1 //===- InstCombinePHI.cpp -------------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the visitPHINode function.
12 //===----------------------------------------------------------------------===//
14 #include "InstCombine.h"
15 #include "llvm/Analysis/InstructionSimplify.h"
16 #include "llvm/Target/TargetData.h"
17 #include "llvm/ADT/SmallPtrSet.h"
18 #include "llvm/ADT/STLExtras.h"
21 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(a,c)]
22 /// and if a/b/c and the add's all have a single use, turn this into a phi
23 /// and a single binop.
24 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
25 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
26 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
27 unsigned Opc = FirstInst->getOpcode();
28 Value *LHSVal = FirstInst->getOperand(0);
29 Value *RHSVal = FirstInst->getOperand(1);
31 const Type *LHSType = LHSVal->getType();
32 const Type *RHSType = RHSVal->getType();
34 // Scan to see if all operands are the same opcode, and all have one use.
35 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
36 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
37 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
38 // Verify type of the LHS matches so we don't fold cmp's of different
39 // types or GEP's with different index types.
40 I->getOperand(0)->getType() != LHSType ||
41 I->getOperand(1)->getType() != RHSType)
44 // If they are CmpInst instructions, check their predicates
45 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
46 if (cast<CmpInst>(I)->getPredicate() !=
47 cast<CmpInst>(FirstInst)->getPredicate())
50 // Keep track of which operand needs a phi node.
51 if (I->getOperand(0) != LHSVal) LHSVal = 0;
52 if (I->getOperand(1) != RHSVal) RHSVal = 0;
55 // If both LHS and RHS would need a PHI, don't do this transformation,
56 // because it would increase the number of PHIs entering the block,
57 // which leads to higher register pressure. This is especially
58 // bad when the PHIs are in the header of a loop.
59 if (!LHSVal && !RHSVal)
62 // Otherwise, this is safe to transform!
64 Value *InLHS = FirstInst->getOperand(0);
65 Value *InRHS = FirstInst->getOperand(1);
66 PHINode *NewLHS = 0, *NewRHS = 0;
68 NewLHS = PHINode::Create(LHSType,
69 FirstInst->getOperand(0)->getName() + ".pn");
70 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
71 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
72 InsertNewInstBefore(NewLHS, PN);
77 NewRHS = PHINode::Create(RHSType,
78 FirstInst->getOperand(1)->getName() + ".pn");
79 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
80 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
81 InsertNewInstBefore(NewRHS, PN);
85 // Add all operands to the new PHIs.
86 if (NewLHS || NewRHS) {
87 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
88 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
90 Value *NewInLHS = InInst->getOperand(0);
91 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
94 Value *NewInRHS = InInst->getOperand(1);
95 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
100 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
101 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
102 CmpInst *CIOp = cast<CmpInst>(FirstInst);
103 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
107 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
108 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
110 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
111 FirstInst->op_end());
112 // This is true if all GEP bases are allocas and if all indices into them are
114 bool AllBasePointersAreAllocas = true;
116 // We don't want to replace this phi if the replacement would require
117 // more than one phi, which leads to higher register pressure. This is
118 // especially bad when the PHIs are in the header of a loop.
119 bool NeededPhi = false;
121 // Scan to see if all operands are the same opcode, and all have one use.
122 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
123 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
124 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() ||
125 GEP->getNumOperands() != FirstInst->getNumOperands())
128 // Keep track of whether or not all GEPs are of alloca pointers.
129 if (AllBasePointersAreAllocas &&
130 (!isa<AllocaInst>(GEP->getOperand(0)) ||
131 !GEP->hasAllConstantIndices()))
132 AllBasePointersAreAllocas = false;
134 // Compare the operand lists.
135 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
136 if (FirstInst->getOperand(op) == GEP->getOperand(op))
139 // Don't merge two GEPs when two operands differ (introducing phi nodes)
140 // if one of the PHIs has a constant for the index. The index may be
141 // substantially cheaper to compute for the constants, so making it a
142 // variable index could pessimize the path. This also handles the case
143 // for struct indices, which must always be constant.
144 if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
145 isa<ConstantInt>(GEP->getOperand(op)))
148 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
151 // If we already needed a PHI for an earlier operand, and another operand
152 // also requires a PHI, we'd be introducing more PHIs than we're
153 // eliminating, which increases register pressure on entry to the PHI's
158 FixedOperands[op] = 0; // Needs a PHI.
163 // If all of the base pointers of the PHI'd GEPs are from allocas, don't
164 // bother doing this transformation. At best, this will just save a bit of
165 // offset calculation, but all the predecessors will have to materialize the
166 // stack address into a register anyway. We'd actually rather *clone* the
167 // load up into the predecessors so that we have a load of a gep of an alloca,
168 // which can usually all be folded into the load.
169 if (AllBasePointersAreAllocas)
172 // Otherwise, this is safe to transform. Insert PHI nodes for each operand
174 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
176 bool HasAnyPHIs = false;
177 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
178 if (FixedOperands[i]) continue; // operand doesn't need a phi.
179 Value *FirstOp = FirstInst->getOperand(i);
180 PHINode *NewPN = PHINode::Create(FirstOp->getType(),
181 FirstOp->getName()+".pn");
182 InsertNewInstBefore(NewPN, PN);
184 NewPN->reserveOperandSpace(e);
185 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
186 OperandPhis[i] = NewPN;
187 FixedOperands[i] = NewPN;
192 // Add all operands to the new PHIs.
194 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
195 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
196 BasicBlock *InBB = PN.getIncomingBlock(i);
198 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
199 if (PHINode *OpPhi = OperandPhis[op])
200 OpPhi->addIncoming(InGEP->getOperand(op), InBB);
204 Value *Base = FixedOperands[0];
205 return cast<GEPOperator>(FirstInst)->isInBounds() ?
206 GetElementPtrInst::CreateInBounds(Base, FixedOperands.begin()+1,
207 FixedOperands.end()) :
208 GetElementPtrInst::Create(Base, FixedOperands.begin()+1,
209 FixedOperands.end());
213 /// isSafeAndProfitableToSinkLoad - Return true if we know that it is safe to
214 /// sink the load out of the block that defines it. This means that it must be
215 /// obvious the value of the load is not changed from the point of the load to
216 /// the end of the block it is in.
218 /// Finally, it is safe, but not profitable, to sink a load targetting a
219 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
221 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
222 BasicBlock::iterator BBI = L, E = L->getParent()->end();
224 for (++BBI; BBI != E; ++BBI)
225 if (BBI->mayWriteToMemory())
228 // Check for non-address taken alloca. If not address-taken already, it isn't
229 // profitable to do this xform.
230 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
231 bool isAddressTaken = false;
232 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
235 if (isa<LoadInst>(U)) continue;
236 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
237 // If storing TO the alloca, then the address isn't taken.
238 if (SI->getOperand(1) == AI) continue;
240 isAddressTaken = true;
244 if (!isAddressTaken && AI->isStaticAlloca())
248 // If this load is a load from a GEP with a constant offset from an alloca,
249 // then we don't want to sink it. In its present form, it will be
250 // load [constant stack offset]. Sinking it will cause us to have to
251 // materialize the stack addresses in each predecessor in a register only to
252 // do a shared load from register in the successor.
253 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))
254 if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))
255 if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
261 Instruction *InstCombiner::FoldPHIArgLoadIntoPHI(PHINode &PN) {
262 LoadInst *FirstLI = cast<LoadInst>(PN.getIncomingValue(0));
264 // When processing loads, we need to propagate two bits of information to the
265 // sunk load: whether it is volatile, and what its alignment is. We currently
266 // don't sink loads when some have their alignment specified and some don't.
267 // visitLoadInst will propagate an alignment onto the load when TD is around,
268 // and if TD isn't around, we can't handle the mixed case.
269 bool isVolatile = FirstLI->isVolatile();
270 unsigned LoadAlignment = FirstLI->getAlignment();
271 unsigned LoadAddrSpace = FirstLI->getPointerAddressSpace();
273 // We can't sink the load if the loaded value could be modified between the
275 if (FirstLI->getParent() != PN.getIncomingBlock(0) ||
276 !isSafeAndProfitableToSinkLoad(FirstLI))
279 // If the PHI is of volatile loads and the load block has multiple
280 // successors, sinking it would remove a load of the volatile value from
281 // the path through the other successor.
283 FirstLI->getParent()->getTerminator()->getNumSuccessors() != 1)
286 // Check to see if all arguments are the same operation.
287 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
288 LoadInst *LI = dyn_cast<LoadInst>(PN.getIncomingValue(i));
289 if (!LI || !LI->hasOneUse())
292 // We can't sink the load if the loaded value could be modified between
293 // the load and the PHI.
294 if (LI->isVolatile() != isVolatile ||
295 LI->getParent() != PN.getIncomingBlock(i) ||
296 LI->getPointerAddressSpace() != LoadAddrSpace ||
297 !isSafeAndProfitableToSinkLoad(LI))
300 // If some of the loads have an alignment specified but not all of them,
301 // we can't do the transformation.
302 if ((LoadAlignment != 0) != (LI->getAlignment() != 0))
305 LoadAlignment = std::min(LoadAlignment, LI->getAlignment());
307 // If the PHI is of volatile loads and the load block has multiple
308 // successors, sinking it would remove a load of the volatile value from
309 // the path through the other successor.
311 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
315 // Okay, they are all the same operation. Create a new PHI node of the
316 // correct type, and PHI together all of the LHS's of the instructions.
317 PHINode *NewPN = PHINode::Create(FirstLI->getOperand(0)->getType(),
319 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
321 Value *InVal = FirstLI->getOperand(0);
322 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
324 // Add all operands to the new PHI.
325 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
326 Value *NewInVal = cast<LoadInst>(PN.getIncomingValue(i))->getOperand(0);
327 if (NewInVal != InVal)
329 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
334 // The new PHI unions all of the same values together. This is really
335 // common, so we handle it intelligently here for compile-time speed.
339 InsertNewInstBefore(NewPN, PN);
343 // If this was a volatile load that we are merging, make sure to loop through
344 // and mark all the input loads as non-volatile. If we don't do this, we will
345 // insert a new volatile load and the old ones will not be deletable.
347 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
348 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
350 return new LoadInst(PhiVal, "", isVolatile, LoadAlignment);
355 /// FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
356 /// operator and they all are only used by the PHI, PHI together their
357 /// inputs, and do the operation once, to the result of the PHI.
358 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
359 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
361 if (isa<GetElementPtrInst>(FirstInst))
362 return FoldPHIArgGEPIntoPHI(PN);
363 if (isa<LoadInst>(FirstInst))
364 return FoldPHIArgLoadIntoPHI(PN);
366 // Scan the instruction, looking for input operations that can be folded away.
367 // If all input operands to the phi are the same instruction (e.g. a cast from
368 // the same type or "+42") we can pull the operation through the PHI, reducing
369 // code size and simplifying code.
370 Constant *ConstantOp = 0;
371 const Type *CastSrcTy = 0;
373 if (isa<CastInst>(FirstInst)) {
374 CastSrcTy = FirstInst->getOperand(0)->getType();
376 // Be careful about transforming integer PHIs. We don't want to pessimize
377 // the code by turning an i32 into an i1293.
378 if (PN.getType()->isIntegerTy() && CastSrcTy->isIntegerTy()) {
379 if (!ShouldChangeType(PN.getType(), CastSrcTy))
382 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
383 // Can fold binop, compare or shift here if the RHS is a constant,
384 // otherwise call FoldPHIArgBinOpIntoPHI.
385 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
387 return FoldPHIArgBinOpIntoPHI(PN);
389 return 0; // Cannot fold this operation.
392 // Check to see if all arguments are the same operation.
393 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
394 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
395 if (I == 0 || !I->hasOneUse() || !I->isSameOperationAs(FirstInst))
398 if (I->getOperand(0)->getType() != CastSrcTy)
399 return 0; // Cast operation must match.
400 } else if (I->getOperand(1) != ConstantOp) {
405 // Okay, they are all the same operation. Create a new PHI node of the
406 // correct type, and PHI together all of the LHS's of the instructions.
407 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
409 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
411 Value *InVal = FirstInst->getOperand(0);
412 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
414 // Add all operands to the new PHI.
415 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
416 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
417 if (NewInVal != InVal)
419 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
424 // The new PHI unions all of the same values together. This is really
425 // common, so we handle it intelligently here for compile-time speed.
429 InsertNewInstBefore(NewPN, PN);
433 // Insert and return the new operation.
434 if (CastInst *FirstCI = dyn_cast<CastInst>(FirstInst))
435 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
437 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
438 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
440 CmpInst *CIOp = cast<CmpInst>(FirstInst);
441 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
445 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
447 static bool DeadPHICycle(PHINode *PN,
448 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
449 if (PN->use_empty()) return true;
450 if (!PN->hasOneUse()) return false;
452 // Remember this node, and if we find the cycle, return.
453 if (!PotentiallyDeadPHIs.insert(PN))
456 // Don't scan crazily complex things.
457 if (PotentiallyDeadPHIs.size() == 16)
460 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
461 return DeadPHICycle(PU, PotentiallyDeadPHIs);
466 /// PHIsEqualValue - Return true if this phi node is always equal to
467 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
468 /// z = some value; x = phi (y, z); y = phi (x, z)
469 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
470 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
471 // See if we already saw this PHI node.
472 if (!ValueEqualPHIs.insert(PN))
475 // Don't scan crazily complex things.
476 if (ValueEqualPHIs.size() == 16)
479 // Scan the operands to see if they are either phi nodes or are equal to
481 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
482 Value *Op = PN->getIncomingValue(i);
483 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
484 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
486 } else if (Op != NonPhiInVal)
495 struct PHIUsageRecord {
496 unsigned PHIId; // The ID # of the PHI (something determinstic to sort on)
497 unsigned Shift; // The amount shifted.
498 Instruction *Inst; // The trunc instruction.
500 PHIUsageRecord(unsigned pn, unsigned Sh, Instruction *User)
501 : PHIId(pn), Shift(Sh), Inst(User) {}
503 bool operator<(const PHIUsageRecord &RHS) const {
504 if (PHIId < RHS.PHIId) return true;
505 if (PHIId > RHS.PHIId) return false;
506 if (Shift < RHS.Shift) return true;
507 if (Shift > RHS.Shift) return false;
508 return Inst->getType()->getPrimitiveSizeInBits() <
509 RHS.Inst->getType()->getPrimitiveSizeInBits();
513 struct LoweredPHIRecord {
514 PHINode *PN; // The PHI that was lowered.
515 unsigned Shift; // The amount shifted.
516 unsigned Width; // The width extracted.
518 LoweredPHIRecord(PHINode *pn, unsigned Sh, const Type *Ty)
519 : PN(pn), Shift(Sh), Width(Ty->getPrimitiveSizeInBits()) {}
521 // Ctor form used by DenseMap.
522 LoweredPHIRecord(PHINode *pn, unsigned Sh)
523 : PN(pn), Shift(Sh), Width(0) {}
529 struct DenseMapInfo<LoweredPHIRecord> {
530 static inline LoweredPHIRecord getEmptyKey() {
531 return LoweredPHIRecord(0, 0);
533 static inline LoweredPHIRecord getTombstoneKey() {
534 return LoweredPHIRecord(0, 1);
536 static unsigned getHashValue(const LoweredPHIRecord &Val) {
537 return DenseMapInfo<PHINode*>::getHashValue(Val.PN) ^ (Val.Shift>>3) ^
540 static bool isEqual(const LoweredPHIRecord &LHS,
541 const LoweredPHIRecord &RHS) {
542 return LHS.PN == RHS.PN && LHS.Shift == RHS.Shift &&
543 LHS.Width == RHS.Width;
547 struct isPodLike<LoweredPHIRecord> { static const bool value = true; };
551 /// SliceUpIllegalIntegerPHI - This is an integer PHI and we know that it has an
552 /// illegal type: see if it is only used by trunc or trunc(lshr) operations. If
553 /// so, we split the PHI into the various pieces being extracted. This sort of
554 /// thing is introduced when SROA promotes an aggregate to large integer values.
556 /// TODO: The user of the trunc may be an bitcast to float/double/vector or an
557 /// inttoptr. We should produce new PHIs in the right type.
559 Instruction *InstCombiner::SliceUpIllegalIntegerPHI(PHINode &FirstPhi) {
560 // PHIUsers - Keep track of all of the truncated values extracted from a set
561 // of PHIs, along with their offset. These are the things we want to rewrite.
562 SmallVector<PHIUsageRecord, 16> PHIUsers;
564 // PHIs are often mutually cyclic, so we keep track of a whole set of PHI
565 // nodes which are extracted from. PHIsToSlice is a set we use to avoid
566 // revisiting PHIs, PHIsInspected is a ordered list of PHIs that we need to
567 // check the uses of (to ensure they are all extracts).
568 SmallVector<PHINode*, 8> PHIsToSlice;
569 SmallPtrSet<PHINode*, 8> PHIsInspected;
571 PHIsToSlice.push_back(&FirstPhi);
572 PHIsInspected.insert(&FirstPhi);
574 for (unsigned PHIId = 0; PHIId != PHIsToSlice.size(); ++PHIId) {
575 PHINode *PN = PHIsToSlice[PHIId];
577 // Scan the input list of the PHI. If any input is an invoke, and if the
578 // input is defined in the predecessor, then we won't be split the critical
579 // edge which is required to insert a truncate. Because of this, we have to
581 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
582 InvokeInst *II = dyn_cast<InvokeInst>(PN->getIncomingValue(i));
583 if (II == 0) continue;
584 if (II->getParent() != PN->getIncomingBlock(i))
587 // If we have a phi, and if it's directly in the predecessor, then we have
588 // a critical edge where we need to put the truncate. Since we can't
589 // split the edge in instcombine, we have to bail out.
594 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end();
596 Instruction *User = cast<Instruction>(*UI);
598 // If the user is a PHI, inspect its uses recursively.
599 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
600 if (PHIsInspected.insert(UserPN))
601 PHIsToSlice.push_back(UserPN);
605 // Truncates are always ok.
606 if (isa<TruncInst>(User)) {
607 PHIUsers.push_back(PHIUsageRecord(PHIId, 0, User));
611 // Otherwise it must be a lshr which can only be used by one trunc.
612 if (User->getOpcode() != Instruction::LShr ||
613 !User->hasOneUse() || !isa<TruncInst>(User->use_back()) ||
614 !isa<ConstantInt>(User->getOperand(1)))
617 unsigned Shift = cast<ConstantInt>(User->getOperand(1))->getZExtValue();
618 PHIUsers.push_back(PHIUsageRecord(PHIId, Shift, User->use_back()));
622 // If we have no users, they must be all self uses, just nuke the PHI.
623 if (PHIUsers.empty())
624 return ReplaceInstUsesWith(FirstPhi, UndefValue::get(FirstPhi.getType()));
626 // If this phi node is transformable, create new PHIs for all the pieces
627 // extracted out of it. First, sort the users by their offset and size.
628 array_pod_sort(PHIUsers.begin(), PHIUsers.end());
630 DEBUG(errs() << "SLICING UP PHI: " << FirstPhi << '\n';
631 for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i)
632 errs() << "AND USER PHI #" << i << ": " << *PHIsToSlice[i] <<'\n';
635 // PredValues - This is a temporary used when rewriting PHI nodes. It is
636 // hoisted out here to avoid construction/destruction thrashing.
637 DenseMap<BasicBlock*, Value*> PredValues;
639 // ExtractedVals - Each new PHI we introduce is saved here so we don't
640 // introduce redundant PHIs.
641 DenseMap<LoweredPHIRecord, PHINode*> ExtractedVals;
643 for (unsigned UserI = 0, UserE = PHIUsers.size(); UserI != UserE; ++UserI) {
644 unsigned PHIId = PHIUsers[UserI].PHIId;
645 PHINode *PN = PHIsToSlice[PHIId];
646 unsigned Offset = PHIUsers[UserI].Shift;
647 const Type *Ty = PHIUsers[UserI].Inst->getType();
651 // If we've already lowered a user like this, reuse the previously lowered
653 if ((EltPHI = ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)]) == 0) {
655 // Otherwise, Create the new PHI node for this user.
656 EltPHI = PHINode::Create(Ty, PN->getName()+".off"+Twine(Offset), PN);
657 assert(EltPHI->getType() != PN->getType() &&
658 "Truncate didn't shrink phi?");
660 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
661 BasicBlock *Pred = PN->getIncomingBlock(i);
662 Value *&PredVal = PredValues[Pred];
664 // If we already have a value for this predecessor, reuse it.
666 EltPHI->addIncoming(PredVal, Pred);
670 // Handle the PHI self-reuse case.
671 Value *InVal = PN->getIncomingValue(i);
674 EltPHI->addIncoming(PredVal, Pred);
678 if (PHINode *InPHI = dyn_cast<PHINode>(PN)) {
679 // If the incoming value was a PHI, and if it was one of the PHIs we
680 // already rewrote it, just use the lowered value.
681 if (Value *Res = ExtractedVals[LoweredPHIRecord(InPHI, Offset, Ty)]) {
683 EltPHI->addIncoming(PredVal, Pred);
688 // Otherwise, do an extract in the predecessor.
689 Builder->SetInsertPoint(Pred, Pred->getTerminator());
692 Res = Builder->CreateLShr(Res, ConstantInt::get(InVal->getType(),
694 Res = Builder->CreateTrunc(Res, Ty, "extract.t");
696 EltPHI->addIncoming(Res, Pred);
698 // If the incoming value was a PHI, and if it was one of the PHIs we are
699 // rewriting, we will ultimately delete the code we inserted. This
700 // means we need to revisit that PHI to make sure we extract out the
702 if (PHINode *OldInVal = dyn_cast<PHINode>(PN->getIncomingValue(i)))
703 if (PHIsInspected.count(OldInVal)) {
704 unsigned RefPHIId = std::find(PHIsToSlice.begin(),PHIsToSlice.end(),
705 OldInVal)-PHIsToSlice.begin();
706 PHIUsers.push_back(PHIUsageRecord(RefPHIId, Offset,
707 cast<Instruction>(Res)));
713 DEBUG(errs() << " Made element PHI for offset " << Offset << ": "
715 ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)] = EltPHI;
718 // Replace the use of this piece with the PHI node.
719 ReplaceInstUsesWith(*PHIUsers[UserI].Inst, EltPHI);
722 // Replace all the remaining uses of the PHI nodes (self uses and the lshrs)
724 Value *Undef = UndefValue::get(FirstPhi.getType());
725 for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i)
726 ReplaceInstUsesWith(*PHIsToSlice[i], Undef);
727 return ReplaceInstUsesWith(FirstPhi, Undef);
730 // PHINode simplification
732 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
733 // If LCSSA is around, don't mess with Phi nodes
734 if (MustPreserveLCSSA) return 0;
736 if (Value *V = SimplifyInstruction(&PN, TD))
737 return ReplaceInstUsesWith(PN, V);
739 // If all PHI operands are the same operation, pull them through the PHI,
740 // reducing code size.
741 if (isa<Instruction>(PN.getIncomingValue(0)) &&
742 isa<Instruction>(PN.getIncomingValue(1)) &&
743 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
744 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
745 // FIXME: The hasOneUse check will fail for PHIs that use the value more
746 // than themselves more than once.
747 PN.getIncomingValue(0)->hasOneUse())
748 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
751 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
752 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
753 // PHI)... break the cycle.
754 if (PN.hasOneUse()) {
755 Instruction *PHIUser = cast<Instruction>(PN.use_back());
756 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
757 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
758 PotentiallyDeadPHIs.insert(&PN);
759 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
760 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
763 // If this phi has a single use, and if that use just computes a value for
764 // the next iteration of a loop, delete the phi. This occurs with unused
765 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
766 // common case here is good because the only other things that catch this
767 // are induction variable analysis (sometimes) and ADCE, which is only run
769 if (PHIUser->hasOneUse() &&
770 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
771 PHIUser->use_back() == &PN) {
772 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
776 // We sometimes end up with phi cycles that non-obviously end up being the
777 // same value, for example:
778 // z = some value; x = phi (y, z); y = phi (x, z)
779 // where the phi nodes don't necessarily need to be in the same block. Do a
780 // quick check to see if the PHI node only contains a single non-phi value, if
781 // so, scan to see if the phi cycle is actually equal to that value.
783 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
784 // Scan for the first non-phi operand.
785 while (InValNo != NumOperandVals &&
786 isa<PHINode>(PN.getIncomingValue(InValNo)))
789 if (InValNo != NumOperandVals) {
790 Value *NonPhiInVal = PN.getOperand(InValNo);
792 // Scan the rest of the operands to see if there are any conflicts, if so
793 // there is no need to recursively scan other phis.
794 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
795 Value *OpVal = PN.getIncomingValue(InValNo);
796 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
800 // If we scanned over all operands, then we have one unique value plus
801 // phi values. Scan PHI nodes to see if they all merge in each other or
803 if (InValNo == NumOperandVals) {
804 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
805 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
806 return ReplaceInstUsesWith(PN, NonPhiInVal);
811 // If there are multiple PHIs, sort their operands so that they all list
812 // the blocks in the same order. This will help identical PHIs be eliminated
813 // by other passes. Other passes shouldn't depend on this for correctness
815 PHINode *FirstPN = cast<PHINode>(PN.getParent()->begin());
817 for (unsigned i = 0, e = FirstPN->getNumIncomingValues(); i != e; ++i) {
818 BasicBlock *BBA = PN.getIncomingBlock(i);
819 BasicBlock *BBB = FirstPN->getIncomingBlock(i);
821 Value *VA = PN.getIncomingValue(i);
822 unsigned j = PN.getBasicBlockIndex(BBB);
823 Value *VB = PN.getIncomingValue(j);
824 PN.setIncomingBlock(i, BBB);
825 PN.setIncomingValue(i, VB);
826 PN.setIncomingBlock(j, BBA);
827 PN.setIncomingValue(j, VA);
828 // NOTE: Instcombine normally would want us to "return &PN" if we
829 // modified any of the operands of an instruction. However, since we
830 // aren't adding or removing uses (just rearranging them) we don't do
831 // this in this case.
835 // If this is an integer PHI and we know that it has an illegal type, see if
836 // it is only used by trunc or trunc(lshr) operations. If so, we split the
837 // PHI into the various pieces being extracted. This sort of thing is
838 // introduced when SROA promotes an aggregate to a single large integer type.
839 if (PN.getType()->isIntegerTy() && TD &&
840 !TD->isLegalInteger(PN.getType()->getPrimitiveSizeInBits()))
841 if (Instruction *Res = SliceUpIllegalIntegerPHI(PN))