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 "InstCombineInternal.h"
15 #include "llvm/ADT/STLExtras.h"
16 #include "llvm/ADT/SmallPtrSet.h"
17 #include "llvm/Analysis/InstructionSimplify.h"
18 #include "llvm/Transforms/Utils/Local.h"
21 #define DEBUG_TYPE "instcombine"
23 /// If we have something like phi [add (a,b), add(a,c)] and if a/b/c and the
24 /// adds all have a single use, turn this into a phi and a single binop.
25 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
26 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
27 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
28 unsigned Opc = FirstInst->getOpcode();
29 Value *LHSVal = FirstInst->getOperand(0);
30 Value *RHSVal = FirstInst->getOperand(1);
32 Type *LHSType = LHSVal->getType();
33 Type *RHSType = RHSVal->getType();
35 bool isNUW = false, isNSW = false, isExact = false;
36 if (OverflowingBinaryOperator *BO =
37 dyn_cast<OverflowingBinaryOperator>(FirstInst)) {
38 isNUW = BO->hasNoUnsignedWrap();
39 isNSW = BO->hasNoSignedWrap();
40 } else if (PossiblyExactOperator *PEO =
41 dyn_cast<PossiblyExactOperator>(FirstInst))
42 isExact = PEO->isExact();
44 // Scan to see if all operands are the same opcode, and all have one use.
45 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
46 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
47 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
48 // Verify type of the LHS matches so we don't fold cmp's of different
50 I->getOperand(0)->getType() != LHSType ||
51 I->getOperand(1)->getType() != RHSType)
54 // If they are CmpInst instructions, check their predicates
55 if (CmpInst *CI = dyn_cast<CmpInst>(I))
56 if (CI->getPredicate() != cast<CmpInst>(FirstInst)->getPredicate())
60 isNUW = cast<OverflowingBinaryOperator>(I)->hasNoUnsignedWrap();
62 isNSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
64 isExact = cast<PossiblyExactOperator>(I)->isExact();
66 // Keep track of which operand needs a phi node.
67 if (I->getOperand(0) != LHSVal) LHSVal = nullptr;
68 if (I->getOperand(1) != RHSVal) RHSVal = nullptr;
71 // If both LHS and RHS would need a PHI, don't do this transformation,
72 // because it would increase the number of PHIs entering the block,
73 // which leads to higher register pressure. This is especially
74 // bad when the PHIs are in the header of a loop.
75 if (!LHSVal && !RHSVal)
78 // Otherwise, this is safe to transform!
80 Value *InLHS = FirstInst->getOperand(0);
81 Value *InRHS = FirstInst->getOperand(1);
82 PHINode *NewLHS = nullptr, *NewRHS = nullptr;
84 NewLHS = PHINode::Create(LHSType, PN.getNumIncomingValues(),
85 FirstInst->getOperand(0)->getName() + ".pn");
86 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
87 InsertNewInstBefore(NewLHS, PN);
92 NewRHS = PHINode::Create(RHSType, PN.getNumIncomingValues(),
93 FirstInst->getOperand(1)->getName() + ".pn");
94 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
95 InsertNewInstBefore(NewRHS, PN);
99 // Add all operands to the new PHIs.
100 if (NewLHS || NewRHS) {
101 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
102 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
104 Value *NewInLHS = InInst->getOperand(0);
105 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
108 Value *NewInRHS = InInst->getOperand(1);
109 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
114 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst)) {
115 CmpInst *NewCI = CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
117 NewCI->setDebugLoc(FirstInst->getDebugLoc());
121 BinaryOperator *BinOp = cast<BinaryOperator>(FirstInst);
122 BinaryOperator *NewBinOp =
123 BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
124 if (isNUW) NewBinOp->setHasNoUnsignedWrap();
125 if (isNSW) NewBinOp->setHasNoSignedWrap();
126 if (isExact) NewBinOp->setIsExact();
127 NewBinOp->setDebugLoc(FirstInst->getDebugLoc());
131 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
132 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
134 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
135 FirstInst->op_end());
136 // This is true if all GEP bases are allocas and if all indices into them are
138 bool AllBasePointersAreAllocas = true;
140 // We don't want to replace this phi if the replacement would require
141 // more than one phi, which leads to higher register pressure. This is
142 // especially bad when the PHIs are in the header of a loop.
143 bool NeededPhi = false;
145 bool AllInBounds = true;
147 // Scan to see if all operands are the same opcode, and all have one use.
148 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
149 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
150 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() ||
151 GEP->getNumOperands() != FirstInst->getNumOperands())
154 AllInBounds &= GEP->isInBounds();
156 // Keep track of whether or not all GEPs are of alloca pointers.
157 if (AllBasePointersAreAllocas &&
158 (!isa<AllocaInst>(GEP->getOperand(0)) ||
159 !GEP->hasAllConstantIndices()))
160 AllBasePointersAreAllocas = false;
162 // Compare the operand lists.
163 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
164 if (FirstInst->getOperand(op) == GEP->getOperand(op))
167 // Don't merge two GEPs when two operands differ (introducing phi nodes)
168 // if one of the PHIs has a constant for the index. The index may be
169 // substantially cheaper to compute for the constants, so making it a
170 // variable index could pessimize the path. This also handles the case
171 // for struct indices, which must always be constant.
172 if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
173 isa<ConstantInt>(GEP->getOperand(op)))
176 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
179 // If we already needed a PHI for an earlier operand, and another operand
180 // also requires a PHI, we'd be introducing more PHIs than we're
181 // eliminating, which increases register pressure on entry to the PHI's
186 FixedOperands[op] = nullptr; // Needs a PHI.
191 // If all of the base pointers of the PHI'd GEPs are from allocas, don't
192 // bother doing this transformation. At best, this will just save a bit of
193 // offset calculation, but all the predecessors will have to materialize the
194 // stack address into a register anyway. We'd actually rather *clone* the
195 // load up into the predecessors so that we have a load of a gep of an alloca,
196 // which can usually all be folded into the load.
197 if (AllBasePointersAreAllocas)
200 // Otherwise, this is safe to transform. Insert PHI nodes for each operand
202 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
204 bool HasAnyPHIs = false;
205 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
206 if (FixedOperands[i]) continue; // operand doesn't need a phi.
207 Value *FirstOp = FirstInst->getOperand(i);
208 PHINode *NewPN = PHINode::Create(FirstOp->getType(), e,
209 FirstOp->getName()+".pn");
210 InsertNewInstBefore(NewPN, PN);
212 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
213 OperandPhis[i] = NewPN;
214 FixedOperands[i] = NewPN;
219 // Add all operands to the new PHIs.
221 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
222 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
223 BasicBlock *InBB = PN.getIncomingBlock(i);
225 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
226 if (PHINode *OpPhi = OperandPhis[op])
227 OpPhi->addIncoming(InGEP->getOperand(op), InBB);
231 Value *Base = FixedOperands[0];
232 GetElementPtrInst *NewGEP =
233 GetElementPtrInst::Create(FirstInst->getSourceElementType(), Base,
234 makeArrayRef(FixedOperands).slice(1));
235 if (AllInBounds) NewGEP->setIsInBounds();
236 NewGEP->setDebugLoc(FirstInst->getDebugLoc());
241 /// Return true if we know that it is safe to sink the load out of the block
242 /// that defines it. This means that it must be obvious the value of the load is
243 /// not changed from the point of the load to the end of the block it is in.
245 /// Finally, it is safe, but not profitable, to sink a load targeting a
246 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
248 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
249 BasicBlock::iterator BBI = L->getIterator(), E = L->getParent()->end();
251 for (++BBI; BBI != E; ++BBI)
252 if (BBI->mayWriteToMemory())
255 // Check for non-address taken alloca. If not address-taken already, it isn't
256 // profitable to do this xform.
257 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
258 bool isAddressTaken = false;
259 for (User *U : AI->users()) {
260 if (isa<LoadInst>(U)) continue;
261 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
262 // If storing TO the alloca, then the address isn't taken.
263 if (SI->getOperand(1) == AI) continue;
265 isAddressTaken = true;
269 if (!isAddressTaken && AI->isStaticAlloca())
273 // If this load is a load from a GEP with a constant offset from an alloca,
274 // then we don't want to sink it. In its present form, it will be
275 // load [constant stack offset]. Sinking it will cause us to have to
276 // materialize the stack addresses in each predecessor in a register only to
277 // do a shared load from register in the successor.
278 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))
279 if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))
280 if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
286 Instruction *InstCombiner::FoldPHIArgLoadIntoPHI(PHINode &PN) {
287 LoadInst *FirstLI = cast<LoadInst>(PN.getIncomingValue(0));
289 // FIXME: This is overconservative; this transform is allowed in some cases
290 // for atomic operations.
291 if (FirstLI->isAtomic())
294 // When processing loads, we need to propagate two bits of information to the
295 // sunk load: whether it is volatile, and what its alignment is. We currently
296 // don't sink loads when some have their alignment specified and some don't.
297 // visitLoadInst will propagate an alignment onto the load when TD is around,
298 // and if TD isn't around, we can't handle the mixed case.
299 bool isVolatile = FirstLI->isVolatile();
300 unsigned LoadAlignment = FirstLI->getAlignment();
301 unsigned LoadAddrSpace = FirstLI->getPointerAddressSpace();
303 // We can't sink the load if the loaded value could be modified between the
305 if (FirstLI->getParent() != PN.getIncomingBlock(0) ||
306 !isSafeAndProfitableToSinkLoad(FirstLI))
309 // If the PHI is of volatile loads and the load block has multiple
310 // successors, sinking it would remove a load of the volatile value from
311 // the path through the other successor.
313 FirstLI->getParent()->getTerminator()->getNumSuccessors() != 1)
316 // Check to see if all arguments are the same operation.
317 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
318 LoadInst *LI = dyn_cast<LoadInst>(PN.getIncomingValue(i));
319 if (!LI || !LI->hasOneUse())
322 // We can't sink the load if the loaded value could be modified between
323 // the load and the PHI.
324 if (LI->isVolatile() != isVolatile ||
325 LI->getParent() != PN.getIncomingBlock(i) ||
326 LI->getPointerAddressSpace() != LoadAddrSpace ||
327 !isSafeAndProfitableToSinkLoad(LI))
330 // If some of the loads have an alignment specified but not all of them,
331 // we can't do the transformation.
332 if ((LoadAlignment != 0) != (LI->getAlignment() != 0))
335 LoadAlignment = std::min(LoadAlignment, LI->getAlignment());
337 // If the PHI is of volatile loads and the load block has multiple
338 // successors, sinking it would remove a load of the volatile value from
339 // the path through the other successor.
341 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
345 // Okay, they are all the same operation. Create a new PHI node of the
346 // correct type, and PHI together all of the LHS's of the instructions.
347 PHINode *NewPN = PHINode::Create(FirstLI->getOperand(0)->getType(),
348 PN.getNumIncomingValues(),
351 Value *InVal = FirstLI->getOperand(0);
352 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
353 LoadInst *NewLI = new LoadInst(NewPN, "", isVolatile, LoadAlignment);
355 unsigned KnownIDs[] = {
356 LLVMContext::MD_tbaa,
357 LLVMContext::MD_range,
358 LLVMContext::MD_invariant_load,
359 LLVMContext::MD_alias_scope,
360 LLVMContext::MD_noalias,
361 LLVMContext::MD_nonnull,
362 LLVMContext::MD_align,
363 LLVMContext::MD_dereferenceable,
364 LLVMContext::MD_dereferenceable_or_null,
367 for (unsigned ID : KnownIDs)
368 NewLI->setMetadata(ID, FirstLI->getMetadata(ID));
370 // Add all operands to the new PHI and combine TBAA metadata.
371 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
372 LoadInst *LI = cast<LoadInst>(PN.getIncomingValue(i));
373 combineMetadata(NewLI, LI, KnownIDs);
374 Value *NewInVal = LI->getOperand(0);
375 if (NewInVal != InVal)
377 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
381 // The new PHI unions all of the same values together. This is really
382 // common, so we handle it intelligently here for compile-time speed.
383 NewLI->setOperand(0, InVal);
386 InsertNewInstBefore(NewPN, PN);
389 // If this was a volatile load that we are merging, make sure to loop through
390 // and mark all the input loads as non-volatile. If we don't do this, we will
391 // insert a new volatile load and the old ones will not be deletable.
393 for (Value *IncValue : PN.incoming_values())
394 cast<LoadInst>(IncValue)->setVolatile(false);
396 NewLI->setDebugLoc(FirstLI->getDebugLoc());
400 /// TODO: This function could handle other cast types, but then it might
401 /// require special-casing a cast from the 'i1' type. See the comment in
402 /// FoldPHIArgOpIntoPHI() about pessimizing illegal integer types.
403 Instruction *InstCombiner::FoldPHIArgZextsIntoPHI(PHINode &Phi) {
404 // Early exit for the common case of a phi with two operands. These are
405 // handled elsewhere. See the comment below where we check the count of zexts
406 // and constants for more details.
407 unsigned NumIncomingValues = Phi.getNumIncomingValues();
408 if (NumIncomingValues < 3)
411 // Find the narrower type specified by the first zext.
412 Type *NarrowType = nullptr;
413 for (Value *V : Phi.incoming_values()) {
414 if (auto *Zext = dyn_cast<ZExtInst>(V)) {
415 NarrowType = Zext->getSrcTy();
422 // Walk the phi operands checking that we only have zexts or constants that
423 // we can shrink for free. Store the new operands for the new phi.
424 SmallVector<Value *, 4> NewIncoming;
425 unsigned NumZexts = 0;
426 unsigned NumConsts = 0;
427 for (Value *V : Phi.incoming_values()) {
428 if (auto *Zext = dyn_cast<ZExtInst>(V)) {
429 // All zexts must be identical and have one use.
430 if (Zext->getSrcTy() != NarrowType || !Zext->hasOneUse())
432 NewIncoming.push_back(Zext->getOperand(0));
434 } else if (auto *C = dyn_cast<Constant>(V)) {
435 // Make sure that constants can fit in the new type.
436 Constant *Trunc = ConstantExpr::getTrunc(C, NarrowType);
437 if (ConstantExpr::getZExt(Trunc, C->getType()) != C)
439 NewIncoming.push_back(Trunc);
442 // If it's not a cast or a constant, bail out.
447 // The more common cases of a phi with no constant operands or just one
448 // variable operand are handled by FoldPHIArgOpIntoPHI() and FoldOpIntoPhi()
449 // respectively. FoldOpIntoPhi() wants to do the opposite transform that is
450 // performed here. It tries to replicate a cast in the phi operand's basic
451 // block to expose other folding opportunities. Thus, InstCombine will
452 // infinite loop without this check.
453 if (NumConsts == 0 || NumZexts < 2)
456 // All incoming values are zexts or constants that are safe to truncate.
457 // Create a new phi node of the narrow type, phi together all of the new
458 // operands, and zext the result back to the original type.
459 PHINode *NewPhi = PHINode::Create(NarrowType, NumIncomingValues,
460 Phi.getName() + ".shrunk");
461 for (unsigned i = 0; i != NumIncomingValues; ++i)
462 NewPhi->addIncoming(NewIncoming[i], Phi.getIncomingBlock(i));
464 InsertNewInstBefore(NewPhi, Phi);
465 return CastInst::CreateZExtOrBitCast(NewPhi, Phi.getType());
468 /// If all operands to a PHI node are the same "unary" operator and they all are
469 /// only used by the PHI, PHI together their inputs, and do the operation once,
470 /// to the result of the PHI.
471 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
472 // We cannot create a new instruction after the PHI if the terminator is an
473 // EHPad because there is no valid insertion point.
474 if (TerminatorInst *TI = PN.getParent()->getTerminator())
478 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
480 if (isa<GetElementPtrInst>(FirstInst))
481 return FoldPHIArgGEPIntoPHI(PN);
482 if (isa<LoadInst>(FirstInst))
483 return FoldPHIArgLoadIntoPHI(PN);
485 // Scan the instruction, looking for input operations that can be folded away.
486 // If all input operands to the phi are the same instruction (e.g. a cast from
487 // the same type or "+42") we can pull the operation through the PHI, reducing
488 // code size and simplifying code.
489 Constant *ConstantOp = nullptr;
490 Type *CastSrcTy = nullptr;
491 bool isNUW = false, isNSW = false, isExact = false;
493 if (isa<CastInst>(FirstInst)) {
494 CastSrcTy = FirstInst->getOperand(0)->getType();
496 // Be careful about transforming integer PHIs. We don't want to pessimize
497 // the code by turning an i32 into an i1293.
498 if (PN.getType()->isIntegerTy() && CastSrcTy->isIntegerTy()) {
499 if (!ShouldChangeType(PN.getType(), CastSrcTy))
502 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
503 // Can fold binop, compare or shift here if the RHS is a constant,
504 // otherwise call FoldPHIArgBinOpIntoPHI.
505 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
507 return FoldPHIArgBinOpIntoPHI(PN);
509 if (OverflowingBinaryOperator *BO =
510 dyn_cast<OverflowingBinaryOperator>(FirstInst)) {
511 isNUW = BO->hasNoUnsignedWrap();
512 isNSW = BO->hasNoSignedWrap();
513 } else if (PossiblyExactOperator *PEO =
514 dyn_cast<PossiblyExactOperator>(FirstInst))
515 isExact = PEO->isExact();
517 return nullptr; // Cannot fold this operation.
520 // Check to see if all arguments are the same operation.
521 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
522 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
523 if (!I || !I->hasOneUse() || !I->isSameOperationAs(FirstInst))
526 if (I->getOperand(0)->getType() != CastSrcTy)
527 return nullptr; // Cast operation must match.
528 } else if (I->getOperand(1) != ConstantOp) {
533 isNUW = cast<OverflowingBinaryOperator>(I)->hasNoUnsignedWrap();
535 isNSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
537 isExact = cast<PossiblyExactOperator>(I)->isExact();
540 // Okay, they are all the same operation. Create a new PHI node of the
541 // correct type, and PHI together all of the LHS's of the instructions.
542 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
543 PN.getNumIncomingValues(),
546 Value *InVal = FirstInst->getOperand(0);
547 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
549 // Add all operands to the new PHI.
550 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
551 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
552 if (NewInVal != InVal)
554 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
559 // The new PHI unions all of the same values together. This is really
560 // common, so we handle it intelligently here for compile-time speed.
564 InsertNewInstBefore(NewPN, PN);
568 // Insert and return the new operation.
569 if (CastInst *FirstCI = dyn_cast<CastInst>(FirstInst)) {
570 CastInst *NewCI = CastInst::Create(FirstCI->getOpcode(), PhiVal,
572 NewCI->setDebugLoc(FirstInst->getDebugLoc());
576 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst)) {
577 BinOp = BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
578 if (isNUW) BinOp->setHasNoUnsignedWrap();
579 if (isNSW) BinOp->setHasNoSignedWrap();
580 if (isExact) BinOp->setIsExact();
581 BinOp->setDebugLoc(FirstInst->getDebugLoc());
585 CmpInst *CIOp = cast<CmpInst>(FirstInst);
586 CmpInst *NewCI = CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
588 NewCI->setDebugLoc(FirstInst->getDebugLoc());
592 /// Return true if this PHI node is only used by a PHI node cycle that is dead.
593 static bool DeadPHICycle(PHINode *PN,
594 SmallPtrSetImpl<PHINode*> &PotentiallyDeadPHIs) {
595 if (PN->use_empty()) return true;
596 if (!PN->hasOneUse()) return false;
598 // Remember this node, and if we find the cycle, return.
599 if (!PotentiallyDeadPHIs.insert(PN).second)
602 // Don't scan crazily complex things.
603 if (PotentiallyDeadPHIs.size() == 16)
606 if (PHINode *PU = dyn_cast<PHINode>(PN->user_back()))
607 return DeadPHICycle(PU, PotentiallyDeadPHIs);
612 /// Return true if this phi node is always equal to NonPhiInVal.
613 /// This happens with mutually cyclic phi nodes like:
614 /// z = some value; x = phi (y, z); y = phi (x, z)
615 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
616 SmallPtrSetImpl<PHINode*> &ValueEqualPHIs) {
617 // See if we already saw this PHI node.
618 if (!ValueEqualPHIs.insert(PN).second)
621 // Don't scan crazily complex things.
622 if (ValueEqualPHIs.size() == 16)
625 // Scan the operands to see if they are either phi nodes or are equal to
627 for (Value *Op : PN->incoming_values()) {
628 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
629 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
631 } else if (Op != NonPhiInVal)
640 struct PHIUsageRecord {
641 unsigned PHIId; // The ID # of the PHI (something determinstic to sort on)
642 unsigned Shift; // The amount shifted.
643 Instruction *Inst; // The trunc instruction.
645 PHIUsageRecord(unsigned pn, unsigned Sh, Instruction *User)
646 : PHIId(pn), Shift(Sh), Inst(User) {}
648 bool operator<(const PHIUsageRecord &RHS) const {
649 if (PHIId < RHS.PHIId) return true;
650 if (PHIId > RHS.PHIId) return false;
651 if (Shift < RHS.Shift) return true;
652 if (Shift > RHS.Shift) return false;
653 return Inst->getType()->getPrimitiveSizeInBits() <
654 RHS.Inst->getType()->getPrimitiveSizeInBits();
658 struct LoweredPHIRecord {
659 PHINode *PN; // The PHI that was lowered.
660 unsigned Shift; // The amount shifted.
661 unsigned Width; // The width extracted.
663 LoweredPHIRecord(PHINode *pn, unsigned Sh, Type *Ty)
664 : PN(pn), Shift(Sh), Width(Ty->getPrimitiveSizeInBits()) {}
666 // Ctor form used by DenseMap.
667 LoweredPHIRecord(PHINode *pn, unsigned Sh)
668 : PN(pn), Shift(Sh), Width(0) {}
674 struct DenseMapInfo<LoweredPHIRecord> {
675 static inline LoweredPHIRecord getEmptyKey() {
676 return LoweredPHIRecord(nullptr, 0);
678 static inline LoweredPHIRecord getTombstoneKey() {
679 return LoweredPHIRecord(nullptr, 1);
681 static unsigned getHashValue(const LoweredPHIRecord &Val) {
682 return DenseMapInfo<PHINode*>::getHashValue(Val.PN) ^ (Val.Shift>>3) ^
685 static bool isEqual(const LoweredPHIRecord &LHS,
686 const LoweredPHIRecord &RHS) {
687 return LHS.PN == RHS.PN && LHS.Shift == RHS.Shift &&
688 LHS.Width == RHS.Width;
694 /// This is an integer PHI and we know that it has an illegal type: see if it is
695 /// only used by trunc or trunc(lshr) operations. If so, we split the PHI into
696 /// the various pieces being extracted. This sort of thing is introduced when
697 /// SROA promotes an aggregate to large integer values.
699 /// TODO: The user of the trunc may be an bitcast to float/double/vector or an
700 /// inttoptr. We should produce new PHIs in the right type.
702 Instruction *InstCombiner::SliceUpIllegalIntegerPHI(PHINode &FirstPhi) {
703 // PHIUsers - Keep track of all of the truncated values extracted from a set
704 // of PHIs, along with their offset. These are the things we want to rewrite.
705 SmallVector<PHIUsageRecord, 16> PHIUsers;
707 // PHIs are often mutually cyclic, so we keep track of a whole set of PHI
708 // nodes which are extracted from. PHIsToSlice is a set we use to avoid
709 // revisiting PHIs, PHIsInspected is a ordered list of PHIs that we need to
710 // check the uses of (to ensure they are all extracts).
711 SmallVector<PHINode*, 8> PHIsToSlice;
712 SmallPtrSet<PHINode*, 8> PHIsInspected;
714 PHIsToSlice.push_back(&FirstPhi);
715 PHIsInspected.insert(&FirstPhi);
717 for (unsigned PHIId = 0; PHIId != PHIsToSlice.size(); ++PHIId) {
718 PHINode *PN = PHIsToSlice[PHIId];
720 // Scan the input list of the PHI. If any input is an invoke, and if the
721 // input is defined in the predecessor, then we won't be split the critical
722 // edge which is required to insert a truncate. Because of this, we have to
724 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
725 InvokeInst *II = dyn_cast<InvokeInst>(PN->getIncomingValue(i));
727 if (II->getParent() != PN->getIncomingBlock(i))
730 // If we have a phi, and if it's directly in the predecessor, then we have
731 // a critical edge where we need to put the truncate. Since we can't
732 // split the edge in instcombine, we have to bail out.
736 for (User *U : PN->users()) {
737 Instruction *UserI = cast<Instruction>(U);
739 // If the user is a PHI, inspect its uses recursively.
740 if (PHINode *UserPN = dyn_cast<PHINode>(UserI)) {
741 if (PHIsInspected.insert(UserPN).second)
742 PHIsToSlice.push_back(UserPN);
746 // Truncates are always ok.
747 if (isa<TruncInst>(UserI)) {
748 PHIUsers.push_back(PHIUsageRecord(PHIId, 0, UserI));
752 // Otherwise it must be a lshr which can only be used by one trunc.
753 if (UserI->getOpcode() != Instruction::LShr ||
754 !UserI->hasOneUse() || !isa<TruncInst>(UserI->user_back()) ||
755 !isa<ConstantInt>(UserI->getOperand(1)))
758 unsigned Shift = cast<ConstantInt>(UserI->getOperand(1))->getZExtValue();
759 PHIUsers.push_back(PHIUsageRecord(PHIId, Shift, UserI->user_back()));
763 // If we have no users, they must be all self uses, just nuke the PHI.
764 if (PHIUsers.empty())
765 return ReplaceInstUsesWith(FirstPhi, UndefValue::get(FirstPhi.getType()));
767 // If this phi node is transformable, create new PHIs for all the pieces
768 // extracted out of it. First, sort the users by their offset and size.
769 array_pod_sort(PHIUsers.begin(), PHIUsers.end());
771 DEBUG(dbgs() << "SLICING UP PHI: " << FirstPhi << '\n';
772 for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i)
773 dbgs() << "AND USER PHI #" << i << ": " << *PHIsToSlice[i] << '\n';
776 // PredValues - This is a temporary used when rewriting PHI nodes. It is
777 // hoisted out here to avoid construction/destruction thrashing.
778 DenseMap<BasicBlock*, Value*> PredValues;
780 // ExtractedVals - Each new PHI we introduce is saved here so we don't
781 // introduce redundant PHIs.
782 DenseMap<LoweredPHIRecord, PHINode*> ExtractedVals;
784 for (unsigned UserI = 0, UserE = PHIUsers.size(); UserI != UserE; ++UserI) {
785 unsigned PHIId = PHIUsers[UserI].PHIId;
786 PHINode *PN = PHIsToSlice[PHIId];
787 unsigned Offset = PHIUsers[UserI].Shift;
788 Type *Ty = PHIUsers[UserI].Inst->getType();
792 // If we've already lowered a user like this, reuse the previously lowered
794 if ((EltPHI = ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)]) == nullptr) {
796 // Otherwise, Create the new PHI node for this user.
797 EltPHI = PHINode::Create(Ty, PN->getNumIncomingValues(),
798 PN->getName()+".off"+Twine(Offset), PN);
799 assert(EltPHI->getType() != PN->getType() &&
800 "Truncate didn't shrink phi?");
802 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
803 BasicBlock *Pred = PN->getIncomingBlock(i);
804 Value *&PredVal = PredValues[Pred];
806 // If we already have a value for this predecessor, reuse it.
808 EltPHI->addIncoming(PredVal, Pred);
812 // Handle the PHI self-reuse case.
813 Value *InVal = PN->getIncomingValue(i);
816 EltPHI->addIncoming(PredVal, Pred);
820 if (PHINode *InPHI = dyn_cast<PHINode>(PN)) {
821 // If the incoming value was a PHI, and if it was one of the PHIs we
822 // already rewrote it, just use the lowered value.
823 if (Value *Res = ExtractedVals[LoweredPHIRecord(InPHI, Offset, Ty)]) {
825 EltPHI->addIncoming(PredVal, Pred);
830 // Otherwise, do an extract in the predecessor.
831 Builder->SetInsertPoint(Pred->getTerminator());
834 Res = Builder->CreateLShr(Res, ConstantInt::get(InVal->getType(),
836 Res = Builder->CreateTrunc(Res, Ty, "extract.t");
838 EltPHI->addIncoming(Res, Pred);
840 // If the incoming value was a PHI, and if it was one of the PHIs we are
841 // rewriting, we will ultimately delete the code we inserted. This
842 // means we need to revisit that PHI to make sure we extract out the
844 if (PHINode *OldInVal = dyn_cast<PHINode>(PN->getIncomingValue(i)))
845 if (PHIsInspected.count(OldInVal)) {
846 unsigned RefPHIId = std::find(PHIsToSlice.begin(),PHIsToSlice.end(),
847 OldInVal)-PHIsToSlice.begin();
848 PHIUsers.push_back(PHIUsageRecord(RefPHIId, Offset,
849 cast<Instruction>(Res)));
855 DEBUG(dbgs() << " Made element PHI for offset " << Offset << ": "
857 ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)] = EltPHI;
860 // Replace the use of this piece with the PHI node.
861 ReplaceInstUsesWith(*PHIUsers[UserI].Inst, EltPHI);
864 // Replace all the remaining uses of the PHI nodes (self uses and the lshrs)
866 Value *Undef = UndefValue::get(FirstPhi.getType());
867 for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i)
868 ReplaceInstUsesWith(*PHIsToSlice[i], Undef);
869 return ReplaceInstUsesWith(FirstPhi, Undef);
872 // PHINode simplification
874 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
875 if (Value *V = SimplifyInstruction(&PN, DL, TLI, DT, AC))
876 return ReplaceInstUsesWith(PN, V);
878 if (Instruction *Result = FoldPHIArgZextsIntoPHI(PN))
881 // If all PHI operands are the same operation, pull them through the PHI,
882 // reducing code size.
883 if (isa<Instruction>(PN.getIncomingValue(0)) &&
884 isa<Instruction>(PN.getIncomingValue(1)) &&
885 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
886 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
887 // FIXME: The hasOneUse check will fail for PHIs that use the value more
888 // than themselves more than once.
889 PN.getIncomingValue(0)->hasOneUse())
890 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
893 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
894 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
895 // PHI)... break the cycle.
896 if (PN.hasOneUse()) {
897 Instruction *PHIUser = cast<Instruction>(PN.user_back());
898 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
899 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
900 PotentiallyDeadPHIs.insert(&PN);
901 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
902 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
905 // If this phi has a single use, and if that use just computes a value for
906 // the next iteration of a loop, delete the phi. This occurs with unused
907 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
908 // common case here is good because the only other things that catch this
909 // are induction variable analysis (sometimes) and ADCE, which is only run
911 if (PHIUser->hasOneUse() &&
912 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
913 PHIUser->user_back() == &PN) {
914 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
918 // We sometimes end up with phi cycles that non-obviously end up being the
919 // same value, for example:
920 // z = some value; x = phi (y, z); y = phi (x, z)
921 // where the phi nodes don't necessarily need to be in the same block. Do a
922 // quick check to see if the PHI node only contains a single non-phi value, if
923 // so, scan to see if the phi cycle is actually equal to that value.
925 unsigned InValNo = 0, NumIncomingVals = PN.getNumIncomingValues();
926 // Scan for the first non-phi operand.
927 while (InValNo != NumIncomingVals &&
928 isa<PHINode>(PN.getIncomingValue(InValNo)))
931 if (InValNo != NumIncomingVals) {
932 Value *NonPhiInVal = PN.getIncomingValue(InValNo);
934 // Scan the rest of the operands to see if there are any conflicts, if so
935 // there is no need to recursively scan other phis.
936 for (++InValNo; InValNo != NumIncomingVals; ++InValNo) {
937 Value *OpVal = PN.getIncomingValue(InValNo);
938 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
942 // If we scanned over all operands, then we have one unique value plus
943 // phi values. Scan PHI nodes to see if they all merge in each other or
945 if (InValNo == NumIncomingVals) {
946 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
947 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
948 return ReplaceInstUsesWith(PN, NonPhiInVal);
953 // If there are multiple PHIs, sort their operands so that they all list
954 // the blocks in the same order. This will help identical PHIs be eliminated
955 // by other passes. Other passes shouldn't depend on this for correctness
957 PHINode *FirstPN = cast<PHINode>(PN.getParent()->begin());
959 for (unsigned i = 0, e = FirstPN->getNumIncomingValues(); i != e; ++i) {
960 BasicBlock *BBA = PN.getIncomingBlock(i);
961 BasicBlock *BBB = FirstPN->getIncomingBlock(i);
963 Value *VA = PN.getIncomingValue(i);
964 unsigned j = PN.getBasicBlockIndex(BBB);
965 Value *VB = PN.getIncomingValue(j);
966 PN.setIncomingBlock(i, BBB);
967 PN.setIncomingValue(i, VB);
968 PN.setIncomingBlock(j, BBA);
969 PN.setIncomingValue(j, VA);
970 // NOTE: Instcombine normally would want us to "return &PN" if we
971 // modified any of the operands of an instruction. However, since we
972 // aren't adding or removing uses (just rearranging them) we don't do
973 // this in this case.
977 // If this is an integer PHI and we know that it has an illegal type, see if
978 // it is only used by trunc or trunc(lshr) operations. If so, we split the
979 // PHI into the various pieces being extracted. This sort of thing is
980 // introduced when SROA promotes an aggregate to a single large integer type.
981 if (PN.getType()->isIntegerTy() &&
982 !DL.isLegalInteger(PN.getType()->getPrimitiveSizeInBits()))
983 if (Instruction *Res = SliceUpIllegalIntegerPHI(PN))