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 the 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 nothing
234 if (CI->hasOneUse() && (CI->getOpcode() != Instruction::BitCast)) {
235 Value *EE = Builder->CreateExtractElement(CI->getOperand(0),
236 EI.getIndexOperand());
237 Worklist.AddValue(EE);
238 return CastInst::Create(CI->getOpcode(), EE, EI.getType());
240 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
241 if (SI->hasOneUse()) {
242 // TODO: For a select on vectors, it might be useful to do this if it
243 // has multiple extractelement uses. For vector select, that seems to
244 // fight the vectorizer.
246 // If we are extracting an element from a vector select or a select on
247 // vectors, a select on the scalars extracted from the vector arguments.
248 Value *TrueVal = SI->getTrueValue();
249 Value *FalseVal = SI->getFalseValue();
251 Value *Cond = SI->getCondition();
252 if (Cond->getType()->isVectorTy()) {
253 Cond = Builder->CreateExtractElement(Cond,
254 EI.getIndexOperand(),
255 Cond->getName() + ".elt");
259 = Builder->CreateExtractElement(TrueVal,
260 EI.getIndexOperand(),
261 TrueVal->getName() + ".elt");
264 = Builder->CreateExtractElement(FalseVal,
265 EI.getIndexOperand(),
266 FalseVal->getName() + ".elt");
267 return SelectInst::Create(Cond,
270 SI->getName() + ".elt");
277 /// If V is a shuffle of values that ONLY returns elements from either LHS or
278 /// RHS, return the shuffle mask and true. Otherwise, return false.
279 static bool collectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
280 SmallVectorImpl<Constant*> &Mask) {
281 assert(LHS->getType() == RHS->getType() &&
282 "Invalid CollectSingleShuffleElements");
283 unsigned NumElts = V->getType()->getVectorNumElements();
285 if (isa<UndefValue>(V)) {
286 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
291 for (unsigned i = 0; i != NumElts; ++i)
292 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
297 for (unsigned i = 0; i != NumElts; ++i)
298 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()),
303 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
304 // If this is an insert of an extract from some other vector, include it.
305 Value *VecOp = IEI->getOperand(0);
306 Value *ScalarOp = IEI->getOperand(1);
307 Value *IdxOp = IEI->getOperand(2);
309 if (!isa<ConstantInt>(IdxOp))
311 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
313 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
314 // We can handle this if the vector we are inserting into is
316 if (collectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
317 // If so, update the mask to reflect the inserted undef.
318 Mask[InsertedIdx] = UndefValue::get(Type::getInt32Ty(V->getContext()));
321 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
322 if (isa<ConstantInt>(EI->getOperand(1))) {
323 unsigned ExtractedIdx =
324 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
325 unsigned NumLHSElts = LHS->getType()->getVectorNumElements();
327 // This must be extracting from either LHS or RHS.
328 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
329 // We can handle this if the vector we are inserting into is
331 if (collectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
332 // If so, update the mask to reflect the inserted value.
333 if (EI->getOperand(0) == LHS) {
334 Mask[InsertedIdx % NumElts] =
335 ConstantInt::get(Type::getInt32Ty(V->getContext()),
338 assert(EI->getOperand(0) == RHS);
339 Mask[InsertedIdx % NumElts] =
340 ConstantInt::get(Type::getInt32Ty(V->getContext()),
341 ExtractedIdx + NumLHSElts);
354 /// We are building a shuffle to create V, which is a sequence of insertelement,
355 /// extractelement pairs. If PermittedRHS is set, then we must either use it or
356 /// not rely on the second vector source. Return a std::pair containing the
357 /// left and right vectors of the proposed shuffle (or 0), and set the Mask
358 /// parameter as required.
360 /// Note: we intentionally don't try to fold earlier shuffles since they have
361 /// often been chosen carefully to be efficiently implementable on the target.
362 typedef std::pair<Value *, Value *> ShuffleOps;
364 static ShuffleOps collectShuffleElements(Value *V,
365 SmallVectorImpl<Constant *> &Mask,
366 Value *PermittedRHS) {
367 assert(V->getType()->isVectorTy() && "Invalid shuffle!");
368 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
370 if (isa<UndefValue>(V)) {
371 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
372 return std::make_pair(
373 PermittedRHS ? UndefValue::get(PermittedRHS->getType()) : V, nullptr);
376 if (isa<ConstantAggregateZero>(V)) {
377 Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(V->getContext()),0));
378 return std::make_pair(V, nullptr);
381 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
382 // If this is an insert of an extract from some other vector, include it.
383 Value *VecOp = IEI->getOperand(0);
384 Value *ScalarOp = IEI->getOperand(1);
385 Value *IdxOp = IEI->getOperand(2);
387 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
388 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) {
389 unsigned ExtractedIdx =
390 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
391 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
393 // Either the extracted from or inserted into vector must be RHSVec,
394 // otherwise we'd end up with a shuffle of three inputs.
395 if (EI->getOperand(0) == PermittedRHS || PermittedRHS == nullptr) {
396 Value *RHS = EI->getOperand(0);
397 ShuffleOps LR = collectShuffleElements(VecOp, Mask, RHS);
398 assert(LR.second == nullptr || LR.second == RHS);
400 if (LR.first->getType() != RHS->getType()) {
401 // We tried our best, but we can't find anything compatible with RHS
402 // further up the chain. Return a trivial shuffle.
403 for (unsigned i = 0; i < NumElts; ++i)
404 Mask[i] = ConstantInt::get(Type::getInt32Ty(V->getContext()), i);
405 return std::make_pair(V, nullptr);
408 unsigned NumLHSElts = RHS->getType()->getVectorNumElements();
409 Mask[InsertedIdx % NumElts] =
410 ConstantInt::get(Type::getInt32Ty(V->getContext()),
411 NumLHSElts+ExtractedIdx);
412 return std::make_pair(LR.first, RHS);
415 if (VecOp == PermittedRHS) {
416 // We've gone as far as we can: anything on the other side of the
417 // extractelement will already have been converted into a shuffle.
418 unsigned NumLHSElts =
419 EI->getOperand(0)->getType()->getVectorNumElements();
420 for (unsigned i = 0; i != NumElts; ++i)
421 Mask.push_back(ConstantInt::get(
422 Type::getInt32Ty(V->getContext()),
423 i == InsertedIdx ? ExtractedIdx : NumLHSElts + i));
424 return std::make_pair(EI->getOperand(0), PermittedRHS);
427 // If this insertelement is a chain that comes from exactly these two
428 // vectors, return the vector and the effective shuffle.
429 if (EI->getOperand(0)->getType() == PermittedRHS->getType() &&
430 collectSingleShuffleElements(IEI, EI->getOperand(0), PermittedRHS,
432 return std::make_pair(EI->getOperand(0), PermittedRHS);
437 // Otherwise, can't do anything fancy. Return an identity vector.
438 for (unsigned i = 0; i != NumElts; ++i)
439 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
440 return std::make_pair(V, nullptr);
443 /// Try to find redundant insertvalue instructions, like the following ones:
444 /// %0 = insertvalue { i8, i32 } undef, i8 %x, 0
445 /// %1 = insertvalue { i8, i32 } %0, i8 %y, 0
446 /// Here the second instruction inserts values at the same indices, as the
447 /// first one, making the first one redundant.
448 /// It should be transformed to:
449 /// %0 = insertvalue { i8, i32 } undef, i8 %y, 0
450 Instruction *InstCombiner::visitInsertValueInst(InsertValueInst &I) {
451 bool IsRedundant = false;
452 ArrayRef<unsigned int> FirstIndices = I.getIndices();
454 // If there is a chain of insertvalue instructions (each of them except the
455 // last one has only one use and it's another insertvalue insn from this
456 // chain), check if any of the 'children' uses the same indices as the first
457 // instruction. In this case, the first one is redundant.
460 while (V->hasOneUse() && Depth < 10) {
461 User *U = V->user_back();
462 auto UserInsInst = dyn_cast<InsertValueInst>(U);
463 if (!UserInsInst || U->getOperand(0) != V)
465 if (UserInsInst->getIndices() == FirstIndices) {
474 return ReplaceInstUsesWith(I, I.getOperand(0));
478 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
479 Value *VecOp = IE.getOperand(0);
480 Value *ScalarOp = IE.getOperand(1);
481 Value *IdxOp = IE.getOperand(2);
483 // Inserting an undef or into an undefined place, remove this.
484 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
485 ReplaceInstUsesWith(IE, VecOp);
487 // If the inserted element was extracted from some other vector, and if the
488 // indexes are constant, try to turn this into a shufflevector operation.
489 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
490 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) {
491 unsigned NumInsertVectorElts = IE.getType()->getNumElements();
492 unsigned NumExtractVectorElts =
493 EI->getOperand(0)->getType()->getVectorNumElements();
494 unsigned ExtractedIdx =
495 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
496 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
498 if (ExtractedIdx >= NumExtractVectorElts) // Out of range extract.
499 return ReplaceInstUsesWith(IE, VecOp);
501 if (InsertedIdx >= NumInsertVectorElts) // Out of range insert.
502 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
504 // If we are extracting a value from a vector, then inserting it right
505 // back into the same place, just use the input vector.
506 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
507 return ReplaceInstUsesWith(IE, VecOp);
509 // If this insertelement isn't used by some other insertelement, turn it
510 // (and any insertelements it points to), into one big shuffle.
511 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.user_back())) {
512 SmallVector<Constant*, 16> Mask;
513 ShuffleOps LR = collectShuffleElements(&IE, Mask, nullptr);
515 // The proposed shuffle may be trivial, in which case we shouldn't
516 // perform the combine.
517 if (LR.first != &IE && LR.second != &IE) {
518 // We now have a shuffle of LHS, RHS, Mask.
519 if (LR.second == nullptr)
520 LR.second = UndefValue::get(LR.first->getType());
521 return new ShuffleVectorInst(LR.first, LR.second,
522 ConstantVector::get(Mask));
528 unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements();
529 APInt UndefElts(VWidth, 0);
530 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
531 if (Value *V = SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts)) {
533 return ReplaceInstUsesWith(IE, V);
540 /// Return true if we can evaluate the specified expression tree if the vector
541 /// elements were shuffled in a different order.
542 static bool CanEvaluateShuffled(Value *V, ArrayRef<int> Mask,
543 unsigned Depth = 5) {
544 // We can always reorder the elements of a constant.
545 if (isa<Constant>(V))
548 // We won't reorder vector arguments. No IPO here.
549 Instruction *I = dyn_cast<Instruction>(V);
550 if (!I) return false;
552 // Two users may expect different orders of the elements. Don't try it.
556 if (Depth == 0) return false;
558 switch (I->getOpcode()) {
559 case Instruction::Add:
560 case Instruction::FAdd:
561 case Instruction::Sub:
562 case Instruction::FSub:
563 case Instruction::Mul:
564 case Instruction::FMul:
565 case Instruction::UDiv:
566 case Instruction::SDiv:
567 case Instruction::FDiv:
568 case Instruction::URem:
569 case Instruction::SRem:
570 case Instruction::FRem:
571 case Instruction::Shl:
572 case Instruction::LShr:
573 case Instruction::AShr:
574 case Instruction::And:
575 case Instruction::Or:
576 case Instruction::Xor:
577 case Instruction::ICmp:
578 case Instruction::FCmp:
579 case Instruction::Trunc:
580 case Instruction::ZExt:
581 case Instruction::SExt:
582 case Instruction::FPToUI:
583 case Instruction::FPToSI:
584 case Instruction::UIToFP:
585 case Instruction::SIToFP:
586 case Instruction::FPTrunc:
587 case Instruction::FPExt:
588 case Instruction::GetElementPtr: {
589 for (Value *Operand : I->operands()) {
590 if (!CanEvaluateShuffled(Operand, Mask, Depth-1))
595 case Instruction::InsertElement: {
596 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(2));
597 if (!CI) return false;
598 int ElementNumber = CI->getLimitedValue();
600 // Verify that 'CI' does not occur twice in Mask. A single 'insertelement'
601 // can't put an element into multiple indices.
602 bool SeenOnce = false;
603 for (int i = 0, e = Mask.size(); i != e; ++i) {
604 if (Mask[i] == ElementNumber) {
610 return CanEvaluateShuffled(I->getOperand(0), Mask, Depth-1);
616 /// Rebuild a new instruction just like 'I' but with the new operands given.
617 /// In the event of type mismatch, the type of the operands is correct.
618 static Value *buildNew(Instruction *I, ArrayRef<Value*> NewOps) {
619 // We don't want to use the IRBuilder here because we want the replacement
620 // instructions to appear next to 'I', not the builder's insertion point.
621 switch (I->getOpcode()) {
622 case Instruction::Add:
623 case Instruction::FAdd:
624 case Instruction::Sub:
625 case Instruction::FSub:
626 case Instruction::Mul:
627 case Instruction::FMul:
628 case Instruction::UDiv:
629 case Instruction::SDiv:
630 case Instruction::FDiv:
631 case Instruction::URem:
632 case Instruction::SRem:
633 case Instruction::FRem:
634 case Instruction::Shl:
635 case Instruction::LShr:
636 case Instruction::AShr:
637 case Instruction::And:
638 case Instruction::Or:
639 case Instruction::Xor: {
640 BinaryOperator *BO = cast<BinaryOperator>(I);
641 assert(NewOps.size() == 2 && "binary operator with #ops != 2");
642 BinaryOperator *New =
643 BinaryOperator::Create(cast<BinaryOperator>(I)->getOpcode(),
644 NewOps[0], NewOps[1], "", BO);
645 if (isa<OverflowingBinaryOperator>(BO)) {
646 New->setHasNoUnsignedWrap(BO->hasNoUnsignedWrap());
647 New->setHasNoSignedWrap(BO->hasNoSignedWrap());
649 if (isa<PossiblyExactOperator>(BO)) {
650 New->setIsExact(BO->isExact());
652 if (isa<FPMathOperator>(BO))
653 New->copyFastMathFlags(I);
656 case Instruction::ICmp:
657 assert(NewOps.size() == 2 && "icmp with #ops != 2");
658 return new ICmpInst(I, cast<ICmpInst>(I)->getPredicate(),
659 NewOps[0], NewOps[1]);
660 case Instruction::FCmp:
661 assert(NewOps.size() == 2 && "fcmp with #ops != 2");
662 return new FCmpInst(I, cast<FCmpInst>(I)->getPredicate(),
663 NewOps[0], NewOps[1]);
664 case Instruction::Trunc:
665 case Instruction::ZExt:
666 case Instruction::SExt:
667 case Instruction::FPToUI:
668 case Instruction::FPToSI:
669 case Instruction::UIToFP:
670 case Instruction::SIToFP:
671 case Instruction::FPTrunc:
672 case Instruction::FPExt: {
673 // It's possible that the mask has a different number of elements from
674 // the original cast. We recompute the destination type to match the mask.
676 VectorType::get(I->getType()->getScalarType(),
677 NewOps[0]->getType()->getVectorNumElements());
678 assert(NewOps.size() == 1 && "cast with #ops != 1");
679 return CastInst::Create(cast<CastInst>(I)->getOpcode(), NewOps[0], DestTy,
682 case Instruction::GetElementPtr: {
683 Value *Ptr = NewOps[0];
684 ArrayRef<Value*> Idx = NewOps.slice(1);
685 GetElementPtrInst *GEP = GetElementPtrInst::Create(
686 cast<GetElementPtrInst>(I)->getSourceElementType(), Ptr, Idx, "", I);
687 GEP->setIsInBounds(cast<GetElementPtrInst>(I)->isInBounds());
691 llvm_unreachable("failed to rebuild vector instructions");
695 InstCombiner::EvaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask) {
696 // Mask.size() does not need to be equal to the number of vector elements.
698 assert(V->getType()->isVectorTy() && "can't reorder non-vector elements");
699 if (isa<UndefValue>(V)) {
700 return UndefValue::get(VectorType::get(V->getType()->getScalarType(),
703 if (isa<ConstantAggregateZero>(V)) {
704 return ConstantAggregateZero::get(
705 VectorType::get(V->getType()->getScalarType(),
708 if (Constant *C = dyn_cast<Constant>(V)) {
709 SmallVector<Constant *, 16> MaskValues;
710 for (int i = 0, e = Mask.size(); i != e; ++i) {
712 MaskValues.push_back(UndefValue::get(Builder->getInt32Ty()));
714 MaskValues.push_back(Builder->getInt32(Mask[i]));
716 return ConstantExpr::getShuffleVector(C, UndefValue::get(C->getType()),
717 ConstantVector::get(MaskValues));
720 Instruction *I = cast<Instruction>(V);
721 switch (I->getOpcode()) {
722 case Instruction::Add:
723 case Instruction::FAdd:
724 case Instruction::Sub:
725 case Instruction::FSub:
726 case Instruction::Mul:
727 case Instruction::FMul:
728 case Instruction::UDiv:
729 case Instruction::SDiv:
730 case Instruction::FDiv:
731 case Instruction::URem:
732 case Instruction::SRem:
733 case Instruction::FRem:
734 case Instruction::Shl:
735 case Instruction::LShr:
736 case Instruction::AShr:
737 case Instruction::And:
738 case Instruction::Or:
739 case Instruction::Xor:
740 case Instruction::ICmp:
741 case Instruction::FCmp:
742 case Instruction::Trunc:
743 case Instruction::ZExt:
744 case Instruction::SExt:
745 case Instruction::FPToUI:
746 case Instruction::FPToSI:
747 case Instruction::UIToFP:
748 case Instruction::SIToFP:
749 case Instruction::FPTrunc:
750 case Instruction::FPExt:
751 case Instruction::Select:
752 case Instruction::GetElementPtr: {
753 SmallVector<Value*, 8> NewOps;
754 bool NeedsRebuild = (Mask.size() != I->getType()->getVectorNumElements());
755 for (int i = 0, e = I->getNumOperands(); i != e; ++i) {
756 Value *V = EvaluateInDifferentElementOrder(I->getOperand(i), Mask);
758 NeedsRebuild |= (V != I->getOperand(i));
761 return buildNew(I, NewOps);
765 case Instruction::InsertElement: {
766 int Element = cast<ConstantInt>(I->getOperand(2))->getLimitedValue();
768 // The insertelement was inserting at Element. Figure out which element
769 // that becomes after shuffling. The answer is guaranteed to be unique
770 // by CanEvaluateShuffled.
773 for (int e = Mask.size(); Index != e; ++Index) {
774 if (Mask[Index] == Element) {
780 // If element is not in Mask, no need to handle the operand 1 (element to
781 // be inserted). Just evaluate values in operand 0 according to Mask.
783 return EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
785 Value *V = EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
786 return InsertElementInst::Create(V, I->getOperand(1),
787 Builder->getInt32(Index), "", I);
790 llvm_unreachable("failed to reorder elements of vector instruction!");
793 static void recognizeIdentityMask(const SmallVectorImpl<int> &Mask,
794 bool &isLHSID, bool &isRHSID) {
795 isLHSID = isRHSID = true;
797 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
798 if (Mask[i] < 0) continue; // Ignore undef values.
799 // Is this an identity shuffle of the LHS value?
800 isLHSID &= (Mask[i] == (int)i);
802 // Is this an identity shuffle of the RHS value?
803 isRHSID &= (Mask[i]-e == i);
807 // Returns true if the shuffle is extracting a contiguous range of values from
809 // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
810 // Input: |AA|BB|CC|DD|EE|FF|GG|HH|II|JJ|KK|LL|MM|NN|OO|PP|
811 // Shuffles to: |EE|FF|GG|HH|
813 static bool isShuffleExtractingFromLHS(ShuffleVectorInst &SVI,
814 SmallVector<int, 16> &Mask) {
816 cast<VectorType>(SVI.getOperand(0)->getType())->getNumElements();
817 unsigned MaskElems = Mask.size();
818 unsigned BegIdx = Mask.front();
819 unsigned EndIdx = Mask.back();
820 if (BegIdx > EndIdx || EndIdx >= LHSElems || EndIdx - BegIdx != MaskElems - 1)
822 for (unsigned I = 0; I != MaskElems; ++I)
823 if (static_cast<unsigned>(Mask[I]) != BegIdx + I)
828 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
829 Value *LHS = SVI.getOperand(0);
830 Value *RHS = SVI.getOperand(1);
831 SmallVector<int, 16> Mask = SVI.getShuffleMask();
832 Type *Int32Ty = Type::getInt32Ty(SVI.getContext());
834 bool MadeChange = false;
836 // Undefined shuffle mask -> undefined value.
837 if (isa<UndefValue>(SVI.getOperand(2)))
838 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
840 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
842 APInt UndefElts(VWidth, 0);
843 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
844 if (Value *V = SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
846 return ReplaceInstUsesWith(SVI, V);
847 LHS = SVI.getOperand(0);
848 RHS = SVI.getOperand(1);
852 unsigned LHSWidth = cast<VectorType>(LHS->getType())->getNumElements();
854 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
855 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
856 if (LHS == RHS || isa<UndefValue>(LHS)) {
857 if (isa<UndefValue>(LHS) && LHS == RHS) {
858 // shuffle(undef,undef,mask) -> undef.
859 Value *Result = (VWidth == LHSWidth)
860 ? LHS : UndefValue::get(SVI.getType());
861 return ReplaceInstUsesWith(SVI, Result);
864 // Remap any references to RHS to use LHS.
865 SmallVector<Constant*, 16> Elts;
866 for (unsigned i = 0, e = LHSWidth; i != VWidth; ++i) {
868 Elts.push_back(UndefValue::get(Int32Ty));
872 if ((Mask[i] >= (int)e && isa<UndefValue>(RHS)) ||
873 (Mask[i] < (int)e && isa<UndefValue>(LHS))) {
874 Mask[i] = -1; // Turn into undef.
875 Elts.push_back(UndefValue::get(Int32Ty));
877 Mask[i] = Mask[i] % e; // Force to LHS.
878 Elts.push_back(ConstantInt::get(Int32Ty, Mask[i]));
881 SVI.setOperand(0, SVI.getOperand(1));
882 SVI.setOperand(1, UndefValue::get(RHS->getType()));
883 SVI.setOperand(2, ConstantVector::get(Elts));
884 LHS = SVI.getOperand(0);
885 RHS = SVI.getOperand(1);
889 if (VWidth == LHSWidth) {
890 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
891 bool isLHSID, isRHSID;
892 recognizeIdentityMask(Mask, isLHSID, isRHSID);
894 // Eliminate identity shuffles.
895 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
896 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
899 if (isa<UndefValue>(RHS) && CanEvaluateShuffled(LHS, Mask)) {
900 Value *V = EvaluateInDifferentElementOrder(LHS, Mask);
901 return ReplaceInstUsesWith(SVI, V);
904 // SROA generates shuffle+bitcast when the extracted sub-vector is bitcast to
905 // a non-vector type. We can instead bitcast the original vector followed by
906 // an extract of the desired element:
908 // %sroa = shufflevector <16 x i8> %in, <16 x i8> undef,
909 // <4 x i32> <i32 0, i32 1, i32 2, i32 3>
910 // %1 = bitcast <4 x i8> %sroa to i32
912 // %bc = bitcast <16 x i8> %in to <4 x i32>
913 // %ext = extractelement <4 x i32> %bc, i32 0
915 // If the shuffle is extracting a contiguous range of values from the input
916 // vector then each use which is a bitcast of the extracted size can be
917 // replaced. This will work if the vector types are compatible, and the begin
918 // index is aligned to a value in the casted vector type. If the begin index
919 // isn't aligned then we can shuffle the original vector (keeping the same
920 // vector type) before extracting.
922 // This code will bail out if the target type is fundamentally incompatible
923 // with vectors of the source type.
925 // Example of <16 x i8>, target type i32:
926 // Index range [4,8): v-----------v Will work.
927 // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
928 // <16 x i8>: | | | | | | | | | | | | | | | | |
929 // <4 x i32>: | | | | |
930 // +-----------+-----------+-----------+-----------+
931 // Index range [6,10): ^-----------^ Needs an extra shuffle.
932 // Target type i40: ^--------------^ Won't work, bail.
933 if (isShuffleExtractingFromLHS(SVI, Mask)) {
935 unsigned MaskElems = Mask.size();
936 unsigned BegIdx = Mask.front();
937 VectorType *SrcTy = cast<VectorType>(V->getType());
938 unsigned VecBitWidth = SrcTy->getBitWidth();
939 unsigned SrcElemBitWidth = DL.getTypeSizeInBits(SrcTy->getElementType());
940 assert(SrcElemBitWidth && "vector elements must have a bitwidth");
941 unsigned SrcNumElems = SrcTy->getNumElements();
942 SmallVector<BitCastInst *, 8> BCs;
943 DenseMap<Type *, Value *> NewBCs;
944 for (User *U : SVI.users())
945 if (BitCastInst *BC = dyn_cast<BitCastInst>(U))
946 if (!BC->use_empty())
947 // Only visit bitcasts that weren't previously handled.
949 for (BitCastInst *BC : BCs) {
950 Type *TgtTy = BC->getDestTy();
951 unsigned TgtElemBitWidth = DL.getTypeSizeInBits(TgtTy);
952 if (!TgtElemBitWidth)
954 unsigned TgtNumElems = VecBitWidth / TgtElemBitWidth;
955 bool VecBitWidthsEqual = VecBitWidth == TgtNumElems * TgtElemBitWidth;
956 bool BegIsAligned = 0 == ((SrcElemBitWidth * BegIdx) % TgtElemBitWidth);
957 if (!VecBitWidthsEqual)
959 if (!VectorType::isValidElementType(TgtTy))
961 VectorType *CastSrcTy = VectorType::get(TgtTy, TgtNumElems);
963 // Shuffle the input so [0,NumElements) contains the output, and
964 // [NumElems,SrcNumElems) is undef.
965 SmallVector<Constant *, 16> ShuffleMask(SrcNumElems,
966 UndefValue::get(Int32Ty));
967 for (unsigned I = 0, E = MaskElems, Idx = BegIdx; I != E; ++Idx, ++I)
968 ShuffleMask[I] = ConstantInt::get(Int32Ty, Idx);
969 V = Builder->CreateShuffleVector(V, UndefValue::get(V->getType()),
970 ConstantVector::get(ShuffleMask),
971 SVI.getName() + ".extract");
974 unsigned SrcElemsPerTgtElem = TgtElemBitWidth / SrcElemBitWidth;
975 assert(SrcElemsPerTgtElem);
976 BegIdx /= SrcElemsPerTgtElem;
977 bool BCAlreadyExists = NewBCs.find(CastSrcTy) != NewBCs.end();
981 : Builder->CreateBitCast(V, CastSrcTy, SVI.getName() + ".bc");
982 if (!BCAlreadyExists)
983 NewBCs[CastSrcTy] = NewBC;
984 auto *Ext = Builder->CreateExtractElement(
985 NewBC, ConstantInt::get(Int32Ty, BegIdx), SVI.getName() + ".extract");
986 // The shufflevector isn't being replaced: the bitcast that used it
987 // is. InstCombine will visit the newly-created instructions.
988 ReplaceInstUsesWith(*BC, Ext);
993 // If the LHS is a shufflevector itself, see if we can combine it with this
994 // one without producing an unusual shuffle.
995 // Cases that might be simplified:
997 // x1=shuffle(v1,v2,mask1)
998 // x=shuffle(x1,undef,mask)
1000 // x=shuffle(v1,undef,newMask)
1001 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : -1
1003 // x1=shuffle(v1,undef,mask1)
1004 // x=shuffle(x1,x2,mask)
1005 // where v1.size() == mask1.size()
1007 // x=shuffle(v1,x2,newMask)
1008 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : mask[i]
1010 // x2=shuffle(v2,undef,mask2)
1011 // x=shuffle(x1,x2,mask)
1012 // where v2.size() == mask2.size()
1014 // x=shuffle(x1,v2,newMask)
1015 // newMask[i] = (mask[i] < x1.size())
1016 // ? mask[i] : mask2[mask[i]-x1.size()]+x1.size()
1018 // x1=shuffle(v1,undef,mask1)
1019 // x2=shuffle(v2,undef,mask2)
1020 // x=shuffle(x1,x2,mask)
1021 // where v1.size() == v2.size()
1023 // x=shuffle(v1,v2,newMask)
1024 // newMask[i] = (mask[i] < x1.size())
1025 // ? mask1[mask[i]] : mask2[mask[i]-x1.size()]+v1.size()
1027 // Here we are really conservative:
1028 // we are absolutely afraid of producing a shuffle mask not in the input
1029 // program, because the code gen may not be smart enough to turn a merged
1030 // shuffle into two specific shuffles: it may produce worse code. As such,
1031 // we only merge two shuffles if the result is either a splat or one of the
1032 // input shuffle masks. In this case, merging the shuffles just removes
1033 // one instruction, which we know is safe. This is good for things like
1034 // turning: (splat(splat)) -> splat, or
1035 // merge(V[0..n], V[n+1..2n]) -> V[0..2n]
1036 ShuffleVectorInst* LHSShuffle = dyn_cast<ShuffleVectorInst>(LHS);
1037 ShuffleVectorInst* RHSShuffle = dyn_cast<ShuffleVectorInst>(RHS);
1039 if (!isa<UndefValue>(LHSShuffle->getOperand(1)) && !isa<UndefValue>(RHS))
1040 LHSShuffle = nullptr;
1042 if (!isa<UndefValue>(RHSShuffle->getOperand(1)))
1043 RHSShuffle = nullptr;
1044 if (!LHSShuffle && !RHSShuffle)
1045 return MadeChange ? &SVI : nullptr;
1047 Value* LHSOp0 = nullptr;
1048 Value* LHSOp1 = nullptr;
1049 Value* RHSOp0 = nullptr;
1050 unsigned LHSOp0Width = 0;
1051 unsigned RHSOp0Width = 0;
1053 LHSOp0 = LHSShuffle->getOperand(0);
1054 LHSOp1 = LHSShuffle->getOperand(1);
1055 LHSOp0Width = cast<VectorType>(LHSOp0->getType())->getNumElements();
1058 RHSOp0 = RHSShuffle->getOperand(0);
1059 RHSOp0Width = cast<VectorType>(RHSOp0->getType())->getNumElements();
1061 Value* newLHS = LHS;
1062 Value* newRHS = RHS;
1065 if (isa<UndefValue>(RHS)) {
1070 else if (LHSOp0Width == LHSWidth) {
1075 if (RHSShuffle && RHSOp0Width == LHSWidth) {
1079 if (LHSOp0 == RHSOp0) {
1084 if (newLHS == LHS && newRHS == RHS)
1085 return MadeChange ? &SVI : nullptr;
1087 SmallVector<int, 16> LHSMask;
1088 SmallVector<int, 16> RHSMask;
1090 LHSMask = LHSShuffle->getShuffleMask();
1091 if (RHSShuffle && newRHS != RHS)
1092 RHSMask = RHSShuffle->getShuffleMask();
1094 unsigned newLHSWidth = (newLHS != LHS) ? LHSOp0Width : LHSWidth;
1095 SmallVector<int, 16> newMask;
1096 bool isSplat = true;
1098 // Create a new mask for the new ShuffleVectorInst so that the new
1099 // ShuffleVectorInst is equivalent to the original one.
1100 for (unsigned i = 0; i < VWidth; ++i) {
1103 // This element is an undef value.
1105 } else if (Mask[i] < (int)LHSWidth) {
1106 // This element is from left hand side vector operand.
1108 // If LHS is going to be replaced (case 1, 2, or 4), calculate the
1109 // new mask value for the element.
1110 if (newLHS != LHS) {
1111 eltMask = LHSMask[Mask[i]];
1112 // If the value selected is an undef value, explicitly specify it
1113 // with a -1 mask value.
1114 if (eltMask >= (int)LHSOp0Width && isa<UndefValue>(LHSOp1))
1119 // This element is from right hand side vector operand
1121 // If the value selected is an undef value, explicitly specify it
1122 // with a -1 mask value. (case 1)
1123 if (isa<UndefValue>(RHS))
1125 // If RHS is going to be replaced (case 3 or 4), calculate the
1126 // new mask value for the element.
1127 else if (newRHS != RHS) {
1128 eltMask = RHSMask[Mask[i]-LHSWidth];
1129 // If the value selected is an undef value, explicitly specify it
1130 // with a -1 mask value.
1131 if (eltMask >= (int)RHSOp0Width) {
1132 assert(isa<UndefValue>(RHSShuffle->getOperand(1))
1133 && "should have been check above");
1137 eltMask = Mask[i]-LHSWidth;
1139 // If LHS's width is changed, shift the mask value accordingly.
1140 // If newRHS == NULL, i.e. LHSOp0 == RHSOp0, we want to remap any
1141 // references from RHSOp0 to LHSOp0, so we don't need to shift the mask.
1142 // If newRHS == newLHS, we want to remap any references from newRHS to
1143 // newLHS so that we can properly identify splats that may occur due to
1144 // obfuscation across the two vectors.
1145 if (eltMask >= 0 && newRHS != nullptr && newLHS != newRHS)
1146 eltMask += newLHSWidth;
1149 // Check if this could still be a splat.
1151 if (SplatElt >= 0 && SplatElt != eltMask)
1156 newMask.push_back(eltMask);
1159 // If the result mask is equal to one of the original shuffle masks,
1160 // or is a splat, do the replacement.
1161 if (isSplat || newMask == LHSMask || newMask == RHSMask || newMask == Mask) {
1162 SmallVector<Constant*, 16> Elts;
1163 for (unsigned i = 0, e = newMask.size(); i != e; ++i) {
1164 if (newMask[i] < 0) {
1165 Elts.push_back(UndefValue::get(Int32Ty));
1167 Elts.push_back(ConstantInt::get(Int32Ty, newMask[i]));
1171 newRHS = UndefValue::get(newLHS->getType());
1172 return new ShuffleVectorInst(newLHS, newRHS, ConstantVector::get(Elts));
1175 // If the result mask is an identity, replace uses of this instruction with
1176 // corresponding argument.
1177 bool isLHSID, isRHSID;
1178 recognizeIdentityMask(newMask, isLHSID, isRHSID);
1179 if (isLHSID && VWidth == LHSOp0Width) return ReplaceInstUsesWith(SVI, newLHS);
1180 if (isRHSID && VWidth == RHSOp0Width) return ReplaceInstUsesWith(SVI, newRHS);
1182 return MadeChange ? &SVI : nullptr;