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/Target/TargetData.h"
16 #include "llvm/ADT/SmallPtrSet.h"
17 #include "llvm/ADT/STLExtras.h"
20 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(a,c)]
21 /// and if a/b/c and the add's all have a single use, turn this into a phi
22 /// and a single binop.
23 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
24 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
25 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
26 unsigned Opc = FirstInst->getOpcode();
27 Value *LHSVal = FirstInst->getOperand(0);
28 Value *RHSVal = FirstInst->getOperand(1);
30 const Type *LHSType = LHSVal->getType();
31 const Type *RHSType = RHSVal->getType();
33 // Scan to see if all operands are the same opcode, and all have one use.
34 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
35 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
36 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
37 // Verify type of the LHS matches so we don't fold cmp's of different
38 // types or GEP's with different index types.
39 I->getOperand(0)->getType() != LHSType ||
40 I->getOperand(1)->getType() != RHSType)
43 // If they are CmpInst instructions, check their predicates
44 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
45 if (cast<CmpInst>(I)->getPredicate() !=
46 cast<CmpInst>(FirstInst)->getPredicate())
49 // Keep track of which operand needs a phi node.
50 if (I->getOperand(0) != LHSVal) LHSVal = 0;
51 if (I->getOperand(1) != RHSVal) RHSVal = 0;
54 // If both LHS and RHS would need a PHI, don't do this transformation,
55 // because it would increase the number of PHIs entering the block,
56 // which leads to higher register pressure. This is especially
57 // bad when the PHIs are in the header of a loop.
58 if (!LHSVal && !RHSVal)
61 // Otherwise, this is safe to transform!
63 Value *InLHS = FirstInst->getOperand(0);
64 Value *InRHS = FirstInst->getOperand(1);
65 PHINode *NewLHS = 0, *NewRHS = 0;
67 NewLHS = PHINode::Create(LHSType,
68 FirstInst->getOperand(0)->getName() + ".pn");
69 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
70 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
71 InsertNewInstBefore(NewLHS, PN);
76 NewRHS = PHINode::Create(RHSType,
77 FirstInst->getOperand(1)->getName() + ".pn");
78 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
79 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
80 InsertNewInstBefore(NewRHS, PN);
84 // Add all operands to the new PHIs.
85 if (NewLHS || NewRHS) {
86 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
87 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
89 Value *NewInLHS = InInst->getOperand(0);
90 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
93 Value *NewInRHS = InInst->getOperand(1);
94 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
99 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
100 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
101 CmpInst *CIOp = cast<CmpInst>(FirstInst);
102 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
106 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
107 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
109 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
110 FirstInst->op_end());
111 // This is true if all GEP bases are allocas and if all indices into them are
113 bool AllBasePointersAreAllocas = true;
115 // We don't want to replace this phi if the replacement would require
116 // more than one phi, which leads to higher register pressure. This is
117 // especially bad when the PHIs are in the header of a loop.
118 bool NeededPhi = false;
120 // Scan to see if all operands are the same opcode, and all have one use.
121 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
122 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
123 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() ||
124 GEP->getNumOperands() != FirstInst->getNumOperands())
127 // Keep track of whether or not all GEPs are of alloca pointers.
128 if (AllBasePointersAreAllocas &&
129 (!isa<AllocaInst>(GEP->getOperand(0)) ||
130 !GEP->hasAllConstantIndices()))
131 AllBasePointersAreAllocas = false;
133 // Compare the operand lists.
134 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
135 if (FirstInst->getOperand(op) == GEP->getOperand(op))
138 // Don't merge two GEPs when two operands differ (introducing phi nodes)
139 // if one of the PHIs has a constant for the index. The index may be
140 // substantially cheaper to compute for the constants, so making it a
141 // variable index could pessimize the path. This also handles the case
142 // for struct indices, which must always be constant.
143 if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
144 isa<ConstantInt>(GEP->getOperand(op)))
147 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
150 // If we already needed a PHI for an earlier operand, and another operand
151 // also requires a PHI, we'd be introducing more PHIs than we're
152 // eliminating, which increases register pressure on entry to the PHI's
157 FixedOperands[op] = 0; // Needs a PHI.
162 // If all of the base pointers of the PHI'd GEPs are from allocas, don't
163 // bother doing this transformation. At best, this will just save a bit of
164 // offset calculation, but all the predecessors will have to materialize the
165 // stack address into a register anyway. We'd actually rather *clone* the
166 // load up into the predecessors so that we have a load of a gep of an alloca,
167 // which can usually all be folded into the load.
168 if (AllBasePointersAreAllocas)
171 // Otherwise, this is safe to transform. Insert PHI nodes for each operand
173 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
175 bool HasAnyPHIs = false;
176 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
177 if (FixedOperands[i]) continue; // operand doesn't need a phi.
178 Value *FirstOp = FirstInst->getOperand(i);
179 PHINode *NewPN = PHINode::Create(FirstOp->getType(),
180 FirstOp->getName()+".pn");
181 InsertNewInstBefore(NewPN, PN);
183 NewPN->reserveOperandSpace(e);
184 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
185 OperandPhis[i] = NewPN;
186 FixedOperands[i] = NewPN;
191 // Add all operands to the new PHIs.
193 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
194 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
195 BasicBlock *InBB = PN.getIncomingBlock(i);
197 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
198 if (PHINode *OpPhi = OperandPhis[op])
199 OpPhi->addIncoming(InGEP->getOperand(op), InBB);
203 Value *Base = FixedOperands[0];
204 return cast<GEPOperator>(FirstInst)->isInBounds() ?
205 GetElementPtrInst::CreateInBounds(Base, FixedOperands.begin()+1,
206 FixedOperands.end()) :
207 GetElementPtrInst::Create(Base, FixedOperands.begin()+1,
208 FixedOperands.end());
212 /// isSafeAndProfitableToSinkLoad - Return true if we know that it is safe to
213 /// sink the load out of the block that defines it. This means that it must be
214 /// obvious the value of the load is not changed from the point of the load to
215 /// the end of the block it is in.
217 /// Finally, it is safe, but not profitable, to sink a load targetting a
218 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
220 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
221 BasicBlock::iterator BBI = L, E = L->getParent()->end();
223 for (++BBI; BBI != E; ++BBI)
224 if (BBI->mayWriteToMemory())
227 // Check for non-address taken alloca. If not address-taken already, it isn't
228 // profitable to do this xform.
229 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
230 bool isAddressTaken = false;
231 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
234 if (isa<LoadInst>(U)) continue;
235 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
236 // If storing TO the alloca, then the address isn't taken.
237 if (SI->getOperand(1) == AI) continue;
239 isAddressTaken = true;
243 if (!isAddressTaken && AI->isStaticAlloca())
247 // If this load is a load from a GEP with a constant offset from an alloca,
248 // then we don't want to sink it. In its present form, it will be
249 // load [constant stack offset]. Sinking it will cause us to have to
250 // materialize the stack addresses in each predecessor in a register only to
251 // do a shared load from register in the successor.
252 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))
253 if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))
254 if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
260 Instruction *InstCombiner::FoldPHIArgLoadIntoPHI(PHINode &PN) {
261 LoadInst *FirstLI = cast<LoadInst>(PN.getIncomingValue(0));
263 // When processing loads, we need to propagate two bits of information to the
264 // sunk load: whether it is volatile, and what its alignment is. We currently
265 // don't sink loads when some have their alignment specified and some don't.
266 // visitLoadInst will propagate an alignment onto the load when TD is around,
267 // and if TD isn't around, we can't handle the mixed case.
268 bool isVolatile = FirstLI->isVolatile();
269 unsigned LoadAlignment = FirstLI->getAlignment();
270 unsigned LoadAddrSpace = FirstLI->getPointerAddressSpace();
272 // We can't sink the load if the loaded value could be modified between the
274 if (FirstLI->getParent() != PN.getIncomingBlock(0) ||
275 !isSafeAndProfitableToSinkLoad(FirstLI))
278 // If the PHI is of volatile loads and the load block has multiple
279 // successors, sinking it would remove a load of the volatile value from
280 // the path through the other successor.
282 FirstLI->getParent()->getTerminator()->getNumSuccessors() != 1)
285 // Check to see if all arguments are the same operation.
286 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
287 LoadInst *LI = dyn_cast<LoadInst>(PN.getIncomingValue(i));
288 if (!LI || !LI->hasOneUse())
291 // We can't sink the load if the loaded value could be modified between
292 // the load and the PHI.
293 if (LI->isVolatile() != isVolatile ||
294 LI->getParent() != PN.getIncomingBlock(i) ||
295 LI->getPointerAddressSpace() != LoadAddrSpace ||
296 !isSafeAndProfitableToSinkLoad(LI))
299 // If some of the loads have an alignment specified but not all of them,
300 // we can't do the transformation.
301 if ((LoadAlignment != 0) != (LI->getAlignment() != 0))
304 LoadAlignment = std::min(LoadAlignment, LI->getAlignment());
306 // If the PHI is of volatile loads and the load block has multiple
307 // successors, sinking it would remove a load of the volatile value from
308 // the path through the other successor.
310 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
314 // Okay, they are all the same operation. Create a new PHI node of the
315 // correct type, and PHI together all of the LHS's of the instructions.
316 PHINode *NewPN = PHINode::Create(FirstLI->getOperand(0)->getType(),
318 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
320 Value *InVal = FirstLI->getOperand(0);
321 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
323 // Add all operands to the new PHI.
324 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
325 Value *NewInVal = cast<LoadInst>(PN.getIncomingValue(i))->getOperand(0);
326 if (NewInVal != InVal)
328 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
333 // The new PHI unions all of the same values together. This is really
334 // common, so we handle it intelligently here for compile-time speed.
338 InsertNewInstBefore(NewPN, PN);
342 // If this was a volatile load that we are merging, make sure to loop through
343 // and mark all the input loads as non-volatile. If we don't do this, we will
344 // insert a new volatile load and the old ones will not be deletable.
346 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
347 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
349 return new LoadInst(PhiVal, "", isVolatile, LoadAlignment);
354 /// FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
355 /// operator and they all are only used by the PHI, PHI together their
356 /// inputs, and do the operation once, to the result of the PHI.
357 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
358 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
360 if (isa<GetElementPtrInst>(FirstInst))
361 return FoldPHIArgGEPIntoPHI(PN);
362 if (isa<LoadInst>(FirstInst))
363 return FoldPHIArgLoadIntoPHI(PN);
365 // Scan the instruction, looking for input operations that can be folded away.
366 // If all input operands to the phi are the same instruction (e.g. a cast from
367 // the same type or "+42") we can pull the operation through the PHI, reducing
368 // code size and simplifying code.
369 Constant *ConstantOp = 0;
370 const Type *CastSrcTy = 0;
372 if (isa<CastInst>(FirstInst)) {
373 CastSrcTy = FirstInst->getOperand(0)->getType();
375 // Be careful about transforming integer PHIs. We don't want to pessimize
376 // the code by turning an i32 into an i1293.
377 if (PN.getType()->isIntegerTy() && CastSrcTy->isIntegerTy()) {
378 if (!ShouldChangeType(PN.getType(), CastSrcTy))
381 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
382 // Can fold binop, compare or shift here if the RHS is a constant,
383 // otherwise call FoldPHIArgBinOpIntoPHI.
384 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
386 return FoldPHIArgBinOpIntoPHI(PN);
388 return 0; // Cannot fold this operation.
391 // Check to see if all arguments are the same operation.
392 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
393 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
394 if (I == 0 || !I->hasOneUse() || !I->isSameOperationAs(FirstInst))
397 if (I->getOperand(0)->getType() != CastSrcTy)
398 return 0; // Cast operation must match.
399 } else if (I->getOperand(1) != ConstantOp) {
404 // Okay, they are all the same operation. Create a new PHI node of the
405 // correct type, and PHI together all of the LHS's of the instructions.
406 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
408 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
410 Value *InVal = FirstInst->getOperand(0);
411 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
413 // Add all operands to the new PHI.
414 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
415 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
416 if (NewInVal != InVal)
418 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
423 // The new PHI unions all of the same values together. This is really
424 // common, so we handle it intelligently here for compile-time speed.
428 InsertNewInstBefore(NewPN, PN);
432 // Insert and return the new operation.
433 if (CastInst *FirstCI = dyn_cast<CastInst>(FirstInst))
434 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
436 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
437 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
439 CmpInst *CIOp = cast<CmpInst>(FirstInst);
440 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
444 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
446 static bool DeadPHICycle(PHINode *PN,
447 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
448 if (PN->use_empty()) return true;
449 if (!PN->hasOneUse()) return false;
451 // Remember this node, and if we find the cycle, return.
452 if (!PotentiallyDeadPHIs.insert(PN))
455 // Don't scan crazily complex things.
456 if (PotentiallyDeadPHIs.size() == 16)
459 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
460 return DeadPHICycle(PU, PotentiallyDeadPHIs);
465 /// PHIsEqualValue - Return true if this phi node is always equal to
466 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
467 /// z = some value; x = phi (y, z); y = phi (x, z)
468 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
469 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
470 // See if we already saw this PHI node.
471 if (!ValueEqualPHIs.insert(PN))
474 // Don't scan crazily complex things.
475 if (ValueEqualPHIs.size() == 16)
478 // Scan the operands to see if they are either phi nodes or are equal to
480 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
481 Value *Op = PN->getIncomingValue(i);
482 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
483 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
485 } else if (Op != NonPhiInVal)
494 struct PHIUsageRecord {
495 unsigned PHIId; // The ID # of the PHI (something determinstic to sort on)
496 unsigned Shift; // The amount shifted.
497 Instruction *Inst; // The trunc instruction.
499 PHIUsageRecord(unsigned pn, unsigned Sh, Instruction *User)
500 : PHIId(pn), Shift(Sh), Inst(User) {}
502 bool operator<(const PHIUsageRecord &RHS) const {
503 if (PHIId < RHS.PHIId) return true;
504 if (PHIId > RHS.PHIId) return false;
505 if (Shift < RHS.Shift) return true;
506 if (Shift > RHS.Shift) return false;
507 return Inst->getType()->getPrimitiveSizeInBits() <
508 RHS.Inst->getType()->getPrimitiveSizeInBits();
512 struct LoweredPHIRecord {
513 PHINode *PN; // The PHI that was lowered.
514 unsigned Shift; // The amount shifted.
515 unsigned Width; // The width extracted.
517 LoweredPHIRecord(PHINode *pn, unsigned Sh, const Type *Ty)
518 : PN(pn), Shift(Sh), Width(Ty->getPrimitiveSizeInBits()) {}
520 // Ctor form used by DenseMap.
521 LoweredPHIRecord(PHINode *pn, unsigned Sh)
522 : PN(pn), Shift(Sh), Width(0) {}
528 struct DenseMapInfo<LoweredPHIRecord> {
529 static inline LoweredPHIRecord getEmptyKey() {
530 return LoweredPHIRecord(0, 0);
532 static inline LoweredPHIRecord getTombstoneKey() {
533 return LoweredPHIRecord(0, 1);
535 static unsigned getHashValue(const LoweredPHIRecord &Val) {
536 return DenseMapInfo<PHINode*>::getHashValue(Val.PN) ^ (Val.Shift>>3) ^
539 static bool isEqual(const LoweredPHIRecord &LHS,
540 const LoweredPHIRecord &RHS) {
541 return LHS.PN == RHS.PN && LHS.Shift == RHS.Shift &&
542 LHS.Width == RHS.Width;
546 struct isPodLike<LoweredPHIRecord> { static const bool value = true; };
550 /// SliceUpIllegalIntegerPHI - This is an integer PHI and we know that it has an
551 /// illegal type: see if it is only used by trunc or trunc(lshr) operations. If
552 /// so, we split the PHI into the various pieces being extracted. This sort of
553 /// thing is introduced when SROA promotes an aggregate to large integer values.
555 /// TODO: The user of the trunc may be an bitcast to float/double/vector or an
556 /// inttoptr. We should produce new PHIs in the right type.
558 Instruction *InstCombiner::SliceUpIllegalIntegerPHI(PHINode &FirstPhi) {
559 // PHIUsers - Keep track of all of the truncated values extracted from a set
560 // of PHIs, along with their offset. These are the things we want to rewrite.
561 SmallVector<PHIUsageRecord, 16> PHIUsers;
563 // PHIs are often mutually cyclic, so we keep track of a whole set of PHI
564 // nodes which are extracted from. PHIsToSlice is a set we use to avoid
565 // revisiting PHIs, PHIsInspected is a ordered list of PHIs that we need to
566 // check the uses of (to ensure they are all extracts).
567 SmallVector<PHINode*, 8> PHIsToSlice;
568 SmallPtrSet<PHINode*, 8> PHIsInspected;
570 PHIsToSlice.push_back(&FirstPhi);
571 PHIsInspected.insert(&FirstPhi);
573 for (unsigned PHIId = 0; PHIId != PHIsToSlice.size(); ++PHIId) {
574 PHINode *PN = PHIsToSlice[PHIId];
576 // Scan the input list of the PHI. If any input is an invoke, and if the
577 // input is defined in the predecessor, then we won't be split the critical
578 // edge which is required to insert a truncate. Because of this, we have to
580 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
581 InvokeInst *II = dyn_cast<InvokeInst>(PN->getIncomingValue(i));
582 if (II == 0) continue;
583 if (II->getParent() != PN->getIncomingBlock(i))
586 // If we have a phi, and if it's directly in the predecessor, then we have
587 // a critical edge where we need to put the truncate. Since we can't
588 // split the edge in instcombine, we have to bail out.
593 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end();
595 Instruction *User = cast<Instruction>(*UI);
597 // If the user is a PHI, inspect its uses recursively.
598 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
599 if (PHIsInspected.insert(UserPN))
600 PHIsToSlice.push_back(UserPN);
604 // Truncates are always ok.
605 if (isa<TruncInst>(User)) {
606 PHIUsers.push_back(PHIUsageRecord(PHIId, 0, User));
610 // Otherwise it must be a lshr which can only be used by one trunc.
611 if (User->getOpcode() != Instruction::LShr ||
612 !User->hasOneUse() || !isa<TruncInst>(User->use_back()) ||
613 !isa<ConstantInt>(User->getOperand(1)))
616 unsigned Shift = cast<ConstantInt>(User->getOperand(1))->getZExtValue();
617 PHIUsers.push_back(PHIUsageRecord(PHIId, Shift, User->use_back()));
621 // If we have no users, they must be all self uses, just nuke the PHI.
622 if (PHIUsers.empty())
623 return ReplaceInstUsesWith(FirstPhi, UndefValue::get(FirstPhi.getType()));
625 // If this phi node is transformable, create new PHIs for all the pieces
626 // extracted out of it. First, sort the users by their offset and size.
627 array_pod_sort(PHIUsers.begin(), PHIUsers.end());
629 DEBUG(errs() << "SLICING UP PHI: " << FirstPhi << '\n';
630 for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i)
631 errs() << "AND USER PHI #" << i << ": " << *PHIsToSlice[i] <<'\n';
634 // PredValues - This is a temporary used when rewriting PHI nodes. It is
635 // hoisted out here to avoid construction/destruction thrashing.
636 DenseMap<BasicBlock*, Value*> PredValues;
638 // ExtractedVals - Each new PHI we introduce is saved here so we don't
639 // introduce redundant PHIs.
640 DenseMap<LoweredPHIRecord, PHINode*> ExtractedVals;
642 for (unsigned UserI = 0, UserE = PHIUsers.size(); UserI != UserE; ++UserI) {
643 unsigned PHIId = PHIUsers[UserI].PHIId;
644 PHINode *PN = PHIsToSlice[PHIId];
645 unsigned Offset = PHIUsers[UserI].Shift;
646 const Type *Ty = PHIUsers[UserI].Inst->getType();
650 // If we've already lowered a user like this, reuse the previously lowered
652 if ((EltPHI = ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)]) == 0) {
654 // Otherwise, Create the new PHI node for this user.
655 EltPHI = PHINode::Create(Ty, PN->getName()+".off"+Twine(Offset), PN);
656 assert(EltPHI->getType() != PN->getType() &&
657 "Truncate didn't shrink phi?");
659 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
660 BasicBlock *Pred = PN->getIncomingBlock(i);
661 Value *&PredVal = PredValues[Pred];
663 // If we already have a value for this predecessor, reuse it.
665 EltPHI->addIncoming(PredVal, Pred);
669 // Handle the PHI self-reuse case.
670 Value *InVal = PN->getIncomingValue(i);
673 EltPHI->addIncoming(PredVal, Pred);
677 if (PHINode *InPHI = dyn_cast<PHINode>(PN)) {
678 // If the incoming value was a PHI, and if it was one of the PHIs we
679 // already rewrote it, just use the lowered value.
680 if (Value *Res = ExtractedVals[LoweredPHIRecord(InPHI, Offset, Ty)]) {
682 EltPHI->addIncoming(PredVal, Pred);
687 // Otherwise, do an extract in the predecessor.
688 Builder->SetInsertPoint(Pred, Pred->getTerminator());
691 Res = Builder->CreateLShr(Res, ConstantInt::get(InVal->getType(),
693 Res = Builder->CreateTrunc(Res, Ty, "extract.t");
695 EltPHI->addIncoming(Res, Pred);
697 // If the incoming value was a PHI, and if it was one of the PHIs we are
698 // rewriting, we will ultimately delete the code we inserted. This
699 // means we need to revisit that PHI to make sure we extract out the
701 if (PHINode *OldInVal = dyn_cast<PHINode>(PN->getIncomingValue(i)))
702 if (PHIsInspected.count(OldInVal)) {
703 unsigned RefPHIId = std::find(PHIsToSlice.begin(),PHIsToSlice.end(),
704 OldInVal)-PHIsToSlice.begin();
705 PHIUsers.push_back(PHIUsageRecord(RefPHIId, Offset,
706 cast<Instruction>(Res)));
712 DEBUG(errs() << " Made element PHI for offset " << Offset << ": "
714 ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)] = EltPHI;
717 // Replace the use of this piece with the PHI node.
718 ReplaceInstUsesWith(*PHIUsers[UserI].Inst, EltPHI);
721 // Replace all the remaining uses of the PHI nodes (self uses and the lshrs)
723 Value *Undef = UndefValue::get(FirstPhi.getType());
724 for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i)
725 ReplaceInstUsesWith(*PHIsToSlice[i], Undef);
726 return ReplaceInstUsesWith(FirstPhi, Undef);
729 // PHINode simplification
731 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
732 // If LCSSA is around, don't mess with Phi nodes
733 if (MustPreserveLCSSA) return 0;
735 if (Value *V = PN.hasConstantValue())
736 return ReplaceInstUsesWith(PN, V);
738 // If all PHI operands are the same operation, pull them through the PHI,
739 // reducing code size.
740 if (isa<Instruction>(PN.getIncomingValue(0)) &&
741 isa<Instruction>(PN.getIncomingValue(1)) &&
742 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
743 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
744 // FIXME: The hasOneUse check will fail for PHIs that use the value more
745 // than themselves more than once.
746 PN.getIncomingValue(0)->hasOneUse())
747 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
750 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
751 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
752 // PHI)... break the cycle.
753 if (PN.hasOneUse()) {
754 Instruction *PHIUser = cast<Instruction>(PN.use_back());
755 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
756 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
757 PotentiallyDeadPHIs.insert(&PN);
758 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
759 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
762 // If this phi has a single use, and if that use just computes a value for
763 // the next iteration of a loop, delete the phi. This occurs with unused
764 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
765 // common case here is good because the only other things that catch this
766 // are induction variable analysis (sometimes) and ADCE, which is only run
768 if (PHIUser->hasOneUse() &&
769 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
770 PHIUser->use_back() == &PN) {
771 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
775 // We sometimes end up with phi cycles that non-obviously end up being the
776 // same value, for example:
777 // z = some value; x = phi (y, z); y = phi (x, z)
778 // where the phi nodes don't necessarily need to be in the same block. Do a
779 // quick check to see if the PHI node only contains a single non-phi value, if
780 // so, scan to see if the phi cycle is actually equal to that value.
782 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
783 // Scan for the first non-phi operand.
784 while (InValNo != NumOperandVals &&
785 isa<PHINode>(PN.getIncomingValue(InValNo)))
788 if (InValNo != NumOperandVals) {
789 Value *NonPhiInVal = PN.getOperand(InValNo);
791 // Scan the rest of the operands to see if there are any conflicts, if so
792 // there is no need to recursively scan other phis.
793 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
794 Value *OpVal = PN.getIncomingValue(InValNo);
795 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
799 // If we scanned over all operands, then we have one unique value plus
800 // phi values. Scan PHI nodes to see if they all merge in each other or
802 if (InValNo == NumOperandVals) {
803 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
804 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
805 return ReplaceInstUsesWith(PN, NonPhiInVal);
810 // If there are multiple PHIs, sort their operands so that they all list
811 // the blocks in the same order. This will help identical PHIs be eliminated
812 // by other passes. Other passes shouldn't depend on this for correctness
814 PHINode *FirstPN = cast<PHINode>(PN.getParent()->begin());
816 for (unsigned i = 0, e = FirstPN->getNumIncomingValues(); i != e; ++i) {
817 BasicBlock *BBA = PN.getIncomingBlock(i);
818 BasicBlock *BBB = FirstPN->getIncomingBlock(i);
820 Value *VA = PN.getIncomingValue(i);
821 unsigned j = PN.getBasicBlockIndex(BBB);
822 Value *VB = PN.getIncomingValue(j);
823 PN.setIncomingBlock(i, BBB);
824 PN.setIncomingValue(i, VB);
825 PN.setIncomingBlock(j, BBA);
826 PN.setIncomingValue(j, VA);
827 // NOTE: Instcombine normally would want us to "return &PN" if we
828 // modified any of the operands of an instruction. However, since we
829 // aren't adding or removing uses (just rearranging them) we don't do
830 // this in this case.
834 // If this is an integer PHI and we know that it has an illegal type, see if
835 // it is only used by trunc or trunc(lshr) operations. If so, we split the
836 // PHI into the various pieces being extracted. This sort of thing is
837 // introduced when SROA promotes an aggregate to a single large integer type.
838 if (PN.getType()->isIntegerTy() && TD &&
839 !TD->isLegalInteger(PN.getType()->getPrimitiveSizeInBits()))
840 if (Instruction *Res = SliceUpIllegalIntegerPHI(PN))