33 }
34 Instruction *I = dyn_cast<Instruction>(V);
35 if (!I) return false;
36
37 // Insert element gets simplified to the inserted element or is deleted if
38 // this is constant idx extract element and its a constant idx insertelt.
39 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
40 isa<ConstantInt>(I->getOperand(2)))
41 return true;
42 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
43 return true;
44 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
45 if (BO->hasOneUse() &&
46 (CheapToScalarize(BO->getOperand(0), isConstant) ||
47 CheapToScalarize(BO->getOperand(1), isConstant)))
48 return true;
49 if (CmpInst *CI = dyn_cast<CmpInst>(I))
50 if (CI->hasOneUse() &&
51 (CheapToScalarize(CI->getOperand(0), isConstant) ||
52 CheapToScalarize(CI->getOperand(1), isConstant)))
53 return true;
54
55 return false;
56}
57
58/// FindScalarElement - Given a vector and an element number, see if the scalar
59/// value is already around as a register, for example if it were inserted then
60/// extracted from the vector.
61static Value *FindScalarElement(Value *V, unsigned EltNo) {
62 assert(V->getType()->isVectorTy() && "Not looking at a vector?");
63 VectorType *VTy = cast<VectorType>(V->getType());
64 unsigned Width = VTy->getNumElements();
65 if (EltNo >= Width) // Out of range access.
66 return UndefValue::get(VTy->getElementType());
67
68 if (Constant *C = dyn_cast<Constant>(V))
69 return C->getAggregateElement(EltNo);
70
71 if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
72 // If this is an insert to a variable element, we don't know what it is.
73 if (!isa<ConstantInt>(III->getOperand(2)))
74 return 0;
75 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
76
77 // If this is an insert to the element we are looking for, return the
78 // inserted value.
79 if (EltNo == IIElt)
80 return III->getOperand(1);
81
82 // Otherwise, the insertelement doesn't modify the value, recurse on its
83 // vector input.
84 return FindScalarElement(III->getOperand(0), EltNo);
85 }
86
87 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
88 unsigned LHSWidth = SVI->getOperand(0)->getType()->getVectorNumElements();
89 int InEl = SVI->getMaskValue(EltNo);
90 if (InEl < 0)
91 return UndefValue::get(VTy->getElementType());
92 if (InEl < (int)LHSWidth)
93 return FindScalarElement(SVI->getOperand(0), InEl);
94 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth);
95 }
96
97 // Extract a value from a vector add operation with a constant zero.
98 Value *Val = 0; Constant *Con = 0;
99 if (match(V, m_Add(m_Value(Val), m_Constant(Con)))) {
100 if (Con->getAggregateElement(EltNo)->isNullValue())
101 return FindScalarElement(Val, EltNo);
102 }
103
104 // Otherwise, we don't know.
105 return 0;
106}
107
108// If we have a PHI node with a vector type that has only 2 uses: feed
109// itself and be an operand of extractelement at a constant location,
110// try to replace the PHI of the vector type with a PHI of a scalar type.
111Instruction *InstCombiner::scalarizePHI(ExtractElementInst &EI, PHINode *PN) {
112 // Verify that the PHI node has exactly 2 uses. Otherwise return NULL.
113 if (!PN->hasNUses(2))
114 return NULL;
115
116 // If so, it's known at this point that one operand is PHI and the other is
117 // an extractelement node. Find the PHI user that is not the extractelement
118 // node.
119 Value::use_iterator iu = PN->use_begin();
120 Instruction *PHIUser = dyn_cast<Instruction>(*iu);
121 if (PHIUser == cast<Instruction>(&EI))
122 PHIUser = cast<Instruction>(*(++iu));
123
124 // Verify that this PHI user has one use, which is the PHI itself,
125 // and that it is a binary operation which is cheap to scalarize.
126 // otherwise return NULL.
127 if (!PHIUser->hasOneUse() || !(PHIUser->use_back() == PN) ||
128 !(isa<BinaryOperator>(PHIUser)) || !CheapToScalarize(PHIUser, true))
129 return NULL;
130
131 // Create a scalar PHI node that will replace the vector PHI node
132 // just before the current PHI node.
133 PHINode *scalarPHI = cast<PHINode>(InsertNewInstWith(
134 PHINode::Create(EI.getType(), PN->getNumIncomingValues(), ""), *PN));
135 // Scalarize each PHI operand.
136 for (unsigned i = 0; i < PN->getNumIncomingValues(); i++) {
137 Value *PHIInVal = PN->getIncomingValue(i);
138 BasicBlock *inBB = PN->getIncomingBlock(i);
139 Value *Elt = EI.getIndexOperand();
140 // If the operand is the PHI induction variable:
141 if (PHIInVal == PHIUser) {
142 // Scalarize the binary operation. Its first operand is the
143 // scalar PHI and the second operand is extracted from the other
144 // vector operand.
145 BinaryOperator *B0 = cast<BinaryOperator>(PHIUser);
146 unsigned opId = (B0->getOperand(0) == PN) ? 1 : 0;
147 Value *Op = InsertNewInstWith(
148 ExtractElementInst::Create(B0->getOperand(opId), Elt,
149 B0->getOperand(opId)->getName() + ".Elt"),
150 *B0);
151 Value *newPHIUser = InsertNewInstWith(
152 BinaryOperator::Create(B0->getOpcode(), scalarPHI, Op), *B0);
153 scalarPHI->addIncoming(newPHIUser, inBB);
154 } else {
155 // Scalarize PHI input:
156 Instruction *newEI = ExtractElementInst::Create(PHIInVal, Elt, "");
157 // Insert the new instruction into the predecessor basic block.
158 Instruction *pos = dyn_cast<Instruction>(PHIInVal);
159 BasicBlock::iterator InsertPos;
160 if (pos && !isa<PHINode>(pos)) {
161 InsertPos = pos;
162 ++InsertPos;
163 } else {
164 InsertPos = inBB->getFirstInsertionPt();
165 }
166
167 InsertNewInstWith(newEI, *InsertPos);
168
169 scalarPHI->addIncoming(newEI, inBB);
170 }
171 }
172 return ReplaceInstUsesWith(EI, scalarPHI);
173}
174
175Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
176 // If vector val is constant with all elements the same, replace EI with
177 // that element. We handle a known element # below.
178 if (Constant *C = dyn_cast<Constant>(EI.getOperand(0)))
179 if (CheapToScalarize(C, false))
180 return ReplaceInstUsesWith(EI, C->getAggregateElement(0U));
181
182 // If extracting a specified index from the vector, see if we can recursively
183 // find a previously computed scalar that was inserted into the vector.
184 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
185 unsigned IndexVal = IdxC->getZExtValue();
186 unsigned VectorWidth = EI.getVectorOperandType()->getNumElements();
187
188 // If this is extracting an invalid index, turn this into undef, to avoid
189 // crashing the code below.
190 if (IndexVal >= VectorWidth)
191 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
192
193 // This instruction only demands the single element from the input vector.
194 // If the input vector has a single use, simplify it based on this use
195 // property.
196 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
197 APInt UndefElts(VectorWidth, 0);
198 APInt DemandedMask(VectorWidth, 0);
199 DemandedMask.setBit(IndexVal);
200 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
201 DemandedMask, UndefElts)) {
202 EI.setOperand(0, V);
203 return &EI;
204 }
205 }
206
207 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
208 return ReplaceInstUsesWith(EI, Elt);
209
210 // If the this extractelement is directly using a bitcast from a vector of
211 // the same number of elements, see if we can find the source element from
212 // it. In this case, we will end up needing to bitcast the scalars.
213 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
214 if (VectorType *VT = dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
215 if (VT->getNumElements() == VectorWidth)
216 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
217 return new BitCastInst(Elt, EI.getType());
218 }
219
220 // If there's a vector PHI feeding a scalar use through this extractelement
221 // instruction, try to scalarize the PHI.
222 if (PHINode *PN = dyn_cast<PHINode>(EI.getOperand(0))) {
223 Instruction *scalarPHI = scalarizePHI(EI, PN);
224 if (scalarPHI)
225 return scalarPHI;
226 }
227 }
228
229 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
230 // Push extractelement into predecessor operation if legal and
231 // profitable to do so
232 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
233 if (I->hasOneUse() &&
234 CheapToScalarize(BO, isa<ConstantInt>(EI.getOperand(1)))) {
235 Value *newEI0 =
236 Builder->CreateExtractElement(BO->getOperand(0), EI.getOperand(1),
237 EI.getName()+".lhs");
238 Value *newEI1 =
239 Builder->CreateExtractElement(BO->getOperand(1), EI.getOperand(1),
240 EI.getName()+".rhs");
241 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
242 }
243 } else if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
244 // Extracting the inserted element?
245 if (IE->getOperand(2) == EI.getOperand(1))
246 return ReplaceInstUsesWith(EI, IE->getOperand(1));
247 // If the inserted and extracted elements are constants, they must not
248 // be the same value, extract from the pre-inserted value instead.
249 if (isa<Constant>(IE->getOperand(2)) && isa<Constant>(EI.getOperand(1))) {
250 Worklist.AddValue(EI.getOperand(0));
251 EI.setOperand(0, IE->getOperand(0));
252 return &EI;
253 }
254 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
255 // If this is extracting an element from a shufflevector, figure out where
256 // it came from and extract from the appropriate input element instead.
257 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
258 int SrcIdx = SVI->getMaskValue(Elt->getZExtValue());
259 Value *Src;
260 unsigned LHSWidth =
261 SVI->getOperand(0)->getType()->getVectorNumElements();
262
263 if (SrcIdx < 0)
264 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
265 if (SrcIdx < (int)LHSWidth)
266 Src = SVI->getOperand(0);
267 else {
268 SrcIdx -= LHSWidth;
269 Src = SVI->getOperand(1);
270 }
271 Type *Int32Ty = Type::getInt32Ty(EI.getContext());
272 return ExtractElementInst::Create(Src,
273 ConstantInt::get(Int32Ty,
274 SrcIdx, false));
275 }
276 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
277 // Canonicalize extractelement(cast) -> cast(extractelement)
278 // bitcasts can change the number of vector elements and they cost nothing
279 if (CI->hasOneUse() && (CI->getOpcode() != Instruction::BitCast)) {
280 Value *EE = Builder->CreateExtractElement(CI->getOperand(0),
281 EI.getIndexOperand());
282 Worklist.AddValue(EE);
283 return CastInst::Create(CI->getOpcode(), EE, EI.getType());
284 }
285 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
286 if (SI->hasOneUse()) {
287 // TODO: For a select on vectors, it might be useful to do this if it
288 // has multiple extractelement uses. For vector select, that seems to
289 // fight the vectorizer.
290
291 // If we are extracting an element from a vector select or a select on
292 // vectors, a select on the scalars extracted from the vector arguments.
293 Value *TrueVal = SI->getTrueValue();
294 Value *FalseVal = SI->getFalseValue();
295
296 Value *Cond = SI->getCondition();
297 if (Cond->getType()->isVectorTy()) {
298 Cond = Builder->CreateExtractElement(Cond,
299 EI.getIndexOperand(),
300 Cond->getName() + ".elt");
301 }
302
303 Value *V1Elem
304 = Builder->CreateExtractElement(TrueVal,
305 EI.getIndexOperand(),
306 TrueVal->getName() + ".elt");
307
308 Value *V2Elem
309 = Builder->CreateExtractElement(FalseVal,
310 EI.getIndexOperand(),
311 FalseVal->getName() + ".elt");
312 return SelectInst::Create(Cond,
313 V1Elem,
314 V2Elem,
315 SI->getName() + ".elt");
316 }
317 }
318 }
319 return 0;
320}
321
322/// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
323/// elements from either LHS or RHS, return the shuffle mask and true.
324/// Otherwise, return false.
325static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
326 SmallVectorImpl<Constant*> &Mask) {
327 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
328 "Invalid CollectSingleShuffleElements");
329 unsigned NumElts = V->getType()->getVectorNumElements();
330
331 if (isa<UndefValue>(V)) {
332 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
333 return true;
334 }
335
336 if (V == LHS) {
337 for (unsigned i = 0; i != NumElts; ++i)
338 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
339 return true;
340 }
341
342 if (V == RHS) {
343 for (unsigned i = 0; i != NumElts; ++i)
344 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()),
345 i+NumElts));
346 return true;
347 }
348
349 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
350 // If this is an insert of an extract from some other vector, include it.
351 Value *VecOp = IEI->getOperand(0);
352 Value *ScalarOp = IEI->getOperand(1);
353 Value *IdxOp = IEI->getOperand(2);
354
355 if (!isa<ConstantInt>(IdxOp))
356 return false;
357 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
358
359 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
360 // Okay, we can handle this if the vector we are insertinting into is
361 // transitively ok.
362 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
363 // If so, update the mask to reflect the inserted undef.
364 Mask[InsertedIdx] = UndefValue::get(Type::getInt32Ty(V->getContext()));
365 return true;
366 }
367 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
368 if (isa<ConstantInt>(EI->getOperand(1)) &&
369 EI->getOperand(0)->getType() == V->getType()) {
370 unsigned ExtractedIdx =
371 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
372
373 // This must be extracting from either LHS or RHS.
374 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
375 // Okay, we can handle this if the vector we are insertinting into is
376 // transitively ok.
377 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
378 // If so, update the mask to reflect the inserted value.
379 if (EI->getOperand(0) == LHS) {
380 Mask[InsertedIdx % NumElts] =
381 ConstantInt::get(Type::getInt32Ty(V->getContext()),
382 ExtractedIdx);
383 } else {
384 assert(EI->getOperand(0) == RHS);
385 Mask[InsertedIdx % NumElts] =
386 ConstantInt::get(Type::getInt32Ty(V->getContext()),
387 ExtractedIdx+NumElts);
388 }
389 return true;
390 }
391 }
392 }
393 }
394 }
395 // TODO: Handle shufflevector here!
396
397 return false;
398}
399
400/// CollectShuffleElements - We are building a shuffle of V, using RHS as the
401/// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
402/// that computes V and the LHS value of the shuffle.
403static Value *CollectShuffleElements(Value *V, SmallVectorImpl<Constant*> &Mask,
404 Value *&RHS) {
405 assert(V->getType()->isVectorTy() &&
406 (RHS == 0 || V->getType() == RHS->getType()) &&
407 "Invalid shuffle!");
408 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
409
410 if (isa<UndefValue>(V)) {
411 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
412 return V;
413 }
414
415 if (isa<ConstantAggregateZero>(V)) {
416 Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(V->getContext()),0));
417 return V;
418 }
419
420 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
421 // If this is an insert of an extract from some other vector, include it.
422 Value *VecOp = IEI->getOperand(0);
423 Value *ScalarOp = IEI->getOperand(1);
424 Value *IdxOp = IEI->getOperand(2);
425
426 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
427 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
428 EI->getOperand(0)->getType() == V->getType()) {
429 unsigned ExtractedIdx =
430 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
431 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
432
433 // Either the extracted from or inserted into vector must be RHSVec,
434 // otherwise we'd end up with a shuffle of three inputs.
435 if (EI->getOperand(0) == RHS || RHS == 0) {
436 RHS = EI->getOperand(0);
437 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
438 Mask[InsertedIdx % NumElts] =
439 ConstantInt::get(Type::getInt32Ty(V->getContext()),
440 NumElts+ExtractedIdx);
441 return V;
442 }
443
444 if (VecOp == RHS) {
445 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
446 // Update Mask to reflect that `ScalarOp' has been inserted at
447 // position `InsertedIdx' within the vector returned by IEI.
448 Mask[InsertedIdx % NumElts] = Mask[ExtractedIdx];
449
450 // Everything but the extracted element is replaced with the RHS.
451 for (unsigned i = 0; i != NumElts; ++i) {
452 if (i != InsertedIdx)
453 Mask[i] = ConstantInt::get(Type::getInt32Ty(V->getContext()),
454 NumElts+i);
455 }
456 return V;
457 }
458
459 // If this insertelement is a chain that comes from exactly these two
460 // vectors, return the vector and the effective shuffle.
461 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
462 return EI->getOperand(0);
463 }
464 }
465 }
466 // TODO: Handle shufflevector here!
467
468 // Otherwise, can't do anything fancy. Return an identity vector.
469 for (unsigned i = 0; i != NumElts; ++i)
470 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
471 return V;
472}
473
474Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
475 Value *VecOp = IE.getOperand(0);
476 Value *ScalarOp = IE.getOperand(1);
477 Value *IdxOp = IE.getOperand(2);
478
479 // Inserting an undef or into an undefined place, remove this.
480 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
481 ReplaceInstUsesWith(IE, VecOp);
482
483 // If the inserted element was extracted from some other vector, and if the
484 // indexes are constant, try to turn this into a shufflevector operation.
485 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
486 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
487 EI->getOperand(0)->getType() == IE.getType()) {
488 unsigned NumVectorElts = IE.getType()->getNumElements();
489 unsigned ExtractedIdx =
490 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
491 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
492
493 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
494 return ReplaceInstUsesWith(IE, VecOp);
495
496 if (InsertedIdx >= NumVectorElts) // Out of range insert.
497 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
498
499 // If we are extracting a value from a vector, then inserting it right
500 // back into the same place, just use the input vector.
501 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
502 return ReplaceInstUsesWith(IE, VecOp);
503
504 // If this insertelement isn't used by some other insertelement, turn it
505 // (and any insertelements it points to), into one big shuffle.
506 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
507 SmallVector<Constant*, 16> Mask;
508 Value *RHS = 0;
509 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
510 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
511 // We now have a shuffle of LHS, RHS, Mask.
512 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
513 }
514 }
515 }
516
517 unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements();
518 APInt UndefElts(VWidth, 0);
519 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
520 if (Value *V = SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts)) {
521 if (V != &IE)
522 return ReplaceInstUsesWith(IE, V);
523 return &IE;
524 }
525
526 return 0;
527}
528
529/// Return true if we can evaluate the specified expression tree if the vector
530/// elements were shuffled in a different order.
531static bool CanEvaluateShuffled(Value *V, ArrayRef<int> Mask,
532 unsigned Depth = 5) {
533 // We can always reorder the elements of a constant.
534 if (isa<Constant>(V))
535 return true;
536
537 // We won't reorder vector arguments. No IPO here.
538 Instruction *I = dyn_cast<Instruction>(V);
539 if (!I) return false;
540
541 // Two users may expect different orders of the elements. Don't try it.
542 if (!I->hasOneUse())
543 return false;
544
545 if (Depth == 0) return false;
546
547 switch (I->getOpcode()) {
548 case Instruction::Add:
549 case Instruction::FAdd:
550 case Instruction::Sub:
551 case Instruction::FSub:
552 case Instruction::Mul:
553 case Instruction::FMul:
554 case Instruction::UDiv:
555 case Instruction::SDiv:
556 case Instruction::FDiv:
557 case Instruction::URem:
558 case Instruction::SRem:
559 case Instruction::FRem:
560 case Instruction::Shl:
561 case Instruction::LShr:
562 case Instruction::AShr:
563 case Instruction::And:
564 case Instruction::Or:
565 case Instruction::Xor:
566 case Instruction::ICmp:
567 case Instruction::FCmp:
568 case Instruction::Trunc:
569 case Instruction::ZExt:
570 case Instruction::SExt:
571 case Instruction::FPToUI:
572 case Instruction::FPToSI:
573 case Instruction::UIToFP:
574 case Instruction::SIToFP:
575 case Instruction::FPTrunc:
576 case Instruction::FPExt:
577 case Instruction::GetElementPtr: {
578 for (int i = 0, e = I->getNumOperands(); i != e; ++i) {
579 if (!CanEvaluateShuffled(I->getOperand(i), Mask, Depth-1))
580 return false;
581 }
582 return true;
583 }
584 case Instruction::InsertElement: {
585 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(2));
586 if (!CI) return false;
587 int ElementNumber = CI->getLimitedValue();
588
589 // Verify that 'CI' does not occur twice in Mask. A single 'insertelement'
590 // can't put an element into multiple indices.
591 bool SeenOnce = false;
592 for (int i = 0, e = Mask.size(); i != e; ++i) {
593 if (Mask[i] == ElementNumber) {
594 if (SeenOnce)
595 return false;
596 SeenOnce = true;
597 }
598 }
599 return CanEvaluateShuffled(I->getOperand(0), Mask, Depth-1);
600 }
601 }
602 return false;
603}
604
605/// Rebuild a new instruction just like 'I' but with the new operands given.
606/// In the event of type mismatch, the type of the operands is correct.
607static Value *BuildNew(Instruction *I, ArrayRef<Value*> NewOps) {
608 // We don't want to use the IRBuilder here because we want the replacement
609 // instructions to appear next to 'I', not the builder's insertion point.
610 switch (I->getOpcode()) {
611 case Instruction::Add:
612 case Instruction::FAdd:
613 case Instruction::Sub:
614 case Instruction::FSub:
615 case Instruction::Mul:
616 case Instruction::FMul:
617 case Instruction::UDiv:
618 case Instruction::SDiv:
619 case Instruction::FDiv:
620 case Instruction::URem:
621 case Instruction::SRem:
622 case Instruction::FRem:
623 case Instruction::Shl:
624 case Instruction::LShr:
625 case Instruction::AShr:
626 case Instruction::And:
627 case Instruction::Or:
628 case Instruction::Xor: {
629 BinaryOperator *BO = cast<BinaryOperator>(I);
630 assert(NewOps.size() == 2 && "binary operator with #ops != 2");
631 BinaryOperator *New =
632 BinaryOperator::Create(cast<BinaryOperator>(I)->getOpcode(),
633 NewOps[0], NewOps[1], "", BO);
634 if (isa<OverflowingBinaryOperator>(BO)) {
635 New->setHasNoUnsignedWrap(BO->hasNoUnsignedWrap());
636 New->setHasNoSignedWrap(BO->hasNoSignedWrap());
637 }
638 if (isa<PossiblyExactOperator>(BO)) {
639 New->setIsExact(BO->isExact());
640 }
641 return New;
642 }
643 case Instruction::ICmp:
644 assert(NewOps.size() == 2 && "icmp with #ops != 2");
645 return new ICmpInst(I, cast<ICmpInst>(I)->getPredicate(),
646 NewOps[0], NewOps[1]);
647 case Instruction::FCmp:
648 assert(NewOps.size() == 2 && "fcmp with #ops != 2");
649 return new FCmpInst(I, cast<FCmpInst>(I)->getPredicate(),
650 NewOps[0], NewOps[1]);
651 case Instruction::Trunc:
652 case Instruction::ZExt:
653 case Instruction::SExt:
654 case Instruction::FPToUI:
655 case Instruction::FPToSI:
656 case Instruction::UIToFP:
657 case Instruction::SIToFP:
658 case Instruction::FPTrunc:
659 case Instruction::FPExt: {
660 // It's possible that the mask has a different number of elements from
661 // the original cast. We recompute the destination type to match the mask.
662 Type *DestTy =
663 VectorType::get(I->getType()->getScalarType(),
664 NewOps[0]->getType()->getVectorNumElements());
665 assert(NewOps.size() == 1 && "cast with #ops != 1");
666 return CastInst::Create(cast<CastInst>(I)->getOpcode(), NewOps[0], DestTy,
667 "", I);
668 }
669 case Instruction::GetElementPtr: {
670 Value *Ptr = NewOps[0];
671 ArrayRef<Value*> Idx = NewOps.slice(1);
672 GetElementPtrInst *GEP = GetElementPtrInst::Create(Ptr, Idx, "", I);
673 GEP->setIsInBounds(cast<GetElementPtrInst>(I)->isInBounds());
674 return GEP;
675 }
676 }
677 llvm_unreachable("failed to rebuild vector instructions");
678}
679
680Value *
681InstCombiner::EvaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask) {
682 // Mask.size() does not need to be equal to the number of vector elements.
683
684 assert(V->getType()->isVectorTy() && "can't reorder non-vector elements");
685 if (isa<UndefValue>(V)) {
686 return UndefValue::get(VectorType::get(V->getType()->getScalarType(),
687 Mask.size()));
688 }
689 if (isa<ConstantAggregateZero>(V)) {
690 return ConstantAggregateZero::get(
691 VectorType::get(V->getType()->getScalarType(),
692 Mask.size()));
693 }
694 if (Constant *C = dyn_cast<Constant>(V)) {
695 SmallVector<Constant *, 16> MaskValues;
696 for (int i = 0, e = Mask.size(); i != e; ++i) {
697 if (Mask[i] == -1)
698 MaskValues.push_back(UndefValue::get(Builder->getInt32Ty()));
699 else
700 MaskValues.push_back(Builder->getInt32(Mask[i]));
701 }
702 return ConstantExpr::getShuffleVector(C, UndefValue::get(C->getType()),
703 ConstantVector::get(MaskValues));
704 }
705
706 Instruction *I = cast<Instruction>(V);
707 switch (I->getOpcode()) {
708 case Instruction::Add:
709 case Instruction::FAdd:
710 case Instruction::Sub:
711 case Instruction::FSub:
712 case Instruction::Mul:
713 case Instruction::FMul:
714 case Instruction::UDiv:
715 case Instruction::SDiv:
716 case Instruction::FDiv:
717 case Instruction::URem:
718 case Instruction::SRem:
719 case Instruction::FRem:
720 case Instruction::Shl:
721 case Instruction::LShr:
722 case Instruction::AShr:
723 case Instruction::And:
724 case Instruction::Or:
725 case Instruction::Xor:
726 case Instruction::ICmp:
727 case Instruction::FCmp:
728 case Instruction::Trunc:
729 case Instruction::ZExt:
730 case Instruction::SExt:
731 case Instruction::FPToUI:
732 case Instruction::FPToSI:
733 case Instruction::UIToFP:
734 case Instruction::SIToFP:
735 case Instruction::FPTrunc:
736 case Instruction::FPExt:
737 case Instruction::Select:
738 case Instruction::GetElementPtr: {
739 SmallVector<Value*, 8> NewOps;
740 bool NeedsRebuild = (Mask.size() != I->getType()->getVectorNumElements());
741 for (int i = 0, e = I->getNumOperands(); i != e; ++i) {
742 Value *V = EvaluateInDifferentElementOrder(I->getOperand(i), Mask);
743 NewOps.push_back(V);
744 NeedsRebuild |= (V != I->getOperand(i));
745 }
746 if (NeedsRebuild) {
747 return BuildNew(I, NewOps);
748 }
749 return I;
750 }
751 case Instruction::InsertElement: {
752 int Element = cast<ConstantInt>(I->getOperand(2))->getLimitedValue();
753
754 // The insertelement was inserting at Element. Figure out which element
755 // that becomes after shuffling. The answer is guaranteed to be unique
756 // by CanEvaluateShuffled.
757 bool Found = false;
758 int Index = 0;
759 for (int e = Mask.size(); Index != e; ++Index) {
760 if (Mask[Index] == Element) {
761 Found = true;
762 break;
763 }
764 }
765
766 if (!Found)
767 return UndefValue::get(
768 VectorType::get(V->getType()->getScalarType(), Mask.size()));
769
770 Value *V = EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
771 return InsertElementInst::Create(V, I->getOperand(1),
772 Builder->getInt32(Index), "", I);
773 }
774 }
775 llvm_unreachable("failed to reorder elements of vector instruction!");
776}
777
778Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
779 Value *LHS = SVI.getOperand(0);
780 Value *RHS = SVI.getOperand(1);
781 SmallVector<int, 16> Mask = SVI.getShuffleMask();
782
783 bool MadeChange = false;
784
785 // Undefined shuffle mask -> undefined value.
786 if (isa<UndefValue>(SVI.getOperand(2)))
787 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
788
789 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
790
791 APInt UndefElts(VWidth, 0);
792 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
793 if (Value *V = SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
794 if (V != &SVI)
795 return ReplaceInstUsesWith(SVI, V);
796 LHS = SVI.getOperand(0);
797 RHS = SVI.getOperand(1);
798 MadeChange = true;
799 }
800
801 unsigned LHSWidth = cast<VectorType>(LHS->getType())->getNumElements();
802
803 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
804 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
805 if (LHS == RHS || isa<UndefValue>(LHS)) {
806 if (isa<UndefValue>(LHS) && LHS == RHS) {
807 // shuffle(undef,undef,mask) -> undef.
808 Value *Result = (VWidth == LHSWidth)
809 ? LHS : UndefValue::get(SVI.getType());
810 return ReplaceInstUsesWith(SVI, Result);
811 }
812
813 // Remap any references to RHS to use LHS.
814 SmallVector<Constant*, 16> Elts;
815 for (unsigned i = 0, e = LHSWidth; i != VWidth; ++i) {
816 if (Mask[i] < 0) {
817 Elts.push_back(UndefValue::get(Type::getInt32Ty(SVI.getContext())));
818 continue;
819 }
820
821 if ((Mask[i] >= (int)e && isa<UndefValue>(RHS)) ||
822 (Mask[i] < (int)e && isa<UndefValue>(LHS))) {
823 Mask[i] = -1; // Turn into undef.
824 Elts.push_back(UndefValue::get(Type::getInt32Ty(SVI.getContext())));
825 } else {
826 Mask[i] = Mask[i] % e; // Force to LHS.
827 Elts.push_back(ConstantInt::get(Type::getInt32Ty(SVI.getContext()),
828 Mask[i]));
829 }
830 }
831 SVI.setOperand(0, SVI.getOperand(1));
832 SVI.setOperand(1, UndefValue::get(RHS->getType()));
833 SVI.setOperand(2, ConstantVector::get(Elts));
834 LHS = SVI.getOperand(0);
835 RHS = SVI.getOperand(1);
836 MadeChange = true;
837 }
838
839 if (VWidth == LHSWidth) {
840 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
841 bool isLHSID = true, isRHSID = true;
842
843 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
844 if (Mask[i] < 0) continue; // Ignore undef values.
845 // Is this an identity shuffle of the LHS value?
846 isLHSID &= (Mask[i] == (int)i);
847
848 // Is this an identity shuffle of the RHS value?
849 isRHSID &= (Mask[i]-e == i);
850 }
851
852 // Eliminate identity shuffles.
853 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
854 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
855 }
856
857 if (isa<UndefValue>(RHS) && CanEvaluateShuffled(LHS, Mask)) {
858 Value *V = EvaluateInDifferentElementOrder(LHS, Mask);
859 return ReplaceInstUsesWith(SVI, V);
860 }
861
862 // If the LHS is a shufflevector itself, see if we can combine it with this
863 // one without producing an unusual shuffle.
864 // Cases that might be simplified:
865 // 1.
866 // x1=shuffle(v1,v2,mask1)
867 // x=shuffle(x1,undef,mask)
868 // ==>
869 // x=shuffle(v1,undef,newMask)
870 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : -1
871 // 2.
872 // x1=shuffle(v1,undef,mask1)
873 // x=shuffle(x1,x2,mask)
874 // where v1.size() == mask1.size()
875 // ==>
876 // x=shuffle(v1,x2,newMask)
877 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : mask[i]
878 // 3.
879 // x2=shuffle(v2,undef,mask2)
880 // x=shuffle(x1,x2,mask)
881 // where v2.size() == mask2.size()
882 // ==>
883 // x=shuffle(x1,v2,newMask)
884 // newMask[i] = (mask[i] < x1.size())
885 // ? mask[i] : mask2[mask[i]-x1.size()]+x1.size()
886 // 4.
887 // x1=shuffle(v1,undef,mask1)
888 // x2=shuffle(v2,undef,mask2)
889 // x=shuffle(x1,x2,mask)
890 // where v1.size() == v2.size()
891 // ==>
892 // x=shuffle(v1,v2,newMask)
893 // newMask[i] = (mask[i] < x1.size())
894 // ? mask1[mask[i]] : mask2[mask[i]-x1.size()]+v1.size()
895 //
896 // Here we are really conservative:
897 // we are absolutely afraid of producing a shuffle mask not in the input
898 // program, because the code gen may not be smart enough to turn a merged
899 // shuffle into two specific shuffles: it may produce worse code. As such,
900 // we only merge two shuffles if the result is either a splat or one of the
901 // input shuffle masks. In this case, merging the shuffles just removes
902 // one instruction, which we know is safe. This is good for things like
903 // turning: (splat(splat)) -> splat, or
904 // merge(V[0..n], V[n+1..2n]) -> V[0..2n]
905 ShuffleVectorInst* LHSShuffle = dyn_cast<ShuffleVectorInst>(LHS);
906 ShuffleVectorInst* RHSShuffle = dyn_cast<ShuffleVectorInst>(RHS);
907 if (LHSShuffle)
908 if (!isa<UndefValue>(LHSShuffle->getOperand(1)) && !isa<UndefValue>(RHS))
909 LHSShuffle = NULL;
910 if (RHSShuffle)
911 if (!isa<UndefValue>(RHSShuffle->getOperand(1)))
912 RHSShuffle = NULL;
913 if (!LHSShuffle && !RHSShuffle)
914 return MadeChange ? &SVI : 0;
915
916 Value* LHSOp0 = NULL;
917 Value* LHSOp1 = NULL;
918 Value* RHSOp0 = NULL;
919 unsigned LHSOp0Width = 0;
920 unsigned RHSOp0Width = 0;
921 if (LHSShuffle) {
922 LHSOp0 = LHSShuffle->getOperand(0);
923 LHSOp1 = LHSShuffle->getOperand(1);
924 LHSOp0Width = cast<VectorType>(LHSOp0->getType())->getNumElements();
925 }
926 if (RHSShuffle) {
927 RHSOp0 = RHSShuffle->getOperand(0);
928 RHSOp0Width = cast<VectorType>(RHSOp0->getType())->getNumElements();
929 }
930 Value* newLHS = LHS;
931 Value* newRHS = RHS;
932 if (LHSShuffle) {
933 // case 1
934 if (isa<UndefValue>(RHS)) {
935 newLHS = LHSOp0;
936 newRHS = LHSOp1;
937 }
938 // case 2 or 4
939 else if (LHSOp0Width == LHSWidth) {
940 newLHS = LHSOp0;
941 }
942 }
943 // case 3 or 4
944 if (RHSShuffle && RHSOp0Width == LHSWidth) {
945 newRHS = RHSOp0;
946 }
947 // case 4
948 if (LHSOp0 == RHSOp0) {
949 newLHS = LHSOp0;
950 newRHS = NULL;
951 }
952
953 if (newLHS == LHS && newRHS == RHS)
954 return MadeChange ? &SVI : 0;
955
956 SmallVector<int, 16> LHSMask;
957 SmallVector<int, 16> RHSMask;
958 if (newLHS != LHS)
959 LHSMask = LHSShuffle->getShuffleMask();
960 if (RHSShuffle && newRHS != RHS)
961 RHSMask = RHSShuffle->getShuffleMask();
962
963 unsigned newLHSWidth = (newLHS != LHS) ? LHSOp0Width : LHSWidth;
964 SmallVector<int, 16> newMask;
965 bool isSplat = true;
966 int SplatElt = -1;
967 // Create a new mask for the new ShuffleVectorInst so that the new
968 // ShuffleVectorInst is equivalent to the original one.
969 for (unsigned i = 0; i < VWidth; ++i) {
970 int eltMask;
971 if (Mask[i] < 0) {
972 // This element is an undef value.
973 eltMask = -1;
974 } else if (Mask[i] < (int)LHSWidth) {
975 // This element is from left hand side vector operand.
976 //
977 // If LHS is going to be replaced (case 1, 2, or 4), calculate the
978 // new mask value for the element.
979 if (newLHS != LHS) {
980 eltMask = LHSMask[Mask[i]];
981 // If the value selected is an undef value, explicitly specify it
982 // with a -1 mask value.
983 if (eltMask >= (int)LHSOp0Width && isa<UndefValue>(LHSOp1))
984 eltMask = -1;
985 } else
986 eltMask = Mask[i];
987 } else {
988 // This element is from right hand side vector operand
989 //
990 // If the value selected is an undef value, explicitly specify it
991 // with a -1 mask value. (case 1)
992 if (isa<UndefValue>(RHS))
993 eltMask = -1;
994 // If RHS is going to be replaced (case 3 or 4), calculate the
995 // new mask value for the element.
996 else if (newRHS != RHS) {
997 eltMask = RHSMask[Mask[i]-LHSWidth];
998 // If the value selected is an undef value, explicitly specify it
999 // with a -1 mask value.
1000 if (eltMask >= (int)RHSOp0Width) {
1001 assert(isa<UndefValue>(RHSShuffle->getOperand(1))
1002 && "should have been check above");
1003 eltMask = -1;
1004 }
1005 } else
1006 eltMask = Mask[i]-LHSWidth;
1007
1008 // If LHS's width is changed, shift the mask value accordingly.
1009 // If newRHS == NULL, i.e. LHSOp0 == RHSOp0, we want to remap any
1010 // references from RHSOp0 to LHSOp0, so we don't need to shift the mask.
1011 // If newRHS == newLHS, we want to remap any references from newRHS to
1012 // newLHS so that we can properly identify splats that may occur due to
1013 // obfuscation accross the two vectors.
1014 if (eltMask >= 0 && newRHS != NULL && newLHS != newRHS)
1015 eltMask += newLHSWidth;
1016 }
1017
1018 // Check if this could still be a splat.
1019 if (eltMask >= 0) {
1020 if (SplatElt >= 0 && SplatElt != eltMask)
1021 isSplat = false;
1022 SplatElt = eltMask;
1023 }
1024
1025 newMask.push_back(eltMask);
1026 }
1027
1028 // If the result mask is equal to one of the original shuffle masks,
1029 // or is a splat, do the replacement.
1030 if (isSplat || newMask == LHSMask || newMask == RHSMask || newMask == Mask) {
1031 SmallVector<Constant*, 16> Elts;
1032 Type *Int32Ty = Type::getInt32Ty(SVI.getContext());
1033 for (unsigned i = 0, e = newMask.size(); i != e; ++i) {
1034 if (newMask[i] < 0) {
1035 Elts.push_back(UndefValue::get(Int32Ty));
1036 } else {
1037 Elts.push_back(ConstantInt::get(Int32Ty, newMask[i]));
1038 }
1039 }
1040 if (newRHS == NULL)
1041 newRHS = UndefValue::get(newLHS->getType());
1042 return new ShuffleVectorInst(newLHS, newRHS, ConstantVector::get(Elts));
1043 }
1044
1045 return MadeChange ? &SVI : 0;
1046}
| 35 }
36 Instruction *I = dyn_cast<Instruction>(V);
37 if (!I) return false;
38
39 // Insert element gets simplified to the inserted element or is deleted if
40 // this is constant idx extract element and its a constant idx insertelt.
41 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
42 isa<ConstantInt>(I->getOperand(2)))
43 return true;
44 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
45 return true;
46 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
47 if (BO->hasOneUse() &&
48 (CheapToScalarize(BO->getOperand(0), isConstant) ||
49 CheapToScalarize(BO->getOperand(1), isConstant)))
50 return true;
51 if (CmpInst *CI = dyn_cast<CmpInst>(I))
52 if (CI->hasOneUse() &&
53 (CheapToScalarize(CI->getOperand(0), isConstant) ||
54 CheapToScalarize(CI->getOperand(1), isConstant)))
55 return true;
56
57 return false;
58}
59
60/// FindScalarElement - Given a vector and an element number, see if the scalar
61/// value is already around as a register, for example if it were inserted then
62/// extracted from the vector.
63static Value *FindScalarElement(Value *V, unsigned EltNo) {
64 assert(V->getType()->isVectorTy() && "Not looking at a vector?");
65 VectorType *VTy = cast<VectorType>(V->getType());
66 unsigned Width = VTy->getNumElements();
67 if (EltNo >= Width) // Out of range access.
68 return UndefValue::get(VTy->getElementType());
69
70 if (Constant *C = dyn_cast<Constant>(V))
71 return C->getAggregateElement(EltNo);
72
73 if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
74 // If this is an insert to a variable element, we don't know what it is.
75 if (!isa<ConstantInt>(III->getOperand(2)))
76 return 0;
77 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
78
79 // If this is an insert to the element we are looking for, return the
80 // inserted value.
81 if (EltNo == IIElt)
82 return III->getOperand(1);
83
84 // Otherwise, the insertelement doesn't modify the value, recurse on its
85 // vector input.
86 return FindScalarElement(III->getOperand(0), EltNo);
87 }
88
89 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
90 unsigned LHSWidth = SVI->getOperand(0)->getType()->getVectorNumElements();
91 int InEl = SVI->getMaskValue(EltNo);
92 if (InEl < 0)
93 return UndefValue::get(VTy->getElementType());
94 if (InEl < (int)LHSWidth)
95 return FindScalarElement(SVI->getOperand(0), InEl);
96 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth);
97 }
98
99 // Extract a value from a vector add operation with a constant zero.
100 Value *Val = 0; Constant *Con = 0;
101 if (match(V, m_Add(m_Value(Val), m_Constant(Con)))) {
102 if (Con->getAggregateElement(EltNo)->isNullValue())
103 return FindScalarElement(Val, EltNo);
104 }
105
106 // Otherwise, we don't know.
107 return 0;
108}
109
110// If we have a PHI node with a vector type that has only 2 uses: feed
111// itself and be an operand of extractelement at a constant location,
112// try to replace the PHI of the vector type with a PHI of a scalar type.
113Instruction *InstCombiner::scalarizePHI(ExtractElementInst &EI, PHINode *PN) {
114 // Verify that the PHI node has exactly 2 uses. Otherwise return NULL.
115 if (!PN->hasNUses(2))
116 return NULL;
117
118 // If so, it's known at this point that one operand is PHI and the other is
119 // an extractelement node. Find the PHI user that is not the extractelement
120 // node.
121 Value::use_iterator iu = PN->use_begin();
122 Instruction *PHIUser = dyn_cast<Instruction>(*iu);
123 if (PHIUser == cast<Instruction>(&EI))
124 PHIUser = cast<Instruction>(*(++iu));
125
126 // Verify that this PHI user has one use, which is the PHI itself,
127 // and that it is a binary operation which is cheap to scalarize.
128 // otherwise return NULL.
129 if (!PHIUser->hasOneUse() || !(PHIUser->use_back() == PN) ||
130 !(isa<BinaryOperator>(PHIUser)) || !CheapToScalarize(PHIUser, true))
131 return NULL;
132
133 // Create a scalar PHI node that will replace the vector PHI node
134 // just before the current PHI node.
135 PHINode *scalarPHI = cast<PHINode>(InsertNewInstWith(
136 PHINode::Create(EI.getType(), PN->getNumIncomingValues(), ""), *PN));
137 // Scalarize each PHI operand.
138 for (unsigned i = 0; i < PN->getNumIncomingValues(); i++) {
139 Value *PHIInVal = PN->getIncomingValue(i);
140 BasicBlock *inBB = PN->getIncomingBlock(i);
141 Value *Elt = EI.getIndexOperand();
142 // If the operand is the PHI induction variable:
143 if (PHIInVal == PHIUser) {
144 // Scalarize the binary operation. Its first operand is the
145 // scalar PHI and the second operand is extracted from the other
146 // vector operand.
147 BinaryOperator *B0 = cast<BinaryOperator>(PHIUser);
148 unsigned opId = (B0->getOperand(0) == PN) ? 1 : 0;
149 Value *Op = InsertNewInstWith(
150 ExtractElementInst::Create(B0->getOperand(opId), Elt,
151 B0->getOperand(opId)->getName() + ".Elt"),
152 *B0);
153 Value *newPHIUser = InsertNewInstWith(
154 BinaryOperator::Create(B0->getOpcode(), scalarPHI, Op), *B0);
155 scalarPHI->addIncoming(newPHIUser, inBB);
156 } else {
157 // Scalarize PHI input:
158 Instruction *newEI = ExtractElementInst::Create(PHIInVal, Elt, "");
159 // Insert the new instruction into the predecessor basic block.
160 Instruction *pos = dyn_cast<Instruction>(PHIInVal);
161 BasicBlock::iterator InsertPos;
162 if (pos && !isa<PHINode>(pos)) {
163 InsertPos = pos;
164 ++InsertPos;
165 } else {
166 InsertPos = inBB->getFirstInsertionPt();
167 }
168
169 InsertNewInstWith(newEI, *InsertPos);
170
171 scalarPHI->addIncoming(newEI, inBB);
172 }
173 }
174 return ReplaceInstUsesWith(EI, scalarPHI);
175}
176
177Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
178 // If vector val is constant with all elements the same, replace EI with
179 // that element. We handle a known element # below.
180 if (Constant *C = dyn_cast<Constant>(EI.getOperand(0)))
181 if (CheapToScalarize(C, false))
182 return ReplaceInstUsesWith(EI, C->getAggregateElement(0U));
183
184 // If extracting a specified index from the vector, see if we can recursively
185 // find a previously computed scalar that was inserted into the vector.
186 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
187 unsigned IndexVal = IdxC->getZExtValue();
188 unsigned VectorWidth = EI.getVectorOperandType()->getNumElements();
189
190 // If this is extracting an invalid index, turn this into undef, to avoid
191 // crashing the code below.
192 if (IndexVal >= VectorWidth)
193 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
194
195 // This instruction only demands the single element from the input vector.
196 // If the input vector has a single use, simplify it based on this use
197 // property.
198 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
199 APInt UndefElts(VectorWidth, 0);
200 APInt DemandedMask(VectorWidth, 0);
201 DemandedMask.setBit(IndexVal);
202 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
203 DemandedMask, UndefElts)) {
204 EI.setOperand(0, V);
205 return &EI;
206 }
207 }
208
209 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
210 return ReplaceInstUsesWith(EI, Elt);
211
212 // If the this extractelement is directly using a bitcast from a vector of
213 // the same number of elements, see if we can find the source element from
214 // it. In this case, we will end up needing to bitcast the scalars.
215 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
216 if (VectorType *VT = dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
217 if (VT->getNumElements() == VectorWidth)
218 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
219 return new BitCastInst(Elt, EI.getType());
220 }
221
222 // If there's a vector PHI feeding a scalar use through this extractelement
223 // instruction, try to scalarize the PHI.
224 if (PHINode *PN = dyn_cast<PHINode>(EI.getOperand(0))) {
225 Instruction *scalarPHI = scalarizePHI(EI, PN);
226 if (scalarPHI)
227 return scalarPHI;
228 }
229 }
230
231 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
232 // Push extractelement into predecessor operation if legal and
233 // profitable to do so
234 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
235 if (I->hasOneUse() &&
236 CheapToScalarize(BO, isa<ConstantInt>(EI.getOperand(1)))) {
237 Value *newEI0 =
238 Builder->CreateExtractElement(BO->getOperand(0), EI.getOperand(1),
239 EI.getName()+".lhs");
240 Value *newEI1 =
241 Builder->CreateExtractElement(BO->getOperand(1), EI.getOperand(1),
242 EI.getName()+".rhs");
243 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
244 }
245 } else if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
246 // Extracting the inserted element?
247 if (IE->getOperand(2) == EI.getOperand(1))
248 return ReplaceInstUsesWith(EI, IE->getOperand(1));
249 // If the inserted and extracted elements are constants, they must not
250 // be the same value, extract from the pre-inserted value instead.
251 if (isa<Constant>(IE->getOperand(2)) && isa<Constant>(EI.getOperand(1))) {
252 Worklist.AddValue(EI.getOperand(0));
253 EI.setOperand(0, IE->getOperand(0));
254 return &EI;
255 }
256 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
257 // If this is extracting an element from a shufflevector, figure out where
258 // it came from and extract from the appropriate input element instead.
259 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
260 int SrcIdx = SVI->getMaskValue(Elt->getZExtValue());
261 Value *Src;
262 unsigned LHSWidth =
263 SVI->getOperand(0)->getType()->getVectorNumElements();
264
265 if (SrcIdx < 0)
266 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
267 if (SrcIdx < (int)LHSWidth)
268 Src = SVI->getOperand(0);
269 else {
270 SrcIdx -= LHSWidth;
271 Src = SVI->getOperand(1);
272 }
273 Type *Int32Ty = Type::getInt32Ty(EI.getContext());
274 return ExtractElementInst::Create(Src,
275 ConstantInt::get(Int32Ty,
276 SrcIdx, false));
277 }
278 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
279 // Canonicalize extractelement(cast) -> cast(extractelement)
280 // bitcasts can change the number of vector elements and they cost nothing
281 if (CI->hasOneUse() && (CI->getOpcode() != Instruction::BitCast)) {
282 Value *EE = Builder->CreateExtractElement(CI->getOperand(0),
283 EI.getIndexOperand());
284 Worklist.AddValue(EE);
285 return CastInst::Create(CI->getOpcode(), EE, EI.getType());
286 }
287 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
288 if (SI->hasOneUse()) {
289 // TODO: For a select on vectors, it might be useful to do this if it
290 // has multiple extractelement uses. For vector select, that seems to
291 // fight the vectorizer.
292
293 // If we are extracting an element from a vector select or a select on
294 // vectors, a select on the scalars extracted from the vector arguments.
295 Value *TrueVal = SI->getTrueValue();
296 Value *FalseVal = SI->getFalseValue();
297
298 Value *Cond = SI->getCondition();
299 if (Cond->getType()->isVectorTy()) {
300 Cond = Builder->CreateExtractElement(Cond,
301 EI.getIndexOperand(),
302 Cond->getName() + ".elt");
303 }
304
305 Value *V1Elem
306 = Builder->CreateExtractElement(TrueVal,
307 EI.getIndexOperand(),
308 TrueVal->getName() + ".elt");
309
310 Value *V2Elem
311 = Builder->CreateExtractElement(FalseVal,
312 EI.getIndexOperand(),
313 FalseVal->getName() + ".elt");
314 return SelectInst::Create(Cond,
315 V1Elem,
316 V2Elem,
317 SI->getName() + ".elt");
318 }
319 }
320 }
321 return 0;
322}
323
324/// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
325/// elements from either LHS or RHS, return the shuffle mask and true.
326/// Otherwise, return false.
327static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
328 SmallVectorImpl<Constant*> &Mask) {
329 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
330 "Invalid CollectSingleShuffleElements");
331 unsigned NumElts = V->getType()->getVectorNumElements();
332
333 if (isa<UndefValue>(V)) {
334 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
335 return true;
336 }
337
338 if (V == LHS) {
339 for (unsigned i = 0; i != NumElts; ++i)
340 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
341 return true;
342 }
343
344 if (V == RHS) {
345 for (unsigned i = 0; i != NumElts; ++i)
346 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()),
347 i+NumElts));
348 return true;
349 }
350
351 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
352 // If this is an insert of an extract from some other vector, include it.
353 Value *VecOp = IEI->getOperand(0);
354 Value *ScalarOp = IEI->getOperand(1);
355 Value *IdxOp = IEI->getOperand(2);
356
357 if (!isa<ConstantInt>(IdxOp))
358 return false;
359 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
360
361 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
362 // Okay, we can handle this if the vector we are insertinting into is
363 // transitively ok.
364 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
365 // If so, update the mask to reflect the inserted undef.
366 Mask[InsertedIdx] = UndefValue::get(Type::getInt32Ty(V->getContext()));
367 return true;
368 }
369 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
370 if (isa<ConstantInt>(EI->getOperand(1)) &&
371 EI->getOperand(0)->getType() == V->getType()) {
372 unsigned ExtractedIdx =
373 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
374
375 // This must be extracting from either LHS or RHS.
376 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
377 // Okay, we can handle this if the vector we are insertinting into is
378 // transitively ok.
379 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
380 // If so, update the mask to reflect the inserted value.
381 if (EI->getOperand(0) == LHS) {
382 Mask[InsertedIdx % NumElts] =
383 ConstantInt::get(Type::getInt32Ty(V->getContext()),
384 ExtractedIdx);
385 } else {
386 assert(EI->getOperand(0) == RHS);
387 Mask[InsertedIdx % NumElts] =
388 ConstantInt::get(Type::getInt32Ty(V->getContext()),
389 ExtractedIdx+NumElts);
390 }
391 return true;
392 }
393 }
394 }
395 }
396 }
397 // TODO: Handle shufflevector here!
398
399 return false;
400}
401
402/// CollectShuffleElements - We are building a shuffle of V, using RHS as the
403/// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
404/// that computes V and the LHS value of the shuffle.
405static Value *CollectShuffleElements(Value *V, SmallVectorImpl<Constant*> &Mask,
406 Value *&RHS) {
407 assert(V->getType()->isVectorTy() &&
408 (RHS == 0 || V->getType() == RHS->getType()) &&
409 "Invalid shuffle!");
410 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
411
412 if (isa<UndefValue>(V)) {
413 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
414 return V;
415 }
416
417 if (isa<ConstantAggregateZero>(V)) {
418 Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(V->getContext()),0));
419 return V;
420 }
421
422 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
423 // If this is an insert of an extract from some other vector, include it.
424 Value *VecOp = IEI->getOperand(0);
425 Value *ScalarOp = IEI->getOperand(1);
426 Value *IdxOp = IEI->getOperand(2);
427
428 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
429 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
430 EI->getOperand(0)->getType() == V->getType()) {
431 unsigned ExtractedIdx =
432 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
433 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
434
435 // Either the extracted from or inserted into vector must be RHSVec,
436 // otherwise we'd end up with a shuffle of three inputs.
437 if (EI->getOperand(0) == RHS || RHS == 0) {
438 RHS = EI->getOperand(0);
439 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
440 Mask[InsertedIdx % NumElts] =
441 ConstantInt::get(Type::getInt32Ty(V->getContext()),
442 NumElts+ExtractedIdx);
443 return V;
444 }
445
446 if (VecOp == RHS) {
447 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
448 // Update Mask to reflect that `ScalarOp' has been inserted at
449 // position `InsertedIdx' within the vector returned by IEI.
450 Mask[InsertedIdx % NumElts] = Mask[ExtractedIdx];
451
452 // Everything but the extracted element is replaced with the RHS.
453 for (unsigned i = 0; i != NumElts; ++i) {
454 if (i != InsertedIdx)
455 Mask[i] = ConstantInt::get(Type::getInt32Ty(V->getContext()),
456 NumElts+i);
457 }
458 return V;
459 }
460
461 // If this insertelement is a chain that comes from exactly these two
462 // vectors, return the vector and the effective shuffle.
463 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
464 return EI->getOperand(0);
465 }
466 }
467 }
468 // TODO: Handle shufflevector here!
469
470 // Otherwise, can't do anything fancy. Return an identity vector.
471 for (unsigned i = 0; i != NumElts; ++i)
472 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
473 return V;
474}
475
476Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
477 Value *VecOp = IE.getOperand(0);
478 Value *ScalarOp = IE.getOperand(1);
479 Value *IdxOp = IE.getOperand(2);
480
481 // Inserting an undef or into an undefined place, remove this.
482 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
483 ReplaceInstUsesWith(IE, VecOp);
484
485 // If the inserted element was extracted from some other vector, and if the
486 // indexes are constant, try to turn this into a shufflevector operation.
487 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
488 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
489 EI->getOperand(0)->getType() == IE.getType()) {
490 unsigned NumVectorElts = IE.getType()->getNumElements();
491 unsigned ExtractedIdx =
492 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
493 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
494
495 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
496 return ReplaceInstUsesWith(IE, VecOp);
497
498 if (InsertedIdx >= NumVectorElts) // Out of range insert.
499 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
500
501 // If we are extracting a value from a vector, then inserting it right
502 // back into the same place, just use the input vector.
503 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
504 return ReplaceInstUsesWith(IE, VecOp);
505
506 // If this insertelement isn't used by some other insertelement, turn it
507 // (and any insertelements it points to), into one big shuffle.
508 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
509 SmallVector<Constant*, 16> Mask;
510 Value *RHS = 0;
511 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
512 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
513 // We now have a shuffle of LHS, RHS, Mask.
514 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
515 }
516 }
517 }
518
519 unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements();
520 APInt UndefElts(VWidth, 0);
521 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
522 if (Value *V = SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts)) {
523 if (V != &IE)
524 return ReplaceInstUsesWith(IE, V);
525 return &IE;
526 }
527
528 return 0;
529}
530
531/// Return true if we can evaluate the specified expression tree if the vector
532/// elements were shuffled in a different order.
533static bool CanEvaluateShuffled(Value *V, ArrayRef<int> Mask,
534 unsigned Depth = 5) {
535 // We can always reorder the elements of a constant.
536 if (isa<Constant>(V))
537 return true;
538
539 // We won't reorder vector arguments. No IPO here.
540 Instruction *I = dyn_cast<Instruction>(V);
541 if (!I) return false;
542
543 // Two users may expect different orders of the elements. Don't try it.
544 if (!I->hasOneUse())
545 return false;
546
547 if (Depth == 0) return false;
548
549 switch (I->getOpcode()) {
550 case Instruction::Add:
551 case Instruction::FAdd:
552 case Instruction::Sub:
553 case Instruction::FSub:
554 case Instruction::Mul:
555 case Instruction::FMul:
556 case Instruction::UDiv:
557 case Instruction::SDiv:
558 case Instruction::FDiv:
559 case Instruction::URem:
560 case Instruction::SRem:
561 case Instruction::FRem:
562 case Instruction::Shl:
563 case Instruction::LShr:
564 case Instruction::AShr:
565 case Instruction::And:
566 case Instruction::Or:
567 case Instruction::Xor:
568 case Instruction::ICmp:
569 case Instruction::FCmp:
570 case Instruction::Trunc:
571 case Instruction::ZExt:
572 case Instruction::SExt:
573 case Instruction::FPToUI:
574 case Instruction::FPToSI:
575 case Instruction::UIToFP:
576 case Instruction::SIToFP:
577 case Instruction::FPTrunc:
578 case Instruction::FPExt:
579 case Instruction::GetElementPtr: {
580 for (int i = 0, e = I->getNumOperands(); i != e; ++i) {
581 if (!CanEvaluateShuffled(I->getOperand(i), Mask, Depth-1))
582 return false;
583 }
584 return true;
585 }
586 case Instruction::InsertElement: {
587 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(2));
588 if (!CI) return false;
589 int ElementNumber = CI->getLimitedValue();
590
591 // Verify that 'CI' does not occur twice in Mask. A single 'insertelement'
592 // can't put an element into multiple indices.
593 bool SeenOnce = false;
594 for (int i = 0, e = Mask.size(); i != e; ++i) {
595 if (Mask[i] == ElementNumber) {
596 if (SeenOnce)
597 return false;
598 SeenOnce = true;
599 }
600 }
601 return CanEvaluateShuffled(I->getOperand(0), Mask, Depth-1);
602 }
603 }
604 return false;
605}
606
607/// Rebuild a new instruction just like 'I' but with the new operands given.
608/// In the event of type mismatch, the type of the operands is correct.
609static Value *BuildNew(Instruction *I, ArrayRef<Value*> NewOps) {
610 // We don't want to use the IRBuilder here because we want the replacement
611 // instructions to appear next to 'I', not the builder's insertion point.
612 switch (I->getOpcode()) {
613 case Instruction::Add:
614 case Instruction::FAdd:
615 case Instruction::Sub:
616 case Instruction::FSub:
617 case Instruction::Mul:
618 case Instruction::FMul:
619 case Instruction::UDiv:
620 case Instruction::SDiv:
621 case Instruction::FDiv:
622 case Instruction::URem:
623 case Instruction::SRem:
624 case Instruction::FRem:
625 case Instruction::Shl:
626 case Instruction::LShr:
627 case Instruction::AShr:
628 case Instruction::And:
629 case Instruction::Or:
630 case Instruction::Xor: {
631 BinaryOperator *BO = cast<BinaryOperator>(I);
632 assert(NewOps.size() == 2 && "binary operator with #ops != 2");
633 BinaryOperator *New =
634 BinaryOperator::Create(cast<BinaryOperator>(I)->getOpcode(),
635 NewOps[0], NewOps[1], "", BO);
636 if (isa<OverflowingBinaryOperator>(BO)) {
637 New->setHasNoUnsignedWrap(BO->hasNoUnsignedWrap());
638 New->setHasNoSignedWrap(BO->hasNoSignedWrap());
639 }
640 if (isa<PossiblyExactOperator>(BO)) {
641 New->setIsExact(BO->isExact());
642 }
643 return New;
644 }
645 case Instruction::ICmp:
646 assert(NewOps.size() == 2 && "icmp with #ops != 2");
647 return new ICmpInst(I, cast<ICmpInst>(I)->getPredicate(),
648 NewOps[0], NewOps[1]);
649 case Instruction::FCmp:
650 assert(NewOps.size() == 2 && "fcmp with #ops != 2");
651 return new FCmpInst(I, cast<FCmpInst>(I)->getPredicate(),
652 NewOps[0], NewOps[1]);
653 case Instruction::Trunc:
654 case Instruction::ZExt:
655 case Instruction::SExt:
656 case Instruction::FPToUI:
657 case Instruction::FPToSI:
658 case Instruction::UIToFP:
659 case Instruction::SIToFP:
660 case Instruction::FPTrunc:
661 case Instruction::FPExt: {
662 // It's possible that the mask has a different number of elements from
663 // the original cast. We recompute the destination type to match the mask.
664 Type *DestTy =
665 VectorType::get(I->getType()->getScalarType(),
666 NewOps[0]->getType()->getVectorNumElements());
667 assert(NewOps.size() == 1 && "cast with #ops != 1");
668 return CastInst::Create(cast<CastInst>(I)->getOpcode(), NewOps[0], DestTy,
669 "", I);
670 }
671 case Instruction::GetElementPtr: {
672 Value *Ptr = NewOps[0];
673 ArrayRef<Value*> Idx = NewOps.slice(1);
674 GetElementPtrInst *GEP = GetElementPtrInst::Create(Ptr, Idx, "", I);
675 GEP->setIsInBounds(cast<GetElementPtrInst>(I)->isInBounds());
676 return GEP;
677 }
678 }
679 llvm_unreachable("failed to rebuild vector instructions");
680}
681
682Value *
683InstCombiner::EvaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask) {
684 // Mask.size() does not need to be equal to the number of vector elements.
685
686 assert(V->getType()->isVectorTy() && "can't reorder non-vector elements");
687 if (isa<UndefValue>(V)) {
688 return UndefValue::get(VectorType::get(V->getType()->getScalarType(),
689 Mask.size()));
690 }
691 if (isa<ConstantAggregateZero>(V)) {
692 return ConstantAggregateZero::get(
693 VectorType::get(V->getType()->getScalarType(),
694 Mask.size()));
695 }
696 if (Constant *C = dyn_cast<Constant>(V)) {
697 SmallVector<Constant *, 16> MaskValues;
698 for (int i = 0, e = Mask.size(); i != e; ++i) {
699 if (Mask[i] == -1)
700 MaskValues.push_back(UndefValue::get(Builder->getInt32Ty()));
701 else
702 MaskValues.push_back(Builder->getInt32(Mask[i]));
703 }
704 return ConstantExpr::getShuffleVector(C, UndefValue::get(C->getType()),
705 ConstantVector::get(MaskValues));
706 }
707
708 Instruction *I = cast<Instruction>(V);
709 switch (I->getOpcode()) {
710 case Instruction::Add:
711 case Instruction::FAdd:
712 case Instruction::Sub:
713 case Instruction::FSub:
714 case Instruction::Mul:
715 case Instruction::FMul:
716 case Instruction::UDiv:
717 case Instruction::SDiv:
718 case Instruction::FDiv:
719 case Instruction::URem:
720 case Instruction::SRem:
721 case Instruction::FRem:
722 case Instruction::Shl:
723 case Instruction::LShr:
724 case Instruction::AShr:
725 case Instruction::And:
726 case Instruction::Or:
727 case Instruction::Xor:
728 case Instruction::ICmp:
729 case Instruction::FCmp:
730 case Instruction::Trunc:
731 case Instruction::ZExt:
732 case Instruction::SExt:
733 case Instruction::FPToUI:
734 case Instruction::FPToSI:
735 case Instruction::UIToFP:
736 case Instruction::SIToFP:
737 case Instruction::FPTrunc:
738 case Instruction::FPExt:
739 case Instruction::Select:
740 case Instruction::GetElementPtr: {
741 SmallVector<Value*, 8> NewOps;
742 bool NeedsRebuild = (Mask.size() != I->getType()->getVectorNumElements());
743 for (int i = 0, e = I->getNumOperands(); i != e; ++i) {
744 Value *V = EvaluateInDifferentElementOrder(I->getOperand(i), Mask);
745 NewOps.push_back(V);
746 NeedsRebuild |= (V != I->getOperand(i));
747 }
748 if (NeedsRebuild) {
749 return BuildNew(I, NewOps);
750 }
751 return I;
752 }
753 case Instruction::InsertElement: {
754 int Element = cast<ConstantInt>(I->getOperand(2))->getLimitedValue();
755
756 // The insertelement was inserting at Element. Figure out which element
757 // that becomes after shuffling. The answer is guaranteed to be unique
758 // by CanEvaluateShuffled.
759 bool Found = false;
760 int Index = 0;
761 for (int e = Mask.size(); Index != e; ++Index) {
762 if (Mask[Index] == Element) {
763 Found = true;
764 break;
765 }
766 }
767
768 if (!Found)
769 return UndefValue::get(
770 VectorType::get(V->getType()->getScalarType(), Mask.size()));
771
772 Value *V = EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
773 return InsertElementInst::Create(V, I->getOperand(1),
774 Builder->getInt32(Index), "", I);
775 }
776 }
777 llvm_unreachable("failed to reorder elements of vector instruction!");
778}
779
780Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
781 Value *LHS = SVI.getOperand(0);
782 Value *RHS = SVI.getOperand(1);
783 SmallVector<int, 16> Mask = SVI.getShuffleMask();
784
785 bool MadeChange = false;
786
787 // Undefined shuffle mask -> undefined value.
788 if (isa<UndefValue>(SVI.getOperand(2)))
789 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
790
791 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
792
793 APInt UndefElts(VWidth, 0);
794 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
795 if (Value *V = SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
796 if (V != &SVI)
797 return ReplaceInstUsesWith(SVI, V);
798 LHS = SVI.getOperand(0);
799 RHS = SVI.getOperand(1);
800 MadeChange = true;
801 }
802
803 unsigned LHSWidth = cast<VectorType>(LHS->getType())->getNumElements();
804
805 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
806 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
807 if (LHS == RHS || isa<UndefValue>(LHS)) {
808 if (isa<UndefValue>(LHS) && LHS == RHS) {
809 // shuffle(undef,undef,mask) -> undef.
810 Value *Result = (VWidth == LHSWidth)
811 ? LHS : UndefValue::get(SVI.getType());
812 return ReplaceInstUsesWith(SVI, Result);
813 }
814
815 // Remap any references to RHS to use LHS.
816 SmallVector<Constant*, 16> Elts;
817 for (unsigned i = 0, e = LHSWidth; i != VWidth; ++i) {
818 if (Mask[i] < 0) {
819 Elts.push_back(UndefValue::get(Type::getInt32Ty(SVI.getContext())));
820 continue;
821 }
822
823 if ((Mask[i] >= (int)e && isa<UndefValue>(RHS)) ||
824 (Mask[i] < (int)e && isa<UndefValue>(LHS))) {
825 Mask[i] = -1; // Turn into undef.
826 Elts.push_back(UndefValue::get(Type::getInt32Ty(SVI.getContext())));
827 } else {
828 Mask[i] = Mask[i] % e; // Force to LHS.
829 Elts.push_back(ConstantInt::get(Type::getInt32Ty(SVI.getContext()),
830 Mask[i]));
831 }
832 }
833 SVI.setOperand(0, SVI.getOperand(1));
834 SVI.setOperand(1, UndefValue::get(RHS->getType()));
835 SVI.setOperand(2, ConstantVector::get(Elts));
836 LHS = SVI.getOperand(0);
837 RHS = SVI.getOperand(1);
838 MadeChange = true;
839 }
840
841 if (VWidth == LHSWidth) {
842 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
843 bool isLHSID = true, isRHSID = true;
844
845 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
846 if (Mask[i] < 0) continue; // Ignore undef values.
847 // Is this an identity shuffle of the LHS value?
848 isLHSID &= (Mask[i] == (int)i);
849
850 // Is this an identity shuffle of the RHS value?
851 isRHSID &= (Mask[i]-e == i);
852 }
853
854 // Eliminate identity shuffles.
855 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
856 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
857 }
858
859 if (isa<UndefValue>(RHS) && CanEvaluateShuffled(LHS, Mask)) {
860 Value *V = EvaluateInDifferentElementOrder(LHS, Mask);
861 return ReplaceInstUsesWith(SVI, V);
862 }
863
864 // If the LHS is a shufflevector itself, see if we can combine it with this
865 // one without producing an unusual shuffle.
866 // Cases that might be simplified:
867 // 1.
868 // x1=shuffle(v1,v2,mask1)
869 // x=shuffle(x1,undef,mask)
870 // ==>
871 // x=shuffle(v1,undef,newMask)
872 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : -1
873 // 2.
874 // x1=shuffle(v1,undef,mask1)
875 // x=shuffle(x1,x2,mask)
876 // where v1.size() == mask1.size()
877 // ==>
878 // x=shuffle(v1,x2,newMask)
879 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : mask[i]
880 // 3.
881 // x2=shuffle(v2,undef,mask2)
882 // x=shuffle(x1,x2,mask)
883 // where v2.size() == mask2.size()
884 // ==>
885 // x=shuffle(x1,v2,newMask)
886 // newMask[i] = (mask[i] < x1.size())
887 // ? mask[i] : mask2[mask[i]-x1.size()]+x1.size()
888 // 4.
889 // x1=shuffle(v1,undef,mask1)
890 // x2=shuffle(v2,undef,mask2)
891 // x=shuffle(x1,x2,mask)
892 // where v1.size() == v2.size()
893 // ==>
894 // x=shuffle(v1,v2,newMask)
895 // newMask[i] = (mask[i] < x1.size())
896 // ? mask1[mask[i]] : mask2[mask[i]-x1.size()]+v1.size()
897 //
898 // Here we are really conservative:
899 // we are absolutely afraid of producing a shuffle mask not in the input
900 // program, because the code gen may not be smart enough to turn a merged
901 // shuffle into two specific shuffles: it may produce worse code. As such,
902 // we only merge two shuffles if the result is either a splat or one of the
903 // input shuffle masks. In this case, merging the shuffles just removes
904 // one instruction, which we know is safe. This is good for things like
905 // turning: (splat(splat)) -> splat, or
906 // merge(V[0..n], V[n+1..2n]) -> V[0..2n]
907 ShuffleVectorInst* LHSShuffle = dyn_cast<ShuffleVectorInst>(LHS);
908 ShuffleVectorInst* RHSShuffle = dyn_cast<ShuffleVectorInst>(RHS);
909 if (LHSShuffle)
910 if (!isa<UndefValue>(LHSShuffle->getOperand(1)) && !isa<UndefValue>(RHS))
911 LHSShuffle = NULL;
912 if (RHSShuffle)
913 if (!isa<UndefValue>(RHSShuffle->getOperand(1)))
914 RHSShuffle = NULL;
915 if (!LHSShuffle && !RHSShuffle)
916 return MadeChange ? &SVI : 0;
917
918 Value* LHSOp0 = NULL;
919 Value* LHSOp1 = NULL;
920 Value* RHSOp0 = NULL;
921 unsigned LHSOp0Width = 0;
922 unsigned RHSOp0Width = 0;
923 if (LHSShuffle) {
924 LHSOp0 = LHSShuffle->getOperand(0);
925 LHSOp1 = LHSShuffle->getOperand(1);
926 LHSOp0Width = cast<VectorType>(LHSOp0->getType())->getNumElements();
927 }
928 if (RHSShuffle) {
929 RHSOp0 = RHSShuffle->getOperand(0);
930 RHSOp0Width = cast<VectorType>(RHSOp0->getType())->getNumElements();
931 }
932 Value* newLHS = LHS;
933 Value* newRHS = RHS;
934 if (LHSShuffle) {
935 // case 1
936 if (isa<UndefValue>(RHS)) {
937 newLHS = LHSOp0;
938 newRHS = LHSOp1;
939 }
940 // case 2 or 4
941 else if (LHSOp0Width == LHSWidth) {
942 newLHS = LHSOp0;
943 }
944 }
945 // case 3 or 4
946 if (RHSShuffle && RHSOp0Width == LHSWidth) {
947 newRHS = RHSOp0;
948 }
949 // case 4
950 if (LHSOp0 == RHSOp0) {
951 newLHS = LHSOp0;
952 newRHS = NULL;
953 }
954
955 if (newLHS == LHS && newRHS == RHS)
956 return MadeChange ? &SVI : 0;
957
958 SmallVector<int, 16> LHSMask;
959 SmallVector<int, 16> RHSMask;
960 if (newLHS != LHS)
961 LHSMask = LHSShuffle->getShuffleMask();
962 if (RHSShuffle && newRHS != RHS)
963 RHSMask = RHSShuffle->getShuffleMask();
964
965 unsigned newLHSWidth = (newLHS != LHS) ? LHSOp0Width : LHSWidth;
966 SmallVector<int, 16> newMask;
967 bool isSplat = true;
968 int SplatElt = -1;
969 // Create a new mask for the new ShuffleVectorInst so that the new
970 // ShuffleVectorInst is equivalent to the original one.
971 for (unsigned i = 0; i < VWidth; ++i) {
972 int eltMask;
973 if (Mask[i] < 0) {
974 // This element is an undef value.
975 eltMask = -1;
976 } else if (Mask[i] < (int)LHSWidth) {
977 // This element is from left hand side vector operand.
978 //
979 // If LHS is going to be replaced (case 1, 2, or 4), calculate the
980 // new mask value for the element.
981 if (newLHS != LHS) {
982 eltMask = LHSMask[Mask[i]];
983 // If the value selected is an undef value, explicitly specify it
984 // with a -1 mask value.
985 if (eltMask >= (int)LHSOp0Width && isa<UndefValue>(LHSOp1))
986 eltMask = -1;
987 } else
988 eltMask = Mask[i];
989 } else {
990 // This element is from right hand side vector operand
991 //
992 // If the value selected is an undef value, explicitly specify it
993 // with a -1 mask value. (case 1)
994 if (isa<UndefValue>(RHS))
995 eltMask = -1;
996 // If RHS is going to be replaced (case 3 or 4), calculate the
997 // new mask value for the element.
998 else if (newRHS != RHS) {
999 eltMask = RHSMask[Mask[i]-LHSWidth];
1000 // If the value selected is an undef value, explicitly specify it
1001 // with a -1 mask value.
1002 if (eltMask >= (int)RHSOp0Width) {
1003 assert(isa<UndefValue>(RHSShuffle->getOperand(1))
1004 && "should have been check above");
1005 eltMask = -1;
1006 }
1007 } else
1008 eltMask = Mask[i]-LHSWidth;
1009
1010 // If LHS's width is changed, shift the mask value accordingly.
1011 // If newRHS == NULL, i.e. LHSOp0 == RHSOp0, we want to remap any
1012 // references from RHSOp0 to LHSOp0, so we don't need to shift the mask.
1013 // If newRHS == newLHS, we want to remap any references from newRHS to
1014 // newLHS so that we can properly identify splats that may occur due to
1015 // obfuscation accross the two vectors.
1016 if (eltMask >= 0 && newRHS != NULL && newLHS != newRHS)
1017 eltMask += newLHSWidth;
1018 }
1019
1020 // Check if this could still be a splat.
1021 if (eltMask >= 0) {
1022 if (SplatElt >= 0 && SplatElt != eltMask)
1023 isSplat = false;
1024 SplatElt = eltMask;
1025 }
1026
1027 newMask.push_back(eltMask);
1028 }
1029
1030 // If the result mask is equal to one of the original shuffle masks,
1031 // or is a splat, do the replacement.
1032 if (isSplat || newMask == LHSMask || newMask == RHSMask || newMask == Mask) {
1033 SmallVector<Constant*, 16> Elts;
1034 Type *Int32Ty = Type::getInt32Ty(SVI.getContext());
1035 for (unsigned i = 0, e = newMask.size(); i != e; ++i) {
1036 if (newMask[i] < 0) {
1037 Elts.push_back(UndefValue::get(Int32Ty));
1038 } else {
1039 Elts.push_back(ConstantInt::get(Int32Ty, newMask[i]));
1040 }
1041 }
1042 if (newRHS == NULL)
1043 newRHS = UndefValue::get(newLHS->getType());
1044 return new ShuffleVectorInst(newLHS, newRHS, ConstantVector::get(Elts));
1045 }
1046
1047 return MadeChange ? &SVI : 0;
1048}
|