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 is only used to 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 SmallVector<Instruction *, 2> Extracts;
70 // The users we want the PHI to have are:
71 // 1) The EI ExtractElement (we already know this)
72 // 2) Possibly more ExtractElements with the same index.
73 // 3) Another operand, which will feed back into the PHI.
74 Instruction *PHIUser = nullptr;
75 for (auto U : PN->users()) {
76 if (ExtractElementInst *EU = dyn_cast<ExtractElementInst>(U)) {
77 if (EI.getIndexOperand() == EU->getIndexOperand())
78 Extracts.push_back(EU);
81 } else if (!PHIUser) {
82 PHIUser = cast<Instruction>(U);
91 // Verify that this PHI user has one use, which is the PHI itself,
92 // and that it is a binary operation which is cheap to scalarize.
93 // otherwise return NULL.
94 if (!PHIUser->hasOneUse() || !(PHIUser->user_back() == PN) ||
95 !(isa<BinaryOperator>(PHIUser)) || !cheapToScalarize(PHIUser, true))
98 // Create a scalar PHI node that will replace the vector PHI node
99 // just before the current PHI node.
100 PHINode *scalarPHI = cast<PHINode>(InsertNewInstWith(
101 PHINode::Create(EI.getType(), PN->getNumIncomingValues(), ""), *PN));
102 // Scalarize each PHI operand.
103 for (unsigned i = 0; i < PN->getNumIncomingValues(); i++) {
104 Value *PHIInVal = PN->getIncomingValue(i);
105 BasicBlock *inBB = PN->getIncomingBlock(i);
106 Value *Elt = EI.getIndexOperand();
107 // If the operand is the PHI induction variable:
108 if (PHIInVal == PHIUser) {
109 // Scalarize the binary operation. Its first operand is the
110 // scalar PHI, and the second operand is extracted from the other
112 BinaryOperator *B0 = cast<BinaryOperator>(PHIUser);
113 unsigned opId = (B0->getOperand(0) == PN) ? 1 : 0;
114 Value *Op = InsertNewInstWith(
115 ExtractElementInst::Create(B0->getOperand(opId), Elt,
116 B0->getOperand(opId)->getName() + ".Elt"),
118 Value *newPHIUser = InsertNewInstWith(
119 BinaryOperator::CreateWithCopiedFlags(B0->getOpcode(),
120 scalarPHI, Op, B0), *B0);
121 scalarPHI->addIncoming(newPHIUser, inBB);
123 // Scalarize PHI input:
124 Instruction *newEI = ExtractElementInst::Create(PHIInVal, Elt, "");
125 // Insert the new instruction into the predecessor basic block.
126 Instruction *pos = dyn_cast<Instruction>(PHIInVal);
127 BasicBlock::iterator InsertPos;
128 if (pos && !isa<PHINode>(pos)) {
129 InsertPos = ++pos->getIterator();
131 InsertPos = inBB->getFirstInsertionPt();
134 InsertNewInstWith(newEI, *InsertPos);
136 scalarPHI->addIncoming(newEI, inBB);
140 for (auto E : Extracts)
141 replaceInstUsesWith(*E, scalarPHI);
146 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
147 if (Value *V = SimplifyExtractElementInst(
148 EI.getVectorOperand(), EI.getIndexOperand(), DL, &TLI, &DT, &AC))
149 return replaceInstUsesWith(EI, V);
151 // If vector val is constant with all elements the same, replace EI with
152 // that element. We handle a known element # below.
153 if (Constant *C = dyn_cast<Constant>(EI.getOperand(0)))
154 if (cheapToScalarize(C, false))
155 return replaceInstUsesWith(EI, C->getAggregateElement(0U));
157 // If extracting a specified index from the vector, see if we can recursively
158 // find a previously computed scalar that was inserted into the vector.
159 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
160 unsigned IndexVal = IdxC->getZExtValue();
161 unsigned VectorWidth = EI.getVectorOperandType()->getNumElements();
163 // InstSimplify handles cases where the index is invalid.
164 assert(IndexVal < VectorWidth);
166 // This instruction only demands the single element from the input vector.
167 // If the input vector has a single use, simplify it based on this use
169 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
170 APInt UndefElts(VectorWidth, 0);
171 APInt DemandedMask(VectorWidth, 0);
172 DemandedMask.setBit(IndexVal);
173 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0), DemandedMask,
180 // If this extractelement is directly using a bitcast from a vector of
181 // the same number of elements, see if we can find the source element from
182 // it. In this case, we will end up needing to bitcast the scalars.
183 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
184 if (VectorType *VT = dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
185 if (VT->getNumElements() == VectorWidth)
186 if (Value *Elt = findScalarElement(BCI->getOperand(0), IndexVal))
187 return new BitCastInst(Elt, EI.getType());
190 // If there's a vector PHI feeding a scalar use through this extractelement
191 // instruction, try to scalarize the PHI.
192 if (PHINode *PN = dyn_cast<PHINode>(EI.getOperand(0))) {
193 Instruction *scalarPHI = scalarizePHI(EI, PN);
199 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
200 // Push extractelement into predecessor operation if legal and
201 // profitable to do so.
202 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
203 if (I->hasOneUse() &&
204 cheapToScalarize(BO, isa<ConstantInt>(EI.getOperand(1)))) {
206 Builder->CreateExtractElement(BO->getOperand(0), EI.getOperand(1),
207 EI.getName()+".lhs");
209 Builder->CreateExtractElement(BO->getOperand(1), EI.getOperand(1),
210 EI.getName()+".rhs");
211 return BinaryOperator::CreateWithCopiedFlags(BO->getOpcode(),
214 } else if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
215 // Extracting the inserted element?
216 if (IE->getOperand(2) == EI.getOperand(1))
217 return replaceInstUsesWith(EI, IE->getOperand(1));
218 // If the inserted and extracted elements are constants, they must not
219 // be the same value, extract from the pre-inserted value instead.
220 if (isa<Constant>(IE->getOperand(2)) && isa<Constant>(EI.getOperand(1))) {
221 Worklist.AddValue(EI.getOperand(0));
222 EI.setOperand(0, IE->getOperand(0));
225 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
226 // If this is extracting an element from a shufflevector, figure out where
227 // it came from and extract from the appropriate input element instead.
228 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
229 int SrcIdx = SVI->getMaskValue(Elt->getZExtValue());
232 SVI->getOperand(0)->getType()->getVectorNumElements();
235 return replaceInstUsesWith(EI, UndefValue::get(EI.getType()));
236 if (SrcIdx < (int)LHSWidth)
237 Src = SVI->getOperand(0);
240 Src = SVI->getOperand(1);
242 Type *Int32Ty = Type::getInt32Ty(EI.getContext());
243 return ExtractElementInst::Create(Src,
244 ConstantInt::get(Int32Ty,
247 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
248 // Canonicalize extractelement(cast) -> cast(extractelement).
249 // Bitcasts can change the number of vector elements, and they cost
251 if (CI->hasOneUse() && (CI->getOpcode() != Instruction::BitCast)) {
252 Value *EE = Builder->CreateExtractElement(CI->getOperand(0),
253 EI.getIndexOperand());
254 Worklist.AddValue(EE);
255 return CastInst::Create(CI->getOpcode(), EE, EI.getType());
257 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
258 if (SI->hasOneUse()) {
259 // TODO: For a select on vectors, it might be useful to do this if it
260 // has multiple extractelement uses. For vector select, that seems to
261 // fight the vectorizer.
263 // If we are extracting an element from a vector select or a select on
264 // vectors, create a select on the scalars extracted from the vector
266 Value *TrueVal = SI->getTrueValue();
267 Value *FalseVal = SI->getFalseValue();
269 Value *Cond = SI->getCondition();
270 if (Cond->getType()->isVectorTy()) {
271 Cond = Builder->CreateExtractElement(Cond,
272 EI.getIndexOperand(),
273 Cond->getName() + ".elt");
277 = Builder->CreateExtractElement(TrueVal,
278 EI.getIndexOperand(),
279 TrueVal->getName() + ".elt");
282 = Builder->CreateExtractElement(FalseVal,
283 EI.getIndexOperand(),
284 FalseVal->getName() + ".elt");
285 return SelectInst::Create(Cond,
288 SI->getName() + ".elt");
295 /// If V is a shuffle of values that ONLY returns elements from either LHS or
296 /// RHS, return the shuffle mask and true. Otherwise, return false.
297 static bool collectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
298 SmallVectorImpl<Constant*> &Mask) {
299 assert(LHS->getType() == RHS->getType() &&
300 "Invalid CollectSingleShuffleElements");
301 unsigned NumElts = V->getType()->getVectorNumElements();
303 if (isa<UndefValue>(V)) {
304 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
309 for (unsigned i = 0; i != NumElts; ++i)
310 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
315 for (unsigned i = 0; i != NumElts; ++i)
316 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()),
321 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
322 // If this is an insert of an extract from some other vector, include it.
323 Value *VecOp = IEI->getOperand(0);
324 Value *ScalarOp = IEI->getOperand(1);
325 Value *IdxOp = IEI->getOperand(2);
327 if (!isa<ConstantInt>(IdxOp))
329 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
331 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
332 // We can handle this if the vector we are inserting into is
334 if (collectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
335 // If so, update the mask to reflect the inserted undef.
336 Mask[InsertedIdx] = UndefValue::get(Type::getInt32Ty(V->getContext()));
339 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
340 if (isa<ConstantInt>(EI->getOperand(1))) {
341 unsigned ExtractedIdx =
342 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
343 unsigned NumLHSElts = LHS->getType()->getVectorNumElements();
345 // This must be extracting from either LHS or RHS.
346 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
347 // We can handle this if the vector we are inserting into is
349 if (collectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
350 // If so, update the mask to reflect the inserted value.
351 if (EI->getOperand(0) == LHS) {
352 Mask[InsertedIdx % NumElts] =
353 ConstantInt::get(Type::getInt32Ty(V->getContext()),
356 assert(EI->getOperand(0) == RHS);
357 Mask[InsertedIdx % NumElts] =
358 ConstantInt::get(Type::getInt32Ty(V->getContext()),
359 ExtractedIdx + NumLHSElts);
371 /// If we have insertion into a vector that is wider than the vector that we
372 /// are extracting from, try to widen the source vector to allow a single
373 /// shufflevector to replace one or more insert/extract pairs.
374 static void replaceExtractElements(InsertElementInst *InsElt,
375 ExtractElementInst *ExtElt,
377 VectorType *InsVecType = InsElt->getType();
378 VectorType *ExtVecType = ExtElt->getVectorOperandType();
379 unsigned NumInsElts = InsVecType->getVectorNumElements();
380 unsigned NumExtElts = ExtVecType->getVectorNumElements();
382 // The inserted-to vector must be wider than the extracted-from vector.
383 if (InsVecType->getElementType() != ExtVecType->getElementType() ||
384 NumExtElts >= NumInsElts)
387 // Create a shuffle mask to widen the extended-from vector using undefined
388 // values. The mask selects all of the values of the original vector followed
389 // by as many undefined values as needed to create a vector of the same length
390 // as the inserted-to vector.
391 SmallVector<Constant *, 16> ExtendMask;
392 IntegerType *IntType = Type::getInt32Ty(InsElt->getContext());
393 for (unsigned i = 0; i < NumExtElts; ++i)
394 ExtendMask.push_back(ConstantInt::get(IntType, i));
395 for (unsigned i = NumExtElts; i < NumInsElts; ++i)
396 ExtendMask.push_back(UndefValue::get(IntType));
398 Value *ExtVecOp = ExtElt->getVectorOperand();
399 auto *ExtVecOpInst = dyn_cast<Instruction>(ExtVecOp);
400 BasicBlock *InsertionBlock = (ExtVecOpInst && !isa<PHINode>(ExtVecOpInst))
401 ? ExtVecOpInst->getParent()
402 : ExtElt->getParent();
404 // TODO: This restriction matches the basic block check below when creating
405 // new extractelement instructions. If that limitation is removed, this one
406 // could also be removed. But for now, we just bail out to ensure that we
407 // will replace the extractelement instruction that is feeding our
408 // insertelement instruction. This allows the insertelement to then be
409 // replaced by a shufflevector. If the insertelement is not replaced, we can
410 // induce infinite looping because there's an optimization for extractelement
411 // that will delete our widening shuffle. This would trigger another attempt
412 // here to create that shuffle, and we spin forever.
413 if (InsertionBlock != InsElt->getParent())
416 // TODO: This restriction matches the check in visitInsertElementInst() and
417 // prevents an infinite loop caused by not turning the extract/insert pair
418 // into a shuffle. We really should not need either check, but we're lacking
419 // folds for shufflevectors because we're afraid to generate shuffle masks
420 // that the backend can't handle.
421 if (InsElt->hasOneUse() && isa<InsertElementInst>(InsElt->user_back()))
424 auto *WideVec = new ShuffleVectorInst(ExtVecOp, UndefValue::get(ExtVecType),
425 ConstantVector::get(ExtendMask));
427 // Insert the new shuffle after the vector operand of the extract is defined
428 // (as long as it's not a PHI) or at the start of the basic block of the
429 // extract, so any subsequent extracts in the same basic block can use it.
430 // TODO: Insert before the earliest ExtractElementInst that is replaced.
431 if (ExtVecOpInst && !isa<PHINode>(ExtVecOpInst))
432 WideVec->insertAfter(ExtVecOpInst);
434 IC.InsertNewInstWith(WideVec, *ExtElt->getParent()->getFirstInsertionPt());
436 // Replace extracts from the original narrow vector with extracts from the new
438 for (User *U : ExtVecOp->users()) {
439 ExtractElementInst *OldExt = dyn_cast<ExtractElementInst>(U);
440 if (!OldExt || OldExt->getParent() != WideVec->getParent())
442 auto *NewExt = ExtractElementInst::Create(WideVec, OldExt->getOperand(1));
443 NewExt->insertAfter(WideVec);
444 IC.replaceInstUsesWith(*OldExt, NewExt);
448 /// We are building a shuffle to create V, which is a sequence of insertelement,
449 /// extractelement pairs. If PermittedRHS is set, then we must either use it or
450 /// not rely on the second vector source. Return a std::pair containing the
451 /// left and right vectors of the proposed shuffle (or 0), and set the Mask
452 /// parameter as required.
454 /// Note: we intentionally don't try to fold earlier shuffles since they have
455 /// often been chosen carefully to be efficiently implementable on the target.
456 typedef std::pair<Value *, Value *> ShuffleOps;
458 static ShuffleOps collectShuffleElements(Value *V,
459 SmallVectorImpl<Constant *> &Mask,
462 assert(V->getType()->isVectorTy() && "Invalid shuffle!");
463 unsigned NumElts = V->getType()->getVectorNumElements();
465 if (isa<UndefValue>(V)) {
466 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
467 return std::make_pair(
468 PermittedRHS ? UndefValue::get(PermittedRHS->getType()) : V, nullptr);
471 if (isa<ConstantAggregateZero>(V)) {
472 Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(V->getContext()),0));
473 return std::make_pair(V, nullptr);
476 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
477 // If this is an insert of an extract from some other vector, include it.
478 Value *VecOp = IEI->getOperand(0);
479 Value *ScalarOp = IEI->getOperand(1);
480 Value *IdxOp = IEI->getOperand(2);
482 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
483 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) {
484 unsigned ExtractedIdx =
485 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
486 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
488 // Either the extracted from or inserted into vector must be RHSVec,
489 // otherwise we'd end up with a shuffle of three inputs.
490 if (EI->getOperand(0) == PermittedRHS || PermittedRHS == nullptr) {
491 Value *RHS = EI->getOperand(0);
492 ShuffleOps LR = collectShuffleElements(VecOp, Mask, RHS, IC);
493 assert(LR.second == nullptr || LR.second == RHS);
495 if (LR.first->getType() != RHS->getType()) {
496 // Although we are giving up for now, see if we can create extracts
497 // that match the inserts for another round of combining.
498 replaceExtractElements(IEI, EI, IC);
500 // We tried our best, but we can't find anything compatible with RHS
501 // further up the chain. Return a trivial shuffle.
502 for (unsigned i = 0; i < NumElts; ++i)
503 Mask[i] = ConstantInt::get(Type::getInt32Ty(V->getContext()), i);
504 return std::make_pair(V, nullptr);
507 unsigned NumLHSElts = RHS->getType()->getVectorNumElements();
508 Mask[InsertedIdx % NumElts] =
509 ConstantInt::get(Type::getInt32Ty(V->getContext()),
510 NumLHSElts+ExtractedIdx);
511 return std::make_pair(LR.first, RHS);
514 if (VecOp == PermittedRHS) {
515 // We've gone as far as we can: anything on the other side of the
516 // extractelement will already have been converted into a shuffle.
517 unsigned NumLHSElts =
518 EI->getOperand(0)->getType()->getVectorNumElements();
519 for (unsigned i = 0; i != NumElts; ++i)
520 Mask.push_back(ConstantInt::get(
521 Type::getInt32Ty(V->getContext()),
522 i == InsertedIdx ? ExtractedIdx : NumLHSElts + i));
523 return std::make_pair(EI->getOperand(0), PermittedRHS);
526 // If this insertelement is a chain that comes from exactly these two
527 // vectors, return the vector and the effective shuffle.
528 if (EI->getOperand(0)->getType() == PermittedRHS->getType() &&
529 collectSingleShuffleElements(IEI, EI->getOperand(0), PermittedRHS,
531 return std::make_pair(EI->getOperand(0), PermittedRHS);
536 // Otherwise, we can't do anything fancy. Return an identity vector.
537 for (unsigned i = 0; i != NumElts; ++i)
538 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
539 return std::make_pair(V, nullptr);
542 /// Try to find redundant insertvalue instructions, like the following ones:
543 /// %0 = insertvalue { i8, i32 } undef, i8 %x, 0
544 /// %1 = insertvalue { i8, i32 } %0, i8 %y, 0
545 /// Here the second instruction inserts values at the same indices, as the
546 /// first one, making the first one redundant.
547 /// It should be transformed to:
548 /// %0 = insertvalue { i8, i32 } undef, i8 %y, 0
549 Instruction *InstCombiner::visitInsertValueInst(InsertValueInst &I) {
550 bool IsRedundant = false;
551 ArrayRef<unsigned int> FirstIndices = I.getIndices();
553 // If there is a chain of insertvalue instructions (each of them except the
554 // last one has only one use and it's another insertvalue insn from this
555 // chain), check if any of the 'children' uses the same indices as the first
556 // instruction. In this case, the first one is redundant.
559 while (V->hasOneUse() && Depth < 10) {
560 User *U = V->user_back();
561 auto UserInsInst = dyn_cast<InsertValueInst>(U);
562 if (!UserInsInst || U->getOperand(0) != V)
564 if (UserInsInst->getIndices() == FirstIndices) {
573 return replaceInstUsesWith(I, I.getOperand(0));
577 static bool isShuffleEquivalentToSelect(ShuffleVectorInst &Shuf) {
578 int MaskSize = Shuf.getMask()->getType()->getVectorNumElements();
579 int VecSize = Shuf.getOperand(0)->getType()->getVectorNumElements();
581 // A vector select does not change the size of the operands.
582 if (MaskSize != VecSize)
585 // Each mask element must be undefined or choose a vector element from one of
586 // the source operands without crossing vector lanes.
587 for (int i = 0; i != MaskSize; ++i) {
588 int Elt = Shuf.getMaskValue(i);
589 if (Elt != -1 && Elt != i && Elt != i + VecSize)
596 // Turn a chain of inserts that splats a value into a canonical insert + shuffle
598 // insertelt(insertelt(insertelt(insertelt X, %k, 0), %k, 1), %k, 2) ... ->
599 // shufflevector(insertelt(X, %k, 0), undef, zero)
600 static Instruction *foldInsSequenceIntoBroadcast(InsertElementInst &InsElt) {
601 // We are interested in the last insert in a chain. So, if this insert
602 // has a single user, and that user is an insert, bail.
603 if (InsElt.hasOneUse() && isa<InsertElementInst>(InsElt.user_back()))
606 VectorType *VT = cast<VectorType>(InsElt.getType());
607 int NumElements = VT->getNumElements();
609 // Do not try to do this for a one-element vector, since that's a nop,
610 // and will cause an inf-loop.
611 if (NumElements == 1)
614 Value *SplatVal = InsElt.getOperand(1);
615 InsertElementInst *CurrIE = &InsElt;
616 SmallVector<bool, 16> ElementPresent(NumElements, false);
618 // Walk the chain backwards, keeping track of which indices we inserted into,
619 // until we hit something that isn't an insert of the splatted value.
621 ConstantInt *Idx = dyn_cast<ConstantInt>(CurrIE->getOperand(2));
622 if (!Idx || CurrIE->getOperand(1) != SplatVal)
625 // Check none of the intermediate steps have any additional uses.
626 if ((CurrIE != &InsElt) && !CurrIE->hasOneUse())
629 ElementPresent[Idx->getZExtValue()] = true;
630 CurrIE = dyn_cast<InsertElementInst>(CurrIE->getOperand(0));
633 // Make sure we've seen an insert into every element.
634 if (llvm::any_of(ElementPresent, [](bool Present) { return !Present; }))
637 // All right, create the insert + shuffle.
638 Instruction *InsertFirst = InsertElementInst::Create(
639 UndefValue::get(VT), SplatVal,
640 ConstantInt::get(Type::getInt32Ty(InsElt.getContext()), 0), "", &InsElt);
642 Constant *ZeroMask = ConstantAggregateZero::get(
643 VectorType::get(Type::getInt32Ty(InsElt.getContext()), NumElements));
645 return new ShuffleVectorInst(InsertFirst, UndefValue::get(VT), ZeroMask);
648 /// insertelt (shufflevector X, CVec, Mask|insertelt X, C1, CIndex1), C, CIndex
649 /// --> shufflevector X, CVec', Mask'
650 static Instruction *foldConstantInsEltIntoShuffle(InsertElementInst &InsElt) {
651 auto *Inst = dyn_cast<Instruction>(InsElt.getOperand(0));
652 // Bail out if the parent has more than one use. In that case, we'd be
653 // replacing the insertelt with a shuffle, and that's not a clear win.
654 if (!Inst || !Inst->hasOneUse())
656 if (auto *Shuf = dyn_cast<ShuffleVectorInst>(InsElt.getOperand(0))) {
657 // The shuffle must have a constant vector operand. The insertelt must have
658 // a constant scalar being inserted at a constant position in the vector.
659 Constant *ShufConstVec, *InsEltScalar;
660 uint64_t InsEltIndex;
661 if (!match(Shuf->getOperand(1), m_Constant(ShufConstVec)) ||
662 !match(InsElt.getOperand(1), m_Constant(InsEltScalar)) ||
663 !match(InsElt.getOperand(2), m_ConstantInt(InsEltIndex)))
666 // Adding an element to an arbitrary shuffle could be expensive, but a
667 // shuffle that selects elements from vectors without crossing lanes is
669 // If we're just adding a constant into that shuffle, it will still be
671 if (!isShuffleEquivalentToSelect(*Shuf))
674 // From the above 'select' check, we know that the mask has the same number
675 // of elements as the vector input operands. We also know that each constant
676 // input element is used in its lane and can not be used more than once by
677 // the shuffle. Therefore, replace the constant in the shuffle's constant
678 // vector with the insertelt constant. Replace the constant in the shuffle's
679 // mask vector with the insertelt index plus the length of the vector
680 // (because the constant vector operand of a shuffle is always the 2nd
682 Constant *Mask = Shuf->getMask();
683 unsigned NumElts = Mask->getType()->getVectorNumElements();
684 SmallVector<Constant *, 16> NewShufElts(NumElts);
685 SmallVector<Constant *, 16> NewMaskElts(NumElts);
686 for (unsigned I = 0; I != NumElts; ++I) {
687 if (I == InsEltIndex) {
688 NewShufElts[I] = InsEltScalar;
689 Type *Int32Ty = Type::getInt32Ty(Shuf->getContext());
690 NewMaskElts[I] = ConstantInt::get(Int32Ty, InsEltIndex + NumElts);
692 // Copy over the existing values.
693 NewShufElts[I] = ShufConstVec->getAggregateElement(I);
694 NewMaskElts[I] = Mask->getAggregateElement(I);
698 // Create new operands for a shuffle that includes the constant of the
699 // original insertelt. The old shuffle will be dead now.
700 return new ShuffleVectorInst(Shuf->getOperand(0),
701 ConstantVector::get(NewShufElts),
702 ConstantVector::get(NewMaskElts));
703 } else if (auto *IEI = dyn_cast<InsertElementInst>(Inst)) {
704 // Transform sequences of insertelements ops with constant data/indexes into
705 // a single shuffle op.
706 unsigned NumElts = InsElt.getType()->getNumElements();
708 uint64_t InsertIdx[2];
710 if (!match(InsElt.getOperand(2), m_ConstantInt(InsertIdx[0])) ||
711 !match(InsElt.getOperand(1), m_Constant(Val[0])) ||
712 !match(IEI->getOperand(2), m_ConstantInt(InsertIdx[1])) ||
713 !match(IEI->getOperand(1), m_Constant(Val[1])))
715 SmallVector<Constant *, 16> Values(NumElts);
716 SmallVector<Constant *, 16> Mask(NumElts);
717 auto ValI = std::begin(Val);
718 // Generate new constant vector and mask.
719 // We have 2 values/masks from the insertelements instructions. Insert them
720 // into new value/mask vectors.
721 for (uint64_t I : InsertIdx) {
725 Mask[I] = ConstantInt::get(Type::getInt32Ty(InsElt.getContext()),
730 // Remaining values are filled with 'undef' values.
731 for (unsigned I = 0; I < NumElts; ++I) {
734 Values[I] = UndefValue::get(InsElt.getType()->getElementType());
735 Mask[I] = ConstantInt::get(Type::getInt32Ty(InsElt.getContext()), I);
738 // Create new operands for a shuffle that includes the constant of the
739 // original insertelt.
740 return new ShuffleVectorInst(IEI->getOperand(0),
741 ConstantVector::get(Values),
742 ConstantVector::get(Mask));
747 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
748 Value *VecOp = IE.getOperand(0);
749 Value *ScalarOp = IE.getOperand(1);
750 Value *IdxOp = IE.getOperand(2);
752 // Inserting an undef or into an undefined place, remove this.
753 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
754 replaceInstUsesWith(IE, VecOp);
756 // If the inserted element was extracted from some other vector, and if the
757 // indexes are constant, try to turn this into a shufflevector operation.
758 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
759 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) {
760 unsigned NumInsertVectorElts = IE.getType()->getNumElements();
761 unsigned NumExtractVectorElts =
762 EI->getOperand(0)->getType()->getVectorNumElements();
763 unsigned ExtractedIdx =
764 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
765 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
767 if (ExtractedIdx >= NumExtractVectorElts) // Out of range extract.
768 return replaceInstUsesWith(IE, VecOp);
770 if (InsertedIdx >= NumInsertVectorElts) // Out of range insert.
771 return replaceInstUsesWith(IE, UndefValue::get(IE.getType()));
773 // If we are extracting a value from a vector, then inserting it right
774 // back into the same place, just use the input vector.
775 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
776 return replaceInstUsesWith(IE, VecOp);
778 // If this insertelement isn't used by some other insertelement, turn it
779 // (and any insertelements it points to), into one big shuffle.
780 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.user_back())) {
781 SmallVector<Constant*, 16> Mask;
782 ShuffleOps LR = collectShuffleElements(&IE, Mask, nullptr, *this);
784 // The proposed shuffle may be trivial, in which case we shouldn't
785 // perform the combine.
786 if (LR.first != &IE && LR.second != &IE) {
787 // We now have a shuffle of LHS, RHS, Mask.
788 if (LR.second == nullptr)
789 LR.second = UndefValue::get(LR.first->getType());
790 return new ShuffleVectorInst(LR.first, LR.second,
791 ConstantVector::get(Mask));
797 unsigned VWidth = VecOp->getType()->getVectorNumElements();
798 APInt UndefElts(VWidth, 0);
799 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
800 if (Value *V = SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts)) {
802 return replaceInstUsesWith(IE, V);
806 if (Instruction *Shuf = foldConstantInsEltIntoShuffle(IE))
809 // Turn a sequence of inserts that broadcasts a scalar into a single
810 // insert + shufflevector.
811 if (Instruction *Broadcast = foldInsSequenceIntoBroadcast(IE))
817 /// Return true if we can evaluate the specified expression tree if the vector
818 /// elements were shuffled in a different order.
819 static bool CanEvaluateShuffled(Value *V, ArrayRef<int> Mask,
820 unsigned Depth = 5) {
821 // We can always reorder the elements of a constant.
822 if (isa<Constant>(V))
825 // We won't reorder vector arguments. No IPO here.
826 Instruction *I = dyn_cast<Instruction>(V);
827 if (!I) return false;
829 // Two users may expect different orders of the elements. Don't try it.
833 if (Depth == 0) return false;
835 switch (I->getOpcode()) {
836 case Instruction::Add:
837 case Instruction::FAdd:
838 case Instruction::Sub:
839 case Instruction::FSub:
840 case Instruction::Mul:
841 case Instruction::FMul:
842 case Instruction::UDiv:
843 case Instruction::SDiv:
844 case Instruction::FDiv:
845 case Instruction::URem:
846 case Instruction::SRem:
847 case Instruction::FRem:
848 case Instruction::Shl:
849 case Instruction::LShr:
850 case Instruction::AShr:
851 case Instruction::And:
852 case Instruction::Or:
853 case Instruction::Xor:
854 case Instruction::ICmp:
855 case Instruction::FCmp:
856 case Instruction::Trunc:
857 case Instruction::ZExt:
858 case Instruction::SExt:
859 case Instruction::FPToUI:
860 case Instruction::FPToSI:
861 case Instruction::UIToFP:
862 case Instruction::SIToFP:
863 case Instruction::FPTrunc:
864 case Instruction::FPExt:
865 case Instruction::GetElementPtr: {
866 for (Value *Operand : I->operands()) {
867 if (!CanEvaluateShuffled(Operand, Mask, Depth-1))
872 case Instruction::InsertElement: {
873 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(2));
874 if (!CI) return false;
875 int ElementNumber = CI->getLimitedValue();
877 // Verify that 'CI' does not occur twice in Mask. A single 'insertelement'
878 // can't put an element into multiple indices.
879 bool SeenOnce = false;
880 for (int i = 0, e = Mask.size(); i != e; ++i) {
881 if (Mask[i] == ElementNumber) {
887 return CanEvaluateShuffled(I->getOperand(0), Mask, Depth-1);
893 /// Rebuild a new instruction just like 'I' but with the new operands given.
894 /// In the event of type mismatch, the type of the operands is correct.
895 static Value *buildNew(Instruction *I, ArrayRef<Value*> NewOps) {
896 // We don't want to use the IRBuilder here because we want the replacement
897 // instructions to appear next to 'I', not the builder's insertion point.
898 switch (I->getOpcode()) {
899 case Instruction::Add:
900 case Instruction::FAdd:
901 case Instruction::Sub:
902 case Instruction::FSub:
903 case Instruction::Mul:
904 case Instruction::FMul:
905 case Instruction::UDiv:
906 case Instruction::SDiv:
907 case Instruction::FDiv:
908 case Instruction::URem:
909 case Instruction::SRem:
910 case Instruction::FRem:
911 case Instruction::Shl:
912 case Instruction::LShr:
913 case Instruction::AShr:
914 case Instruction::And:
915 case Instruction::Or:
916 case Instruction::Xor: {
917 BinaryOperator *BO = cast<BinaryOperator>(I);
918 assert(NewOps.size() == 2 && "binary operator with #ops != 2");
919 BinaryOperator *New =
920 BinaryOperator::Create(cast<BinaryOperator>(I)->getOpcode(),
921 NewOps[0], NewOps[1], "", BO);
922 if (isa<OverflowingBinaryOperator>(BO)) {
923 New->setHasNoUnsignedWrap(BO->hasNoUnsignedWrap());
924 New->setHasNoSignedWrap(BO->hasNoSignedWrap());
926 if (isa<PossiblyExactOperator>(BO)) {
927 New->setIsExact(BO->isExact());
929 if (isa<FPMathOperator>(BO))
930 New->copyFastMathFlags(I);
933 case Instruction::ICmp:
934 assert(NewOps.size() == 2 && "icmp with #ops != 2");
935 return new ICmpInst(I, cast<ICmpInst>(I)->getPredicate(),
936 NewOps[0], NewOps[1]);
937 case Instruction::FCmp:
938 assert(NewOps.size() == 2 && "fcmp with #ops != 2");
939 return new FCmpInst(I, cast<FCmpInst>(I)->getPredicate(),
940 NewOps[0], NewOps[1]);
941 case Instruction::Trunc:
942 case Instruction::ZExt:
943 case Instruction::SExt:
944 case Instruction::FPToUI:
945 case Instruction::FPToSI:
946 case Instruction::UIToFP:
947 case Instruction::SIToFP:
948 case Instruction::FPTrunc:
949 case Instruction::FPExt: {
950 // It's possible that the mask has a different number of elements from
951 // the original cast. We recompute the destination type to match the mask.
953 VectorType::get(I->getType()->getScalarType(),
954 NewOps[0]->getType()->getVectorNumElements());
955 assert(NewOps.size() == 1 && "cast with #ops != 1");
956 return CastInst::Create(cast<CastInst>(I)->getOpcode(), NewOps[0], DestTy,
959 case Instruction::GetElementPtr: {
960 Value *Ptr = NewOps[0];
961 ArrayRef<Value*> Idx = NewOps.slice(1);
962 GetElementPtrInst *GEP = GetElementPtrInst::Create(
963 cast<GetElementPtrInst>(I)->getSourceElementType(), Ptr, Idx, "", I);
964 GEP->setIsInBounds(cast<GetElementPtrInst>(I)->isInBounds());
968 llvm_unreachable("failed to rebuild vector instructions");
972 InstCombiner::EvaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask) {
973 // Mask.size() does not need to be equal to the number of vector elements.
975 assert(V->getType()->isVectorTy() && "can't reorder non-vector elements");
976 if (isa<UndefValue>(V)) {
977 return UndefValue::get(VectorType::get(V->getType()->getScalarType(),
980 if (isa<ConstantAggregateZero>(V)) {
981 return ConstantAggregateZero::get(
982 VectorType::get(V->getType()->getScalarType(),
985 if (Constant *C = dyn_cast<Constant>(V)) {
986 SmallVector<Constant *, 16> MaskValues;
987 for (int i = 0, e = Mask.size(); i != e; ++i) {
989 MaskValues.push_back(UndefValue::get(Builder->getInt32Ty()));
991 MaskValues.push_back(Builder->getInt32(Mask[i]));
993 return ConstantExpr::getShuffleVector(C, UndefValue::get(C->getType()),
994 ConstantVector::get(MaskValues));
997 Instruction *I = cast<Instruction>(V);
998 switch (I->getOpcode()) {
999 case Instruction::Add:
1000 case Instruction::FAdd:
1001 case Instruction::Sub:
1002 case Instruction::FSub:
1003 case Instruction::Mul:
1004 case Instruction::FMul:
1005 case Instruction::UDiv:
1006 case Instruction::SDiv:
1007 case Instruction::FDiv:
1008 case Instruction::URem:
1009 case Instruction::SRem:
1010 case Instruction::FRem:
1011 case Instruction::Shl:
1012 case Instruction::LShr:
1013 case Instruction::AShr:
1014 case Instruction::And:
1015 case Instruction::Or:
1016 case Instruction::Xor:
1017 case Instruction::ICmp:
1018 case Instruction::FCmp:
1019 case Instruction::Trunc:
1020 case Instruction::ZExt:
1021 case Instruction::SExt:
1022 case Instruction::FPToUI:
1023 case Instruction::FPToSI:
1024 case Instruction::UIToFP:
1025 case Instruction::SIToFP:
1026 case Instruction::FPTrunc:
1027 case Instruction::FPExt:
1028 case Instruction::Select:
1029 case Instruction::GetElementPtr: {
1030 SmallVector<Value*, 8> NewOps;
1031 bool NeedsRebuild = (Mask.size() != I->getType()->getVectorNumElements());
1032 for (int i = 0, e = I->getNumOperands(); i != e; ++i) {
1033 Value *V = EvaluateInDifferentElementOrder(I->getOperand(i), Mask);
1034 NewOps.push_back(V);
1035 NeedsRebuild |= (V != I->getOperand(i));
1038 return buildNew(I, NewOps);
1042 case Instruction::InsertElement: {
1043 int Element = cast<ConstantInt>(I->getOperand(2))->getLimitedValue();
1045 // The insertelement was inserting at Element. Figure out which element
1046 // that becomes after shuffling. The answer is guaranteed to be unique
1047 // by CanEvaluateShuffled.
1050 for (int e = Mask.size(); Index != e; ++Index) {
1051 if (Mask[Index] == Element) {
1057 // If element is not in Mask, no need to handle the operand 1 (element to
1058 // be inserted). Just evaluate values in operand 0 according to Mask.
1060 return EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
1062 Value *V = EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
1063 return InsertElementInst::Create(V, I->getOperand(1),
1064 Builder->getInt32(Index), "", I);
1067 llvm_unreachable("failed to reorder elements of vector instruction!");
1070 static void recognizeIdentityMask(const SmallVectorImpl<int> &Mask,
1071 bool &isLHSID, bool &isRHSID) {
1072 isLHSID = isRHSID = true;
1074 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
1075 if (Mask[i] < 0) continue; // Ignore undef values.
1076 // Is this an identity shuffle of the LHS value?
1077 isLHSID &= (Mask[i] == (int)i);
1079 // Is this an identity shuffle of the RHS value?
1080 isRHSID &= (Mask[i]-e == i);
1084 // Returns true if the shuffle is extracting a contiguous range of values from
1085 // LHS, for example:
1086 // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
1087 // Input: |AA|BB|CC|DD|EE|FF|GG|HH|II|JJ|KK|LL|MM|NN|OO|PP|
1088 // Shuffles to: |EE|FF|GG|HH|
1090 static bool isShuffleExtractingFromLHS(ShuffleVectorInst &SVI,
1091 SmallVector<int, 16> &Mask) {
1092 unsigned LHSElems = SVI.getOperand(0)->getType()->getVectorNumElements();
1093 unsigned MaskElems = Mask.size();
1094 unsigned BegIdx = Mask.front();
1095 unsigned EndIdx = Mask.back();
1096 if (BegIdx > EndIdx || EndIdx >= LHSElems || EndIdx - BegIdx != MaskElems - 1)
1098 for (unsigned I = 0; I != MaskElems; ++I)
1099 if (static_cast<unsigned>(Mask[I]) != BegIdx + I)
1104 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
1105 Value *LHS = SVI.getOperand(0);
1106 Value *RHS = SVI.getOperand(1);
1107 SmallVector<int, 16> Mask = SVI.getShuffleMask();
1108 Type *Int32Ty = Type::getInt32Ty(SVI.getContext());
1110 bool MadeChange = false;
1112 // Undefined shuffle mask -> undefined value.
1113 if (isa<UndefValue>(SVI.getOperand(2)))
1114 return replaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
1116 unsigned VWidth = SVI.getType()->getVectorNumElements();
1118 APInt UndefElts(VWidth, 0);
1119 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
1120 if (Value *V = SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
1122 return replaceInstUsesWith(SVI, V);
1123 LHS = SVI.getOperand(0);
1124 RHS = SVI.getOperand(1);
1128 unsigned LHSWidth = LHS->getType()->getVectorNumElements();
1130 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
1131 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
1132 if (LHS == RHS || isa<UndefValue>(LHS)) {
1133 if (isa<UndefValue>(LHS) && LHS == RHS) {
1134 // shuffle(undef,undef,mask) -> undef.
1135 Value *Result = (VWidth == LHSWidth)
1136 ? LHS : UndefValue::get(SVI.getType());
1137 return replaceInstUsesWith(SVI, Result);
1140 // Remap any references to RHS to use LHS.
1141 SmallVector<Constant*, 16> Elts;
1142 for (unsigned i = 0, e = LHSWidth; i != VWidth; ++i) {
1144 Elts.push_back(UndefValue::get(Int32Ty));
1148 if ((Mask[i] >= (int)e && isa<UndefValue>(RHS)) ||
1149 (Mask[i] < (int)e && isa<UndefValue>(LHS))) {
1150 Mask[i] = -1; // Turn into undef.
1151 Elts.push_back(UndefValue::get(Int32Ty));
1153 Mask[i] = Mask[i] % e; // Force to LHS.
1154 Elts.push_back(ConstantInt::get(Int32Ty, Mask[i]));
1157 SVI.setOperand(0, SVI.getOperand(1));
1158 SVI.setOperand(1, UndefValue::get(RHS->getType()));
1159 SVI.setOperand(2, ConstantVector::get(Elts));
1160 LHS = SVI.getOperand(0);
1161 RHS = SVI.getOperand(1);
1165 if (VWidth == LHSWidth) {
1166 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
1167 bool isLHSID, isRHSID;
1168 recognizeIdentityMask(Mask, isLHSID, isRHSID);
1170 // Eliminate identity shuffles.
1171 if (isLHSID) return replaceInstUsesWith(SVI, LHS);
1172 if (isRHSID) return replaceInstUsesWith(SVI, RHS);
1175 if (isa<UndefValue>(RHS) && CanEvaluateShuffled(LHS, Mask)) {
1176 Value *V = EvaluateInDifferentElementOrder(LHS, Mask);
1177 return replaceInstUsesWith(SVI, V);
1180 // SROA generates shuffle+bitcast when the extracted sub-vector is bitcast to
1181 // a non-vector type. We can instead bitcast the original vector followed by
1182 // an extract of the desired element:
1184 // %sroa = shufflevector <16 x i8> %in, <16 x i8> undef,
1185 // <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1186 // %1 = bitcast <4 x i8> %sroa to i32
1188 // %bc = bitcast <16 x i8> %in to <4 x i32>
1189 // %ext = extractelement <4 x i32> %bc, i32 0
1191 // If the shuffle is extracting a contiguous range of values from the input
1192 // vector then each use which is a bitcast of the extracted size can be
1193 // replaced. This will work if the vector types are compatible, and the begin
1194 // index is aligned to a value in the casted vector type. If the begin index
1195 // isn't aligned then we can shuffle the original vector (keeping the same
1196 // vector type) before extracting.
1198 // This code will bail out if the target type is fundamentally incompatible
1199 // with vectors of the source type.
1201 // Example of <16 x i8>, target type i32:
1202 // Index range [4,8): v-----------v Will work.
1203 // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
1204 // <16 x i8>: | | | | | | | | | | | | | | | | |
1205 // <4 x i32>: | | | | |
1206 // +-----------+-----------+-----------+-----------+
1207 // Index range [6,10): ^-----------^ Needs an extra shuffle.
1208 // Target type i40: ^--------------^ Won't work, bail.
1209 if (isShuffleExtractingFromLHS(SVI, Mask)) {
1211 unsigned MaskElems = Mask.size();
1212 unsigned BegIdx = Mask.front();
1213 VectorType *SrcTy = cast<VectorType>(V->getType());
1214 unsigned VecBitWidth = SrcTy->getBitWidth();
1215 unsigned SrcElemBitWidth = DL.getTypeSizeInBits(SrcTy->getElementType());
1216 assert(SrcElemBitWidth && "vector elements must have a bitwidth");
1217 unsigned SrcNumElems = SrcTy->getNumElements();
1218 SmallVector<BitCastInst *, 8> BCs;
1219 DenseMap<Type *, Value *> NewBCs;
1220 for (User *U : SVI.users())
1221 if (BitCastInst *BC = dyn_cast<BitCastInst>(U))
1222 if (!BC->use_empty())
1223 // Only visit bitcasts that weren't previously handled.
1225 for (BitCastInst *BC : BCs) {
1226 Type *TgtTy = BC->getDestTy();
1227 unsigned TgtElemBitWidth = DL.getTypeSizeInBits(TgtTy);
1228 if (!TgtElemBitWidth)
1230 unsigned TgtNumElems = VecBitWidth / TgtElemBitWidth;
1231 bool VecBitWidthsEqual = VecBitWidth == TgtNumElems * TgtElemBitWidth;
1232 bool BegIsAligned = 0 == ((SrcElemBitWidth * BegIdx) % TgtElemBitWidth);
1233 if (!VecBitWidthsEqual)
1235 if (!VectorType::isValidElementType(TgtTy))
1237 VectorType *CastSrcTy = VectorType::get(TgtTy, TgtNumElems);
1238 if (!BegIsAligned) {
1239 // Shuffle the input so [0,NumElements) contains the output, and
1240 // [NumElems,SrcNumElems) is undef.
1241 SmallVector<Constant *, 16> ShuffleMask(SrcNumElems,
1242 UndefValue::get(Int32Ty));
1243 for (unsigned I = 0, E = MaskElems, Idx = BegIdx; I != E; ++Idx, ++I)
1244 ShuffleMask[I] = ConstantInt::get(Int32Ty, Idx);
1245 V = Builder->CreateShuffleVector(V, UndefValue::get(V->getType()),
1246 ConstantVector::get(ShuffleMask),
1247 SVI.getName() + ".extract");
1250 unsigned SrcElemsPerTgtElem = TgtElemBitWidth / SrcElemBitWidth;
1251 assert(SrcElemsPerTgtElem);
1252 BegIdx /= SrcElemsPerTgtElem;
1253 bool BCAlreadyExists = NewBCs.find(CastSrcTy) != NewBCs.end();
1257 : Builder->CreateBitCast(V, CastSrcTy, SVI.getName() + ".bc");
1258 if (!BCAlreadyExists)
1259 NewBCs[CastSrcTy] = NewBC;
1260 auto *Ext = Builder->CreateExtractElement(
1261 NewBC, ConstantInt::get(Int32Ty, BegIdx), SVI.getName() + ".extract");
1262 // The shufflevector isn't being replaced: the bitcast that used it
1263 // is. InstCombine will visit the newly-created instructions.
1264 replaceInstUsesWith(*BC, Ext);
1269 // If the LHS is a shufflevector itself, see if we can combine it with this
1270 // one without producing an unusual shuffle.
1271 // Cases that might be simplified:
1273 // x1=shuffle(v1,v2,mask1)
1274 // x=shuffle(x1,undef,mask)
1276 // x=shuffle(v1,undef,newMask)
1277 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : -1
1279 // x1=shuffle(v1,undef,mask1)
1280 // x=shuffle(x1,x2,mask)
1281 // where v1.size() == mask1.size()
1283 // x=shuffle(v1,x2,newMask)
1284 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : mask[i]
1286 // x2=shuffle(v2,undef,mask2)
1287 // x=shuffle(x1,x2,mask)
1288 // where v2.size() == mask2.size()
1290 // x=shuffle(x1,v2,newMask)
1291 // newMask[i] = (mask[i] < x1.size())
1292 // ? mask[i] : mask2[mask[i]-x1.size()]+x1.size()
1294 // x1=shuffle(v1,undef,mask1)
1295 // x2=shuffle(v2,undef,mask2)
1296 // x=shuffle(x1,x2,mask)
1297 // where v1.size() == v2.size()
1299 // x=shuffle(v1,v2,newMask)
1300 // newMask[i] = (mask[i] < x1.size())
1301 // ? mask1[mask[i]] : mask2[mask[i]-x1.size()]+v1.size()
1303 // Here we are really conservative:
1304 // we are absolutely afraid of producing a shuffle mask not in the input
1305 // program, because the code gen may not be smart enough to turn a merged
1306 // shuffle into two specific shuffles: it may produce worse code. As such,
1307 // we only merge two shuffles if the result is either a splat or one of the
1308 // input shuffle masks. In this case, merging the shuffles just removes
1309 // one instruction, which we know is safe. This is good for things like
1310 // turning: (splat(splat)) -> splat, or
1311 // merge(V[0..n], V[n+1..2n]) -> V[0..2n]
1312 ShuffleVectorInst* LHSShuffle = dyn_cast<ShuffleVectorInst>(LHS);
1313 ShuffleVectorInst* RHSShuffle = dyn_cast<ShuffleVectorInst>(RHS);
1315 if (!isa<UndefValue>(LHSShuffle->getOperand(1)) && !isa<UndefValue>(RHS))
1316 LHSShuffle = nullptr;
1318 if (!isa<UndefValue>(RHSShuffle->getOperand(1)))
1319 RHSShuffle = nullptr;
1320 if (!LHSShuffle && !RHSShuffle)
1321 return MadeChange ? &SVI : nullptr;
1323 Value* LHSOp0 = nullptr;
1324 Value* LHSOp1 = nullptr;
1325 Value* RHSOp0 = nullptr;
1326 unsigned LHSOp0Width = 0;
1327 unsigned RHSOp0Width = 0;
1329 LHSOp0 = LHSShuffle->getOperand(0);
1330 LHSOp1 = LHSShuffle->getOperand(1);
1331 LHSOp0Width = LHSOp0->getType()->getVectorNumElements();
1334 RHSOp0 = RHSShuffle->getOperand(0);
1335 RHSOp0Width = RHSOp0->getType()->getVectorNumElements();
1337 Value* newLHS = LHS;
1338 Value* newRHS = RHS;
1341 if (isa<UndefValue>(RHS)) {
1346 else if (LHSOp0Width == LHSWidth) {
1351 if (RHSShuffle && RHSOp0Width == LHSWidth) {
1355 if (LHSOp0 == RHSOp0) {
1360 if (newLHS == LHS && newRHS == RHS)
1361 return MadeChange ? &SVI : nullptr;
1363 SmallVector<int, 16> LHSMask;
1364 SmallVector<int, 16> RHSMask;
1366 LHSMask = LHSShuffle->getShuffleMask();
1367 if (RHSShuffle && newRHS != RHS)
1368 RHSMask = RHSShuffle->getShuffleMask();
1370 unsigned newLHSWidth = (newLHS != LHS) ? LHSOp0Width : LHSWidth;
1371 SmallVector<int, 16> newMask;
1372 bool isSplat = true;
1374 // Create a new mask for the new ShuffleVectorInst so that the new
1375 // ShuffleVectorInst is equivalent to the original one.
1376 for (unsigned i = 0; i < VWidth; ++i) {
1379 // This element is an undef value.
1381 } else if (Mask[i] < (int)LHSWidth) {
1382 // This element is from left hand side vector operand.
1384 // If LHS is going to be replaced (case 1, 2, or 4), calculate the
1385 // new mask value for the element.
1386 if (newLHS != LHS) {
1387 eltMask = LHSMask[Mask[i]];
1388 // If the value selected is an undef value, explicitly specify it
1389 // with a -1 mask value.
1390 if (eltMask >= (int)LHSOp0Width && isa<UndefValue>(LHSOp1))
1395 // This element is from right hand side vector operand
1397 // If the value selected is an undef value, explicitly specify it
1398 // with a -1 mask value. (case 1)
1399 if (isa<UndefValue>(RHS))
1401 // If RHS is going to be replaced (case 3 or 4), calculate the
1402 // new mask value for the element.
1403 else if (newRHS != RHS) {
1404 eltMask = RHSMask[Mask[i]-LHSWidth];
1405 // If the value selected is an undef value, explicitly specify it
1406 // with a -1 mask value.
1407 if (eltMask >= (int)RHSOp0Width) {
1408 assert(isa<UndefValue>(RHSShuffle->getOperand(1))
1409 && "should have been check above");
1413 eltMask = Mask[i]-LHSWidth;
1415 // If LHS's width is changed, shift the mask value accordingly.
1416 // If newRHS == NULL, i.e. LHSOp0 == RHSOp0, we want to remap any
1417 // references from RHSOp0 to LHSOp0, so we don't need to shift the mask.
1418 // If newRHS == newLHS, we want to remap any references from newRHS to
1419 // newLHS so that we can properly identify splats that may occur due to
1420 // obfuscation across the two vectors.
1421 if (eltMask >= 0 && newRHS != nullptr && newLHS != newRHS)
1422 eltMask += newLHSWidth;
1425 // Check if this could still be a splat.
1427 if (SplatElt >= 0 && SplatElt != eltMask)
1432 newMask.push_back(eltMask);
1435 // If the result mask is equal to one of the original shuffle masks,
1436 // or is a splat, do the replacement.
1437 if (isSplat || newMask == LHSMask || newMask == RHSMask || newMask == Mask) {
1438 SmallVector<Constant*, 16> Elts;
1439 for (unsigned i = 0, e = newMask.size(); i != e; ++i) {
1440 if (newMask[i] < 0) {
1441 Elts.push_back(UndefValue::get(Int32Ty));
1443 Elts.push_back(ConstantInt::get(Int32Ty, newMask[i]));
1447 newRHS = UndefValue::get(newLHS->getType());
1448 return new ShuffleVectorInst(newLHS, newRHS, ConstantVector::get(Elts));
1451 // If the result mask is an identity, replace uses of this instruction with
1452 // corresponding argument.
1453 bool isLHSID, isRHSID;
1454 recognizeIdentityMask(newMask, isLHSID, isRHSID);
1455 if (isLHSID && VWidth == LHSOp0Width) return replaceInstUsesWith(SVI, newLHS);
1456 if (isRHSID && VWidth == RHSOp0Width) return replaceInstUsesWith(SVI, newRHS);
1458 return MadeChange ? &SVI : nullptr;