shuffle masks. In this case, merging the shuffles just removes
903 // one instruction, which we know is safe. This is good for things like
904 // turning: (splat(splat)) -> splat, or
905 // merge(V[0..n], V[n+1..2n]) -> V[0..2n]
906 ShuffleVectorInst* LHSShuffle = dyn_cast<ShuffleVectorInst>(LHS);
907 ShuffleVectorInst* RHSShuffle = dyn_cast<ShuffleVectorInst>(RHS);
909 if (!isa<UndefValue>(LHSShuffle->getOperand(1)) && !isa<UndefValue>(RHS))
912 if (!isa<UndefValue>(RHSShuffle->getOperand(1)))
914 if (!LHSShuffle && !RHSShuffle)
915 return MadeChange ? &SVI : 0;
917 Value* LHSOp0 = NULL;
918 Value* LHSOp1 = NULL;
919 Value* RHSOp0 = NULL;
920 unsigned LHSOp0Width = 0;
921 unsigned RHSOp0Width = 0;
923 LHSOp0 = LHSShuffle->getOperand(0);
924 LHSOp1 = LHSShuffle->getOperand(1);
925 LHSOp0Width = cast<VectorType>(LHSOp0->getType())->getNumElements();
928 RHSOp0 = RHSShuffle->getOperand(0);
929 RHSOp0Width = cast<VectorType>(RHSOp0->getType())->getNumElements();
935 if (isa<UndefValue>(RHS)) {
940 else if (LHSOp0Width == LHSWidth) {
945 if (RHSShuffle && RHSOp0Width == LHSWidth) {
949 if (LHSOp0 == RHSOp0) {
954 if (newLHS == LHS && newRHS == RHS)
955 return MadeChange ? &SVI : 0;
957 SmallVector<int, 16> LHSMask;
958 SmallVector<int, 16> RHSMask;
960 LHSMask = LHSShuffle->getShuffleMask();
961 if (RHSShuffle && newRHS != RHS)
962 RHSMask = RHSShuffle->getShuffleMask();
964 unsigned newLHSWidth = (newLHS != LHS) ? LHSOp0Width : LHSWidth;
965 SmallVector<int, 16> newMask;
968 // Create a new mask for the new ShuffleVectorInst so that the new
969 // ShuffleVectorInst is equivalent to the original one.
970 for (unsigned i = 0; i < VWidth; ++i) {
973 // This element is an undef value.
975 } else if (Mask[i] < (int)LHSWidth) {
976 // This element is from left hand side vector operand.
978 // If LHS is going to be replaced (case 1, 2, or 4), calculate the
979 // new mask value for the element.
981 eltMask = LHSMask[Mask[i]];
982 // If the value selected is an undef value, explicitly specify it
983 // with a -1 mask value.
984 if (eltMask >= (int)LHSOp0Width && isa<UndefValue>(LHSOp1))
989 // This element is from right hand side vector operand
991 // If the value selected is an undef value, explicitly specify it
992 // with a -1 mask value. (case 1)
993 if (isa<UndefValue>(RHS))
995 // If RHS is going to be replaced (case 3 or 4), calculate the
996 // new mask value for the element.
997 else if (newRHS != RHS) {
998 eltMask = RHSMask[Mask[i]-LHSWidth];
999 // If the value selected is an undef value, explicitly specify it
1000 // with a -1 mask value.
1001 if (eltMask >= (int)RHSOp0Width) {
1002 assert(isa<UndefValue>(RHSShuffle->getOperand(1))
1003 && "should have been check above");
1007 eltMask = Mask[i]-LHSWidth;
1009 // If LHS's width is changed, shift the mask value accordingly.
1010 // If newRHS == NULL, i.e. LHSOp0 == RHSOp0, we want to remap any
1011 // references from RHSOp0 to LHSOp0, so we don't need to shift the mask.
1012 // If newRHS == newLHS, we want to remap any references from newRHS to
1013 // newLHS so that we can properly identify splats that may occur due to
1014 // obfuscation accross the two vectors.
1015 if (eltMask >= 0 && newRHS != NULL && newLHS != newRHS)
1016 eltMask += newLHSWidth;
1019 // Check if this could still be a splat.
1021 if (SplatElt >= 0 && SplatElt != eltMask)
1026 newMask.push_back(eltMask);
1029 // If the result mask is equal to one of the original shuffle masks,
1030 // or is a splat, do the replacement.
1031 if (isSplat || newMask == LHSMask || newMask == RHSMask || newMask == Mask) {
1032 SmallVector<Constant*, 16> Elts;
1033 Type *Int32Ty = Type::getInt32Ty(SVI.getContext());
1034 for (unsigned i = 0, e = newMask.size(); i != e; ++i) {
1035 if (newMask[i] < 0) {
1036 Elts.push_back(UndefValue::get(Int32Ty));
1038 Elts.push_back(ConstantInt::get(Int32Ty, newMask[i]));
1042 newRHS = UndefValue::get(newLHS->getType());
1043 return new ShuffleVectorInst(newLHS, newRHS, ConstantVector::get(Elts));
1046 return MadeChange ? &SVI : 0;