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 "InstCombineInternal.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/Analysis/InstructionSimplify.h"
18 #include "llvm/Analysis/VectorUtils.h"
19 #include "llvm/IR/PatternMatch.h"
21 using namespace PatternMatch;
23 #define DEBUG_TYPE "instcombine"
25 /// Return true if the value is cheaper to scalarize than it is to leave as a
26 /// vector operation. isConstant indicates whether we're extracting one known
27 /// element. If false we're extracting a variable index.
28 static bool cheapToScalarize(Value *V, bool isConstant) {
29 if (Constant *C = dyn_cast<Constant>(V)) {
30 if (isConstant) return true;
32 // If all elts are the same, we can extract it and use any of the values.
33 if (Constant *Op0 = C->getAggregateElement(0U)) {
34 for (unsigned i = 1, e = V->getType()->getVectorNumElements(); i != e;
36 if (C->getAggregateElement(i) != Op0)
41 Instruction *I = dyn_cast<Instruction>(V);
44 // Insert element gets simplified to the inserted element or is deleted if
45 // this is constant idx extract element and its a constant idx insertelt.
46 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
47 isa<ConstantInt>(I->getOperand(2)))
49 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
51 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
52 if (BO->hasOneUse() &&
53 (cheapToScalarize(BO->getOperand(0), isConstant) ||
54 cheapToScalarize(BO->getOperand(1), isConstant)))
56 if (CmpInst *CI = dyn_cast<CmpInst>(I))
57 if (CI->hasOneUse() &&
58 (cheapToScalarize(CI->getOperand(0), isConstant) ||
59 cheapToScalarize(CI->getOperand(1), isConstant)))
65 // If we have a PHI node with a vector type that has only 2 uses: feed
66 // itself and be an operand of extractelement at a constant location,
67 // try to replace the PHI of the vector type with a PHI of a scalar type.
68 Instruction *InstCombiner::scalarizePHI(ExtractElementInst &EI, PHINode *PN) {
69 // Verify that the PHI node has exactly 2 uses. Otherwise return NULL.
73 // If so, it's known at this point that one operand is PHI and the other is
74 // an extractelement node. Find the PHI user that is not the extractelement
76 auto iu = PN->user_begin();
77 Instruction *PHIUser = dyn_cast<Instruction>(*iu);
78 if (PHIUser == cast<Instruction>(&EI))
79 PHIUser = cast<Instruction>(*(++iu));
81 // Verify that this PHI user has one use, which is the PHI itself,
82 // and that it is a binary operation which is cheap to scalarize.
83 // otherwise return NULL.
84 if (!PHIUser->hasOneUse() || !(PHIUser->user_back() == PN) ||
85 !(isa<BinaryOperator>(PHIUser)) || !cheapToScalarize(PHIUser, true))
88 // Create a scalar PHI node that will replace the vector PHI node
89 // just before the current PHI node.
90 PHINode *scalarPHI = cast<PHINode>(InsertNewInstWith(
91 PHINode::Create(EI.getType(), PN->getNumIncomingValues(), ""), *PN));
92 // Scalarize each PHI operand.
93 for (unsigned i = 0; i < PN->getNumIncomingValues(); i++) {
94 Value *PHIInVal = PN->getIncomingValue(i);
95 BasicBlock *inBB = PN->getIncomingBlock(i);
96 Value *Elt = EI.getIndexOperand();
97 // If the operand is the PHI induction variable:
98 if (PHIInVal == PHIUser) {
99 // Scalarize the binary operation. Its first operand is the
100 // scalar PHI, and the second operand is extracted from the other
102 BinaryOperator *B0 = cast<BinaryOperator>(PHIUser);
103 unsigned opId = (B0->getOperand(0) == PN) ? 1 : 0;
104 Value *Op = InsertNewInstWith(
105 ExtractElementInst::Create(B0->getOperand(opId), Elt,
106 B0->getOperand(opId)->getName() + ".Elt"),
108 Value *newPHIUser = InsertNewInstWith(
109 BinaryOperator::Create(B0->getOpcode(), scalarPHI, Op), *B0);
110 scalarPHI->addIncoming(newPHIUser, inBB);
112 // Scalarize PHI input:
113 Instruction *newEI = ExtractElementInst::Create(PHIInVal, Elt, "");
114 // Insert the new instruction into the predecessor basic block.
115 Instruction *pos = dyn_cast<Instruction>(PHIInVal);
116 BasicBlock::iterator InsertPos;
117 if (pos && !isa<PHINode>(pos)) {
118 InsertPos = ++pos->getIterator();
120 InsertPos = inBB->getFirstInsertionPt();
123 InsertNewInstWith(newEI, *InsertPos);
125 scalarPHI->addIncoming(newEI, inBB);
128 return ReplaceInstUsesWith(EI, scalarPHI);
131 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
132 if (Value *V = SimplifyExtractElementInst(
133 EI.getVectorOperand(), EI.getIndexOperand(), DL, TLI, DT, AC))
134 return ReplaceInstUsesWith(EI, V);
136 // If vector val is constant with all elements the same, replace EI with
137 // that element. We handle a known element # below.
138 if (Constant *C = dyn_cast<Constant>(EI.getOperand(0)))
139 if (cheapToScalarize(C, false))
140 return ReplaceInstUsesWith(EI, C->getAggregateElement(0U));
142 // If extracting a specified index from the vector, see if we can recursively
143 // find a previously computed scalar that was inserted into the vector.
144 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
145 unsigned IndexVal = IdxC->getZExtValue();
146 unsigned VectorWidth = EI.getVectorOperandType()->getNumElements();
148 // InstSimplify handles cases where the index is invalid.
149 assert(IndexVal < VectorWidth);
151 // This instruction only demands the single element from the input vector.
152 // If the input vector has a single use, simplify it based on this use
154 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
155 APInt UndefElts(VectorWidth, 0);
156 APInt DemandedMask(VectorWidth, 0);
157 DemandedMask.setBit(IndexVal);
158 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0), DemandedMask,
165 // If this extractelement is directly using a bitcast from a vector of
166 // the same number of elements, see if we can find the source element from
167 // it. In this case, we will end up needing to bitcast the scalars.
168 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
169 if (VectorType *VT = dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
170 if (VT->getNumElements() == VectorWidth)
171 if (Value *Elt = findScalarElement(BCI->getOperand(0), IndexVal))
172 return new BitCastInst(Elt, EI.getType());
175 // If there's a vector PHI feeding a scalar use through this extractelement
176 // instruction, try to scalarize the PHI.
177 if (PHINode *PN = dyn_cast<PHINode>(EI.getOperand(0))) {
178 Instruction *scalarPHI = scalarizePHI(EI, PN);
184 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
185 // Push extractelement into predecessor operation if legal and
186 // profitable to do so.
187 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
188 if (I->hasOneUse() &&
189 cheapToScalarize(BO, isa<ConstantInt>(EI.getOperand(1)))) {
191 Builder->CreateExtractElement(BO->getOperand(0), EI.getOperand(1),
192 EI.getName()+".lhs");
194 Builder->CreateExtractElement(BO->getOperand(1), EI.getOperand(1),
195 EI.getName()+".rhs");
196 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
198 } else if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
199 // Extracting the inserted element?
200 if (IE->getOperand(2) == EI.getOperand(1))
201 return ReplaceInstUsesWith(EI, IE->getOperand(1));
202 // If the inserted and extracted elements are constants, they must not
203 // be the same value, extract from the pre-inserted value instead.
204 if (isa<Constant>(IE->getOperand(2)) && isa<Constant>(EI.getOperand(1))) {
205 Worklist.AddValue(EI.getOperand(0));
206 EI.setOperand(0, IE->getOperand(0));
209 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
210 // If this is extracting an element from a shufflevector, figure out where
211 // it came from and extract from the appropriate input element instead.
212 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
213 int SrcIdx = SVI->getMaskValue(Elt->getZExtValue());
216 SVI->getOperand(0)->getType()->getVectorNumElements();
219 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
220 if (SrcIdx < (int)LHSWidth)
221 Src = SVI->getOperand(0);
224 Src = SVI->getOperand(1);
226 Type *Int32Ty = Type::getInt32Ty(EI.getContext());
227 return ExtractElementInst::Create(Src,
228 ConstantInt::get(Int32Ty,
231 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
232 // Canonicalize extractelement(cast) -> cast(extractelement).
233 // Bitcasts can change the number of vector elements, and they cost
235 if (CI->hasOneUse() && (CI->getOpcode() != Instruction::BitCast)) {
236 Value *EE = Builder->CreateExtractElement(CI->getOperand(0),
237 EI.getIndexOperand());
238 Worklist.AddValue(EE);
239 return CastInst::Create(CI->getOpcode(), EE, EI.getType());
241 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
242 if (SI->hasOneUse()) {
243 // TODO: For a select on vectors, it might be useful to do this if it
244 // has multiple extractelement uses. For vector select, that seems to
245 // fight the vectorizer.
247 // If we are extracting an element from a vector select or a select on
248 // vectors, create a select on the scalars extracted from the vector
250 Value *TrueVal = SI->getTrueValue();
251 Value *FalseVal = SI->getFalseValue();
253 Value *Cond = SI->getCondition();
254 if (Cond->getType()->isVectorTy()) {
255 Cond = Builder->CreateExtractElement(Cond,
256 EI.getIndexOperand(),
257 Cond->getName() + ".elt");
261 = Builder->CreateExtractElement(TrueVal,
262 EI.getIndexOperand(),
263 TrueVal->getName() + ".elt");
266 = Builder->CreateExtractElement(FalseVal,
267 EI.getIndexOperand(),
268 FalseVal->getName() + ".elt");
269 return SelectInst::Create(Cond,
272 SI->getName() + ".elt");
279 /// If V is a shuffle of values that ONLY returns elements from either LHS or
280 /// RHS, return the shuffle mask and true. Otherwise, return false.
281 static bool collectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
282 SmallVectorImpl<Constant*> &Mask) {
283 assert(LHS->getType() == RHS->getType() &&
284 "Invalid CollectSingleShuffleElements");
285 unsigned NumElts = V->getType()->getVectorNumElements();
287 if (isa<UndefValue>(V)) {
288 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
293 for (unsigned i = 0; i != NumElts; ++i)
294 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
299 for (unsigned i = 0; i != NumElts; ++i)
300 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()),
305 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
306 // If this is an insert of an extract from some other vector, include it.
307 Value *VecOp = IEI->getOperand(0);
308 Value *ScalarOp = IEI->getOperand(1);
309 Value *IdxOp = IEI->getOperand(2);
311 if (!isa<ConstantInt>(IdxOp))
313 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
315 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
316 // We can handle this if the vector we are inserting into is
318 if (collectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
319 // If so, update the mask to reflect the inserted undef.
320 Mask[InsertedIdx] = UndefValue::get(Type::getInt32Ty(V->getContext()));
323 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
324 if (isa<ConstantInt>(EI->getOperand(1))) {
325 unsigned ExtractedIdx =
326 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
327 unsigned NumLHSElts = LHS->getType()->getVectorNumElements();
329 // This must be extracting from either LHS or RHS.
330 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
331 // We can handle this if the vector we are inserting into is
333 if (collectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
334 // If so, update the mask to reflect the inserted value.
335 if (EI->getOperand(0) == LHS) {
336 Mask[InsertedIdx % NumElts] =
337 ConstantInt::get(Type::getInt32Ty(V->getContext()),
340 assert(EI->getOperand(0) == RHS);
341 Mask[InsertedIdx % NumElts] =
342 ConstantInt::get(Type::getInt32Ty(V->getContext()),
343 ExtractedIdx + NumLHSElts);
355 /// If we have insertion into a vector that is wider than the vector that we
356 /// are extracting from, try to widen the source vector to allow a single
357 /// shufflevector to replace one or more insert/extract pairs.
358 static void replaceExtractElements(InsertElementInst *InsElt,
359 ExtractElementInst *ExtElt,
361 VectorType *InsVecType = InsElt->getType();
362 VectorType *ExtVecType = ExtElt->getVectorOperandType();
363 unsigned NumInsElts = InsVecType->getVectorNumElements();
364 unsigned NumExtElts = ExtVecType->getVectorNumElements();
366 // The inserted-to vector must be wider than the extracted-from vector.
367 if (InsVecType->getElementType() != ExtVecType->getElementType() ||
368 NumExtElts >= NumInsElts)
371 // Create a shuffle mask to widen the extended-from vector using undefined
372 // values. The mask selects all of the values of the original vector followed
373 // by as many undefined values as needed to create a vector of the same length
374 // as the inserted-to vector.
375 SmallVector<Constant *, 16> ExtendMask;
376 IntegerType *IntType = Type::getInt32Ty(InsElt->getContext());
377 for (unsigned i = 0; i < NumExtElts; ++i)
378 ExtendMask.push_back(ConstantInt::get(IntType, i));
379 for (unsigned i = NumExtElts; i < NumInsElts; ++i)
380 ExtendMask.push_back(UndefValue::get(IntType));
382 Value *ExtVecOp = ExtElt->getVectorOperand();
383 auto *WideVec = new ShuffleVectorInst(ExtVecOp, UndefValue::get(ExtVecType),
384 ConstantVector::get(ExtendMask));
386 // Insert the new shuffle after the vector operand of the extract is defined
387 // or at the start of the basic block, so any subsequent extracts can use it.
388 bool ReplaceAllExtUsers;
389 if (auto *ExtVecOpInst = dyn_cast<Instruction>(ExtVecOp)) {
390 WideVec->insertAfter(ExtVecOpInst);
391 ReplaceAllExtUsers = true;
393 // TODO: Insert at start of function, so it's always safe to replace all?
394 IC.InsertNewInstWith(WideVec, *ExtElt->getParent()->getFirstInsertionPt());
395 ReplaceAllExtUsers = false;
398 // Replace extracts from the original narrow vector with extracts from the new
400 for (User *U : ExtVecOp->users()) {
401 ExtractElementInst *OldExt = dyn_cast<ExtractElementInst>(U);
403 (!ReplaceAllExtUsers && OldExt->getParent() != WideVec->getParent()))
405 auto *NewExt = ExtractElementInst::Create(WideVec, OldExt->getOperand(1));
406 NewExt->insertAfter(WideVec);
407 IC.ReplaceInstUsesWith(*OldExt, NewExt);
411 /// We are building a shuffle to create V, which is a sequence of insertelement,
412 /// extractelement pairs. If PermittedRHS is set, then we must either use it or
413 /// not rely on the second vector source. Return a std::pair containing the
414 /// left and right vectors of the proposed shuffle (or 0), and set the Mask
415 /// parameter as required.
417 /// Note: we intentionally don't try to fold earlier shuffles since they have
418 /// often been chosen carefully to be efficiently implementable on the target.
419 typedef std::pair<Value *, Value *> ShuffleOps;
421 static ShuffleOps collectShuffleElements(Value *V,
422 SmallVectorImpl<Constant *> &Mask,
425 assert(V->getType()->isVectorTy() && "Invalid shuffle!");
426 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
428 if (isa<UndefValue>(V)) {
429 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
430 return std::make_pair(
431 PermittedRHS ? UndefValue::get(PermittedRHS->getType()) : V, nullptr);
434 if (isa<ConstantAggregateZero>(V)) {
435 Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(V->getContext()),0));
436 return std::make_pair(V, nullptr);
439 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
440 // If this is an insert of an extract from some other vector, include it.
441 Value *VecOp = IEI->getOperand(0);
442 Value *ScalarOp = IEI->getOperand(1);
443 Value *IdxOp = IEI->getOperand(2);
445 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
446 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) {
447 unsigned ExtractedIdx =
448 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
449 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
451 // Either the extracted from or inserted into vector must be RHSVec,
452 // otherwise we'd end up with a shuffle of three inputs.
453 if (EI->getOperand(0) == PermittedRHS || PermittedRHS == nullptr) {
454 Value *RHS = EI->getOperand(0);
455 ShuffleOps LR = collectShuffleElements(VecOp, Mask, RHS, IC);
456 assert(LR.second == nullptr || LR.second == RHS);
458 if (LR.first->getType() != RHS->getType()) {
459 // Although we are giving up for now, see if we can create extracts
460 // that match the inserts for another round of combining.
461 replaceExtractElements(IEI, EI, IC);
463 // We tried our best, but we can't find anything compatible with RHS
464 // further up the chain. Return a trivial shuffle.
465 for (unsigned i = 0; i < NumElts; ++i)
466 Mask[i] = ConstantInt::get(Type::getInt32Ty(V->getContext()), i);
467 return std::make_pair(V, nullptr);
470 unsigned NumLHSElts = RHS->getType()->getVectorNumElements();
471 Mask[InsertedIdx % NumElts] =
472 ConstantInt::get(Type::getInt32Ty(V->getContext()),
473 NumLHSElts+ExtractedIdx);
474 return std::make_pair(LR.first, RHS);
477 if (VecOp == PermittedRHS) {
478 // We've gone as far as we can: anything on the other side of the
479 // extractelement will already have been converted into a shuffle.
480 unsigned NumLHSElts =
481 EI->getOperand(0)->getType()->getVectorNumElements();
482 for (unsigned i = 0; i != NumElts; ++i)
483 Mask.push_back(ConstantInt::get(
484 Type::getInt32Ty(V->getContext()),
485 i == InsertedIdx ? ExtractedIdx : NumLHSElts + i));
486 return std::make_pair(EI->getOperand(0), PermittedRHS);
489 // If this insertelement is a chain that comes from exactly these two
490 // vectors, return the vector and the effective shuffle.
491 if (EI->getOperand(0)->getType() == PermittedRHS->getType() &&
492 collectSingleShuffleElements(IEI, EI->getOperand(0), PermittedRHS,
494 return std::make_pair(EI->getOperand(0), PermittedRHS);
499 // Otherwise, we can't do anything fancy. Return an identity vector.
500 for (unsigned i = 0; i != NumElts; ++i)
501 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
502 return std::make_pair(V, nullptr);
505 /// Try to find redundant insertvalue instructions, like the following ones:
506 /// %0 = insertvalue { i8, i32 } undef, i8 %x, 0
507 /// %1 = insertvalue { i8, i32 } %0, i8 %y, 0
508 /// Here the second instruction inserts values at the same indices, as the
509 /// first one, making the first one redundant.
510 /// It should be transformed to:
511 /// %0 = insertvalue { i8, i32 } undef, i8 %y, 0
512 Instruction *InstCombiner::visitInsertValueInst(InsertValueInst &I) {
513 bool IsRedundant = false;
514 ArrayRef<unsigned int> FirstIndices = I.getIndices();
516 // If there is a chain of insertvalue instructions (each of them except the
517 // last one has only one use and it's another insertvalue insn from this
518 // chain), check if any of the 'children' uses the same indices as the first
519 // instruction. In this case, the first one is redundant.
522 while (V->hasOneUse() && Depth < 10) {
523 User *U = V->user_back();
524 auto UserInsInst = dyn_cast<InsertValueInst>(U);
525 if (!UserInsInst || U->getOperand(0) != V)
527 if (UserInsInst->getIndices() == FirstIndices) {
536 return ReplaceInstUsesWith(I, I.getOperand(0));
540 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
541 Value *VecOp = IE.getOperand(0);
542 Value *ScalarOp = IE.getOperand(1);
543 Value *IdxOp = IE.getOperand(2);
545 // Inserting an undef or into an undefined place, remove this.
546 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
547 ReplaceInstUsesWith(IE, VecOp);
549 // If the inserted element was extracted from some other vector, and if the
550 // indexes are constant, try to turn this into a shufflevector operation.
551 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
552 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) {
553 unsigned NumInsertVectorElts = IE.getType()->getNumElements();
554 unsigned NumExtractVectorElts =
555 EI->getOperand(0)->getType()->getVectorNumElements();
556 unsigned ExtractedIdx =
557 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
558 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
560 if (ExtractedIdx >= NumExtractVectorElts) // Out of range extract.
561 return ReplaceInstUsesWith(IE, VecOp);
563 if (InsertedIdx >= NumInsertVectorElts) // Out of range insert.
564 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
566 // If we are extracting a value from a vector, then inserting it right
567 // back into the same place, just use the input vector.
568 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
569 return ReplaceInstUsesWith(IE, VecOp);
571 // If this insertelement isn't used by some other insertelement, turn it
572 // (and any insertelements it points to), into one big shuffle.
573 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.user_back())) {
574 SmallVector<Constant*, 16> Mask;
575 ShuffleOps LR = collectShuffleElements(&IE, Mask, nullptr, *this);
577 // The proposed shuffle may be trivial, in which case we shouldn't
578 // perform the combine.
579 if (LR.first != &IE && LR.second != &IE) {
580 // We now have a shuffle of LHS, RHS, Mask.
581 if (LR.second == nullptr)
582 LR.second = UndefValue::get(LR.first->getType());
583 return new ShuffleVectorInst(LR.first, LR.second,
584 ConstantVector::get(Mask));
590 unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements();
591 APInt UndefElts(VWidth, 0);
592 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
593 if (Value *V = SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts)) {
595 return ReplaceInstUsesWith(IE, V);
602 /// Return true if we can evaluate the specified expression tree if the vector
603 /// elements were shuffled in a different order.
604 static bool CanEvaluateShuffled(Value *V, ArrayRef<int> Mask,
605 unsigned Depth = 5) {
606 // We can always reorder the elements of a constant.
607 if (isa<Constant>(V))
610 // We won't reorder vector arguments. No IPO here.
611 Instruction *I = dyn_cast<Instruction>(V);
612 if (!I) return false;
614 // Two users may expect different orders of the elements. Don't try it.
618 if (Depth == 0) return false;
620 switch (I->getOpcode()) {
621 case Instruction::Add:
622 case Instruction::FAdd:
623 case Instruction::Sub:
624 case Instruction::FSub:
625 case Instruction::Mul:
626 case Instruction::FMul:
627 case Instruction::UDiv:
628 case Instruction::SDiv:
629 case Instruction::FDiv:
630 case Instruction::URem:
631 case Instruction::SRem:
632 case Instruction::FRem:
633 case Instruction::Shl:
634 case Instruction::LShr:
635 case Instruction::AShr:
636 case Instruction::And:
637 case Instruction::Or:
638 case Instruction::Xor:
639 case Instruction::ICmp:
640 case Instruction::FCmp:
641 case Instruction::Trunc:
642 case Instruction::ZExt:
643 case Instruction::SExt:
644 case Instruction::FPToUI:
645 case Instruction::FPToSI:
646 case Instruction::UIToFP:
647 case Instruction::SIToFP:
648 case Instruction::FPTrunc:
649 case Instruction::FPExt:
650 case Instruction::GetElementPtr: {
651 for (Value *Operand : I->operands()) {
652 if (!CanEvaluateShuffled(Operand, Mask, Depth-1))
657 case Instruction::InsertElement: {
658 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(2));
659 if (!CI) return false;
660 int ElementNumber = CI->getLimitedValue();
662 // Verify that 'CI' does not occur twice in Mask. A single 'insertelement'
663 // can't put an element into multiple indices.
664 bool SeenOnce = false;
665 for (int i = 0, e = Mask.size(); i != e; ++i) {
666 if (Mask[i] == ElementNumber) {
672 return CanEvaluateShuffled(I->getOperand(0), Mask, Depth-1);
678 /// Rebuild a new instruction just like 'I' but with the new operands given.
679 /// In the event of type mismatch, the type of the operands is correct.
680 static Value *buildNew(Instruction *I, ArrayRef<Value*> NewOps) {
681 // We don't want to use the IRBuilder here because we want the replacement
682 // instructions to appear next to 'I', not the builder's insertion point.
683 switch (I->getOpcode()) {
684 case Instruction::Add:
685 case Instruction::FAdd:
686 case Instruction::Sub:
687 case Instruction::FSub:
688 case Instruction::Mul:
689 case Instruction::FMul:
690 case Instruction::UDiv:
691 case Instruction::SDiv:
692 case Instruction::FDiv:
693 case Instruction::URem:
694 case Instruction::SRem:
695 case Instruction::FRem:
696 case Instruction::Shl:
697 case Instruction::LShr:
698 case Instruction::AShr:
699 case Instruction::And:
700 case Instruction::Or:
701 case Instruction::Xor: {
702 BinaryOperator *BO = cast<BinaryOperator>(I);
703 assert(NewOps.size() == 2 && "binary operator with #ops != 2");
704 BinaryOperator *New =
705 BinaryOperator::Create(cast<BinaryOperator>(I)->getOpcode(),
706 NewOps[0], NewOps[1], "", BO);
707 if (isa<OverflowingBinaryOperator>(BO)) {
708 New->setHasNoUnsignedWrap(BO->hasNoUnsignedWrap());
709 New->setHasNoSignedWrap(BO->hasNoSignedWrap());
711 if (isa<PossiblyExactOperator>(BO)) {
712 New->setIsExact(BO->isExact());
714 if (isa<FPMathOperator>(BO))
715 New->copyFastMathFlags(I);
718 case Instruction::ICmp:
719 assert(NewOps.size() == 2 && "icmp with #ops != 2");
720 return new ICmpInst(I, cast<ICmpInst>(I)->getPredicate(),
721 NewOps[0], NewOps[1]);
722 case Instruction::FCmp:
723 assert(NewOps.size() == 2 && "fcmp with #ops != 2");
724 return new FCmpInst(I, cast<FCmpInst>(I)->getPredicate(),
725 NewOps[0], NewOps[1]);
726 case Instruction::Trunc:
727 case Instruction::ZExt:
728 case Instruction::SExt:
729 case Instruction::FPToUI:
730 case Instruction::FPToSI:
731 case Instruction::UIToFP:
732 case Instruction::SIToFP:
733 case Instruction::FPTrunc:
734 case Instruction::FPExt: {
735 // It's possible that the mask has a different number of elements from
736 // the original cast. We recompute the destination type to match the mask.
738 VectorType::get(I->getType()->getScalarType(),
739 NewOps[0]->getType()->getVectorNumElements());
740 assert(NewOps.size() == 1 && "cast with #ops != 1");
741 return CastInst::Create(cast<CastInst>(I)->getOpcode(), NewOps[0], DestTy,
744 case Instruction::GetElementPtr: {
745 Value *Ptr = NewOps[0];
746 ArrayRef<Value*> Idx = NewOps.slice(1);
747 GetElementPtrInst *GEP = GetElementPtrInst::Create(
748 cast<GetElementPtrInst>(I)->getSourceElementType(), Ptr, Idx, "", I);
749 GEP->setIsInBounds(cast<GetElementPtrInst>(I)->isInBounds());
753 llvm_unreachable("failed to rebuild vector instructions");
757 InstCombiner::EvaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask) {
758 // Mask.size() does not need to be equal to the number of vector elements.
760 assert(V->getType()->isVectorTy() && "can't reorder non-vector elements");
761 if (isa<UndefValue>(V)) {
762 return UndefValue::get(VectorType::get(V->getType()->getScalarType(),
765 if (isa<ConstantAggregateZero>(V)) {
766 return ConstantAggregateZero::get(
767 VectorType::get(V->getType()->getScalarType(),
770 if (Constant *C = dyn_cast<Constant>(V)) {
771 SmallVector<Constant *, 16> MaskValues;
772 for (int i = 0, e = Mask.size(); i != e; ++i) {
774 MaskValues.push_back(UndefValue::get(Builder->getInt32Ty()));
776 MaskValues.push_back(Builder->getInt32(Mask[i]));
778 return ConstantExpr::getShuffleVector(C, UndefValue::get(C->getType()),
779 ConstantVector::get(MaskValues));
782 Instruction *I = cast<Instruction>(V);
783 switch (I->getOpcode()) {
784 case Instruction::Add:
785 case Instruction::FAdd:
786 case Instruction::Sub:
787 case Instruction::FSub:
788 case Instruction::Mul:
789 case Instruction::FMul:
790 case Instruction::UDiv:
791 case Instruction::SDiv:
792 case Instruction::FDiv:
793 case Instruction::URem:
794 case Instruction::SRem:
795 case Instruction::FRem:
796 case Instruction::Shl:
797 case Instruction::LShr:
798 case Instruction::AShr:
799 case Instruction::And:
800 case Instruction::Or:
801 case Instruction::Xor:
802 case Instruction::ICmp:
803 case Instruction::FCmp:
804 case Instruction::Trunc:
805 case Instruction::ZExt:
806 case Instruction::SExt:
807 case Instruction::FPToUI:
808 case Instruction::FPToSI:
809 case Instruction::UIToFP:
810 case Instruction::SIToFP:
811 case Instruction::FPTrunc:
812 case Instruction::FPExt:
813 case Instruction::Select:
814 case Instruction::GetElementPtr: {
815 SmallVector<Value*, 8> NewOps;
816 bool NeedsRebuild = (Mask.size() != I->getType()->getVectorNumElements());
817 for (int i = 0, e = I->getNumOperands(); i != e; ++i) {
818 Value *V = EvaluateInDifferentElementOrder(I->getOperand(i), Mask);
820 NeedsRebuild |= (V != I->getOperand(i));
823 return buildNew(I, NewOps);
827 case Instruction::InsertElement: {
828 int Element = cast<ConstantInt>(I->getOperand(2))->getLimitedValue();
830 // The insertelement was inserting at Element. Figure out which element
831 // that becomes after shuffling. The answer is guaranteed to be unique
832 // by CanEvaluateShuffled.
835 for (int e = Mask.size(); Index != e; ++Index) {
836 if (Mask[Index] == Element) {
842 // If element is not in Mask, no need to handle the operand 1 (element to
843 // be inserted). Just evaluate values in operand 0 according to Mask.
845 return EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
847 Value *V = EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
848 return InsertElementInst::Create(V, I->getOperand(1),
849 Builder->getInt32(Index), "", I);
852 llvm_unreachable("failed to reorder elements of vector instruction!");
855 static void recognizeIdentityMask(const SmallVectorImpl<int> &Mask,
856 bool &isLHSID, bool &isRHSID) {
857 isLHSID = isRHSID = true;
859 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
860 if (Mask[i] < 0) continue; // Ignore undef values.
861 // Is this an identity shuffle of the LHS value?
862 isLHSID &= (Mask[i] == (int)i);
864 // Is this an identity shuffle of the RHS value?
865 isRHSID &= (Mask[i]-e == i);
869 // Returns true if the shuffle is extracting a contiguous range of values from
871 // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
872 // Input: |AA|BB|CC|DD|EE|FF|GG|HH|II|JJ|KK|LL|MM|NN|OO|PP|
873 // Shuffles to: |EE|FF|GG|HH|
875 static bool isShuffleExtractingFromLHS(ShuffleVectorInst &SVI,
876 SmallVector<int, 16> &Mask) {
878 cast<VectorType>(SVI.getOperand(0)->getType())->getNumElements();
879 unsigned MaskElems = Mask.size();
880 unsigned BegIdx = Mask.front();
881 unsigned EndIdx = Mask.back();
882 if (BegIdx > EndIdx || EndIdx >= LHSElems || EndIdx - BegIdx != MaskElems - 1)
884 for (unsigned I = 0; I != MaskElems; ++I)
885 if (static_cast<unsigned>(Mask[I]) != BegIdx + I)
890 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
891 Value *LHS = SVI.getOperand(0);
892 Value *RHS = SVI.getOperand(1);
893 SmallVector<int, 16> Mask = SVI.getShuffleMask();
894 Type *Int32Ty = Type::getInt32Ty(SVI.getContext());
896 bool MadeChange = false;
898 // Undefined shuffle mask -> undefined value.
899 if (isa<UndefValue>(SVI.getOperand(2)))
900 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
902 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
904 APInt UndefElts(VWidth, 0);
905 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
906 if (Value *V = SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
908 return ReplaceInstUsesWith(SVI, V);
909 LHS = SVI.getOperand(0);
910 RHS = SVI.getOperand(1);
914 unsigned LHSWidth = cast<VectorType>(LHS->getType())->getNumElements();
916 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
917 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
918 if (LHS == RHS || isa<UndefValue>(LHS)) {
919 if (isa<UndefValue>(LHS) && LHS == RHS) {
920 // shuffle(undef,undef,mask) -> undef.
921 Value *Result = (VWidth == LHSWidth)
922 ? LHS : UndefValue::get(SVI.getType());
923 return ReplaceInstUsesWith(SVI, Result);
926 // Remap any references to RHS to use LHS.
927 SmallVector<Constant*, 16> Elts;
928 for (unsigned i = 0, e = LHSWidth; i != VWidth; ++i) {
930 Elts.push_back(UndefValue::get(Int32Ty));
934 if ((Mask[i] >= (int)e && isa<UndefValue>(RHS)) ||
935 (Mask[i] < (int)e && isa<UndefValue>(LHS))) {
936 Mask[i] = -1; // Turn into undef.
937 Elts.push_back(UndefValue::get(Int32Ty));
939 Mask[i] = Mask[i] % e; // Force to LHS.
940 Elts.push_back(ConstantInt::get(Int32Ty, Mask[i]));
943 SVI.setOperand(0, SVI.getOperand(1));
944 SVI.setOperand(1, UndefValue::get(RHS->getType()));
945 SVI.setOperand(2, ConstantVector::get(Elts));
946 LHS = SVI.getOperand(0);
947 RHS = SVI.getOperand(1);
951 if (VWidth == LHSWidth) {
952 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
953 bool isLHSID, isRHSID;
954 recognizeIdentityMask(Mask, isLHSID, isRHSID);
956 // Eliminate identity shuffles.
957 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
958 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
961 if (isa<UndefValue>(RHS) && CanEvaluateShuffled(LHS, Mask)) {
962 Value *V = EvaluateInDifferentElementOrder(LHS, Mask);
963 return ReplaceInstUsesWith(SVI, V);
966 // SROA generates shuffle+bitcast when the extracted sub-vector is bitcast to
967 // a non-vector type. We can instead bitcast the original vector followed by
968 // an extract of the desired element:
970 // %sroa = shufflevector <16 x i8> %in, <16 x i8> undef,
971 // <4 x i32> <i32 0, i32 1, i32 2, i32 3>
972 // %1 = bitcast <4 x i8> %sroa to i32
974 // %bc = bitcast <16 x i8> %in to <4 x i32>
975 // %ext = extractelement <4 x i32> %bc, i32 0
977 // If the shuffle is extracting a contiguous range of values from the input
978 // vector then each use which is a bitcast of the extracted size can be
979 // replaced. This will work if the vector types are compatible, and the begin
980 // index is aligned to a value in the casted vector type. If the begin index
981 // isn't aligned then we can shuffle the original vector (keeping the same
982 // vector type) before extracting.
984 // This code will bail out if the target type is fundamentally incompatible
985 // with vectors of the source type.
987 // Example of <16 x i8>, target type i32:
988 // Index range [4,8): v-----------v Will work.
989 // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
990 // <16 x i8>: | | | | | | | | | | | | | | | | |
991 // <4 x i32>: | | | | |
992 // +-----------+-----------+-----------+-----------+
993 // Index range [6,10): ^-----------^ Needs an extra shuffle.
994 // Target type i40: ^--------------^ Won't work, bail.
995 if (isShuffleExtractingFromLHS(SVI, Mask)) {
997 unsigned MaskElems = Mask.size();
998 unsigned BegIdx = Mask.front();
999 VectorType *SrcTy = cast<VectorType>(V->getType());
1000 unsigned VecBitWidth = SrcTy->getBitWidth();
1001 unsigned SrcElemBitWidth = DL.getTypeSizeInBits(SrcTy->getElementType());
1002 assert(SrcElemBitWidth && "vector elements must have a bitwidth");
1003 unsigned SrcNumElems = SrcTy->getNumElements();
1004 SmallVector<BitCastInst *, 8> BCs;
1005 DenseMap<Type *, Value *> NewBCs;
1006 for (User *U : SVI.users())
1007 if (BitCastInst *BC = dyn_cast<BitCastInst>(U))
1008 if (!BC->use_empty())
1009 // Only visit bitcasts that weren't previously handled.
1011 for (BitCastInst *BC : BCs) {
1012 Type *TgtTy = BC->getDestTy();
1013 unsigned TgtElemBitWidth = DL.getTypeSizeInBits(TgtTy);
1014 if (!TgtElemBitWidth)
1016 unsigned TgtNumElems = VecBitWidth / TgtElemBitWidth;
1017 bool VecBitWidthsEqual = VecBitWidth == TgtNumElems * TgtElemBitWidth;
1018 bool BegIsAligned = 0 == ((SrcElemBitWidth * BegIdx) % TgtElemBitWidth);
1019 if (!VecBitWidthsEqual)
1021 if (!VectorType::isValidElementType(TgtTy))
1023 VectorType *CastSrcTy = VectorType::get(TgtTy, TgtNumElems);
1024 if (!BegIsAligned) {
1025 // Shuffle the input so [0,NumElements) contains the output, and
1026 // [NumElems,SrcNumElems) is undef.
1027 SmallVector<Constant *, 16> ShuffleMask(SrcNumElems,
1028 UndefValue::get(Int32Ty));
1029 for (unsigned I = 0, E = MaskElems, Idx = BegIdx; I != E; ++Idx, ++I)
1030 ShuffleMask[I] = ConstantInt::get(Int32Ty, Idx);
1031 V = Builder->CreateShuffleVector(V, UndefValue::get(V->getType()),
1032 ConstantVector::get(ShuffleMask),
1033 SVI.getName() + ".extract");
1036 unsigned SrcElemsPerTgtElem = TgtElemBitWidth / SrcElemBitWidth;
1037 assert(SrcElemsPerTgtElem);
1038 BegIdx /= SrcElemsPerTgtElem;
1039 bool BCAlreadyExists = NewBCs.find(CastSrcTy) != NewBCs.end();
1043 : Builder->CreateBitCast(V, CastSrcTy, SVI.getName() + ".bc");
1044 if (!BCAlreadyExists)
1045 NewBCs[CastSrcTy] = NewBC;
1046 auto *Ext = Builder->CreateExtractElement(
1047 NewBC, ConstantInt::get(Int32Ty, BegIdx), SVI.getName() + ".extract");
1048 // The shufflevector isn't being replaced: the bitcast that used it
1049 // is. InstCombine will visit the newly-created instructions.
1050 ReplaceInstUsesWith(*BC, Ext);
1055 // If the LHS is a shufflevector itself, see if we can combine it with this
1056 // one without producing an unusual shuffle.
1057 // Cases that might be simplified:
1059 // x1=shuffle(v1,v2,mask1)
1060 // x=shuffle(x1,undef,mask)
1062 // x=shuffle(v1,undef,newMask)
1063 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : -1
1065 // x1=shuffle(v1,undef,mask1)
1066 // x=shuffle(x1,x2,mask)
1067 // where v1.size() == mask1.size()
1069 // x=shuffle(v1,x2,newMask)
1070 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : mask[i]
1072 // x2=shuffle(v2,undef,mask2)
1073 // x=shuffle(x1,x2,mask)
1074 // where v2.size() == mask2.size()
1076 // x=shuffle(x1,v2,newMask)
1077 // newMask[i] = (mask[i] < x1.size())
1078 // ? mask[i] : mask2[mask[i]-x1.size()]+x1.size()
1080 // x1=shuffle(v1,undef,mask1)
1081 // x2=shuffle(v2,undef,mask2)
1082 // x=shuffle(x1,x2,mask)
1083 // where v1.size() == v2.size()
1085 // x=shuffle(v1,v2,newMask)
1086 // newMask[i] = (mask[i] < x1.size())
1087 // ? mask1[mask[i]] : mask2[mask[i]-x1.size()]+v1.size()
1089 // Here we are really conservative:
1090 // we are absolutely afraid of producing a shuffle mask not in the input
1091 // program, because the code gen may not be smart enough to turn a merged
1092 // shuffle into two specific shuffles: it may produce worse code. As such,
1093 // we only merge two shuffles if the result is either a splat or one of the
1094 // input shuffle masks. In this case, merging the shuffles just removes
1095 // one instruction, which we know is safe. This is good for things like
1096 // turning: (splat(splat)) -> splat, or
1097 // merge(V[0..n], V[n+1..2n]) -> V[0..2n]
1098 ShuffleVectorInst* LHSShuffle = dyn_cast<ShuffleVectorInst>(LHS);
1099 ShuffleVectorInst* RHSShuffle = dyn_cast<ShuffleVectorInst>(RHS);
1101 if (!isa<UndefValue>(LHSShuffle->getOperand(1)) && !isa<UndefValue>(RHS))
1102 LHSShuffle = nullptr;
1104 if (!isa<UndefValue>(RHSShuffle->getOperand(1)))
1105 RHSShuffle = nullptr;
1106 if (!LHSShuffle && !RHSShuffle)
1107 return MadeChange ? &SVI : nullptr;
1109 Value* LHSOp0 = nullptr;
1110 Value* LHSOp1 = nullptr;
1111 Value* RHSOp0 = nullptr;
1112 unsigned LHSOp0Width = 0;
1113 unsigned RHSOp0Width = 0;
1115 LHSOp0 = LHSShuffle->getOperand(0);
1116 LHSOp1 = LHSShuffle->getOperand(1);
1117 LHSOp0Width = cast<VectorType>(LHSOp0->getType())->getNumElements();
1120 RHSOp0 = RHSShuffle->getOperand(0);
1121 RHSOp0Width = cast<VectorType>(RHSOp0->getType())->getNumElements();
1123 Value* newLHS = LHS;
1124 Value* newRHS = RHS;
1127 if (isa<UndefValue>(RHS)) {
1132 else if (LHSOp0Width == LHSWidth) {
1137 if (RHSShuffle && RHSOp0Width == LHSWidth) {
1141 if (LHSOp0 == RHSOp0) {
1146 if (newLHS == LHS && newRHS == RHS)
1147 return MadeChange ? &SVI : nullptr;
1149 SmallVector<int, 16> LHSMask;
1150 SmallVector<int, 16> RHSMask;
1152 LHSMask = LHSShuffle->getShuffleMask();
1153 if (RHSShuffle && newRHS != RHS)
1154 RHSMask = RHSShuffle->getShuffleMask();
1156 unsigned newLHSWidth = (newLHS != LHS) ? LHSOp0Width : LHSWidth;
1157 SmallVector<int, 16> newMask;
1158 bool isSplat = true;
1160 // Create a new mask for the new ShuffleVectorInst so that the new
1161 // ShuffleVectorInst is equivalent to the original one.
1162 for (unsigned i = 0; i < VWidth; ++i) {
1165 // This element is an undef value.
1167 } else if (Mask[i] < (int)LHSWidth) {
1168 // This element is from left hand side vector operand.
1170 // If LHS is going to be replaced (case 1, 2, or 4), calculate the
1171 // new mask value for the element.
1172 if (newLHS != LHS) {
1173 eltMask = LHSMask[Mask[i]];
1174 // If the value selected is an undef value, explicitly specify it
1175 // with a -1 mask value.
1176 if (eltMask >= (int)LHSOp0Width && isa<UndefValue>(LHSOp1))
1181 // This element is from right hand side vector operand
1183 // If the value selected is an undef value, explicitly specify it
1184 // with a -1 mask value. (case 1)
1185 if (isa<UndefValue>(RHS))
1187 // If RHS is going to be replaced (case 3 or 4), calculate the
1188 // new mask value for the element.
1189 else if (newRHS != RHS) {
1190 eltMask = RHSMask[Mask[i]-LHSWidth];
1191 // If the value selected is an undef value, explicitly specify it
1192 // with a -1 mask value.
1193 if (eltMask >= (int)RHSOp0Width) {
1194 assert(isa<UndefValue>(RHSShuffle->getOperand(1))
1195 && "should have been check above");
1199 eltMask = Mask[i]-LHSWidth;
1201 // If LHS's width is changed, shift the mask value accordingly.
1202 // If newRHS == NULL, i.e. LHSOp0 == RHSOp0, we want to remap any
1203 // references from RHSOp0 to LHSOp0, so we don't need to shift the mask.
1204 // If newRHS == newLHS, we want to remap any references from newRHS to
1205 // newLHS so that we can properly identify splats that may occur due to
1206 // obfuscation across the two vectors.
1207 if (eltMask >= 0 && newRHS != nullptr && newLHS != newRHS)
1208 eltMask += newLHSWidth;
1211 // Check if this could still be a splat.
1213 if (SplatElt >= 0 && SplatElt != eltMask)
1218 newMask.push_back(eltMask);
1221 // If the result mask is equal to one of the original shuffle masks,
1222 // or is a splat, do the replacement.
1223 if (isSplat || newMask == LHSMask || newMask == RHSMask || newMask == Mask) {
1224 SmallVector<Constant*, 16> Elts;
1225 for (unsigned i = 0, e = newMask.size(); i != e; ++i) {
1226 if (newMask[i] < 0) {
1227 Elts.push_back(UndefValue::get(Int32Ty));
1229 Elts.push_back(ConstantInt::get(Int32Ty, newMask[i]));
1233 newRHS = UndefValue::get(newLHS->getType());
1234 return new ShuffleVectorInst(newLHS, newRHS, ConstantVector::get(Elts));
1237 // If the result mask is an identity, replace uses of this instruction with
1238 // corresponding argument.
1239 bool isLHSID, isRHSID;
1240 recognizeIdentityMask(newMask, isLHSID, isRHSID);
1241 if (isLHSID && VWidth == LHSOp0Width) return ReplaceInstUsesWith(SVI, newLHS);
1242 if (isRHSID && VWidth == RHSOp0Width) return ReplaceInstUsesWith(SVI, newRHS);
1244 return MadeChange ? &SVI : nullptr;