• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===- InstCombineVectorOps.cpp -------------------------------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements instcombine for ExtractElement, InsertElement and
11 // ShuffleVector.
12 //
13 //===----------------------------------------------------------------------===//
14 
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"
20 using namespace llvm;
21 using namespace PatternMatch;
22 
23 #define DEBUG_TYPE "instcombine"
24 
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.
cheapToScalarize(Value * V,bool isConstant)28 static bool cheapToScalarize(Value *V, bool isConstant) {
29   if (Constant *C = dyn_cast<Constant>(V)) {
30     if (isConstant) return true;
31 
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;
35            ++i)
36         if (C->getAggregateElement(i) != Op0)
37           return false;
38       return true;
39     }
40   }
41   Instruction *I = dyn_cast<Instruction>(V);
42   if (!I) return false;
43 
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)))
48     return true;
49   if (I->getOpcode() == Instruction::Load && I->hasOneUse())
50     return true;
51   if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
52     if (BO->hasOneUse() &&
53         (cheapToScalarize(BO->getOperand(0), isConstant) ||
54          cheapToScalarize(BO->getOperand(1), isConstant)))
55       return true;
56   if (CmpInst *CI = dyn_cast<CmpInst>(I))
57     if (CI->hasOneUse() &&
58         (cheapToScalarize(CI->getOperand(0), isConstant) ||
59          cheapToScalarize(CI->getOperand(1), isConstant)))
60       return true;
61 
62   return false;
63 }
64 
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.
scalarizePHI(ExtractElementInst & EI,PHINode * PN)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);
79       else
80         return nullptr;
81     } else if (!PHIUser) {
82       PHIUser = cast<Instruction>(U);
83     } else {
84       return nullptr;
85     }
86   }
87 
88   if (!PHIUser)
89     return nullptr;
90 
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))
96     return nullptr;
97 
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
111       // vector operand.
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"),
117           *B0);
118       Value *newPHIUser = InsertNewInstWith(
119           BinaryOperator::CreateWithCopiedFlags(B0->getOpcode(),
120                                                 scalarPHI, Op, B0), *B0);
121       scalarPHI->addIncoming(newPHIUser, inBB);
122     } else {
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();
130       } else {
131         InsertPos = inBB->getFirstInsertionPt();
132       }
133 
134       InsertNewInstWith(newEI, *InsertPos);
135 
136       scalarPHI->addIncoming(newEI, inBB);
137     }
138   }
139 
140   for (auto E : Extracts)
141     replaceInstUsesWith(*E, scalarPHI);
142 
143   return &EI;
144 }
145 
visitExtractElementInst(ExtractElementInst & EI)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);
150 
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));
156 
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();
162 
163     // InstSimplify handles cases where the index is invalid.
164     assert(IndexVal < VectorWidth);
165 
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
168     // property.
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,
174                                                 UndefElts)) {
175         EI.setOperand(0, V);
176         return &EI;
177       }
178     }
179 
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());
188     }
189 
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);
194       if (scalarPHI)
195         return scalarPHI;
196     }
197   }
198 
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)))) {
205         Value *newEI0 =
206           Builder->CreateExtractElement(BO->getOperand(0), EI.getOperand(1),
207                                         EI.getName()+".lhs");
208         Value *newEI1 =
209           Builder->CreateExtractElement(BO->getOperand(1), EI.getOperand(1),
210                                         EI.getName()+".rhs");
211         return BinaryOperator::CreateWithCopiedFlags(BO->getOpcode(),
212                                                      newEI0, newEI1, BO);
213       }
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));
223         return &EI;
224       }
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());
230         Value *Src;
231         unsigned LHSWidth =
232           SVI->getOperand(0)->getType()->getVectorNumElements();
233 
234         if (SrcIdx < 0)
235           return replaceInstUsesWith(EI, UndefValue::get(EI.getType()));
236         if (SrcIdx < (int)LHSWidth)
237           Src = SVI->getOperand(0);
238         else {
239           SrcIdx -= LHSWidth;
240           Src = SVI->getOperand(1);
241         }
242         Type *Int32Ty = Type::getInt32Ty(EI.getContext());
243         return ExtractElementInst::Create(Src,
244                                           ConstantInt::get(Int32Ty,
245                                                            SrcIdx, false));
246       }
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
250       // nothing.
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());
256       }
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.
262 
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
265         // arguments.
266         Value *TrueVal = SI->getTrueValue();
267         Value *FalseVal = SI->getFalseValue();
268 
269         Value *Cond = SI->getCondition();
270         if (Cond->getType()->isVectorTy()) {
271           Cond = Builder->CreateExtractElement(Cond,
272                                                EI.getIndexOperand(),
273                                                Cond->getName() + ".elt");
274         }
275 
276         Value *V1Elem
277           = Builder->CreateExtractElement(TrueVal,
278                                           EI.getIndexOperand(),
279                                           TrueVal->getName() + ".elt");
280 
281         Value *V2Elem
282           = Builder->CreateExtractElement(FalseVal,
283                                           EI.getIndexOperand(),
284                                           FalseVal->getName() + ".elt");
285         return SelectInst::Create(Cond,
286                                   V1Elem,
287                                   V2Elem,
288                                   SI->getName() + ".elt");
289       }
290     }
291   }
292   return nullptr;
293 }
294 
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.
collectSingleShuffleElements(Value * V,Value * LHS,Value * RHS,SmallVectorImpl<Constant * > & Mask)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();
302 
303   if (isa<UndefValue>(V)) {
304     Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
305     return true;
306   }
307 
308   if (V == LHS) {
309     for (unsigned i = 0; i != NumElts; ++i)
310       Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
311     return true;
312   }
313 
314   if (V == RHS) {
315     for (unsigned i = 0; i != NumElts; ++i)
316       Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()),
317                                       i+NumElts));
318     return true;
319   }
320 
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);
326 
327     if (!isa<ConstantInt>(IdxOp))
328       return false;
329     unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
330 
331     if (isa<UndefValue>(ScalarOp)) {  // inserting undef into vector.
332       // We can handle this if the vector we are inserting into is
333       // transitively ok.
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()));
337         return true;
338       }
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();
344 
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
348           // transitively ok.
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()),
354                                ExtractedIdx);
355             } else {
356               assert(EI->getOperand(0) == RHS);
357               Mask[InsertedIdx % NumElts] =
358               ConstantInt::get(Type::getInt32Ty(V->getContext()),
359                                ExtractedIdx + NumLHSElts);
360             }
361             return true;
362           }
363         }
364       }
365     }
366   }
367 
368   return false;
369 }
370 
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.
replaceExtractElements(InsertElementInst * InsElt,ExtractElementInst * ExtElt,InstCombiner & IC)374 static void replaceExtractElements(InsertElementInst *InsElt,
375                                    ExtractElementInst *ExtElt,
376                                    InstCombiner &IC) {
377   VectorType *InsVecType = InsElt->getType();
378   VectorType *ExtVecType = ExtElt->getVectorOperandType();
379   unsigned NumInsElts = InsVecType->getVectorNumElements();
380   unsigned NumExtElts = ExtVecType->getVectorNumElements();
381 
382   // The inserted-to vector must be wider than the extracted-from vector.
383   if (InsVecType->getElementType() != ExtVecType->getElementType() ||
384       NumExtElts >= NumInsElts)
385     return;
386 
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));
397 
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();
403 
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())
414     return;
415 
416   auto *WideVec = new ShuffleVectorInst(ExtVecOp, UndefValue::get(ExtVecType),
417                                         ConstantVector::get(ExtendMask));
418 
419   // Insert the new shuffle after the vector operand of the extract is defined
420   // (as long as it's not a PHI) or at the start of the basic block of the
421   // extract, so any subsequent extracts in the same basic block can use it.
422   // TODO: Insert before the earliest ExtractElementInst that is replaced.
423   if (ExtVecOpInst && !isa<PHINode>(ExtVecOpInst))
424     WideVec->insertAfter(ExtVecOpInst);
425   else
426     IC.InsertNewInstWith(WideVec, *ExtElt->getParent()->getFirstInsertionPt());
427 
428   // Replace extracts from the original narrow vector with extracts from the new
429   // wide vector.
430   for (User *U : ExtVecOp->users()) {
431     ExtractElementInst *OldExt = dyn_cast<ExtractElementInst>(U);
432     if (!OldExt || OldExt->getParent() != WideVec->getParent())
433       continue;
434     auto *NewExt = ExtractElementInst::Create(WideVec, OldExt->getOperand(1));
435     NewExt->insertAfter(WideVec);
436     IC.replaceInstUsesWith(*OldExt, NewExt);
437   }
438 }
439 
440 /// We are building a shuffle to create V, which is a sequence of insertelement,
441 /// extractelement pairs. If PermittedRHS is set, then we must either use it or
442 /// not rely on the second vector source. Return a std::pair containing the
443 /// left and right vectors of the proposed shuffle (or 0), and set the Mask
444 /// parameter as required.
445 ///
446 /// Note: we intentionally don't try to fold earlier shuffles since they have
447 /// often been chosen carefully to be efficiently implementable on the target.
448 typedef std::pair<Value *, Value *> ShuffleOps;
449 
collectShuffleElements(Value * V,SmallVectorImpl<Constant * > & Mask,Value * PermittedRHS,InstCombiner & IC)450 static ShuffleOps collectShuffleElements(Value *V,
451                                          SmallVectorImpl<Constant *> &Mask,
452                                          Value *PermittedRHS,
453                                          InstCombiner &IC) {
454   assert(V->getType()->isVectorTy() && "Invalid shuffle!");
455   unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
456 
457   if (isa<UndefValue>(V)) {
458     Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
459     return std::make_pair(
460         PermittedRHS ? UndefValue::get(PermittedRHS->getType()) : V, nullptr);
461   }
462 
463   if (isa<ConstantAggregateZero>(V)) {
464     Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(V->getContext()),0));
465     return std::make_pair(V, nullptr);
466   }
467 
468   if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
469     // If this is an insert of an extract from some other vector, include it.
470     Value *VecOp    = IEI->getOperand(0);
471     Value *ScalarOp = IEI->getOperand(1);
472     Value *IdxOp    = IEI->getOperand(2);
473 
474     if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
475       if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) {
476         unsigned ExtractedIdx =
477           cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
478         unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
479 
480         // Either the extracted from or inserted into vector must be RHSVec,
481         // otherwise we'd end up with a shuffle of three inputs.
482         if (EI->getOperand(0) == PermittedRHS || PermittedRHS == nullptr) {
483           Value *RHS = EI->getOperand(0);
484           ShuffleOps LR = collectShuffleElements(VecOp, Mask, RHS, IC);
485           assert(LR.second == nullptr || LR.second == RHS);
486 
487           if (LR.first->getType() != RHS->getType()) {
488             // Although we are giving up for now, see if we can create extracts
489             // that match the inserts for another round of combining.
490             replaceExtractElements(IEI, EI, IC);
491 
492             // We tried our best, but we can't find anything compatible with RHS
493             // further up the chain. Return a trivial shuffle.
494             for (unsigned i = 0; i < NumElts; ++i)
495               Mask[i] = ConstantInt::get(Type::getInt32Ty(V->getContext()), i);
496             return std::make_pair(V, nullptr);
497           }
498 
499           unsigned NumLHSElts = RHS->getType()->getVectorNumElements();
500           Mask[InsertedIdx % NumElts] =
501             ConstantInt::get(Type::getInt32Ty(V->getContext()),
502                              NumLHSElts+ExtractedIdx);
503           return std::make_pair(LR.first, RHS);
504         }
505 
506         if (VecOp == PermittedRHS) {
507           // We've gone as far as we can: anything on the other side of the
508           // extractelement will already have been converted into a shuffle.
509           unsigned NumLHSElts =
510               EI->getOperand(0)->getType()->getVectorNumElements();
511           for (unsigned i = 0; i != NumElts; ++i)
512             Mask.push_back(ConstantInt::get(
513                 Type::getInt32Ty(V->getContext()),
514                 i == InsertedIdx ? ExtractedIdx : NumLHSElts + i));
515           return std::make_pair(EI->getOperand(0), PermittedRHS);
516         }
517 
518         // If this insertelement is a chain that comes from exactly these two
519         // vectors, return the vector and the effective shuffle.
520         if (EI->getOperand(0)->getType() == PermittedRHS->getType() &&
521             collectSingleShuffleElements(IEI, EI->getOperand(0), PermittedRHS,
522                                          Mask))
523           return std::make_pair(EI->getOperand(0), PermittedRHS);
524       }
525     }
526   }
527 
528   // Otherwise, we can't do anything fancy. Return an identity vector.
529   for (unsigned i = 0; i != NumElts; ++i)
530     Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
531   return std::make_pair(V, nullptr);
532 }
533 
534 /// Try to find redundant insertvalue instructions, like the following ones:
535 ///  %0 = insertvalue { i8, i32 } undef, i8 %x, 0
536 ///  %1 = insertvalue { i8, i32 } %0,    i8 %y, 0
537 /// Here the second instruction inserts values at the same indices, as the
538 /// first one, making the first one redundant.
539 /// It should be transformed to:
540 ///  %0 = insertvalue { i8, i32 } undef, i8 %y, 0
visitInsertValueInst(InsertValueInst & I)541 Instruction *InstCombiner::visitInsertValueInst(InsertValueInst &I) {
542   bool IsRedundant = false;
543   ArrayRef<unsigned int> FirstIndices = I.getIndices();
544 
545   // If there is a chain of insertvalue instructions (each of them except the
546   // last one has only one use and it's another insertvalue insn from this
547   // chain), check if any of the 'children' uses the same indices as the first
548   // instruction. In this case, the first one is redundant.
549   Value *V = &I;
550   unsigned Depth = 0;
551   while (V->hasOneUse() && Depth < 10) {
552     User *U = V->user_back();
553     auto UserInsInst = dyn_cast<InsertValueInst>(U);
554     if (!UserInsInst || U->getOperand(0) != V)
555       break;
556     if (UserInsInst->getIndices() == FirstIndices) {
557       IsRedundant = true;
558       break;
559     }
560     V = UserInsInst;
561     Depth++;
562   }
563 
564   if (IsRedundant)
565     return replaceInstUsesWith(I, I.getOperand(0));
566   return nullptr;
567 }
568 
visitInsertElementInst(InsertElementInst & IE)569 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
570   Value *VecOp    = IE.getOperand(0);
571   Value *ScalarOp = IE.getOperand(1);
572   Value *IdxOp    = IE.getOperand(2);
573 
574   // Inserting an undef or into an undefined place, remove this.
575   if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
576     replaceInstUsesWith(IE, VecOp);
577 
578   // If the inserted element was extracted from some other vector, and if the
579   // indexes are constant, try to turn this into a shufflevector operation.
580   if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
581     if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) {
582       unsigned NumInsertVectorElts = IE.getType()->getNumElements();
583       unsigned NumExtractVectorElts =
584           EI->getOperand(0)->getType()->getVectorNumElements();
585       unsigned ExtractedIdx =
586         cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
587       unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
588 
589       if (ExtractedIdx >= NumExtractVectorElts) // Out of range extract.
590         return replaceInstUsesWith(IE, VecOp);
591 
592       if (InsertedIdx >= NumInsertVectorElts)  // Out of range insert.
593         return replaceInstUsesWith(IE, UndefValue::get(IE.getType()));
594 
595       // If we are extracting a value from a vector, then inserting it right
596       // back into the same place, just use the input vector.
597       if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
598         return replaceInstUsesWith(IE, VecOp);
599 
600       // If this insertelement isn't used by some other insertelement, turn it
601       // (and any insertelements it points to), into one big shuffle.
602       if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.user_back())) {
603         SmallVector<Constant*, 16> Mask;
604         ShuffleOps LR = collectShuffleElements(&IE, Mask, nullptr, *this);
605 
606         // The proposed shuffle may be trivial, in which case we shouldn't
607         // perform the combine.
608         if (LR.first != &IE && LR.second != &IE) {
609           // We now have a shuffle of LHS, RHS, Mask.
610           if (LR.second == nullptr)
611             LR.second = UndefValue::get(LR.first->getType());
612           return new ShuffleVectorInst(LR.first, LR.second,
613                                        ConstantVector::get(Mask));
614         }
615       }
616     }
617   }
618 
619   unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements();
620   APInt UndefElts(VWidth, 0);
621   APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
622   if (Value *V = SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts)) {
623     if (V != &IE)
624       return replaceInstUsesWith(IE, V);
625     return &IE;
626   }
627 
628   return nullptr;
629 }
630 
631 /// Return true if we can evaluate the specified expression tree if the vector
632 /// elements were shuffled in a different order.
CanEvaluateShuffled(Value * V,ArrayRef<int> Mask,unsigned Depth=5)633 static bool CanEvaluateShuffled(Value *V, ArrayRef<int> Mask,
634                                 unsigned Depth = 5) {
635   // We can always reorder the elements of a constant.
636   if (isa<Constant>(V))
637     return true;
638 
639   // We won't reorder vector arguments. No IPO here.
640   Instruction *I = dyn_cast<Instruction>(V);
641   if (!I) return false;
642 
643   // Two users may expect different orders of the elements. Don't try it.
644   if (!I->hasOneUse())
645     return false;
646 
647   if (Depth == 0) return false;
648 
649   switch (I->getOpcode()) {
650     case Instruction::Add:
651     case Instruction::FAdd:
652     case Instruction::Sub:
653     case Instruction::FSub:
654     case Instruction::Mul:
655     case Instruction::FMul:
656     case Instruction::UDiv:
657     case Instruction::SDiv:
658     case Instruction::FDiv:
659     case Instruction::URem:
660     case Instruction::SRem:
661     case Instruction::FRem:
662     case Instruction::Shl:
663     case Instruction::LShr:
664     case Instruction::AShr:
665     case Instruction::And:
666     case Instruction::Or:
667     case Instruction::Xor:
668     case Instruction::ICmp:
669     case Instruction::FCmp:
670     case Instruction::Trunc:
671     case Instruction::ZExt:
672     case Instruction::SExt:
673     case Instruction::FPToUI:
674     case Instruction::FPToSI:
675     case Instruction::UIToFP:
676     case Instruction::SIToFP:
677     case Instruction::FPTrunc:
678     case Instruction::FPExt:
679     case Instruction::GetElementPtr: {
680       for (Value *Operand : I->operands()) {
681         if (!CanEvaluateShuffled(Operand, Mask, Depth-1))
682           return false;
683       }
684       return true;
685     }
686     case Instruction::InsertElement: {
687       ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(2));
688       if (!CI) return false;
689       int ElementNumber = CI->getLimitedValue();
690 
691       // Verify that 'CI' does not occur twice in Mask. A single 'insertelement'
692       // can't put an element into multiple indices.
693       bool SeenOnce = false;
694       for (int i = 0, e = Mask.size(); i != e; ++i) {
695         if (Mask[i] == ElementNumber) {
696           if (SeenOnce)
697             return false;
698           SeenOnce = true;
699         }
700       }
701       return CanEvaluateShuffled(I->getOperand(0), Mask, Depth-1);
702     }
703   }
704   return false;
705 }
706 
707 /// Rebuild a new instruction just like 'I' but with the new operands given.
708 /// In the event of type mismatch, the type of the operands is correct.
buildNew(Instruction * I,ArrayRef<Value * > NewOps)709 static Value *buildNew(Instruction *I, ArrayRef<Value*> NewOps) {
710   // We don't want to use the IRBuilder here because we want the replacement
711   // instructions to appear next to 'I', not the builder's insertion point.
712   switch (I->getOpcode()) {
713     case Instruction::Add:
714     case Instruction::FAdd:
715     case Instruction::Sub:
716     case Instruction::FSub:
717     case Instruction::Mul:
718     case Instruction::FMul:
719     case Instruction::UDiv:
720     case Instruction::SDiv:
721     case Instruction::FDiv:
722     case Instruction::URem:
723     case Instruction::SRem:
724     case Instruction::FRem:
725     case Instruction::Shl:
726     case Instruction::LShr:
727     case Instruction::AShr:
728     case Instruction::And:
729     case Instruction::Or:
730     case Instruction::Xor: {
731       BinaryOperator *BO = cast<BinaryOperator>(I);
732       assert(NewOps.size() == 2 && "binary operator with #ops != 2");
733       BinaryOperator *New =
734           BinaryOperator::Create(cast<BinaryOperator>(I)->getOpcode(),
735                                  NewOps[0], NewOps[1], "", BO);
736       if (isa<OverflowingBinaryOperator>(BO)) {
737         New->setHasNoUnsignedWrap(BO->hasNoUnsignedWrap());
738         New->setHasNoSignedWrap(BO->hasNoSignedWrap());
739       }
740       if (isa<PossiblyExactOperator>(BO)) {
741         New->setIsExact(BO->isExact());
742       }
743       if (isa<FPMathOperator>(BO))
744         New->copyFastMathFlags(I);
745       return New;
746     }
747     case Instruction::ICmp:
748       assert(NewOps.size() == 2 && "icmp with #ops != 2");
749       return new ICmpInst(I, cast<ICmpInst>(I)->getPredicate(),
750                           NewOps[0], NewOps[1]);
751     case Instruction::FCmp:
752       assert(NewOps.size() == 2 && "fcmp with #ops != 2");
753       return new FCmpInst(I, cast<FCmpInst>(I)->getPredicate(),
754                           NewOps[0], NewOps[1]);
755     case Instruction::Trunc:
756     case Instruction::ZExt:
757     case Instruction::SExt:
758     case Instruction::FPToUI:
759     case Instruction::FPToSI:
760     case Instruction::UIToFP:
761     case Instruction::SIToFP:
762     case Instruction::FPTrunc:
763     case Instruction::FPExt: {
764       // It's possible that the mask has a different number of elements from
765       // the original cast. We recompute the destination type to match the mask.
766       Type *DestTy =
767           VectorType::get(I->getType()->getScalarType(),
768                           NewOps[0]->getType()->getVectorNumElements());
769       assert(NewOps.size() == 1 && "cast with #ops != 1");
770       return CastInst::Create(cast<CastInst>(I)->getOpcode(), NewOps[0], DestTy,
771                               "", I);
772     }
773     case Instruction::GetElementPtr: {
774       Value *Ptr = NewOps[0];
775       ArrayRef<Value*> Idx = NewOps.slice(1);
776       GetElementPtrInst *GEP = GetElementPtrInst::Create(
777           cast<GetElementPtrInst>(I)->getSourceElementType(), Ptr, Idx, "", I);
778       GEP->setIsInBounds(cast<GetElementPtrInst>(I)->isInBounds());
779       return GEP;
780     }
781   }
782   llvm_unreachable("failed to rebuild vector instructions");
783 }
784 
785 Value *
EvaluateInDifferentElementOrder(Value * V,ArrayRef<int> Mask)786 InstCombiner::EvaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask) {
787   // Mask.size() does not need to be equal to the number of vector elements.
788 
789   assert(V->getType()->isVectorTy() && "can't reorder non-vector elements");
790   if (isa<UndefValue>(V)) {
791     return UndefValue::get(VectorType::get(V->getType()->getScalarType(),
792                                            Mask.size()));
793   }
794   if (isa<ConstantAggregateZero>(V)) {
795     return ConstantAggregateZero::get(
796                VectorType::get(V->getType()->getScalarType(),
797                                Mask.size()));
798   }
799   if (Constant *C = dyn_cast<Constant>(V)) {
800     SmallVector<Constant *, 16> MaskValues;
801     for (int i = 0, e = Mask.size(); i != e; ++i) {
802       if (Mask[i] == -1)
803         MaskValues.push_back(UndefValue::get(Builder->getInt32Ty()));
804       else
805         MaskValues.push_back(Builder->getInt32(Mask[i]));
806     }
807     return ConstantExpr::getShuffleVector(C, UndefValue::get(C->getType()),
808                                           ConstantVector::get(MaskValues));
809   }
810 
811   Instruction *I = cast<Instruction>(V);
812   switch (I->getOpcode()) {
813     case Instruction::Add:
814     case Instruction::FAdd:
815     case Instruction::Sub:
816     case Instruction::FSub:
817     case Instruction::Mul:
818     case Instruction::FMul:
819     case Instruction::UDiv:
820     case Instruction::SDiv:
821     case Instruction::FDiv:
822     case Instruction::URem:
823     case Instruction::SRem:
824     case Instruction::FRem:
825     case Instruction::Shl:
826     case Instruction::LShr:
827     case Instruction::AShr:
828     case Instruction::And:
829     case Instruction::Or:
830     case Instruction::Xor:
831     case Instruction::ICmp:
832     case Instruction::FCmp:
833     case Instruction::Trunc:
834     case Instruction::ZExt:
835     case Instruction::SExt:
836     case Instruction::FPToUI:
837     case Instruction::FPToSI:
838     case Instruction::UIToFP:
839     case Instruction::SIToFP:
840     case Instruction::FPTrunc:
841     case Instruction::FPExt:
842     case Instruction::Select:
843     case Instruction::GetElementPtr: {
844       SmallVector<Value*, 8> NewOps;
845       bool NeedsRebuild = (Mask.size() != I->getType()->getVectorNumElements());
846       for (int i = 0, e = I->getNumOperands(); i != e; ++i) {
847         Value *V = EvaluateInDifferentElementOrder(I->getOperand(i), Mask);
848         NewOps.push_back(V);
849         NeedsRebuild |= (V != I->getOperand(i));
850       }
851       if (NeedsRebuild) {
852         return buildNew(I, NewOps);
853       }
854       return I;
855     }
856     case Instruction::InsertElement: {
857       int Element = cast<ConstantInt>(I->getOperand(2))->getLimitedValue();
858 
859       // The insertelement was inserting at Element. Figure out which element
860       // that becomes after shuffling. The answer is guaranteed to be unique
861       // by CanEvaluateShuffled.
862       bool Found = false;
863       int Index = 0;
864       for (int e = Mask.size(); Index != e; ++Index) {
865         if (Mask[Index] == Element) {
866           Found = true;
867           break;
868         }
869       }
870 
871       // If element is not in Mask, no need to handle the operand 1 (element to
872       // be inserted). Just evaluate values in operand 0 according to Mask.
873       if (!Found)
874         return EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
875 
876       Value *V = EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
877       return InsertElementInst::Create(V, I->getOperand(1),
878                                        Builder->getInt32(Index), "", I);
879     }
880   }
881   llvm_unreachable("failed to reorder elements of vector instruction!");
882 }
883 
recognizeIdentityMask(const SmallVectorImpl<int> & Mask,bool & isLHSID,bool & isRHSID)884 static void recognizeIdentityMask(const SmallVectorImpl<int> &Mask,
885                                   bool &isLHSID, bool &isRHSID) {
886   isLHSID = isRHSID = true;
887 
888   for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
889     if (Mask[i] < 0) continue;  // Ignore undef values.
890     // Is this an identity shuffle of the LHS value?
891     isLHSID &= (Mask[i] == (int)i);
892 
893     // Is this an identity shuffle of the RHS value?
894     isRHSID &= (Mask[i]-e == i);
895   }
896 }
897 
898 // Returns true if the shuffle is extracting a contiguous range of values from
899 // LHS, for example:
900 //                 +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
901 //   Input:        |AA|BB|CC|DD|EE|FF|GG|HH|II|JJ|KK|LL|MM|NN|OO|PP|
902 //   Shuffles to:  |EE|FF|GG|HH|
903 //                 +--+--+--+--+
isShuffleExtractingFromLHS(ShuffleVectorInst & SVI,SmallVector<int,16> & Mask)904 static bool isShuffleExtractingFromLHS(ShuffleVectorInst &SVI,
905                                        SmallVector<int, 16> &Mask) {
906   unsigned LHSElems =
907       cast<VectorType>(SVI.getOperand(0)->getType())->getNumElements();
908   unsigned MaskElems = Mask.size();
909   unsigned BegIdx = Mask.front();
910   unsigned EndIdx = Mask.back();
911   if (BegIdx > EndIdx || EndIdx >= LHSElems || EndIdx - BegIdx != MaskElems - 1)
912     return false;
913   for (unsigned I = 0; I != MaskElems; ++I)
914     if (static_cast<unsigned>(Mask[I]) != BegIdx + I)
915       return false;
916   return true;
917 }
918 
visitShuffleVectorInst(ShuffleVectorInst & SVI)919 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
920   Value *LHS = SVI.getOperand(0);
921   Value *RHS = SVI.getOperand(1);
922   SmallVector<int, 16> Mask = SVI.getShuffleMask();
923   Type *Int32Ty = Type::getInt32Ty(SVI.getContext());
924 
925   bool MadeChange = false;
926 
927   // Undefined shuffle mask -> undefined value.
928   if (isa<UndefValue>(SVI.getOperand(2)))
929     return replaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
930 
931   unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
932 
933   APInt UndefElts(VWidth, 0);
934   APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
935   if (Value *V = SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
936     if (V != &SVI)
937       return replaceInstUsesWith(SVI, V);
938     LHS = SVI.getOperand(0);
939     RHS = SVI.getOperand(1);
940     MadeChange = true;
941   }
942 
943   unsigned LHSWidth = cast<VectorType>(LHS->getType())->getNumElements();
944 
945   // Canonicalize shuffle(x    ,x,mask) -> shuffle(x, undef,mask')
946   // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
947   if (LHS == RHS || isa<UndefValue>(LHS)) {
948     if (isa<UndefValue>(LHS) && LHS == RHS) {
949       // shuffle(undef,undef,mask) -> undef.
950       Value *Result = (VWidth == LHSWidth)
951                       ? LHS : UndefValue::get(SVI.getType());
952       return replaceInstUsesWith(SVI, Result);
953     }
954 
955     // Remap any references to RHS to use LHS.
956     SmallVector<Constant*, 16> Elts;
957     for (unsigned i = 0, e = LHSWidth; i != VWidth; ++i) {
958       if (Mask[i] < 0) {
959         Elts.push_back(UndefValue::get(Int32Ty));
960         continue;
961       }
962 
963       if ((Mask[i] >= (int)e && isa<UndefValue>(RHS)) ||
964           (Mask[i] <  (int)e && isa<UndefValue>(LHS))) {
965         Mask[i] = -1;     // Turn into undef.
966         Elts.push_back(UndefValue::get(Int32Ty));
967       } else {
968         Mask[i] = Mask[i] % e;  // Force to LHS.
969         Elts.push_back(ConstantInt::get(Int32Ty, Mask[i]));
970       }
971     }
972     SVI.setOperand(0, SVI.getOperand(1));
973     SVI.setOperand(1, UndefValue::get(RHS->getType()));
974     SVI.setOperand(2, ConstantVector::get(Elts));
975     LHS = SVI.getOperand(0);
976     RHS = SVI.getOperand(1);
977     MadeChange = true;
978   }
979 
980   if (VWidth == LHSWidth) {
981     // Analyze the shuffle, are the LHS or RHS and identity shuffles?
982     bool isLHSID, isRHSID;
983     recognizeIdentityMask(Mask, isLHSID, isRHSID);
984 
985     // Eliminate identity shuffles.
986     if (isLHSID) return replaceInstUsesWith(SVI, LHS);
987     if (isRHSID) return replaceInstUsesWith(SVI, RHS);
988   }
989 
990   if (isa<UndefValue>(RHS) && CanEvaluateShuffled(LHS, Mask)) {
991     Value *V = EvaluateInDifferentElementOrder(LHS, Mask);
992     return replaceInstUsesWith(SVI, V);
993   }
994 
995   // SROA generates shuffle+bitcast when the extracted sub-vector is bitcast to
996   // a non-vector type. We can instead bitcast the original vector followed by
997   // an extract of the desired element:
998   //
999   //   %sroa = shufflevector <16 x i8> %in, <16 x i8> undef,
1000   //                         <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1001   //   %1 = bitcast <4 x i8> %sroa to i32
1002   // Becomes:
1003   //   %bc = bitcast <16 x i8> %in to <4 x i32>
1004   //   %ext = extractelement <4 x i32> %bc, i32 0
1005   //
1006   // If the shuffle is extracting a contiguous range of values from the input
1007   // vector then each use which is a bitcast of the extracted size can be
1008   // replaced. This will work if the vector types are compatible, and the begin
1009   // index is aligned to a value in the casted vector type. If the begin index
1010   // isn't aligned then we can shuffle the original vector (keeping the same
1011   // vector type) before extracting.
1012   //
1013   // This code will bail out if the target type is fundamentally incompatible
1014   // with vectors of the source type.
1015   //
1016   // Example of <16 x i8>, target type i32:
1017   // Index range [4,8):         v-----------v Will work.
1018   //                +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
1019   //     <16 x i8>: |  |  |  |  |  |  |  |  |  |  |  |  |  |  |  |  |
1020   //     <4 x i32>: |           |           |           |           |
1021   //                +-----------+-----------+-----------+-----------+
1022   // Index range [6,10):              ^-----------^ Needs an extra shuffle.
1023   // Target type i40:           ^--------------^ Won't work, bail.
1024   if (isShuffleExtractingFromLHS(SVI, Mask)) {
1025     Value *V = LHS;
1026     unsigned MaskElems = Mask.size();
1027     unsigned BegIdx = Mask.front();
1028     VectorType *SrcTy = cast<VectorType>(V->getType());
1029     unsigned VecBitWidth = SrcTy->getBitWidth();
1030     unsigned SrcElemBitWidth = DL.getTypeSizeInBits(SrcTy->getElementType());
1031     assert(SrcElemBitWidth && "vector elements must have a bitwidth");
1032     unsigned SrcNumElems = SrcTy->getNumElements();
1033     SmallVector<BitCastInst *, 8> BCs;
1034     DenseMap<Type *, Value *> NewBCs;
1035     for (User *U : SVI.users())
1036       if (BitCastInst *BC = dyn_cast<BitCastInst>(U))
1037         if (!BC->use_empty())
1038           // Only visit bitcasts that weren't previously handled.
1039           BCs.push_back(BC);
1040     for (BitCastInst *BC : BCs) {
1041       Type *TgtTy = BC->getDestTy();
1042       unsigned TgtElemBitWidth = DL.getTypeSizeInBits(TgtTy);
1043       if (!TgtElemBitWidth)
1044         continue;
1045       unsigned TgtNumElems = VecBitWidth / TgtElemBitWidth;
1046       bool VecBitWidthsEqual = VecBitWidth == TgtNumElems * TgtElemBitWidth;
1047       bool BegIsAligned = 0 == ((SrcElemBitWidth * BegIdx) % TgtElemBitWidth);
1048       if (!VecBitWidthsEqual)
1049         continue;
1050       if (!VectorType::isValidElementType(TgtTy))
1051         continue;
1052       VectorType *CastSrcTy = VectorType::get(TgtTy, TgtNumElems);
1053       if (!BegIsAligned) {
1054         // Shuffle the input so [0,NumElements) contains the output, and
1055         // [NumElems,SrcNumElems) is undef.
1056         SmallVector<Constant *, 16> ShuffleMask(SrcNumElems,
1057                                                 UndefValue::get(Int32Ty));
1058         for (unsigned I = 0, E = MaskElems, Idx = BegIdx; I != E; ++Idx, ++I)
1059           ShuffleMask[I] = ConstantInt::get(Int32Ty, Idx);
1060         V = Builder->CreateShuffleVector(V, UndefValue::get(V->getType()),
1061                                          ConstantVector::get(ShuffleMask),
1062                                          SVI.getName() + ".extract");
1063         BegIdx = 0;
1064       }
1065       unsigned SrcElemsPerTgtElem = TgtElemBitWidth / SrcElemBitWidth;
1066       assert(SrcElemsPerTgtElem);
1067       BegIdx /= SrcElemsPerTgtElem;
1068       bool BCAlreadyExists = NewBCs.find(CastSrcTy) != NewBCs.end();
1069       auto *NewBC =
1070           BCAlreadyExists
1071               ? NewBCs[CastSrcTy]
1072               : Builder->CreateBitCast(V, CastSrcTy, SVI.getName() + ".bc");
1073       if (!BCAlreadyExists)
1074         NewBCs[CastSrcTy] = NewBC;
1075       auto *Ext = Builder->CreateExtractElement(
1076           NewBC, ConstantInt::get(Int32Ty, BegIdx), SVI.getName() + ".extract");
1077       // The shufflevector isn't being replaced: the bitcast that used it
1078       // is. InstCombine will visit the newly-created instructions.
1079       replaceInstUsesWith(*BC, Ext);
1080       MadeChange = true;
1081     }
1082   }
1083 
1084   // If the LHS is a shufflevector itself, see if we can combine it with this
1085   // one without producing an unusual shuffle.
1086   // Cases that might be simplified:
1087   // 1.
1088   // x1=shuffle(v1,v2,mask1)
1089   //  x=shuffle(x1,undef,mask)
1090   //        ==>
1091   //  x=shuffle(v1,undef,newMask)
1092   // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : -1
1093   // 2.
1094   // x1=shuffle(v1,undef,mask1)
1095   //  x=shuffle(x1,x2,mask)
1096   // where v1.size() == mask1.size()
1097   //        ==>
1098   //  x=shuffle(v1,x2,newMask)
1099   // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : mask[i]
1100   // 3.
1101   // x2=shuffle(v2,undef,mask2)
1102   //  x=shuffle(x1,x2,mask)
1103   // where v2.size() == mask2.size()
1104   //        ==>
1105   //  x=shuffle(x1,v2,newMask)
1106   // newMask[i] = (mask[i] < x1.size())
1107   //              ? mask[i] : mask2[mask[i]-x1.size()]+x1.size()
1108   // 4.
1109   // x1=shuffle(v1,undef,mask1)
1110   // x2=shuffle(v2,undef,mask2)
1111   //  x=shuffle(x1,x2,mask)
1112   // where v1.size() == v2.size()
1113   //        ==>
1114   //  x=shuffle(v1,v2,newMask)
1115   // newMask[i] = (mask[i] < x1.size())
1116   //              ? mask1[mask[i]] : mask2[mask[i]-x1.size()]+v1.size()
1117   //
1118   // Here we are really conservative:
1119   // we are absolutely afraid of producing a shuffle mask not in the input
1120   // program, because the code gen may not be smart enough to turn a merged
1121   // shuffle into two specific shuffles: it may produce worse code.  As such,
1122   // we only merge two shuffles if the result is either a splat or one of the
1123   // input shuffle masks.  In this case, merging the shuffles just removes
1124   // one instruction, which we know is safe.  This is good for things like
1125   // turning: (splat(splat)) -> splat, or
1126   // merge(V[0..n], V[n+1..2n]) -> V[0..2n]
1127   ShuffleVectorInst* LHSShuffle = dyn_cast<ShuffleVectorInst>(LHS);
1128   ShuffleVectorInst* RHSShuffle = dyn_cast<ShuffleVectorInst>(RHS);
1129   if (LHSShuffle)
1130     if (!isa<UndefValue>(LHSShuffle->getOperand(1)) && !isa<UndefValue>(RHS))
1131       LHSShuffle = nullptr;
1132   if (RHSShuffle)
1133     if (!isa<UndefValue>(RHSShuffle->getOperand(1)))
1134       RHSShuffle = nullptr;
1135   if (!LHSShuffle && !RHSShuffle)
1136     return MadeChange ? &SVI : nullptr;
1137 
1138   Value* LHSOp0 = nullptr;
1139   Value* LHSOp1 = nullptr;
1140   Value* RHSOp0 = nullptr;
1141   unsigned LHSOp0Width = 0;
1142   unsigned RHSOp0Width = 0;
1143   if (LHSShuffle) {
1144     LHSOp0 = LHSShuffle->getOperand(0);
1145     LHSOp1 = LHSShuffle->getOperand(1);
1146     LHSOp0Width = cast<VectorType>(LHSOp0->getType())->getNumElements();
1147   }
1148   if (RHSShuffle) {
1149     RHSOp0 = RHSShuffle->getOperand(0);
1150     RHSOp0Width = cast<VectorType>(RHSOp0->getType())->getNumElements();
1151   }
1152   Value* newLHS = LHS;
1153   Value* newRHS = RHS;
1154   if (LHSShuffle) {
1155     // case 1
1156     if (isa<UndefValue>(RHS)) {
1157       newLHS = LHSOp0;
1158       newRHS = LHSOp1;
1159     }
1160     // case 2 or 4
1161     else if (LHSOp0Width == LHSWidth) {
1162       newLHS = LHSOp0;
1163     }
1164   }
1165   // case 3 or 4
1166   if (RHSShuffle && RHSOp0Width == LHSWidth) {
1167     newRHS = RHSOp0;
1168   }
1169   // case 4
1170   if (LHSOp0 == RHSOp0) {
1171     newLHS = LHSOp0;
1172     newRHS = nullptr;
1173   }
1174 
1175   if (newLHS == LHS && newRHS == RHS)
1176     return MadeChange ? &SVI : nullptr;
1177 
1178   SmallVector<int, 16> LHSMask;
1179   SmallVector<int, 16> RHSMask;
1180   if (newLHS != LHS)
1181     LHSMask = LHSShuffle->getShuffleMask();
1182   if (RHSShuffle && newRHS != RHS)
1183     RHSMask = RHSShuffle->getShuffleMask();
1184 
1185   unsigned newLHSWidth = (newLHS != LHS) ? LHSOp0Width : LHSWidth;
1186   SmallVector<int, 16> newMask;
1187   bool isSplat = true;
1188   int SplatElt = -1;
1189   // Create a new mask for the new ShuffleVectorInst so that the new
1190   // ShuffleVectorInst is equivalent to the original one.
1191   for (unsigned i = 0; i < VWidth; ++i) {
1192     int eltMask;
1193     if (Mask[i] < 0) {
1194       // This element is an undef value.
1195       eltMask = -1;
1196     } else if (Mask[i] < (int)LHSWidth) {
1197       // This element is from left hand side vector operand.
1198       //
1199       // If LHS is going to be replaced (case 1, 2, or 4), calculate the
1200       // new mask value for the element.
1201       if (newLHS != LHS) {
1202         eltMask = LHSMask[Mask[i]];
1203         // If the value selected is an undef value, explicitly specify it
1204         // with a -1 mask value.
1205         if (eltMask >= (int)LHSOp0Width && isa<UndefValue>(LHSOp1))
1206           eltMask = -1;
1207       } else
1208         eltMask = Mask[i];
1209     } else {
1210       // This element is from right hand side vector operand
1211       //
1212       // If the value selected is an undef value, explicitly specify it
1213       // with a -1 mask value. (case 1)
1214       if (isa<UndefValue>(RHS))
1215         eltMask = -1;
1216       // If RHS is going to be replaced (case 3 or 4), calculate the
1217       // new mask value for the element.
1218       else if (newRHS != RHS) {
1219         eltMask = RHSMask[Mask[i]-LHSWidth];
1220         // If the value selected is an undef value, explicitly specify it
1221         // with a -1 mask value.
1222         if (eltMask >= (int)RHSOp0Width) {
1223           assert(isa<UndefValue>(RHSShuffle->getOperand(1))
1224                  && "should have been check above");
1225           eltMask = -1;
1226         }
1227       } else
1228         eltMask = Mask[i]-LHSWidth;
1229 
1230       // If LHS's width is changed, shift the mask value accordingly.
1231       // If newRHS == NULL, i.e. LHSOp0 == RHSOp0, we want to remap any
1232       // references from RHSOp0 to LHSOp0, so we don't need to shift the mask.
1233       // If newRHS == newLHS, we want to remap any references from newRHS to
1234       // newLHS so that we can properly identify splats that may occur due to
1235       // obfuscation across the two vectors.
1236       if (eltMask >= 0 && newRHS != nullptr && newLHS != newRHS)
1237         eltMask += newLHSWidth;
1238     }
1239 
1240     // Check if this could still be a splat.
1241     if (eltMask >= 0) {
1242       if (SplatElt >= 0 && SplatElt != eltMask)
1243         isSplat = false;
1244       SplatElt = eltMask;
1245     }
1246 
1247     newMask.push_back(eltMask);
1248   }
1249 
1250   // If the result mask is equal to one of the original shuffle masks,
1251   // or is a splat, do the replacement.
1252   if (isSplat || newMask == LHSMask || newMask == RHSMask || newMask == Mask) {
1253     SmallVector<Constant*, 16> Elts;
1254     for (unsigned i = 0, e = newMask.size(); i != e; ++i) {
1255       if (newMask[i] < 0) {
1256         Elts.push_back(UndefValue::get(Int32Ty));
1257       } else {
1258         Elts.push_back(ConstantInt::get(Int32Ty, newMask[i]));
1259       }
1260     }
1261     if (!newRHS)
1262       newRHS = UndefValue::get(newLHS->getType());
1263     return new ShuffleVectorInst(newLHS, newRHS, ConstantVector::get(Elts));
1264   }
1265 
1266   // If the result mask is an identity, replace uses of this instruction with
1267   // corresponding argument.
1268   bool isLHSID, isRHSID;
1269   recognizeIdentityMask(newMask, isLHSID, isRHSID);
1270   if (isLHSID && VWidth == LHSOp0Width) return replaceInstUsesWith(SVI, newLHS);
1271   if (isRHSID && VWidth == RHSOp0Width) return replaceInstUsesWith(SVI, newRHS);
1272 
1273   return MadeChange ? &SVI : nullptr;
1274 }
1275