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 *ExtVecOpInst = dyn_cast<Instruction>(ExtVecOp);
384 BasicBlock *InsertionBlock = (ExtVecOpInst && !isa<PHINode>(ExtVecOpInst))
385 ? ExtVecOpInst->getParent()
386 : ExtElt->getParent();
388 // TODO: This restriction matches the basic block check below when creating
389 // new extractelement instructions. If that limitation is removed, this one
390 // could also be removed. But for now, we just bail out to ensure that we
391 // will replace the extractelement instruction that is feeding our
392 // insertelement instruction. This allows the insertelement to then be
393 // replaced by a shufflevector. If the insertelement is not replaced, we can
394 // induce infinite looping because there's an optimization for extractelement
395 // that will delete our widening shuffle. This would trigger another attempt
396 // here to create that shuffle, and we spin forever.
397 if (InsertionBlock != InsElt->getParent())
400 auto *WideVec = new ShuffleVectorInst(ExtVecOp, UndefValue::get(ExtVecType),
401 ConstantVector::get(ExtendMask));
403 // Insert the new shuffle after the vector operand of the extract is defined
404 // (as long as it's not a PHI) or at the start of the basic block of the
405 // extract, so any subsequent extracts in the same basic block can use it.
406 // TODO: Insert before the earliest ExtractElementInst that is replaced.
407 if (ExtVecOpInst && !isa<PHINode>(ExtVecOpInst))
408 WideVec->insertAfter(ExtVecOpInst);
410 IC.InsertNewInstWith(WideVec, *ExtElt->getParent()->getFirstInsertionPt());
412 // Replace extracts from the original narrow vector with extracts from the new
414 for (User *U : ExtVecOp->users()) {
415 ExtractElementInst *OldExt = dyn_cast<ExtractElementInst>(U);
416 if (!OldExt || OldExt->getParent() != WideVec->getParent())
418 auto *NewExt = ExtractElementInst::Create(WideVec, OldExt->getOperand(1));
419 NewExt->insertAfter(WideVec);
420 IC.ReplaceInstUsesWith(*OldExt, NewExt);
424 /// We are building a shuffle to create V, which is a sequence of insertelement,
425 /// extractelement pairs. If PermittedRHS is set, then we must either use it or
426 /// not rely on the second vector source. Return a std::pair containing the
427 /// left and right vectors of the proposed shuffle (or 0), and set the Mask
428 /// parameter as required.
430 /// Note: we intentionally don't try to fold earlier shuffles since they have
431 /// often been chosen carefully to be efficiently implementable on the target.
432 typedef std::pair<Value *, Value *> ShuffleOps;
434 static ShuffleOps collectShuffleElements(Value *V,
435 SmallVectorImpl<Constant *> &Mask,
438 assert(V->getType()->isVectorTy() && "Invalid shuffle!");
439 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
441 if (isa<UndefValue>(V)) {
442 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
443 return std::make_pair(
444 PermittedRHS ? UndefValue::get(PermittedRHS->getType()) : V, nullptr);
447 if (isa<ConstantAggregateZero>(V)) {
448 Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(V->getContext()),0));
449 return std::make_pair(V, nullptr);
452 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
453 // If this is an insert of an extract from some other vector, include it.
454 Value *VecOp = IEI->getOperand(0);
455 Value *ScalarOp = IEI->getOperand(1);
456 Value *IdxOp = IEI->getOperand(2);
458 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
459 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) {
460 unsigned ExtractedIdx =
461 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
462 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
464 // Either the extracted from or inserted into vector must be RHSVec,
465 // otherwise we'd end up with a shuffle of three inputs.
466 if (EI->getOperand(0) == PermittedRHS || PermittedRHS == nullptr) {
467 Value *RHS = EI->getOperand(0);
468 ShuffleOps LR = collectShuffleElements(VecOp, Mask, RHS, IC);
469 assert(LR.second == nullptr || LR.second == RHS);
471 if (LR.first->getType() != RHS->getType()) {
472 // Although we are giving up for now, see if we can create extracts
473 // that match the inserts for another round of combining.
474 replaceExtractElements(IEI, EI, IC);
476 // We tried our best, but we can't find anything compatible with RHS
477 // further up the chain. Return a trivial shuffle.
478 for (unsigned i = 0; i < NumElts; ++i)
479 Mask[i] = ConstantInt::get(Type::getInt32Ty(V->getContext()), i);
480 return std::make_pair(V, nullptr);
483 unsigned NumLHSElts = RHS->getType()->getVectorNumElements();
484 Mask[InsertedIdx % NumElts] =
485 ConstantInt::get(Type::getInt32Ty(V->getContext()),
486 NumLHSElts+ExtractedIdx);
487 return std::make_pair(LR.first, RHS);
490 if (VecOp == PermittedRHS) {
491 // We've gone as far as we can: anything on the other side of the
492 // extractelement will already have been converted into a shuffle.
493 unsigned NumLHSElts =
494 EI->getOperand(0)->getType()->getVectorNumElements();
495 for (unsigned i = 0; i != NumElts; ++i)
496 Mask.push_back(ConstantInt::get(
497 Type::getInt32Ty(V->getContext()),
498 i == InsertedIdx ? ExtractedIdx : NumLHSElts + i));
499 return std::make_pair(EI->getOperand(0), PermittedRHS);
502 // If this insertelement is a chain that comes from exactly these two
503 // vectors, return the vector and the effective shuffle.
504 if (EI->getOperand(0)->getType() == PermittedRHS->getType() &&
505 collectSingleShuffleElements(IEI, EI->getOperand(0), PermittedRHS,
507 return std::make_pair(EI->getOperand(0), PermittedRHS);
512 // Otherwise, we can't do anything fancy. Return an identity vector.
513 for (unsigned i = 0; i != NumElts; ++i)
514 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
515 return std::make_pair(V, nullptr);
518 /// Try to find redundant insertvalue instructions, like the following ones:
519 /// %0 = insertvalue { i8, i32 } undef, i8 %x, 0
520 /// %1 = insertvalue { i8, i32 } %0, i8 %y, 0
521 /// Here the second instruction inserts values at the same indices, as the
522 /// first one, making the first one redundant.
523 /// It should be transformed to:
524 /// %0 = insertvalue { i8, i32 } undef, i8 %y, 0
525 Instruction *InstCombiner::visitInsertValueInst(InsertValueInst &I) {
526 bool IsRedundant = false;
527 ArrayRef<unsigned int> FirstIndices = I.getIndices();
529 // If there is a chain of insertvalue instructions (each of them except the
530 // last one has only one use and it's another insertvalue insn from this
531 // chain), check if any of the 'children' uses the same indices as the first
532 // instruction. In this case, the first one is redundant.
535 while (V->hasOneUse() && Depth < 10) {
536 User *U = V->user_back();
537 auto UserInsInst = dyn_cast<InsertValueInst>(U);
538 if (!UserInsInst || U->getOperand(0) != V)
540 if (UserInsInst->getIndices() == FirstIndices) {
549 return ReplaceInstUsesWith(I, I.getOperand(0));
553 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
554 Value *VecOp = IE.getOperand(0);
555 Value *ScalarOp = IE.getOperand(1);
556 Value *IdxOp = IE.getOperand(2);
558 // Inserting an undef or into an undefined place, remove this.
559 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
560 ReplaceInstUsesWith(IE, VecOp);
562 // If the inserted element was extracted from some other vector, and if the
563 // indexes are constant, try to turn this into a shufflevector operation.
564 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
565 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) {
566 unsigned NumInsertVectorElts = IE.getType()->getNumElements();
567 unsigned NumExtractVectorElts =
568 EI->getOperand(0)->getType()->getVectorNumElements();
569 unsigned ExtractedIdx =
570 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
571 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
573 if (ExtractedIdx >= NumExtractVectorElts) // Out of range extract.
574 return ReplaceInstUsesWith(IE, VecOp);
576 if (InsertedIdx >= NumInsertVectorElts) // Out of range insert.
577 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
579 // If we are extracting a value from a vector, then inserting it right
580 // back into the same place, just use the input vector.
581 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
582 return ReplaceInstUsesWith(IE, VecOp);
584 // If this insertelement isn't used by some other insertelement, turn it
585 // (and any insertelements it points to), into one big shuffle.
586 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.user_back())) {
587 SmallVector<Constant*, 16> Mask;
588 ShuffleOps LR = collectShuffleElements(&IE, Mask, nullptr, *this);
590 // The proposed shuffle may be trivial, in which case we shouldn't
591 // perform the combine.
592 if (LR.first != &IE && LR.second != &IE) {
593 // We now have a shuffle of LHS, RHS, Mask.
594 if (LR.second == nullptr)
595 LR.second = UndefValue::get(LR.first->getType());
596 return new ShuffleVectorInst(LR.first, LR.second,
597 ConstantVector::get(Mask));
603 unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements();
604 APInt UndefElts(VWidth, 0);
605 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
606 if (Value *V = SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts)) {
608 return ReplaceInstUsesWith(IE, V);
615 /// Return true if we can evaluate the specified expression tree if the vector
616 /// elements were shuffled in a different order.
617 static bool CanEvaluateShuffled(Value *V, ArrayRef<int> Mask,
618 unsigned Depth = 5) {
619 // We can always reorder the elements of a constant.
620 if (isa<Constant>(V))
623 // We won't reorder vector arguments. No IPO here.
624 Instruction *I = dyn_cast<Instruction>(V);
625 if (!I) return false;
627 // Two users may expect different orders of the elements. Don't try it.
631 if (Depth == 0) return false;
633 switch (I->getOpcode()) {
634 case Instruction::Add:
635 case Instruction::FAdd:
636 case Instruction::Sub:
637 case Instruction::FSub:
638 case Instruction::Mul:
639 case Instruction::FMul:
640 case Instruction::UDiv:
641 case Instruction::SDiv:
642 case Instruction::FDiv:
643 case Instruction::URem:
644 case Instruction::SRem:
645 case Instruction::FRem:
646 case Instruction::Shl:
647 case Instruction::LShr:
648 case Instruction::AShr:
649 case Instruction::And:
650 case Instruction::Or:
651 case Instruction::Xor:
652 case Instruction::ICmp:
653 case Instruction::FCmp:
654 case Instruction::Trunc:
655 case Instruction::ZExt:
656 case Instruction::SExt:
657 case Instruction::FPToUI:
658 case Instruction::FPToSI:
659 case Instruction::UIToFP:
660 case Instruction::SIToFP:
661 case Instruction::FPTrunc:
662 case Instruction::FPExt:
663 case Instruction::GetElementPtr: {
664 for (Value *Operand : I->operands()) {
665 if (!CanEvaluateShuffled(Operand, Mask, Depth-1))
670 case Instruction::InsertElement: {
671 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(2));
672 if (!CI) return false;
673 int ElementNumber = CI->getLimitedValue();
675 // Verify that 'CI' does not occur twice in Mask. A single 'insertelement'
676 // can't put an element into multiple indices.
677 bool SeenOnce = false;
678 for (int i = 0, e = Mask.size(); i != e; ++i) {
679 if (Mask[i] == ElementNumber) {
685 return CanEvaluateShuffled(I->getOperand(0), Mask, Depth-1);
691 /// Rebuild a new instruction just like 'I' but with the new operands given.
692 /// In the event of type mismatch, the type of the operands is correct.
693 static Value *buildNew(Instruction *I, ArrayRef<Value*> NewOps) {
694 // We don't want to use the IRBuilder here because we want the replacement
695 // instructions to appear next to 'I', not the builder's insertion point.
696 switch (I->getOpcode()) {
697 case Instruction::Add:
698 case Instruction::FAdd:
699 case Instruction::Sub:
700 case Instruction::FSub:
701 case Instruction::Mul:
702 case Instruction::FMul:
703 case Instruction::UDiv:
704 case Instruction::SDiv:
705 case Instruction::FDiv:
706 case Instruction::URem:
707 case Instruction::SRem:
708 case Instruction::FRem:
709 case Instruction::Shl:
710 case Instruction::LShr:
711 case Instruction::AShr:
712 case Instruction::And:
713 case Instruction::Or:
714 case Instruction::Xor: {
715 BinaryOperator *BO = cast<BinaryOperator>(I);
716 assert(NewOps.size() == 2 && "binary operator with #ops != 2");
717 BinaryOperator *New =
718 BinaryOperator::Create(cast<BinaryOperator>(I)->getOpcode(),
719 NewOps[0], NewOps[1], "", BO);
720 if (isa<OverflowingBinaryOperator>(BO)) {
721 New->setHasNoUnsignedWrap(BO->hasNoUnsignedWrap());
722 New->setHasNoSignedWrap(BO->hasNoSignedWrap());
724 if (isa<PossiblyExactOperator>(BO)) {
725 New->setIsExact(BO->isExact());
727 if (isa<FPMathOperator>(BO))
728 New->copyFastMathFlags(I);
731 case Instruction::ICmp:
732 assert(NewOps.size() == 2 && "icmp with #ops != 2");
733 return new ICmpInst(I, cast<ICmpInst>(I)->getPredicate(),
734 NewOps[0], NewOps[1]);
735 case Instruction::FCmp:
736 assert(NewOps.size() == 2 && "fcmp with #ops != 2");
737 return new FCmpInst(I, cast<FCmpInst>(I)->getPredicate(),
738 NewOps[0], NewOps[1]);
739 case Instruction::Trunc:
740 case Instruction::ZExt:
741 case Instruction::SExt:
742 case Instruction::FPToUI:
743 case Instruction::FPToSI:
744 case Instruction::UIToFP:
745 case Instruction::SIToFP:
746 case Instruction::FPTrunc:
747 case Instruction::FPExt: {
748 // It's possible that the mask has a different number of elements from
749 // the original cast. We recompute the destination type to match the mask.
751 VectorType::get(I->getType()->getScalarType(),
752 NewOps[0]->getType()->getVectorNumElements());
753 assert(NewOps.size() == 1 && "cast with #ops != 1");
754 return CastInst::Create(cast<CastInst>(I)->getOpcode(), NewOps[0], DestTy,
757 case Instruction::GetElementPtr: {
758 Value *Ptr = NewOps[0];
759 ArrayRef<Value*> Idx = NewOps.slice(1);
760 GetElementPtrInst *GEP = GetElementPtrInst::Create(
761 cast<GetElementPtrInst>(I)->getSourceElementType(), Ptr, Idx, "", I);
762 GEP->setIsInBounds(cast<GetElementPtrInst>(I)->isInBounds());
766 llvm_unreachable("failed to rebuild vector instructions");
770 InstCombiner::EvaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask) {
771 // Mask.size() does not need to be equal to the number of vector elements.
773 assert(V->getType()->isVectorTy() && "can't reorder non-vector elements");
774 if (isa<UndefValue>(V)) {
775 return UndefValue::get(VectorType::get(V->getType()->getScalarType(),
778 if (isa<ConstantAggregateZero>(V)) {
779 return ConstantAggregateZero::get(
780 VectorType::get(V->getType()->getScalarType(),
783 if (Constant *C = dyn_cast<Constant>(V)) {
784 SmallVector<Constant *, 16> MaskValues;
785 for (int i = 0, e = Mask.size(); i != e; ++i) {
787 MaskValues.push_back(UndefValue::get(Builder->getInt32Ty()));
789 MaskValues.push_back(Builder->getInt32(Mask[i]));
791 return ConstantExpr::getShuffleVector(C, UndefValue::get(C->getType()),
792 ConstantVector::get(MaskValues));
795 Instruction *I = cast<Instruction>(V);
796 switch (I->getOpcode()) {
797 case Instruction::Add:
798 case Instruction::FAdd:
799 case Instruction::Sub:
800 case Instruction::FSub:
801 case Instruction::Mul:
802 case Instruction::FMul:
803 case Instruction::UDiv:
804 case Instruction::SDiv:
805 case Instruction::FDiv:
806 case Instruction::URem:
807 case Instruction::SRem:
808 case Instruction::FRem:
809 case Instruction::Shl:
810 case Instruction::LShr:
811 case Instruction::AShr:
812 case Instruction::And:
813 case Instruction::Or:
814 case Instruction::Xor:
815 case Instruction::ICmp:
816 case Instruction::FCmp:
817 case Instruction::Trunc:
818 case Instruction::ZExt:
819 case Instruction::SExt:
820 case Instruction::FPToUI:
821 case Instruction::FPToSI:
822 case Instruction::UIToFP:
823 case Instruction::SIToFP:
824 case Instruction::FPTrunc:
825 case Instruction::FPExt:
826 case Instruction::Select:
827 case Instruction::GetElementPtr: {
828 SmallVector<Value*, 8> NewOps;
829 bool NeedsRebuild = (Mask.size() != I->getType()->getVectorNumElements());
830 for (int i = 0, e = I->getNumOperands(); i != e; ++i) {
831 Value *V = EvaluateInDifferentElementOrder(I->getOperand(i), Mask);
833 NeedsRebuild |= (V != I->getOperand(i));
836 return buildNew(I, NewOps);
840 case Instruction::InsertElement: {
841 int Element = cast<ConstantInt>(I->getOperand(2))->getLimitedValue();
843 // The insertelement was inserting at Element. Figure out which element
844 // that becomes after shuffling. The answer is guaranteed to be unique
845 // by CanEvaluateShuffled.
848 for (int e = Mask.size(); Index != e; ++Index) {
849 if (Mask[Index] == Element) {
855 // If element is not in Mask, no need to handle the operand 1 (element to
856 // be inserted). Just evaluate values in operand 0 according to Mask.
858 return EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
860 Value *V = EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
861 return InsertElementInst::Create(V, I->getOperand(1),
862 Builder->getInt32(Index), "", I);
865 llvm_unreachable("failed to reorder elements of vector instruction!");
868 static void recognizeIdentityMask(const SmallVectorImpl<int> &Mask,
869 bool &isLHSID, bool &isRHSID) {
870 isLHSID = isRHSID = true;
872 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
873 if (Mask[i] < 0) continue; // Ignore undef values.
874 // Is this an identity shuffle of the LHS value?
875 isLHSID &= (Mask[i] == (int)i);
877 // Is this an identity shuffle of the RHS value?
878 isRHSID &= (Mask[i]-e == i);
882 // Returns true if the shuffle is extracting a contiguous range of values from
884 // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
885 // Input: |AA|BB|CC|DD|EE|FF|GG|HH|II|JJ|KK|LL|MM|NN|OO|PP|
886 // Shuffles to: |EE|FF|GG|HH|
888 static bool isShuffleExtractingFromLHS(ShuffleVectorInst &SVI,
889 SmallVector<int, 16> &Mask) {
891 cast<VectorType>(SVI.getOperand(0)->getType())->getNumElements();
892 unsigned MaskElems = Mask.size();
893 unsigned BegIdx = Mask.front();
894 unsigned EndIdx = Mask.back();
895 if (BegIdx > EndIdx || EndIdx >= LHSElems || EndIdx - BegIdx != MaskElems - 1)
897 for (unsigned I = 0; I != MaskElems; ++I)
898 if (static_cast<unsigned>(Mask[I]) != BegIdx + I)
903 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
904 Value *LHS = SVI.getOperand(0);
905 Value *RHS = SVI.getOperand(1);
906 SmallVector<int, 16> Mask = SVI.getShuffleMask();
907 Type *Int32Ty = Type::getInt32Ty(SVI.getContext());
909 bool MadeChange = false;
911 // Undefined shuffle mask -> undefined value.
912 if (isa<UndefValue>(SVI.getOperand(2)))
913 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
915 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
917 APInt UndefElts(VWidth, 0);
918 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
919 if (Value *V = SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
921 return ReplaceInstUsesWith(SVI, V);
922 LHS = SVI.getOperand(0);
923 RHS = SVI.getOperand(1);
927 unsigned LHSWidth = cast<VectorType>(LHS->getType())->getNumElements();
929 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
930 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
931 if (LHS == RHS || isa<UndefValue>(LHS)) {
932 if (isa<UndefValue>(LHS) && LHS == RHS) {
933 // shuffle(undef,undef,mask) -> undef.
934 Value *Result = (VWidth == LHSWidth)
935 ? LHS : UndefValue::get(SVI.getType());
936 return ReplaceInstUsesWith(SVI, Result);
939 // Remap any references to RHS to use LHS.
940 SmallVector<Constant*, 16> Elts;
941 for (unsigned i = 0, e = LHSWidth; i != VWidth; ++i) {
943 Elts.push_back(UndefValue::get(Int32Ty));
947 if ((Mask[i] >= (int)e && isa<UndefValue>(RHS)) ||
948 (Mask[i] < (int)e && isa<UndefValue>(LHS))) {
949 Mask[i] = -1; // Turn into undef.
950 Elts.push_back(UndefValue::get(Int32Ty));
952 Mask[i] = Mask[i] % e; // Force to LHS.
953 Elts.push_back(ConstantInt::get(Int32Ty, Mask[i]));
956 SVI.setOperand(0, SVI.getOperand(1));
957 SVI.setOperand(1, UndefValue::get(RHS->getType()));
958 SVI.setOperand(2, ConstantVector::get(Elts));
959 LHS = SVI.getOperand(0);
960 RHS = SVI.getOperand(1);
964 if (VWidth == LHSWidth) {
965 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
966 bool isLHSID, isRHSID;
967 recognizeIdentityMask(Mask, isLHSID, isRHSID);
969 // Eliminate identity shuffles.
970 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
971 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
974 if (isa<UndefValue>(RHS) && CanEvaluateShuffled(LHS, Mask)) {
975 Value *V = EvaluateInDifferentElementOrder(LHS, Mask);
976 return ReplaceInstUsesWith(SVI, V);
979 // SROA generates shuffle+bitcast when the extracted sub-vector is bitcast to
980 // a non-vector type. We can instead bitcast the original vector followed by
981 // an extract of the desired element:
983 // %sroa = shufflevector <16 x i8> %in, <16 x i8> undef,
984 // <4 x i32> <i32 0, i32 1, i32 2, i32 3>
985 // %1 = bitcast <4 x i8> %sroa to i32
987 // %bc = bitcast <16 x i8> %in to <4 x i32>
988 // %ext = extractelement <4 x i32> %bc, i32 0
990 // If the shuffle is extracting a contiguous range of values from the input
991 // vector then each use which is a bitcast of the extracted size can be
992 // replaced. This will work if the vector types are compatible, and the begin
993 // index is aligned to a value in the casted vector type. If the begin index
994 // isn't aligned then we can shuffle the original vector (keeping the same
995 // vector type) before extracting.
997 // This code will bail out if the target type is fundamentally incompatible
998 // with vectors of the source type.
1000 // Example of <16 x i8>, target type i32:
1001 // Index range [4,8): v-----------v Will work.
1002 // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
1003 // <16 x i8>: | | | | | | | | | | | | | | | | |
1004 // <4 x i32>: | | | | |
1005 // +-----------+-----------+-----------+-----------+
1006 // Index range [6,10): ^-----------^ Needs an extra shuffle.
1007 // Target type i40: ^--------------^ Won't work, bail.
1008 if (isShuffleExtractingFromLHS(SVI, Mask)) {
1010 unsigned MaskElems = Mask.size();
1011 unsigned BegIdx = Mask.front();
1012 VectorType *SrcTy = cast<VectorType>(V->getType());
1013 unsigned VecBitWidth = SrcTy->getBitWidth();
1014 unsigned SrcElemBitWidth = DL.getTypeSizeInBits(SrcTy->getElementType());
1015 assert(SrcElemBitWidth && "vector elements must have a bitwidth");
1016 unsigned SrcNumElems = SrcTy->getNumElements();
1017 SmallVector<BitCastInst *, 8> BCs;
1018 DenseMap<Type *, Value *> NewBCs;
1019 for (User *U : SVI.users())
1020 if (BitCastInst *BC = dyn_cast<BitCastInst>(U))
1021 if (!BC->use_empty())
1022 // Only visit bitcasts that weren't previously handled.
1024 for (BitCastInst *BC : BCs) {
1025 Type *TgtTy = BC->getDestTy();
1026 unsigned TgtElemBitWidth = DL.getTypeSizeInBits(TgtTy);
1027 if (!TgtElemBitWidth)
1029 unsigned TgtNumElems = VecBitWidth / TgtElemBitWidth;
1030 bool VecBitWidthsEqual = VecBitWidth == TgtNumElems * TgtElemBitWidth;
1031 bool BegIsAligned = 0 == ((SrcElemBitWidth * BegIdx) % TgtElemBitWidth);
1032 if (!VecBitWidthsEqual)
1034 if (!VectorType::isValidElementType(TgtTy))
1036 VectorType *CastSrcTy = VectorType::get(TgtTy, TgtNumElems);
1037 if (!BegIsAligned) {
1038 // Shuffle the input so [0,NumElements) contains the output, and
1039 // [NumElems,SrcNumElems) is undef.
1040 SmallVector<Constant *, 16> ShuffleMask(SrcNumElems,
1041 UndefValue::get(Int32Ty));
1042 for (unsigned I = 0, E = MaskElems, Idx = BegIdx; I != E; ++Idx, ++I)
1043 ShuffleMask[I] = ConstantInt::get(Int32Ty, Idx);
1044 V = Builder->CreateShuffleVector(V, UndefValue::get(V->getType()),
1045 ConstantVector::get(ShuffleMask),
1046 SVI.getName() + ".extract");
1049 unsigned SrcElemsPerTgtElem = TgtElemBitWidth / SrcElemBitWidth;
1050 assert(SrcElemsPerTgtElem);
1051 BegIdx /= SrcElemsPerTgtElem;
1052 bool BCAlreadyExists = NewBCs.find(CastSrcTy) != NewBCs.end();
1056 : Builder->CreateBitCast(V, CastSrcTy, SVI.getName() + ".bc");
1057 if (!BCAlreadyExists)
1058 NewBCs[CastSrcTy] = NewBC;
1059 auto *Ext = Builder->CreateExtractElement(
1060 NewBC, ConstantInt::get(Int32Ty, BegIdx), SVI.getName() + ".extract");
1061 // The shufflevector isn't being replaced: the bitcast that used it
1062 // is. InstCombine will visit the newly-created instructions.
1063 ReplaceInstUsesWith(*BC, Ext);
1068 // If the LHS is a shufflevector itself, see if we can combine it with this
1069 // one without producing an unusual shuffle.
1070 // Cases that might be simplified:
1072 // x1=shuffle(v1,v2,mask1)
1073 // x=shuffle(x1,undef,mask)
1075 // x=shuffle(v1,undef,newMask)
1076 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : -1
1078 // x1=shuffle(v1,undef,mask1)
1079 // x=shuffle(x1,x2,mask)
1080 // where v1.size() == mask1.size()
1082 // x=shuffle(v1,x2,newMask)
1083 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : mask[i]
1085 // x2=shuffle(v2,undef,mask2)
1086 // x=shuffle(x1,x2,mask)
1087 // where v2.size() == mask2.size()
1089 // x=shuffle(x1,v2,newMask)
1090 // newMask[i] = (mask[i] < x1.size())
1091 // ? mask[i] : mask2[mask[i]-x1.size()]+x1.size()
1093 // x1=shuffle(v1,undef,mask1)
1094 // x2=shuffle(v2,undef,mask2)
1095 // x=shuffle(x1,x2,mask)
1096 // where v1.size() == v2.size()
1098 // x=shuffle(v1,v2,newMask)
1099 // newMask[i] = (mask[i] < x1.size())
1100 // ? mask1[mask[i]] : mask2[mask[i]-x1.size()]+v1.size()
1102 // Here we are really conservative:
1103 // we are absolutely afraid of producing a shuffle mask not in the input
1104 // program, because the code gen may not be smart enough to turn a merged
1105 // shuffle into two specific shuffles: it may produce worse code. As such,
1106 // we only merge two shuffles if the result is either a splat or one of the
1107 // input shuffle masks. In this case, merging the shuffles just removes
1108 // one instruction, which we know is safe. This is good for things like
1109 // turning: (splat(splat)) -> splat, or
1110 // merge(V[0..n], V[n+1..2n]) -> V[0..2n]
1111 ShuffleVectorInst* LHSShuffle = dyn_cast<ShuffleVectorInst>(LHS);
1112 ShuffleVectorInst* RHSShuffle = dyn_cast<ShuffleVectorInst>(RHS);
1114 if (!isa<UndefValue>(LHSShuffle->getOperand(1)) && !isa<UndefValue>(RHS))
1115 LHSShuffle = nullptr;
1117 if (!isa<UndefValue>(RHSShuffle->getOperand(1)))
1118 RHSShuffle = nullptr;
1119 if (!LHSShuffle && !RHSShuffle)
1120 return MadeChange ? &SVI : nullptr;
1122 Value* LHSOp0 = nullptr;
1123 Value* LHSOp1 = nullptr;
1124 Value* RHSOp0 = nullptr;
1125 unsigned LHSOp0Width = 0;
1126 unsigned RHSOp0Width = 0;
1128 LHSOp0 = LHSShuffle->getOperand(0);
1129 LHSOp1 = LHSShuffle->getOperand(1);
1130 LHSOp0Width = cast<VectorType>(LHSOp0->getType())->getNumElements();
1133 RHSOp0 = RHSShuffle->getOperand(0);
1134 RHSOp0Width = cast<VectorType>(RHSOp0->getType())->getNumElements();
1136 Value* newLHS = LHS;
1137 Value* newRHS = RHS;
1140 if (isa<UndefValue>(RHS)) {
1145 else if (LHSOp0Width == LHSWidth) {
1150 if (RHSShuffle && RHSOp0Width == LHSWidth) {
1154 if (LHSOp0 == RHSOp0) {
1159 if (newLHS == LHS && newRHS == RHS)
1160 return MadeChange ? &SVI : nullptr;
1162 SmallVector<int, 16> LHSMask;
1163 SmallVector<int, 16> RHSMask;
1165 LHSMask = LHSShuffle->getShuffleMask();
1166 if (RHSShuffle && newRHS != RHS)
1167 RHSMask = RHSShuffle->getShuffleMask();
1169 unsigned newLHSWidth = (newLHS != LHS) ? LHSOp0Width : LHSWidth;
1170 SmallVector<int, 16> newMask;
1171 bool isSplat = true;
1173 // Create a new mask for the new ShuffleVectorInst so that the new
1174 // ShuffleVectorInst is equivalent to the original one.
1175 for (unsigned i = 0; i < VWidth; ++i) {
1178 // This element is an undef value.
1180 } else if (Mask[i] < (int)LHSWidth) {
1181 // This element is from left hand side vector operand.
1183 // If LHS is going to be replaced (case 1, 2, or 4), calculate the
1184 // new mask value for the element.
1185 if (newLHS != LHS) {
1186 eltMask = LHSMask[Mask[i]];
1187 // If the value selected is an undef value, explicitly specify it
1188 // with a -1 mask value.
1189 if (eltMask >= (int)LHSOp0Width && isa<UndefValue>(LHSOp1))
1194 // This element is from right hand side vector operand
1196 // If the value selected is an undef value, explicitly specify it
1197 // with a -1 mask value. (case 1)
1198 if (isa<UndefValue>(RHS))
1200 // If RHS is going to be replaced (case 3 or 4), calculate the
1201 // new mask value for the element.
1202 else if (newRHS != RHS) {
1203 eltMask = RHSMask[Mask[i]-LHSWidth];
1204 // If the value selected is an undef value, explicitly specify it
1205 // with a -1 mask value.
1206 if (eltMask >= (int)RHSOp0Width) {
1207 assert(isa<UndefValue>(RHSShuffle->getOperand(1))
1208 && "should have been check above");
1212 eltMask = Mask[i]-LHSWidth;
1214 // If LHS's width is changed, shift the mask value accordingly.
1215 // If newRHS == NULL, i.e. LHSOp0 == RHSOp0, we want to remap any
1216 // references from RHSOp0 to LHSOp0, so we don't need to shift the mask.
1217 // If newRHS == newLHS, we want to remap any references from newRHS to
1218 // newLHS so that we can properly identify splats that may occur due to
1219 // obfuscation across the two vectors.
1220 if (eltMask >= 0 && newRHS != nullptr && newLHS != newRHS)
1221 eltMask += newLHSWidth;
1224 // Check if this could still be a splat.
1226 if (SplatElt >= 0 && SplatElt != eltMask)
1231 newMask.push_back(eltMask);
1234 // If the result mask is equal to one of the original shuffle masks,
1235 // or is a splat, do the replacement.
1236 if (isSplat || newMask == LHSMask || newMask == RHSMask || newMask == Mask) {
1237 SmallVector<Constant*, 16> Elts;
1238 for (unsigned i = 0, e = newMask.size(); i != e; ++i) {
1239 if (newMask[i] < 0) {
1240 Elts.push_back(UndefValue::get(Int32Ty));
1242 Elts.push_back(ConstantInt::get(Int32Ty, newMask[i]));
1246 newRHS = UndefValue::get(newLHS->getType());
1247 return new ShuffleVectorInst(newLHS, newRHS, ConstantVector::get(Elts));
1250 // If the result mask is an identity, replace uses of this instruction with
1251 // corresponding argument.
1252 bool isLHSID, isRHSID;
1253 recognizeIdentityMask(newMask, isLHSID, isRHSID);
1254 if (isLHSID && VWidth == LHSOp0Width) return ReplaceInstUsesWith(SVI, newLHS);
1255 if (isRHSID && VWidth == RHSOp0Width) return ReplaceInstUsesWith(SVI, newRHS);
1257 return MadeChange ? &SVI : nullptr;