1 //===- InstCombinePHI.cpp -------------------------------------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements the visitPHINode function.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "InstCombineInternal.h"
15 #include "llvm/ADT/STLExtras.h"
16 #include "llvm/ADT/SmallPtrSet.h"
17 #include "llvm/Analysis/InstructionSimplify.h"
18 using namespace llvm;
19 
20 #define DEBUG_TYPE "instcombine"
21 
22 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(a,c)]
23 /// and if a/b/c and the add's all have a single use, turn this into a phi
24 /// and a single binop.
FoldPHIArgBinOpIntoPHI(PHINode & PN)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);
31 
32   Type *LHSType = LHSVal->getType();
33   Type *RHSType = RHSVal->getType();
34 
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();
43 
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
49         // types.
50         I->getOperand(0)->getType() != LHSType ||
51         I->getOperand(1)->getType() != RHSType)
52       return nullptr;
53 
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())
57         return nullptr;
58 
59     if (isNUW)
60       isNUW = cast<OverflowingBinaryOperator>(I)->hasNoUnsignedWrap();
61     if (isNSW)
62       isNSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
63     if (isExact)
64       isExact = cast<PossiblyExactOperator>(I)->isExact();
65 
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;
69   }
70 
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)
76     return nullptr;
77 
78   // Otherwise, this is safe to transform!
79 
80   Value *InLHS = FirstInst->getOperand(0);
81   Value *InRHS = FirstInst->getOperand(1);
82   PHINode *NewLHS = nullptr, *NewRHS = nullptr;
83   if (!LHSVal) {
84     NewLHS = PHINode::Create(LHSType, PN.getNumIncomingValues(),
85                              FirstInst->getOperand(0)->getName() + ".pn");
86     NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
87     InsertNewInstBefore(NewLHS, PN);
88     LHSVal = NewLHS;
89   }
90 
91   if (!RHSVal) {
92     NewRHS = PHINode::Create(RHSType, PN.getNumIncomingValues(),
93                              FirstInst->getOperand(1)->getName() + ".pn");
94     NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
95     InsertNewInstBefore(NewRHS, PN);
96     RHSVal = NewRHS;
97   }
98 
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));
103       if (NewLHS) {
104         Value *NewInLHS = InInst->getOperand(0);
105         NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
106       }
107       if (NewRHS) {
108         Value *NewInRHS = InInst->getOperand(1);
109         NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
110       }
111     }
112   }
113 
114   if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst)) {
115     CmpInst *NewCI = CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
116                                      LHSVal, RHSVal);
117     NewCI->setDebugLoc(FirstInst->getDebugLoc());
118     return NewCI;
119   }
120 
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());
128   return NewBinOp;
129 }
130 
FoldPHIArgGEPIntoPHI(PHINode & PN)131 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
132   GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
133 
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
137   // constants.
138   bool AllBasePointersAreAllocas = true;
139 
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;
144 
145   bool AllInBounds = true;
146 
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())
152       return nullptr;
153 
154     AllInBounds &= GEP->isInBounds();
155 
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;
161 
162     // Compare the operand lists.
163     for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
164       if (FirstInst->getOperand(op) == GEP->getOperand(op))
165         continue;
166 
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)))
174         return nullptr;
175 
176       if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
177         return nullptr;
178 
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
182       // block.
183       if (NeededPhi)
184         return nullptr;
185 
186       FixedOperands[op] = nullptr;  // Needs a PHI.
187       NeededPhi = true;
188     }
189   }
190 
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)
198     return nullptr;
199 
200   // Otherwise, this is safe to transform.  Insert PHI nodes for each operand
201   // that is variable.
202   SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
203 
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);
211 
212     NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
213     OperandPhis[i] = NewPN;
214     FixedOperands[i] = NewPN;
215     HasAnyPHIs = true;
216   }
217 
218 
219   // Add all operands to the new PHIs.
220   if (HasAnyPHIs) {
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);
224 
225       for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
226         if (PHINode *OpPhi = OperandPhis[op])
227           OpPhi->addIncoming(InGEP->getOperand(op), InBB);
228     }
229   }
230 
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());
237   return NewGEP;
238 }
239 
240 
241 /// isSafeAndProfitableToSinkLoad - Return true if we know that it is safe to
242 /// sink the load out of the block that defines it.  This means that it must be
243 /// obvious the value of the load is not changed from the point of the load to
244 /// the end of the block it is in.
245 ///
246 /// Finally, it is safe, but not profitable, to sink a load targeting a
247 /// non-address-taken alloca.  Doing so will cause us to not promote the alloca
248 /// to a register.
isSafeAndProfitableToSinkLoad(LoadInst * L)249 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
250   BasicBlock::iterator BBI = L, E = L->getParent()->end();
251 
252   for (++BBI; BBI != E; ++BBI)
253     if (BBI->mayWriteToMemory())
254       return false;
255 
256   // Check for non-address taken alloca.  If not address-taken already, it isn't
257   // profitable to do this xform.
258   if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
259     bool isAddressTaken = false;
260     for (User *U : AI->users()) {
261       if (isa<LoadInst>(U)) continue;
262       if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
263         // If storing TO the alloca, then the address isn't taken.
264         if (SI->getOperand(1) == AI) continue;
265       }
266       isAddressTaken = true;
267       break;
268     }
269 
270     if (!isAddressTaken && AI->isStaticAlloca())
271       return false;
272   }
273 
274   // If this load is a load from a GEP with a constant offset from an alloca,
275   // then we don't want to sink it.  In its present form, it will be
276   // load [constant stack offset].  Sinking it will cause us to have to
277   // materialize the stack addresses in each predecessor in a register only to
278   // do a shared load from register in the successor.
279   if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))
280     if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))
281       if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
282         return false;
283 
284   return true;
285 }
286 
FoldPHIArgLoadIntoPHI(PHINode & PN)287 Instruction *InstCombiner::FoldPHIArgLoadIntoPHI(PHINode &PN) {
288   LoadInst *FirstLI = cast<LoadInst>(PN.getIncomingValue(0));
289 
290   // FIXME: This is overconservative; this transform is allowed in some cases
291   // for atomic operations.
292   if (FirstLI->isAtomic())
293     return nullptr;
294 
295   // When processing loads, we need to propagate two bits of information to the
296   // sunk load: whether it is volatile, and what its alignment is.  We currently
297   // don't sink loads when some have their alignment specified and some don't.
298   // visitLoadInst will propagate an alignment onto the load when TD is around,
299   // and if TD isn't around, we can't handle the mixed case.
300   bool isVolatile = FirstLI->isVolatile();
301   unsigned LoadAlignment = FirstLI->getAlignment();
302   unsigned LoadAddrSpace = FirstLI->getPointerAddressSpace();
303 
304   // We can't sink the load if the loaded value could be modified between the
305   // load and the PHI.
306   if (FirstLI->getParent() != PN.getIncomingBlock(0) ||
307       !isSafeAndProfitableToSinkLoad(FirstLI))
308     return nullptr;
309 
310   // If the PHI is of volatile loads and the load block has multiple
311   // successors, sinking it would remove a load of the volatile value from
312   // the path through the other successor.
313   if (isVolatile &&
314       FirstLI->getParent()->getTerminator()->getNumSuccessors() != 1)
315     return nullptr;
316 
317   // Check to see if all arguments are the same operation.
318   for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
319     LoadInst *LI = dyn_cast<LoadInst>(PN.getIncomingValue(i));
320     if (!LI || !LI->hasOneUse())
321       return nullptr;
322 
323     // We can't sink the load if the loaded value could be modified between
324     // the load and the PHI.
325     if (LI->isVolatile() != isVolatile ||
326         LI->getParent() != PN.getIncomingBlock(i) ||
327         LI->getPointerAddressSpace() != LoadAddrSpace ||
328         !isSafeAndProfitableToSinkLoad(LI))
329       return nullptr;
330 
331     // If some of the loads have an alignment specified but not all of them,
332     // we can't do the transformation.
333     if ((LoadAlignment != 0) != (LI->getAlignment() != 0))
334       return nullptr;
335 
336     LoadAlignment = std::min(LoadAlignment, LI->getAlignment());
337 
338     // If the PHI is of volatile loads and the load block has multiple
339     // successors, sinking it would remove a load of the volatile value from
340     // the path through the other successor.
341     if (isVolatile &&
342         LI->getParent()->getTerminator()->getNumSuccessors() != 1)
343       return nullptr;
344   }
345 
346   // Okay, they are all the same operation.  Create a new PHI node of the
347   // correct type, and PHI together all of the LHS's of the instructions.
348   PHINode *NewPN = PHINode::Create(FirstLI->getOperand(0)->getType(),
349                                    PN.getNumIncomingValues(),
350                                    PN.getName()+".in");
351 
352   Value *InVal = FirstLI->getOperand(0);
353   NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
354 
355   // Add all operands to the new PHI.
356   for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
357     Value *NewInVal = cast<LoadInst>(PN.getIncomingValue(i))->getOperand(0);
358     if (NewInVal != InVal)
359       InVal = nullptr;
360     NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
361   }
362 
363   Value *PhiVal;
364   if (InVal) {
365     // The new PHI unions all of the same values together.  This is really
366     // common, so we handle it intelligently here for compile-time speed.
367     PhiVal = InVal;
368     delete NewPN;
369   } else {
370     InsertNewInstBefore(NewPN, PN);
371     PhiVal = NewPN;
372   }
373 
374   // If this was a volatile load that we are merging, make sure to loop through
375   // and mark all the input loads as non-volatile.  If we don't do this, we will
376   // insert a new volatile load and the old ones will not be deletable.
377   if (isVolatile)
378     for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
379       cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
380 
381   LoadInst *NewLI = new LoadInst(PhiVal, "", isVolatile, LoadAlignment);
382   NewLI->setDebugLoc(FirstLI->getDebugLoc());
383   return NewLI;
384 }
385 
386 
387 
388 /// FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
389 /// operator and they all are only used by the PHI, PHI together their
390 /// inputs, and do the operation once, to the result of the PHI.
FoldPHIArgOpIntoPHI(PHINode & PN)391 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
392   Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
393 
394   if (isa<GetElementPtrInst>(FirstInst))
395     return FoldPHIArgGEPIntoPHI(PN);
396   if (isa<LoadInst>(FirstInst))
397     return FoldPHIArgLoadIntoPHI(PN);
398 
399   // Scan the instruction, looking for input operations that can be folded away.
400   // If all input operands to the phi are the same instruction (e.g. a cast from
401   // the same type or "+42") we can pull the operation through the PHI, reducing
402   // code size and simplifying code.
403   Constant *ConstantOp = nullptr;
404   Type *CastSrcTy = nullptr;
405   bool isNUW = false, isNSW = false, isExact = false;
406 
407   if (isa<CastInst>(FirstInst)) {
408     CastSrcTy = FirstInst->getOperand(0)->getType();
409 
410     // Be careful about transforming integer PHIs.  We don't want to pessimize
411     // the code by turning an i32 into an i1293.
412     if (PN.getType()->isIntegerTy() && CastSrcTy->isIntegerTy()) {
413       if (!ShouldChangeType(PN.getType(), CastSrcTy))
414         return nullptr;
415     }
416   } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
417     // Can fold binop, compare or shift here if the RHS is a constant,
418     // otherwise call FoldPHIArgBinOpIntoPHI.
419     ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
420     if (!ConstantOp)
421       return FoldPHIArgBinOpIntoPHI(PN);
422 
423     if (OverflowingBinaryOperator *BO =
424         dyn_cast<OverflowingBinaryOperator>(FirstInst)) {
425       isNUW = BO->hasNoUnsignedWrap();
426       isNSW = BO->hasNoSignedWrap();
427     } else if (PossiblyExactOperator *PEO =
428                dyn_cast<PossiblyExactOperator>(FirstInst))
429       isExact = PEO->isExact();
430   } else {
431     return nullptr;  // Cannot fold this operation.
432   }
433 
434   // Check to see if all arguments are the same operation.
435   for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
436     Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
437     if (!I || !I->hasOneUse() || !I->isSameOperationAs(FirstInst))
438       return nullptr;
439     if (CastSrcTy) {
440       if (I->getOperand(0)->getType() != CastSrcTy)
441         return nullptr;  // Cast operation must match.
442     } else if (I->getOperand(1) != ConstantOp) {
443       return nullptr;
444     }
445 
446     if (isNUW)
447       isNUW = cast<OverflowingBinaryOperator>(I)->hasNoUnsignedWrap();
448     if (isNSW)
449       isNSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
450     if (isExact)
451       isExact = cast<PossiblyExactOperator>(I)->isExact();
452   }
453 
454   // Okay, they are all the same operation.  Create a new PHI node of the
455   // correct type, and PHI together all of the LHS's of the instructions.
456   PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
457                                    PN.getNumIncomingValues(),
458                                    PN.getName()+".in");
459 
460   Value *InVal = FirstInst->getOperand(0);
461   NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
462 
463   // Add all operands to the new PHI.
464   for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
465     Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
466     if (NewInVal != InVal)
467       InVal = nullptr;
468     NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
469   }
470 
471   Value *PhiVal;
472   if (InVal) {
473     // The new PHI unions all of the same values together.  This is really
474     // common, so we handle it intelligently here for compile-time speed.
475     PhiVal = InVal;
476     delete NewPN;
477   } else {
478     InsertNewInstBefore(NewPN, PN);
479     PhiVal = NewPN;
480   }
481 
482   // Insert and return the new operation.
483   if (CastInst *FirstCI = dyn_cast<CastInst>(FirstInst)) {
484     CastInst *NewCI = CastInst::Create(FirstCI->getOpcode(), PhiVal,
485                                        PN.getType());
486     NewCI->setDebugLoc(FirstInst->getDebugLoc());
487     return NewCI;
488   }
489 
490   if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst)) {
491     BinOp = BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
492     if (isNUW) BinOp->setHasNoUnsignedWrap();
493     if (isNSW) BinOp->setHasNoSignedWrap();
494     if (isExact) BinOp->setIsExact();
495     BinOp->setDebugLoc(FirstInst->getDebugLoc());
496     return BinOp;
497   }
498 
499   CmpInst *CIOp = cast<CmpInst>(FirstInst);
500   CmpInst *NewCI = CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
501                                    PhiVal, ConstantOp);
502   NewCI->setDebugLoc(FirstInst->getDebugLoc());
503   return NewCI;
504 }
505 
506 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
507 /// that is dead.
DeadPHICycle(PHINode * PN,SmallPtrSetImpl<PHINode * > & PotentiallyDeadPHIs)508 static bool DeadPHICycle(PHINode *PN,
509                          SmallPtrSetImpl<PHINode*> &PotentiallyDeadPHIs) {
510   if (PN->use_empty()) return true;
511   if (!PN->hasOneUse()) return false;
512 
513   // Remember this node, and if we find the cycle, return.
514   if (!PotentiallyDeadPHIs.insert(PN).second)
515     return true;
516 
517   // Don't scan crazily complex things.
518   if (PotentiallyDeadPHIs.size() == 16)
519     return false;
520 
521   if (PHINode *PU = dyn_cast<PHINode>(PN->user_back()))
522     return DeadPHICycle(PU, PotentiallyDeadPHIs);
523 
524   return false;
525 }
526 
527 /// PHIsEqualValue - Return true if this phi node is always equal to
528 /// NonPhiInVal.  This happens with mutually cyclic phi nodes like:
529 ///   z = some value; x = phi (y, z); y = phi (x, z)
PHIsEqualValue(PHINode * PN,Value * NonPhiInVal,SmallPtrSetImpl<PHINode * > & ValueEqualPHIs)530 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
531                            SmallPtrSetImpl<PHINode*> &ValueEqualPHIs) {
532   // See if we already saw this PHI node.
533   if (!ValueEqualPHIs.insert(PN).second)
534     return true;
535 
536   // Don't scan crazily complex things.
537   if (ValueEqualPHIs.size() == 16)
538     return false;
539 
540   // Scan the operands to see if they are either phi nodes or are equal to
541   // the value.
542   for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
543     Value *Op = PN->getIncomingValue(i);
544     if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
545       if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
546         return false;
547     } else if (Op != NonPhiInVal)
548       return false;
549   }
550 
551   return true;
552 }
553 
554 
555 namespace {
556 struct PHIUsageRecord {
557   unsigned PHIId;     // The ID # of the PHI (something determinstic to sort on)
558   unsigned Shift;     // The amount shifted.
559   Instruction *Inst;  // The trunc instruction.
560 
PHIUsageRecord__anon825d72360111::PHIUsageRecord561   PHIUsageRecord(unsigned pn, unsigned Sh, Instruction *User)
562     : PHIId(pn), Shift(Sh), Inst(User) {}
563 
operator <__anon825d72360111::PHIUsageRecord564   bool operator<(const PHIUsageRecord &RHS) const {
565     if (PHIId < RHS.PHIId) return true;
566     if (PHIId > RHS.PHIId) return false;
567     if (Shift < RHS.Shift) return true;
568     if (Shift > RHS.Shift) return false;
569     return Inst->getType()->getPrimitiveSizeInBits() <
570            RHS.Inst->getType()->getPrimitiveSizeInBits();
571   }
572 };
573 
574 struct LoweredPHIRecord {
575   PHINode *PN;        // The PHI that was lowered.
576   unsigned Shift;     // The amount shifted.
577   unsigned Width;     // The width extracted.
578 
LoweredPHIRecord__anon825d72360111::LoweredPHIRecord579   LoweredPHIRecord(PHINode *pn, unsigned Sh, Type *Ty)
580     : PN(pn), Shift(Sh), Width(Ty->getPrimitiveSizeInBits()) {}
581 
582   // Ctor form used by DenseMap.
LoweredPHIRecord__anon825d72360111::LoweredPHIRecord583   LoweredPHIRecord(PHINode *pn, unsigned Sh)
584     : PN(pn), Shift(Sh), Width(0) {}
585 };
586 }
587 
588 namespace llvm {
589   template<>
590   struct DenseMapInfo<LoweredPHIRecord> {
getEmptyKeyllvm::DenseMapInfo591     static inline LoweredPHIRecord getEmptyKey() {
592       return LoweredPHIRecord(nullptr, 0);
593     }
getTombstoneKeyllvm::DenseMapInfo594     static inline LoweredPHIRecord getTombstoneKey() {
595       return LoweredPHIRecord(nullptr, 1);
596     }
getHashValuellvm::DenseMapInfo597     static unsigned getHashValue(const LoweredPHIRecord &Val) {
598       return DenseMapInfo<PHINode*>::getHashValue(Val.PN) ^ (Val.Shift>>3) ^
599              (Val.Width>>3);
600     }
isEqualllvm::DenseMapInfo601     static bool isEqual(const LoweredPHIRecord &LHS,
602                         const LoweredPHIRecord &RHS) {
603       return LHS.PN == RHS.PN && LHS.Shift == RHS.Shift &&
604              LHS.Width == RHS.Width;
605     }
606   };
607 }
608 
609 
610 /// SliceUpIllegalIntegerPHI - This is an integer PHI and we know that it has an
611 /// illegal type: see if it is only used by trunc or trunc(lshr) operations.  If
612 /// so, we split the PHI into the various pieces being extracted.  This sort of
613 /// thing is introduced when SROA promotes an aggregate to large integer values.
614 ///
615 /// TODO: The user of the trunc may be an bitcast to float/double/vector or an
616 /// inttoptr.  We should produce new PHIs in the right type.
617 ///
SliceUpIllegalIntegerPHI(PHINode & FirstPhi)618 Instruction *InstCombiner::SliceUpIllegalIntegerPHI(PHINode &FirstPhi) {
619   // PHIUsers - Keep track of all of the truncated values extracted from a set
620   // of PHIs, along with their offset.  These are the things we want to rewrite.
621   SmallVector<PHIUsageRecord, 16> PHIUsers;
622 
623   // PHIs are often mutually cyclic, so we keep track of a whole set of PHI
624   // nodes which are extracted from. PHIsToSlice is a set we use to avoid
625   // revisiting PHIs, PHIsInspected is a ordered list of PHIs that we need to
626   // check the uses of (to ensure they are all extracts).
627   SmallVector<PHINode*, 8> PHIsToSlice;
628   SmallPtrSet<PHINode*, 8> PHIsInspected;
629 
630   PHIsToSlice.push_back(&FirstPhi);
631   PHIsInspected.insert(&FirstPhi);
632 
633   for (unsigned PHIId = 0; PHIId != PHIsToSlice.size(); ++PHIId) {
634     PHINode *PN = PHIsToSlice[PHIId];
635 
636     // Scan the input list of the PHI.  If any input is an invoke, and if the
637     // input is defined in the predecessor, then we won't be split the critical
638     // edge which is required to insert a truncate.  Because of this, we have to
639     // bail out.
640     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
641       InvokeInst *II = dyn_cast<InvokeInst>(PN->getIncomingValue(i));
642       if (!II) continue;
643       if (II->getParent() != PN->getIncomingBlock(i))
644         continue;
645 
646       // If we have a phi, and if it's directly in the predecessor, then we have
647       // a critical edge where we need to put the truncate.  Since we can't
648       // split the edge in instcombine, we have to bail out.
649       return nullptr;
650     }
651 
652     for (User *U : PN->users()) {
653       Instruction *UserI = cast<Instruction>(U);
654 
655       // If the user is a PHI, inspect its uses recursively.
656       if (PHINode *UserPN = dyn_cast<PHINode>(UserI)) {
657         if (PHIsInspected.insert(UserPN).second)
658           PHIsToSlice.push_back(UserPN);
659         continue;
660       }
661 
662       // Truncates are always ok.
663       if (isa<TruncInst>(UserI)) {
664         PHIUsers.push_back(PHIUsageRecord(PHIId, 0, UserI));
665         continue;
666       }
667 
668       // Otherwise it must be a lshr which can only be used by one trunc.
669       if (UserI->getOpcode() != Instruction::LShr ||
670           !UserI->hasOneUse() || !isa<TruncInst>(UserI->user_back()) ||
671           !isa<ConstantInt>(UserI->getOperand(1)))
672         return nullptr;
673 
674       unsigned Shift = cast<ConstantInt>(UserI->getOperand(1))->getZExtValue();
675       PHIUsers.push_back(PHIUsageRecord(PHIId, Shift, UserI->user_back()));
676     }
677   }
678 
679   // If we have no users, they must be all self uses, just nuke the PHI.
680   if (PHIUsers.empty())
681     return ReplaceInstUsesWith(FirstPhi, UndefValue::get(FirstPhi.getType()));
682 
683   // If this phi node is transformable, create new PHIs for all the pieces
684   // extracted out of it.  First, sort the users by their offset and size.
685   array_pod_sort(PHIUsers.begin(), PHIUsers.end());
686 
687   DEBUG(dbgs() << "SLICING UP PHI: " << FirstPhi << '\n';
688         for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i)
689           dbgs() << "AND USER PHI #" << i << ": " << *PHIsToSlice[i] << '\n';
690     );
691 
692   // PredValues - This is a temporary used when rewriting PHI nodes.  It is
693   // hoisted out here to avoid construction/destruction thrashing.
694   DenseMap<BasicBlock*, Value*> PredValues;
695 
696   // ExtractedVals - Each new PHI we introduce is saved here so we don't
697   // introduce redundant PHIs.
698   DenseMap<LoweredPHIRecord, PHINode*> ExtractedVals;
699 
700   for (unsigned UserI = 0, UserE = PHIUsers.size(); UserI != UserE; ++UserI) {
701     unsigned PHIId = PHIUsers[UserI].PHIId;
702     PHINode *PN = PHIsToSlice[PHIId];
703     unsigned Offset = PHIUsers[UserI].Shift;
704     Type *Ty = PHIUsers[UserI].Inst->getType();
705 
706     PHINode *EltPHI;
707 
708     // If we've already lowered a user like this, reuse the previously lowered
709     // value.
710     if ((EltPHI = ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)]) == nullptr) {
711 
712       // Otherwise, Create the new PHI node for this user.
713       EltPHI = PHINode::Create(Ty, PN->getNumIncomingValues(),
714                                PN->getName()+".off"+Twine(Offset), PN);
715       assert(EltPHI->getType() != PN->getType() &&
716              "Truncate didn't shrink phi?");
717 
718       for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
719         BasicBlock *Pred = PN->getIncomingBlock(i);
720         Value *&PredVal = PredValues[Pred];
721 
722         // If we already have a value for this predecessor, reuse it.
723         if (PredVal) {
724           EltPHI->addIncoming(PredVal, Pred);
725           continue;
726         }
727 
728         // Handle the PHI self-reuse case.
729         Value *InVal = PN->getIncomingValue(i);
730         if (InVal == PN) {
731           PredVal = EltPHI;
732           EltPHI->addIncoming(PredVal, Pred);
733           continue;
734         }
735 
736         if (PHINode *InPHI = dyn_cast<PHINode>(PN)) {
737           // If the incoming value was a PHI, and if it was one of the PHIs we
738           // already rewrote it, just use the lowered value.
739           if (Value *Res = ExtractedVals[LoweredPHIRecord(InPHI, Offset, Ty)]) {
740             PredVal = Res;
741             EltPHI->addIncoming(PredVal, Pred);
742             continue;
743           }
744         }
745 
746         // Otherwise, do an extract in the predecessor.
747         Builder->SetInsertPoint(Pred, Pred->getTerminator());
748         Value *Res = InVal;
749         if (Offset)
750           Res = Builder->CreateLShr(Res, ConstantInt::get(InVal->getType(),
751                                                           Offset), "extract");
752         Res = Builder->CreateTrunc(Res, Ty, "extract.t");
753         PredVal = Res;
754         EltPHI->addIncoming(Res, Pred);
755 
756         // If the incoming value was a PHI, and if it was one of the PHIs we are
757         // rewriting, we will ultimately delete the code we inserted.  This
758         // means we need to revisit that PHI to make sure we extract out the
759         // needed piece.
760         if (PHINode *OldInVal = dyn_cast<PHINode>(PN->getIncomingValue(i)))
761           if (PHIsInspected.count(OldInVal)) {
762             unsigned RefPHIId = std::find(PHIsToSlice.begin(),PHIsToSlice.end(),
763                                           OldInVal)-PHIsToSlice.begin();
764             PHIUsers.push_back(PHIUsageRecord(RefPHIId, Offset,
765                                               cast<Instruction>(Res)));
766             ++UserE;
767           }
768       }
769       PredValues.clear();
770 
771       DEBUG(dbgs() << "  Made element PHI for offset " << Offset << ": "
772                    << *EltPHI << '\n');
773       ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)] = EltPHI;
774     }
775 
776     // Replace the use of this piece with the PHI node.
777     ReplaceInstUsesWith(*PHIUsers[UserI].Inst, EltPHI);
778   }
779 
780   // Replace all the remaining uses of the PHI nodes (self uses and the lshrs)
781   // with undefs.
782   Value *Undef = UndefValue::get(FirstPhi.getType());
783   for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i)
784     ReplaceInstUsesWith(*PHIsToSlice[i], Undef);
785   return ReplaceInstUsesWith(FirstPhi, Undef);
786 }
787 
788 // PHINode simplification
789 //
visitPHINode(PHINode & PN)790 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
791   if (Value *V = SimplifyInstruction(&PN, DL, TLI, DT, AC))
792     return ReplaceInstUsesWith(PN, V);
793 
794   // If all PHI operands are the same operation, pull them through the PHI,
795   // reducing code size.
796   if (isa<Instruction>(PN.getIncomingValue(0)) &&
797       isa<Instruction>(PN.getIncomingValue(1)) &&
798       cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
799       cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
800       // FIXME: The hasOneUse check will fail for PHIs that use the value more
801       // than themselves more than once.
802       PN.getIncomingValue(0)->hasOneUse())
803     if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
804       return Result;
805 
806   // If this is a trivial cycle in the PHI node graph, remove it.  Basically, if
807   // this PHI only has a single use (a PHI), and if that PHI only has one use (a
808   // PHI)... break the cycle.
809   if (PN.hasOneUse()) {
810     Instruction *PHIUser = cast<Instruction>(PN.user_back());
811     if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
812       SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
813       PotentiallyDeadPHIs.insert(&PN);
814       if (DeadPHICycle(PU, PotentiallyDeadPHIs))
815         return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
816     }
817 
818     // If this phi has a single use, and if that use just computes a value for
819     // the next iteration of a loop, delete the phi.  This occurs with unused
820     // induction variables, e.g. "for (int j = 0; ; ++j);".  Detecting this
821     // common case here is good because the only other things that catch this
822     // are induction variable analysis (sometimes) and ADCE, which is only run
823     // late.
824     if (PHIUser->hasOneUse() &&
825         (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
826         PHIUser->user_back() == &PN) {
827       return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
828     }
829   }
830 
831   // We sometimes end up with phi cycles that non-obviously end up being the
832   // same value, for example:
833   //   z = some value; x = phi (y, z); y = phi (x, z)
834   // where the phi nodes don't necessarily need to be in the same block.  Do a
835   // quick check to see if the PHI node only contains a single non-phi value, if
836   // so, scan to see if the phi cycle is actually equal to that value.
837   {
838     unsigned InValNo = 0, NumIncomingVals = PN.getNumIncomingValues();
839     // Scan for the first non-phi operand.
840     while (InValNo != NumIncomingVals &&
841            isa<PHINode>(PN.getIncomingValue(InValNo)))
842       ++InValNo;
843 
844     if (InValNo != NumIncomingVals) {
845       Value *NonPhiInVal = PN.getIncomingValue(InValNo);
846 
847       // Scan the rest of the operands to see if there are any conflicts, if so
848       // there is no need to recursively scan other phis.
849       for (++InValNo; InValNo != NumIncomingVals; ++InValNo) {
850         Value *OpVal = PN.getIncomingValue(InValNo);
851         if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
852           break;
853       }
854 
855       // If we scanned over all operands, then we have one unique value plus
856       // phi values.  Scan PHI nodes to see if they all merge in each other or
857       // the value.
858       if (InValNo == NumIncomingVals) {
859         SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
860         if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
861           return ReplaceInstUsesWith(PN, NonPhiInVal);
862       }
863     }
864   }
865 
866   // If there are multiple PHIs, sort their operands so that they all list
867   // the blocks in the same order. This will help identical PHIs be eliminated
868   // by other passes. Other passes shouldn't depend on this for correctness
869   // however.
870   PHINode *FirstPN = cast<PHINode>(PN.getParent()->begin());
871   if (&PN != FirstPN)
872     for (unsigned i = 0, e = FirstPN->getNumIncomingValues(); i != e; ++i) {
873       BasicBlock *BBA = PN.getIncomingBlock(i);
874       BasicBlock *BBB = FirstPN->getIncomingBlock(i);
875       if (BBA != BBB) {
876         Value *VA = PN.getIncomingValue(i);
877         unsigned j = PN.getBasicBlockIndex(BBB);
878         Value *VB = PN.getIncomingValue(j);
879         PN.setIncomingBlock(i, BBB);
880         PN.setIncomingValue(i, VB);
881         PN.setIncomingBlock(j, BBA);
882         PN.setIncomingValue(j, VA);
883         // NOTE: Instcombine normally would want us to "return &PN" if we
884         // modified any of the operands of an instruction.  However, since we
885         // aren't adding or removing uses (just rearranging them) we don't do
886         // this in this case.
887       }
888     }
889 
890   // If this is an integer PHI and we know that it has an illegal type, see if
891   // it is only used by trunc or trunc(lshr) operations.  If so, we split the
892   // PHI into the various pieces being extracted.  This sort of thing is
893   // introduced when SROA promotes an aggregate to a single large integer type.
894   if (PN.getType()->isIntegerTy() &&
895       !DL.isLegalInteger(PN.getType()->getPrimitiveSizeInBits()))
896     if (Instruction *Res = SliceUpIllegalIntegerPHI(PN))
897       return Res;
898 
899   return nullptr;
900 }
901