xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/InstCombine/InstCombinePHI.cpp (revision d409305fa3838fb39b38c26fc085fb729b8766d5)
1 //===- InstCombinePHI.cpp -------------------------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the visitPHINode function.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "InstCombineInternal.h"
14 #include "llvm/ADT/STLExtras.h"
15 #include "llvm/ADT/SmallPtrSet.h"
16 #include "llvm/ADT/Statistic.h"
17 #include "llvm/Analysis/InstructionSimplify.h"
18 #include "llvm/Analysis/ValueTracking.h"
19 #include "llvm/IR/PatternMatch.h"
20 #include "llvm/Support/CommandLine.h"
21 #include "llvm/Transforms/InstCombine/InstCombiner.h"
22 #include "llvm/Transforms/Utils/Local.h"
23 
24 using namespace llvm;
25 using namespace llvm::PatternMatch;
26 
27 #define DEBUG_TYPE "instcombine"
28 
29 static cl::opt<unsigned>
30 MaxNumPhis("instcombine-max-num-phis", cl::init(512),
31            cl::desc("Maximum number phis to handle in intptr/ptrint folding"));
32 
33 STATISTIC(NumPHIsOfInsertValues,
34           "Number of phi-of-insertvalue turned into insertvalue-of-phis");
35 STATISTIC(NumPHIsOfExtractValues,
36           "Number of phi-of-extractvalue turned into extractvalue-of-phi");
37 STATISTIC(NumPHICSEs, "Number of PHI's that got CSE'd");
38 
39 /// The PHI arguments will be folded into a single operation with a PHI node
40 /// as input. The debug location of the single operation will be the merged
41 /// locations of the original PHI node arguments.
42 void InstCombinerImpl::PHIArgMergedDebugLoc(Instruction *Inst, PHINode &PN) {
43   auto *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
44   Inst->setDebugLoc(FirstInst->getDebugLoc());
45   // We do not expect a CallInst here, otherwise, N-way merging of DebugLoc
46   // will be inefficient.
47   assert(!isa<CallInst>(Inst));
48 
49   for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
50     auto *I = cast<Instruction>(PN.getIncomingValue(i));
51     Inst->applyMergedLocation(Inst->getDebugLoc(), I->getDebugLoc());
52   }
53 }
54 
55 // Replace Integer typed PHI PN if the PHI's value is used as a pointer value.
56 // If there is an existing pointer typed PHI that produces the same value as PN,
57 // replace PN and the IntToPtr operation with it. Otherwise, synthesize a new
58 // PHI node:
59 //
60 // Case-1:
61 // bb1:
62 //     int_init = PtrToInt(ptr_init)
63 //     br label %bb2
64 // bb2:
65 //    int_val = PHI([int_init, %bb1], [int_val_inc, %bb2]
66 //    ptr_val = PHI([ptr_init, %bb1], [ptr_val_inc, %bb2]
67 //    ptr_val2 = IntToPtr(int_val)
68 //    ...
69 //    use(ptr_val2)
70 //    ptr_val_inc = ...
71 //    inc_val_inc = PtrToInt(ptr_val_inc)
72 //
73 // ==>
74 // bb1:
75 //     br label %bb2
76 // bb2:
77 //    ptr_val = PHI([ptr_init, %bb1], [ptr_val_inc, %bb2]
78 //    ...
79 //    use(ptr_val)
80 //    ptr_val_inc = ...
81 //
82 // Case-2:
83 // bb1:
84 //    int_ptr = BitCast(ptr_ptr)
85 //    int_init = Load(int_ptr)
86 //    br label %bb2
87 // bb2:
88 //    int_val = PHI([int_init, %bb1], [int_val_inc, %bb2]
89 //    ptr_val2 = IntToPtr(int_val)
90 //    ...
91 //    use(ptr_val2)
92 //    ptr_val_inc = ...
93 //    inc_val_inc = PtrToInt(ptr_val_inc)
94 // ==>
95 // bb1:
96 //    ptr_init = Load(ptr_ptr)
97 //    br label %bb2
98 // bb2:
99 //    ptr_val = PHI([ptr_init, %bb1], [ptr_val_inc, %bb2]
100 //    ...
101 //    use(ptr_val)
102 //    ptr_val_inc = ...
103 //    ...
104 //
105 Instruction *InstCombinerImpl::foldIntegerTypedPHI(PHINode &PN) {
106   if (!PN.getType()->isIntegerTy())
107     return nullptr;
108   if (!PN.hasOneUse())
109     return nullptr;
110 
111   auto *IntToPtr = dyn_cast<IntToPtrInst>(PN.user_back());
112   if (!IntToPtr)
113     return nullptr;
114 
115   // Check if the pointer is actually used as pointer:
116   auto HasPointerUse = [](Instruction *IIP) {
117     for (User *U : IIP->users()) {
118       Value *Ptr = nullptr;
119       if (LoadInst *LoadI = dyn_cast<LoadInst>(U)) {
120         Ptr = LoadI->getPointerOperand();
121       } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
122         Ptr = SI->getPointerOperand();
123       } else if (GetElementPtrInst *GI = dyn_cast<GetElementPtrInst>(U)) {
124         Ptr = GI->getPointerOperand();
125       }
126 
127       if (Ptr && Ptr == IIP)
128         return true;
129     }
130     return false;
131   };
132 
133   if (!HasPointerUse(IntToPtr))
134     return nullptr;
135 
136   if (DL.getPointerSizeInBits(IntToPtr->getAddressSpace()) !=
137       DL.getTypeSizeInBits(IntToPtr->getOperand(0)->getType()))
138     return nullptr;
139 
140   SmallVector<Value *, 4> AvailablePtrVals;
141   for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
142     Value *Arg = PN.getIncomingValue(i);
143 
144     // First look backward:
145     if (auto *PI = dyn_cast<PtrToIntInst>(Arg)) {
146       AvailablePtrVals.emplace_back(PI->getOperand(0));
147       continue;
148     }
149 
150     // Next look forward:
151     Value *ArgIntToPtr = nullptr;
152     for (User *U : Arg->users()) {
153       if (isa<IntToPtrInst>(U) && U->getType() == IntToPtr->getType() &&
154           (DT.dominates(cast<Instruction>(U), PN.getIncomingBlock(i)) ||
155            cast<Instruction>(U)->getParent() == PN.getIncomingBlock(i))) {
156         ArgIntToPtr = U;
157         break;
158       }
159     }
160 
161     if (ArgIntToPtr) {
162       AvailablePtrVals.emplace_back(ArgIntToPtr);
163       continue;
164     }
165 
166     // If Arg is defined by a PHI, allow it. This will also create
167     // more opportunities iteratively.
168     if (isa<PHINode>(Arg)) {
169       AvailablePtrVals.emplace_back(Arg);
170       continue;
171     }
172 
173     // For a single use integer load:
174     auto *LoadI = dyn_cast<LoadInst>(Arg);
175     if (!LoadI)
176       return nullptr;
177 
178     if (!LoadI->hasOneUse())
179       return nullptr;
180 
181     // Push the integer typed Load instruction into the available
182     // value set, and fix it up later when the pointer typed PHI
183     // is synthesized.
184     AvailablePtrVals.emplace_back(LoadI);
185   }
186 
187   // Now search for a matching PHI
188   auto *BB = PN.getParent();
189   assert(AvailablePtrVals.size() == PN.getNumIncomingValues() &&
190          "Not enough available ptr typed incoming values");
191   PHINode *MatchingPtrPHI = nullptr;
192   unsigned NumPhis = 0;
193   for (auto II = BB->begin(); II != BB->end(); II++, NumPhis++) {
194     // FIXME: consider handling this in AggressiveInstCombine
195     PHINode *PtrPHI = dyn_cast<PHINode>(II);
196     if (!PtrPHI)
197       break;
198     if (NumPhis > MaxNumPhis)
199       return nullptr;
200     if (PtrPHI == &PN || PtrPHI->getType() != IntToPtr->getType())
201       continue;
202     MatchingPtrPHI = PtrPHI;
203     for (unsigned i = 0; i != PtrPHI->getNumIncomingValues(); ++i) {
204       if (AvailablePtrVals[i] !=
205           PtrPHI->getIncomingValueForBlock(PN.getIncomingBlock(i))) {
206         MatchingPtrPHI = nullptr;
207         break;
208       }
209     }
210 
211     if (MatchingPtrPHI)
212       break;
213   }
214 
215   if (MatchingPtrPHI) {
216     assert(MatchingPtrPHI->getType() == IntToPtr->getType() &&
217            "Phi's Type does not match with IntToPtr");
218     // The PtrToCast + IntToPtr will be simplified later
219     return CastInst::CreateBitOrPointerCast(MatchingPtrPHI,
220                                             IntToPtr->getOperand(0)->getType());
221   }
222 
223   // If it requires a conversion for every PHI operand, do not do it.
224   if (all_of(AvailablePtrVals, [&](Value *V) {
225         return (V->getType() != IntToPtr->getType()) || isa<IntToPtrInst>(V);
226       }))
227     return nullptr;
228 
229   // If any of the operand that requires casting is a terminator
230   // instruction, do not do it. Similarly, do not do the transform if the value
231   // is PHI in a block with no insertion point, for example, a catchswitch
232   // block, since we will not be able to insert a cast after the PHI.
233   if (any_of(AvailablePtrVals, [&](Value *V) {
234         if (V->getType() == IntToPtr->getType())
235           return false;
236         auto *Inst = dyn_cast<Instruction>(V);
237         if (!Inst)
238           return false;
239         if (Inst->isTerminator())
240           return true;
241         auto *BB = Inst->getParent();
242         if (isa<PHINode>(Inst) && BB->getFirstInsertionPt() == BB->end())
243           return true;
244         return false;
245       }))
246     return nullptr;
247 
248   PHINode *NewPtrPHI = PHINode::Create(
249       IntToPtr->getType(), PN.getNumIncomingValues(), PN.getName() + ".ptr");
250 
251   InsertNewInstBefore(NewPtrPHI, PN);
252   SmallDenseMap<Value *, Instruction *> Casts;
253   for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
254     auto *IncomingBB = PN.getIncomingBlock(i);
255     auto *IncomingVal = AvailablePtrVals[i];
256 
257     if (IncomingVal->getType() == IntToPtr->getType()) {
258       NewPtrPHI->addIncoming(IncomingVal, IncomingBB);
259       continue;
260     }
261 
262 #ifndef NDEBUG
263     LoadInst *LoadI = dyn_cast<LoadInst>(IncomingVal);
264     assert((isa<PHINode>(IncomingVal) ||
265             IncomingVal->getType()->isPointerTy() ||
266             (LoadI && LoadI->hasOneUse())) &&
267            "Can not replace LoadInst with multiple uses");
268 #endif
269     // Need to insert a BitCast.
270     // For an integer Load instruction with a single use, the load + IntToPtr
271     // cast will be simplified into a pointer load:
272     // %v = load i64, i64* %a.ip, align 8
273     // %v.cast = inttoptr i64 %v to float **
274     // ==>
275     // %v.ptrp = bitcast i64 * %a.ip to float **
276     // %v.cast = load float *, float ** %v.ptrp, align 8
277     Instruction *&CI = Casts[IncomingVal];
278     if (!CI) {
279       CI = CastInst::CreateBitOrPointerCast(IncomingVal, IntToPtr->getType(),
280                                             IncomingVal->getName() + ".ptr");
281       if (auto *IncomingI = dyn_cast<Instruction>(IncomingVal)) {
282         BasicBlock::iterator InsertPos(IncomingI);
283         InsertPos++;
284         BasicBlock *BB = IncomingI->getParent();
285         if (isa<PHINode>(IncomingI))
286           InsertPos = BB->getFirstInsertionPt();
287         assert(InsertPos != BB->end() && "should have checked above");
288         InsertNewInstBefore(CI, *InsertPos);
289       } else {
290         auto *InsertBB = &IncomingBB->getParent()->getEntryBlock();
291         InsertNewInstBefore(CI, *InsertBB->getFirstInsertionPt());
292       }
293     }
294     NewPtrPHI->addIncoming(CI, IncomingBB);
295   }
296 
297   // The PtrToCast + IntToPtr will be simplified later
298   return CastInst::CreateBitOrPointerCast(NewPtrPHI,
299                                           IntToPtr->getOperand(0)->getType());
300 }
301 
302 /// If we have something like phi [insertvalue(a,b,0), insertvalue(c,d,0)],
303 /// turn this into a phi[a,c] and phi[b,d] and a single insertvalue.
304 Instruction *
305 InstCombinerImpl::foldPHIArgInsertValueInstructionIntoPHI(PHINode &PN) {
306   auto *FirstIVI = cast<InsertValueInst>(PN.getIncomingValue(0));
307 
308   // Scan to see if all operands are `insertvalue`'s with the same indicies,
309   // and all have a single use.
310   for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
311     auto *I = dyn_cast<InsertValueInst>(PN.getIncomingValue(i));
312     if (!I || !I->hasOneUser() || I->getIndices() != FirstIVI->getIndices())
313       return nullptr;
314   }
315 
316   // For each operand of an `insertvalue`
317   std::array<PHINode *, 2> NewOperands;
318   for (int OpIdx : {0, 1}) {
319     auto *&NewOperand = NewOperands[OpIdx];
320     // Create a new PHI node to receive the values the operand has in each
321     // incoming basic block.
322     NewOperand = PHINode::Create(
323         FirstIVI->getOperand(OpIdx)->getType(), PN.getNumIncomingValues(),
324         FirstIVI->getOperand(OpIdx)->getName() + ".pn");
325     // And populate each operand's PHI with said values.
326     for (auto Incoming : zip(PN.blocks(), PN.incoming_values()))
327       NewOperand->addIncoming(
328           cast<InsertValueInst>(std::get<1>(Incoming))->getOperand(OpIdx),
329           std::get<0>(Incoming));
330     InsertNewInstBefore(NewOperand, PN);
331   }
332 
333   // And finally, create `insertvalue` over the newly-formed PHI nodes.
334   auto *NewIVI = InsertValueInst::Create(NewOperands[0], NewOperands[1],
335                                          FirstIVI->getIndices(), PN.getName());
336 
337   PHIArgMergedDebugLoc(NewIVI, PN);
338   ++NumPHIsOfInsertValues;
339   return NewIVI;
340 }
341 
342 /// If we have something like phi [extractvalue(a,0), extractvalue(b,0)],
343 /// turn this into a phi[a,b] and a single extractvalue.
344 Instruction *
345 InstCombinerImpl::foldPHIArgExtractValueInstructionIntoPHI(PHINode &PN) {
346   auto *FirstEVI = cast<ExtractValueInst>(PN.getIncomingValue(0));
347 
348   // Scan to see if all operands are `extractvalue`'s with the same indicies,
349   // and all have a single use.
350   for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
351     auto *I = dyn_cast<ExtractValueInst>(PN.getIncomingValue(i));
352     if (!I || !I->hasOneUser() || I->getIndices() != FirstEVI->getIndices() ||
353         I->getAggregateOperand()->getType() !=
354             FirstEVI->getAggregateOperand()->getType())
355       return nullptr;
356   }
357 
358   // Create a new PHI node to receive the values the aggregate operand has
359   // in each incoming basic block.
360   auto *NewAggregateOperand = PHINode::Create(
361       FirstEVI->getAggregateOperand()->getType(), PN.getNumIncomingValues(),
362       FirstEVI->getAggregateOperand()->getName() + ".pn");
363   // And populate the PHI with said values.
364   for (auto Incoming : zip(PN.blocks(), PN.incoming_values()))
365     NewAggregateOperand->addIncoming(
366         cast<ExtractValueInst>(std::get<1>(Incoming))->getAggregateOperand(),
367         std::get<0>(Incoming));
368   InsertNewInstBefore(NewAggregateOperand, PN);
369 
370   // And finally, create `extractvalue` over the newly-formed PHI nodes.
371   auto *NewEVI = ExtractValueInst::Create(NewAggregateOperand,
372                                           FirstEVI->getIndices(), PN.getName());
373 
374   PHIArgMergedDebugLoc(NewEVI, PN);
375   ++NumPHIsOfExtractValues;
376   return NewEVI;
377 }
378 
379 /// If we have something like phi [add (a,b), add(a,c)] and if a/b/c and the
380 /// adds all have a single user, turn this into a phi and a single binop.
381 Instruction *InstCombinerImpl::foldPHIArgBinOpIntoPHI(PHINode &PN) {
382   Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
383   assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
384   unsigned Opc = FirstInst->getOpcode();
385   Value *LHSVal = FirstInst->getOperand(0);
386   Value *RHSVal = FirstInst->getOperand(1);
387 
388   Type *LHSType = LHSVal->getType();
389   Type *RHSType = RHSVal->getType();
390 
391   // Scan to see if all operands are the same opcode, and all have one user.
392   for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
393     Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
394     if (!I || I->getOpcode() != Opc || !I->hasOneUser() ||
395         // Verify type of the LHS matches so we don't fold cmp's of different
396         // types.
397         I->getOperand(0)->getType() != LHSType ||
398         I->getOperand(1)->getType() != RHSType)
399       return nullptr;
400 
401     // If they are CmpInst instructions, check their predicates
402     if (CmpInst *CI = dyn_cast<CmpInst>(I))
403       if (CI->getPredicate() != cast<CmpInst>(FirstInst)->getPredicate())
404         return nullptr;
405 
406     // Keep track of which operand needs a phi node.
407     if (I->getOperand(0) != LHSVal) LHSVal = nullptr;
408     if (I->getOperand(1) != RHSVal) RHSVal = nullptr;
409   }
410 
411   // If both LHS and RHS would need a PHI, don't do this transformation,
412   // because it would increase the number of PHIs entering the block,
413   // which leads to higher register pressure. This is especially
414   // bad when the PHIs are in the header of a loop.
415   if (!LHSVal && !RHSVal)
416     return nullptr;
417 
418   // Otherwise, this is safe to transform!
419 
420   Value *InLHS = FirstInst->getOperand(0);
421   Value *InRHS = FirstInst->getOperand(1);
422   PHINode *NewLHS = nullptr, *NewRHS = nullptr;
423   if (!LHSVal) {
424     NewLHS = PHINode::Create(LHSType, PN.getNumIncomingValues(),
425                              FirstInst->getOperand(0)->getName() + ".pn");
426     NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
427     InsertNewInstBefore(NewLHS, PN);
428     LHSVal = NewLHS;
429   }
430 
431   if (!RHSVal) {
432     NewRHS = PHINode::Create(RHSType, PN.getNumIncomingValues(),
433                              FirstInst->getOperand(1)->getName() + ".pn");
434     NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
435     InsertNewInstBefore(NewRHS, PN);
436     RHSVal = NewRHS;
437   }
438 
439   // Add all operands to the new PHIs.
440   if (NewLHS || NewRHS) {
441     for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
442       Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
443       if (NewLHS) {
444         Value *NewInLHS = InInst->getOperand(0);
445         NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
446       }
447       if (NewRHS) {
448         Value *NewInRHS = InInst->getOperand(1);
449         NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
450       }
451     }
452   }
453 
454   if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst)) {
455     CmpInst *NewCI = CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
456                                      LHSVal, RHSVal);
457     PHIArgMergedDebugLoc(NewCI, PN);
458     return NewCI;
459   }
460 
461   BinaryOperator *BinOp = cast<BinaryOperator>(FirstInst);
462   BinaryOperator *NewBinOp =
463     BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
464 
465   NewBinOp->copyIRFlags(PN.getIncomingValue(0));
466 
467   for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i)
468     NewBinOp->andIRFlags(PN.getIncomingValue(i));
469 
470   PHIArgMergedDebugLoc(NewBinOp, PN);
471   return NewBinOp;
472 }
473 
474 Instruction *InstCombinerImpl::foldPHIArgGEPIntoPHI(PHINode &PN) {
475   GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
476 
477   SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
478                                         FirstInst->op_end());
479   // This is true if all GEP bases are allocas and if all indices into them are
480   // constants.
481   bool AllBasePointersAreAllocas = true;
482 
483   // We don't want to replace this phi if the replacement would require
484   // more than one phi, which leads to higher register pressure. This is
485   // especially bad when the PHIs are in the header of a loop.
486   bool NeededPhi = false;
487 
488   bool AllInBounds = true;
489 
490   // Scan to see if all operands are the same opcode, and all have one user.
491   for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
492     GetElementPtrInst *GEP =
493         dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
494     if (!GEP || !GEP->hasOneUser() || GEP->getType() != FirstInst->getType() ||
495         GEP->getNumOperands() != FirstInst->getNumOperands())
496       return nullptr;
497 
498     AllInBounds &= GEP->isInBounds();
499 
500     // Keep track of whether or not all GEPs are of alloca pointers.
501     if (AllBasePointersAreAllocas &&
502         (!isa<AllocaInst>(GEP->getOperand(0)) ||
503          !GEP->hasAllConstantIndices()))
504       AllBasePointersAreAllocas = false;
505 
506     // Compare the operand lists.
507     for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
508       if (FirstInst->getOperand(op) == GEP->getOperand(op))
509         continue;
510 
511       // Don't merge two GEPs when two operands differ (introducing phi nodes)
512       // if one of the PHIs has a constant for the index.  The index may be
513       // substantially cheaper to compute for the constants, so making it a
514       // variable index could pessimize the path.  This also handles the case
515       // for struct indices, which must always be constant.
516       if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
517           isa<ConstantInt>(GEP->getOperand(op)))
518         return nullptr;
519 
520       if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
521         return nullptr;
522 
523       // If we already needed a PHI for an earlier operand, and another operand
524       // also requires a PHI, we'd be introducing more PHIs than we're
525       // eliminating, which increases register pressure on entry to the PHI's
526       // block.
527       if (NeededPhi)
528         return nullptr;
529 
530       FixedOperands[op] = nullptr;  // Needs a PHI.
531       NeededPhi = true;
532     }
533   }
534 
535   // If all of the base pointers of the PHI'd GEPs are from allocas, don't
536   // bother doing this transformation.  At best, this will just save a bit of
537   // offset calculation, but all the predecessors will have to materialize the
538   // stack address into a register anyway.  We'd actually rather *clone* the
539   // load up into the predecessors so that we have a load of a gep of an alloca,
540   // which can usually all be folded into the load.
541   if (AllBasePointersAreAllocas)
542     return nullptr;
543 
544   // Otherwise, this is safe to transform.  Insert PHI nodes for each operand
545   // that is variable.
546   SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
547 
548   bool HasAnyPHIs = false;
549   for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
550     if (FixedOperands[i]) continue;  // operand doesn't need a phi.
551     Value *FirstOp = FirstInst->getOperand(i);
552     PHINode *NewPN = PHINode::Create(FirstOp->getType(), e,
553                                      FirstOp->getName()+".pn");
554     InsertNewInstBefore(NewPN, PN);
555 
556     NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
557     OperandPhis[i] = NewPN;
558     FixedOperands[i] = NewPN;
559     HasAnyPHIs = true;
560   }
561 
562 
563   // Add all operands to the new PHIs.
564   if (HasAnyPHIs) {
565     for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
566       GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
567       BasicBlock *InBB = PN.getIncomingBlock(i);
568 
569       for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
570         if (PHINode *OpPhi = OperandPhis[op])
571           OpPhi->addIncoming(InGEP->getOperand(op), InBB);
572     }
573   }
574 
575   Value *Base = FixedOperands[0];
576   GetElementPtrInst *NewGEP =
577       GetElementPtrInst::Create(FirstInst->getSourceElementType(), Base,
578                                 makeArrayRef(FixedOperands).slice(1));
579   if (AllInBounds) NewGEP->setIsInBounds();
580   PHIArgMergedDebugLoc(NewGEP, PN);
581   return NewGEP;
582 }
583 
584 /// Return true if we know that it is safe to sink the load out of the block
585 /// that defines it. This means that it must be obvious the value of the load is
586 /// not changed from the point of the load to the end of the block it is in.
587 ///
588 /// Finally, it is safe, but not profitable, to sink a load targeting a
589 /// non-address-taken alloca.  Doing so will cause us to not promote the alloca
590 /// to a register.
591 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
592   BasicBlock::iterator BBI = L->getIterator(), E = L->getParent()->end();
593 
594   for (++BBI; BBI != E; ++BBI)
595     if (BBI->mayWriteToMemory()) {
596       // Calls that only access inaccessible memory do not block sinking the
597       // load.
598       if (auto *CB = dyn_cast<CallBase>(BBI))
599         if (CB->onlyAccessesInaccessibleMemory())
600           continue;
601       return false;
602     }
603 
604   // Check for non-address taken alloca.  If not address-taken already, it isn't
605   // profitable to do this xform.
606   if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
607     bool isAddressTaken = false;
608     for (User *U : AI->users()) {
609       if (isa<LoadInst>(U)) continue;
610       if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
611         // If storing TO the alloca, then the address isn't taken.
612         if (SI->getOperand(1) == AI) continue;
613       }
614       isAddressTaken = true;
615       break;
616     }
617 
618     if (!isAddressTaken && AI->isStaticAlloca())
619       return false;
620   }
621 
622   // If this load is a load from a GEP with a constant offset from an alloca,
623   // then we don't want to sink it.  In its present form, it will be
624   // load [constant stack offset].  Sinking it will cause us to have to
625   // materialize the stack addresses in each predecessor in a register only to
626   // do a shared load from register in the successor.
627   if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))
628     if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))
629       if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
630         return false;
631 
632   return true;
633 }
634 
635 Instruction *InstCombinerImpl::foldPHIArgLoadIntoPHI(PHINode &PN) {
636   LoadInst *FirstLI = cast<LoadInst>(PN.getIncomingValue(0));
637 
638   // FIXME: This is overconservative; this transform is allowed in some cases
639   // for atomic operations.
640   if (FirstLI->isAtomic())
641     return nullptr;
642 
643   // When processing loads, we need to propagate two bits of information to the
644   // sunk load: whether it is volatile, and what its alignment is.  We currently
645   // don't sink loads when some have their alignment specified and some don't.
646   // visitLoadInst will propagate an alignment onto the load when TD is around,
647   // and if TD isn't around, we can't handle the mixed case.
648   bool isVolatile = FirstLI->isVolatile();
649   Align LoadAlignment = FirstLI->getAlign();
650   unsigned LoadAddrSpace = FirstLI->getPointerAddressSpace();
651 
652   // We can't sink the load if the loaded value could be modified between the
653   // load and the PHI.
654   if (FirstLI->getParent() != PN.getIncomingBlock(0) ||
655       !isSafeAndProfitableToSinkLoad(FirstLI))
656     return nullptr;
657 
658   // If the PHI is of volatile loads and the load block has multiple
659   // successors, sinking it would remove a load of the volatile value from
660   // the path through the other successor.
661   if (isVolatile &&
662       FirstLI->getParent()->getTerminator()->getNumSuccessors() != 1)
663     return nullptr;
664 
665   // Check to see if all arguments are the same operation.
666   for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
667     LoadInst *LI = dyn_cast<LoadInst>(PN.getIncomingValue(i));
668     if (!LI || !LI->hasOneUser())
669       return nullptr;
670 
671     // We can't sink the load if the loaded value could be modified between
672     // the load and the PHI.
673     if (LI->isVolatile() != isVolatile ||
674         LI->getParent() != PN.getIncomingBlock(i) ||
675         LI->getPointerAddressSpace() != LoadAddrSpace ||
676         !isSafeAndProfitableToSinkLoad(LI))
677       return nullptr;
678 
679     LoadAlignment = std::min(LoadAlignment, Align(LI->getAlign()));
680 
681     // If the PHI is of volatile loads and the load block has multiple
682     // successors, sinking it would remove a load of the volatile value from
683     // the path through the other successor.
684     if (isVolatile &&
685         LI->getParent()->getTerminator()->getNumSuccessors() != 1)
686       return nullptr;
687   }
688 
689   // Okay, they are all the same operation.  Create a new PHI node of the
690   // correct type, and PHI together all of the LHS's of the instructions.
691   PHINode *NewPN = PHINode::Create(FirstLI->getOperand(0)->getType(),
692                                    PN.getNumIncomingValues(),
693                                    PN.getName()+".in");
694 
695   Value *InVal = FirstLI->getOperand(0);
696   NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
697   LoadInst *NewLI =
698       new LoadInst(FirstLI->getType(), NewPN, "", isVolatile, LoadAlignment);
699 
700   unsigned KnownIDs[] = {
701     LLVMContext::MD_tbaa,
702     LLVMContext::MD_range,
703     LLVMContext::MD_invariant_load,
704     LLVMContext::MD_alias_scope,
705     LLVMContext::MD_noalias,
706     LLVMContext::MD_nonnull,
707     LLVMContext::MD_align,
708     LLVMContext::MD_dereferenceable,
709     LLVMContext::MD_dereferenceable_or_null,
710     LLVMContext::MD_access_group,
711   };
712 
713   for (unsigned ID : KnownIDs)
714     NewLI->setMetadata(ID, FirstLI->getMetadata(ID));
715 
716   // Add all operands to the new PHI and combine TBAA metadata.
717   for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
718     LoadInst *LI = cast<LoadInst>(PN.getIncomingValue(i));
719     combineMetadata(NewLI, LI, KnownIDs, true);
720     Value *NewInVal = LI->getOperand(0);
721     if (NewInVal != InVal)
722       InVal = nullptr;
723     NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
724   }
725 
726   if (InVal) {
727     // The new PHI unions all of the same values together.  This is really
728     // common, so we handle it intelligently here for compile-time speed.
729     NewLI->setOperand(0, InVal);
730     delete NewPN;
731   } else {
732     InsertNewInstBefore(NewPN, PN);
733   }
734 
735   // If this was a volatile load that we are merging, make sure to loop through
736   // and mark all the input loads as non-volatile.  If we don't do this, we will
737   // insert a new volatile load and the old ones will not be deletable.
738   if (isVolatile)
739     for (Value *IncValue : PN.incoming_values())
740       cast<LoadInst>(IncValue)->setVolatile(false);
741 
742   PHIArgMergedDebugLoc(NewLI, PN);
743   return NewLI;
744 }
745 
746 /// TODO: This function could handle other cast types, but then it might
747 /// require special-casing a cast from the 'i1' type. See the comment in
748 /// FoldPHIArgOpIntoPHI() about pessimizing illegal integer types.
749 Instruction *InstCombinerImpl::foldPHIArgZextsIntoPHI(PHINode &Phi) {
750   // We cannot create a new instruction after the PHI if the terminator is an
751   // EHPad because there is no valid insertion point.
752   if (Instruction *TI = Phi.getParent()->getTerminator())
753     if (TI->isEHPad())
754       return nullptr;
755 
756   // Early exit for the common case of a phi with two operands. These are
757   // handled elsewhere. See the comment below where we check the count of zexts
758   // and constants for more details.
759   unsigned NumIncomingValues = Phi.getNumIncomingValues();
760   if (NumIncomingValues < 3)
761     return nullptr;
762 
763   // Find the narrower type specified by the first zext.
764   Type *NarrowType = nullptr;
765   for (Value *V : Phi.incoming_values()) {
766     if (auto *Zext = dyn_cast<ZExtInst>(V)) {
767       NarrowType = Zext->getSrcTy();
768       break;
769     }
770   }
771   if (!NarrowType)
772     return nullptr;
773 
774   // Walk the phi operands checking that we only have zexts or constants that
775   // we can shrink for free. Store the new operands for the new phi.
776   SmallVector<Value *, 4> NewIncoming;
777   unsigned NumZexts = 0;
778   unsigned NumConsts = 0;
779   for (Value *V : Phi.incoming_values()) {
780     if (auto *Zext = dyn_cast<ZExtInst>(V)) {
781       // All zexts must be identical and have one user.
782       if (Zext->getSrcTy() != NarrowType || !Zext->hasOneUser())
783         return nullptr;
784       NewIncoming.push_back(Zext->getOperand(0));
785       NumZexts++;
786     } else if (auto *C = dyn_cast<Constant>(V)) {
787       // Make sure that constants can fit in the new type.
788       Constant *Trunc = ConstantExpr::getTrunc(C, NarrowType);
789       if (ConstantExpr::getZExt(Trunc, C->getType()) != C)
790         return nullptr;
791       NewIncoming.push_back(Trunc);
792       NumConsts++;
793     } else {
794       // If it's not a cast or a constant, bail out.
795       return nullptr;
796     }
797   }
798 
799   // The more common cases of a phi with no constant operands or just one
800   // variable operand are handled by FoldPHIArgOpIntoPHI() and foldOpIntoPhi()
801   // respectively. foldOpIntoPhi() wants to do the opposite transform that is
802   // performed here. It tries to replicate a cast in the phi operand's basic
803   // block to expose other folding opportunities. Thus, InstCombine will
804   // infinite loop without this check.
805   if (NumConsts == 0 || NumZexts < 2)
806     return nullptr;
807 
808   // All incoming values are zexts or constants that are safe to truncate.
809   // Create a new phi node of the narrow type, phi together all of the new
810   // operands, and zext the result back to the original type.
811   PHINode *NewPhi = PHINode::Create(NarrowType, NumIncomingValues,
812                                     Phi.getName() + ".shrunk");
813   for (unsigned i = 0; i != NumIncomingValues; ++i)
814     NewPhi->addIncoming(NewIncoming[i], Phi.getIncomingBlock(i));
815 
816   InsertNewInstBefore(NewPhi, Phi);
817   return CastInst::CreateZExtOrBitCast(NewPhi, Phi.getType());
818 }
819 
820 /// If all operands to a PHI node are the same "unary" operator and they all are
821 /// only used by the PHI, PHI together their inputs, and do the operation once,
822 /// to the result of the PHI.
823 Instruction *InstCombinerImpl::foldPHIArgOpIntoPHI(PHINode &PN) {
824   // We cannot create a new instruction after the PHI if the terminator is an
825   // EHPad because there is no valid insertion point.
826   if (Instruction *TI = PN.getParent()->getTerminator())
827     if (TI->isEHPad())
828       return nullptr;
829 
830   Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
831 
832   if (isa<GetElementPtrInst>(FirstInst))
833     return foldPHIArgGEPIntoPHI(PN);
834   if (isa<LoadInst>(FirstInst))
835     return foldPHIArgLoadIntoPHI(PN);
836   if (isa<InsertValueInst>(FirstInst))
837     return foldPHIArgInsertValueInstructionIntoPHI(PN);
838   if (isa<ExtractValueInst>(FirstInst))
839     return foldPHIArgExtractValueInstructionIntoPHI(PN);
840 
841   // Scan the instruction, looking for input operations that can be folded away.
842   // If all input operands to the phi are the same instruction (e.g. a cast from
843   // the same type or "+42") we can pull the operation through the PHI, reducing
844   // code size and simplifying code.
845   Constant *ConstantOp = nullptr;
846   Type *CastSrcTy = nullptr;
847 
848   if (isa<CastInst>(FirstInst)) {
849     CastSrcTy = FirstInst->getOperand(0)->getType();
850 
851     // Be careful about transforming integer PHIs.  We don't want to pessimize
852     // the code by turning an i32 into an i1293.
853     if (PN.getType()->isIntegerTy() && CastSrcTy->isIntegerTy()) {
854       if (!shouldChangeType(PN.getType(), CastSrcTy))
855         return nullptr;
856     }
857   } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
858     // Can fold binop, compare or shift here if the RHS is a constant,
859     // otherwise call FoldPHIArgBinOpIntoPHI.
860     ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
861     if (!ConstantOp)
862       return foldPHIArgBinOpIntoPHI(PN);
863   } else {
864     return nullptr;  // Cannot fold this operation.
865   }
866 
867   // Check to see if all arguments are the same operation.
868   for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
869     Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
870     if (!I || !I->hasOneUser() || !I->isSameOperationAs(FirstInst))
871       return nullptr;
872     if (CastSrcTy) {
873       if (I->getOperand(0)->getType() != CastSrcTy)
874         return nullptr;  // Cast operation must match.
875     } else if (I->getOperand(1) != ConstantOp) {
876       return nullptr;
877     }
878   }
879 
880   // Okay, they are all the same operation.  Create a new PHI node of the
881   // correct type, and PHI together all of the LHS's of the instructions.
882   PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
883                                    PN.getNumIncomingValues(),
884                                    PN.getName()+".in");
885 
886   Value *InVal = FirstInst->getOperand(0);
887   NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
888 
889   // Add all operands to the new PHI.
890   for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
891     Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
892     if (NewInVal != InVal)
893       InVal = nullptr;
894     NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
895   }
896 
897   Value *PhiVal;
898   if (InVal) {
899     // The new PHI unions all of the same values together.  This is really
900     // common, so we handle it intelligently here for compile-time speed.
901     PhiVal = InVal;
902     delete NewPN;
903   } else {
904     InsertNewInstBefore(NewPN, PN);
905     PhiVal = NewPN;
906   }
907 
908   // Insert and return the new operation.
909   if (CastInst *FirstCI = dyn_cast<CastInst>(FirstInst)) {
910     CastInst *NewCI = CastInst::Create(FirstCI->getOpcode(), PhiVal,
911                                        PN.getType());
912     PHIArgMergedDebugLoc(NewCI, PN);
913     return NewCI;
914   }
915 
916   if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst)) {
917     BinOp = BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
918     BinOp->copyIRFlags(PN.getIncomingValue(0));
919 
920     for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i)
921       BinOp->andIRFlags(PN.getIncomingValue(i));
922 
923     PHIArgMergedDebugLoc(BinOp, PN);
924     return BinOp;
925   }
926 
927   CmpInst *CIOp = cast<CmpInst>(FirstInst);
928   CmpInst *NewCI = CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
929                                    PhiVal, ConstantOp);
930   PHIArgMergedDebugLoc(NewCI, PN);
931   return NewCI;
932 }
933 
934 /// Return true if this PHI node is only used by a PHI node cycle that is dead.
935 static bool DeadPHICycle(PHINode *PN,
936                          SmallPtrSetImpl<PHINode*> &PotentiallyDeadPHIs) {
937   if (PN->use_empty()) return true;
938   if (!PN->hasOneUse()) return false;
939 
940   // Remember this node, and if we find the cycle, return.
941   if (!PotentiallyDeadPHIs.insert(PN).second)
942     return true;
943 
944   // Don't scan crazily complex things.
945   if (PotentiallyDeadPHIs.size() == 16)
946     return false;
947 
948   if (PHINode *PU = dyn_cast<PHINode>(PN->user_back()))
949     return DeadPHICycle(PU, PotentiallyDeadPHIs);
950 
951   return false;
952 }
953 
954 /// Return true if this phi node is always equal to NonPhiInVal.
955 /// This happens with mutually cyclic phi nodes like:
956 ///   z = some value; x = phi (y, z); y = phi (x, z)
957 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
958                            SmallPtrSetImpl<PHINode*> &ValueEqualPHIs) {
959   // See if we already saw this PHI node.
960   if (!ValueEqualPHIs.insert(PN).second)
961     return true;
962 
963   // Don't scan crazily complex things.
964   if (ValueEqualPHIs.size() == 16)
965     return false;
966 
967   // Scan the operands to see if they are either phi nodes or are equal to
968   // the value.
969   for (Value *Op : PN->incoming_values()) {
970     if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
971       if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
972         return false;
973     } else if (Op != NonPhiInVal)
974       return false;
975   }
976 
977   return true;
978 }
979 
980 /// Return an existing non-zero constant if this phi node has one, otherwise
981 /// return constant 1.
982 static ConstantInt *GetAnyNonZeroConstInt(PHINode &PN) {
983   assert(isa<IntegerType>(PN.getType()) && "Expect only integer type phi");
984   for (Value *V : PN.operands())
985     if (auto *ConstVA = dyn_cast<ConstantInt>(V))
986       if (!ConstVA->isZero())
987         return ConstVA;
988   return ConstantInt::get(cast<IntegerType>(PN.getType()), 1);
989 }
990 
991 namespace {
992 struct PHIUsageRecord {
993   unsigned PHIId;     // The ID # of the PHI (something determinstic to sort on)
994   unsigned Shift;     // The amount shifted.
995   Instruction *Inst;  // The trunc instruction.
996 
997   PHIUsageRecord(unsigned pn, unsigned Sh, Instruction *User)
998     : PHIId(pn), Shift(Sh), Inst(User) {}
999 
1000   bool operator<(const PHIUsageRecord &RHS) const {
1001     if (PHIId < RHS.PHIId) return true;
1002     if (PHIId > RHS.PHIId) return false;
1003     if (Shift < RHS.Shift) return true;
1004     if (Shift > RHS.Shift) return false;
1005     return Inst->getType()->getPrimitiveSizeInBits() <
1006            RHS.Inst->getType()->getPrimitiveSizeInBits();
1007   }
1008 };
1009 
1010 struct LoweredPHIRecord {
1011   PHINode *PN;        // The PHI that was lowered.
1012   unsigned Shift;     // The amount shifted.
1013   unsigned Width;     // The width extracted.
1014 
1015   LoweredPHIRecord(PHINode *pn, unsigned Sh, Type *Ty)
1016     : PN(pn), Shift(Sh), Width(Ty->getPrimitiveSizeInBits()) {}
1017 
1018   // Ctor form used by DenseMap.
1019   LoweredPHIRecord(PHINode *pn, unsigned Sh)
1020     : PN(pn), Shift(Sh), Width(0) {}
1021 };
1022 } // namespace
1023 
1024 namespace llvm {
1025   template<>
1026   struct DenseMapInfo<LoweredPHIRecord> {
1027     static inline LoweredPHIRecord getEmptyKey() {
1028       return LoweredPHIRecord(nullptr, 0);
1029     }
1030     static inline LoweredPHIRecord getTombstoneKey() {
1031       return LoweredPHIRecord(nullptr, 1);
1032     }
1033     static unsigned getHashValue(const LoweredPHIRecord &Val) {
1034       return DenseMapInfo<PHINode*>::getHashValue(Val.PN) ^ (Val.Shift>>3) ^
1035              (Val.Width>>3);
1036     }
1037     static bool isEqual(const LoweredPHIRecord &LHS,
1038                         const LoweredPHIRecord &RHS) {
1039       return LHS.PN == RHS.PN && LHS.Shift == RHS.Shift &&
1040              LHS.Width == RHS.Width;
1041     }
1042   };
1043 } // namespace llvm
1044 
1045 
1046 /// This is an integer PHI and we know that it has an illegal type: see if it is
1047 /// only used by trunc or trunc(lshr) operations. If so, we split the PHI into
1048 /// the various pieces being extracted. This sort of thing is introduced when
1049 /// SROA promotes an aggregate to large integer values.
1050 ///
1051 /// TODO: The user of the trunc may be an bitcast to float/double/vector or an
1052 /// inttoptr.  We should produce new PHIs in the right type.
1053 ///
1054 Instruction *InstCombinerImpl::SliceUpIllegalIntegerPHI(PHINode &FirstPhi) {
1055   // PHIUsers - Keep track of all of the truncated values extracted from a set
1056   // of PHIs, along with their offset.  These are the things we want to rewrite.
1057   SmallVector<PHIUsageRecord, 16> PHIUsers;
1058 
1059   // PHIs are often mutually cyclic, so we keep track of a whole set of PHI
1060   // nodes which are extracted from. PHIsToSlice is a set we use to avoid
1061   // revisiting PHIs, PHIsInspected is a ordered list of PHIs that we need to
1062   // check the uses of (to ensure they are all extracts).
1063   SmallVector<PHINode*, 8> PHIsToSlice;
1064   SmallPtrSet<PHINode*, 8> PHIsInspected;
1065 
1066   PHIsToSlice.push_back(&FirstPhi);
1067   PHIsInspected.insert(&FirstPhi);
1068 
1069   for (unsigned PHIId = 0; PHIId != PHIsToSlice.size(); ++PHIId) {
1070     PHINode *PN = PHIsToSlice[PHIId];
1071 
1072     // Scan the input list of the PHI.  If any input is an invoke, and if the
1073     // input is defined in the predecessor, then we won't be split the critical
1074     // edge which is required to insert a truncate.  Because of this, we have to
1075     // bail out.
1076     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1077       InvokeInst *II = dyn_cast<InvokeInst>(PN->getIncomingValue(i));
1078       if (!II) continue;
1079       if (II->getParent() != PN->getIncomingBlock(i))
1080         continue;
1081 
1082       // If we have a phi, and if it's directly in the predecessor, then we have
1083       // a critical edge where we need to put the truncate.  Since we can't
1084       // split the edge in instcombine, we have to bail out.
1085       return nullptr;
1086     }
1087 
1088     for (User *U : PN->users()) {
1089       Instruction *UserI = cast<Instruction>(U);
1090 
1091       // If the user is a PHI, inspect its uses recursively.
1092       if (PHINode *UserPN = dyn_cast<PHINode>(UserI)) {
1093         if (PHIsInspected.insert(UserPN).second)
1094           PHIsToSlice.push_back(UserPN);
1095         continue;
1096       }
1097 
1098       // Truncates are always ok.
1099       if (isa<TruncInst>(UserI)) {
1100         PHIUsers.push_back(PHIUsageRecord(PHIId, 0, UserI));
1101         continue;
1102       }
1103 
1104       // Otherwise it must be a lshr which can only be used by one trunc.
1105       if (UserI->getOpcode() != Instruction::LShr ||
1106           !UserI->hasOneUse() || !isa<TruncInst>(UserI->user_back()) ||
1107           !isa<ConstantInt>(UserI->getOperand(1)))
1108         return nullptr;
1109 
1110       // Bail on out of range shifts.
1111       unsigned SizeInBits = UserI->getType()->getScalarSizeInBits();
1112       if (cast<ConstantInt>(UserI->getOperand(1))->getValue().uge(SizeInBits))
1113         return nullptr;
1114 
1115       unsigned Shift = cast<ConstantInt>(UserI->getOperand(1))->getZExtValue();
1116       PHIUsers.push_back(PHIUsageRecord(PHIId, Shift, UserI->user_back()));
1117     }
1118   }
1119 
1120   // If we have no users, they must be all self uses, just nuke the PHI.
1121   if (PHIUsers.empty())
1122     return replaceInstUsesWith(FirstPhi, UndefValue::get(FirstPhi.getType()));
1123 
1124   // If this phi node is transformable, create new PHIs for all the pieces
1125   // extracted out of it.  First, sort the users by their offset and size.
1126   array_pod_sort(PHIUsers.begin(), PHIUsers.end());
1127 
1128   LLVM_DEBUG(dbgs() << "SLICING UP PHI: " << FirstPhi << '\n';
1129              for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i) dbgs()
1130              << "AND USER PHI #" << i << ": " << *PHIsToSlice[i] << '\n';);
1131 
1132   // PredValues - This is a temporary used when rewriting PHI nodes.  It is
1133   // hoisted out here to avoid construction/destruction thrashing.
1134   DenseMap<BasicBlock*, Value*> PredValues;
1135 
1136   // ExtractedVals - Each new PHI we introduce is saved here so we don't
1137   // introduce redundant PHIs.
1138   DenseMap<LoweredPHIRecord, PHINode*> ExtractedVals;
1139 
1140   for (unsigned UserI = 0, UserE = PHIUsers.size(); UserI != UserE; ++UserI) {
1141     unsigned PHIId = PHIUsers[UserI].PHIId;
1142     PHINode *PN = PHIsToSlice[PHIId];
1143     unsigned Offset = PHIUsers[UserI].Shift;
1144     Type *Ty = PHIUsers[UserI].Inst->getType();
1145 
1146     PHINode *EltPHI;
1147 
1148     // If we've already lowered a user like this, reuse the previously lowered
1149     // value.
1150     if ((EltPHI = ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)]) == nullptr) {
1151 
1152       // Otherwise, Create the new PHI node for this user.
1153       EltPHI = PHINode::Create(Ty, PN->getNumIncomingValues(),
1154                                PN->getName()+".off"+Twine(Offset), PN);
1155       assert(EltPHI->getType() != PN->getType() &&
1156              "Truncate didn't shrink phi?");
1157 
1158       for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1159         BasicBlock *Pred = PN->getIncomingBlock(i);
1160         Value *&PredVal = PredValues[Pred];
1161 
1162         // If we already have a value for this predecessor, reuse it.
1163         if (PredVal) {
1164           EltPHI->addIncoming(PredVal, Pred);
1165           continue;
1166         }
1167 
1168         // Handle the PHI self-reuse case.
1169         Value *InVal = PN->getIncomingValue(i);
1170         if (InVal == PN) {
1171           PredVal = EltPHI;
1172           EltPHI->addIncoming(PredVal, Pred);
1173           continue;
1174         }
1175 
1176         if (PHINode *InPHI = dyn_cast<PHINode>(PN)) {
1177           // If the incoming value was a PHI, and if it was one of the PHIs we
1178           // already rewrote it, just use the lowered value.
1179           if (Value *Res = ExtractedVals[LoweredPHIRecord(InPHI, Offset, Ty)]) {
1180             PredVal = Res;
1181             EltPHI->addIncoming(PredVal, Pred);
1182             continue;
1183           }
1184         }
1185 
1186         // Otherwise, do an extract in the predecessor.
1187         Builder.SetInsertPoint(Pred->getTerminator());
1188         Value *Res = InVal;
1189         if (Offset)
1190           Res = Builder.CreateLShr(Res, ConstantInt::get(InVal->getType(),
1191                                                           Offset), "extract");
1192         Res = Builder.CreateTrunc(Res, Ty, "extract.t");
1193         PredVal = Res;
1194         EltPHI->addIncoming(Res, Pred);
1195 
1196         // If the incoming value was a PHI, and if it was one of the PHIs we are
1197         // rewriting, we will ultimately delete the code we inserted.  This
1198         // means we need to revisit that PHI to make sure we extract out the
1199         // needed piece.
1200         if (PHINode *OldInVal = dyn_cast<PHINode>(PN->getIncomingValue(i)))
1201           if (PHIsInspected.count(OldInVal)) {
1202             unsigned RefPHIId =
1203                 find(PHIsToSlice, OldInVal) - PHIsToSlice.begin();
1204             PHIUsers.push_back(PHIUsageRecord(RefPHIId, Offset,
1205                                               cast<Instruction>(Res)));
1206             ++UserE;
1207           }
1208       }
1209       PredValues.clear();
1210 
1211       LLVM_DEBUG(dbgs() << "  Made element PHI for offset " << Offset << ": "
1212                         << *EltPHI << '\n');
1213       ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)] = EltPHI;
1214     }
1215 
1216     // Replace the use of this piece with the PHI node.
1217     replaceInstUsesWith(*PHIUsers[UserI].Inst, EltPHI);
1218   }
1219 
1220   // Replace all the remaining uses of the PHI nodes (self uses and the lshrs)
1221   // with undefs.
1222   Value *Undef = UndefValue::get(FirstPhi.getType());
1223   for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i)
1224     replaceInstUsesWith(*PHIsToSlice[i], Undef);
1225   return replaceInstUsesWith(FirstPhi, Undef);
1226 }
1227 
1228 static Value *SimplifyUsingControlFlow(InstCombiner &Self, PHINode &PN,
1229                                        const DominatorTree &DT) {
1230   // Simplify the following patterns:
1231   //       if (cond)
1232   //       /       \
1233   //      ...      ...
1234   //       \       /
1235   //    phi [true] [false]
1236   if (!PN.getType()->isIntegerTy(1))
1237     return nullptr;
1238 
1239   if (PN.getNumOperands() != 2)
1240     return nullptr;
1241 
1242   // Make sure all inputs are constants.
1243   if (!all_of(PN.operands(), [](Value *V) { return isa<ConstantInt>(V); }))
1244     return nullptr;
1245 
1246   BasicBlock *BB = PN.getParent();
1247   // Do not bother with unreachable instructions.
1248   if (!DT.isReachableFromEntry(BB))
1249     return nullptr;
1250 
1251   // Same inputs.
1252   if (PN.getOperand(0) == PN.getOperand(1))
1253     return PN.getOperand(0);
1254 
1255   BasicBlock *TruePred = nullptr, *FalsePred = nullptr;
1256   for (auto *Pred : predecessors(BB)) {
1257     auto *Input = cast<ConstantInt>(PN.getIncomingValueForBlock(Pred));
1258     if (Input->isAllOnesValue())
1259       TruePred = Pred;
1260     else
1261       FalsePred = Pred;
1262   }
1263   assert(TruePred && FalsePred && "Must be!");
1264 
1265   // Check which edge of the dominator dominates the true input. If it is the
1266   // false edge, we should invert the condition.
1267   auto *IDom = DT.getNode(BB)->getIDom()->getBlock();
1268   auto *BI = dyn_cast<BranchInst>(IDom->getTerminator());
1269   if (!BI || BI->isUnconditional())
1270     return nullptr;
1271 
1272   // Check that edges outgoing from the idom's terminators dominate respective
1273   // inputs of the Phi.
1274   BasicBlockEdge TrueOutEdge(IDom, BI->getSuccessor(0));
1275   BasicBlockEdge FalseOutEdge(IDom, BI->getSuccessor(1));
1276 
1277   BasicBlockEdge TrueIncEdge(TruePred, BB);
1278   BasicBlockEdge FalseIncEdge(FalsePred, BB);
1279 
1280   auto *Cond = BI->getCondition();
1281   if (DT.dominates(TrueOutEdge, TrueIncEdge) &&
1282       DT.dominates(FalseOutEdge, FalseIncEdge))
1283     // This Phi is actually equivalent to branching condition of IDom.
1284     return Cond;
1285   else if (DT.dominates(TrueOutEdge, FalseIncEdge) &&
1286            DT.dominates(FalseOutEdge, TrueIncEdge)) {
1287     // This Phi is actually opposite to branching condition of IDom. We invert
1288     // the condition that will potentially open up some opportunities for
1289     // sinking.
1290     auto InsertPt = BB->getFirstInsertionPt();
1291     if (InsertPt != BB->end()) {
1292       Self.Builder.SetInsertPoint(&*InsertPt);
1293       return Self.Builder.CreateNot(Cond);
1294     }
1295   }
1296 
1297   return nullptr;
1298 }
1299 
1300 // PHINode simplification
1301 //
1302 Instruction *InstCombinerImpl::visitPHINode(PHINode &PN) {
1303   if (Value *V = SimplifyInstruction(&PN, SQ.getWithInstruction(&PN)))
1304     return replaceInstUsesWith(PN, V);
1305 
1306   if (Instruction *Result = foldPHIArgZextsIntoPHI(PN))
1307     return Result;
1308 
1309   // If all PHI operands are the same operation, pull them through the PHI,
1310   // reducing code size.
1311   if (isa<Instruction>(PN.getIncomingValue(0)) &&
1312       isa<Instruction>(PN.getIncomingValue(1)) &&
1313       cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
1314           cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
1315       PN.getIncomingValue(0)->hasOneUser())
1316     if (Instruction *Result = foldPHIArgOpIntoPHI(PN))
1317       return Result;
1318 
1319   // If this is a trivial cycle in the PHI node graph, remove it.  Basically, if
1320   // this PHI only has a single use (a PHI), and if that PHI only has one use (a
1321   // PHI)... break the cycle.
1322   if (PN.hasOneUse()) {
1323     if (Instruction *Result = foldIntegerTypedPHI(PN))
1324       return Result;
1325 
1326     Instruction *PHIUser = cast<Instruction>(PN.user_back());
1327     if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
1328       SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
1329       PotentiallyDeadPHIs.insert(&PN);
1330       if (DeadPHICycle(PU, PotentiallyDeadPHIs))
1331         return replaceInstUsesWith(PN, UndefValue::get(PN.getType()));
1332     }
1333 
1334     // If this phi has a single use, and if that use just computes a value for
1335     // the next iteration of a loop, delete the phi.  This occurs with unused
1336     // induction variables, e.g. "for (int j = 0; ; ++j);".  Detecting this
1337     // common case here is good because the only other things that catch this
1338     // are induction variable analysis (sometimes) and ADCE, which is only run
1339     // late.
1340     if (PHIUser->hasOneUse() &&
1341         (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
1342         PHIUser->user_back() == &PN) {
1343       return replaceInstUsesWith(PN, UndefValue::get(PN.getType()));
1344     }
1345     // When a PHI is used only to be compared with zero, it is safe to replace
1346     // an incoming value proved as known nonzero with any non-zero constant.
1347     // For example, in the code below, the incoming value %v can be replaced
1348     // with any non-zero constant based on the fact that the PHI is only used to
1349     // be compared with zero and %v is a known non-zero value:
1350     // %v = select %cond, 1, 2
1351     // %p = phi [%v, BB] ...
1352     //      icmp eq, %p, 0
1353     auto *CmpInst = dyn_cast<ICmpInst>(PHIUser);
1354     // FIXME: To be simple, handle only integer type for now.
1355     if (CmpInst && isa<IntegerType>(PN.getType()) && CmpInst->isEquality() &&
1356         match(CmpInst->getOperand(1), m_Zero())) {
1357       ConstantInt *NonZeroConst = nullptr;
1358       bool MadeChange = false;
1359       for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
1360         Instruction *CtxI = PN.getIncomingBlock(i)->getTerminator();
1361         Value *VA = PN.getIncomingValue(i);
1362         if (isKnownNonZero(VA, DL, 0, &AC, CtxI, &DT)) {
1363           if (!NonZeroConst)
1364             NonZeroConst = GetAnyNonZeroConstInt(PN);
1365 
1366           if (NonZeroConst != VA) {
1367             replaceOperand(PN, i, NonZeroConst);
1368             MadeChange = true;
1369           }
1370         }
1371       }
1372       if (MadeChange)
1373         return &PN;
1374     }
1375   }
1376 
1377   // We sometimes end up with phi cycles that non-obviously end up being the
1378   // same value, for example:
1379   //   z = some value; x = phi (y, z); y = phi (x, z)
1380   // where the phi nodes don't necessarily need to be in the same block.  Do a
1381   // quick check to see if the PHI node only contains a single non-phi value, if
1382   // so, scan to see if the phi cycle is actually equal to that value.
1383   {
1384     unsigned InValNo = 0, NumIncomingVals = PN.getNumIncomingValues();
1385     // Scan for the first non-phi operand.
1386     while (InValNo != NumIncomingVals &&
1387            isa<PHINode>(PN.getIncomingValue(InValNo)))
1388       ++InValNo;
1389 
1390     if (InValNo != NumIncomingVals) {
1391       Value *NonPhiInVal = PN.getIncomingValue(InValNo);
1392 
1393       // Scan the rest of the operands to see if there are any conflicts, if so
1394       // there is no need to recursively scan other phis.
1395       for (++InValNo; InValNo != NumIncomingVals; ++InValNo) {
1396         Value *OpVal = PN.getIncomingValue(InValNo);
1397         if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
1398           break;
1399       }
1400 
1401       // If we scanned over all operands, then we have one unique value plus
1402       // phi values.  Scan PHI nodes to see if they all merge in each other or
1403       // the value.
1404       if (InValNo == NumIncomingVals) {
1405         SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
1406         if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
1407           return replaceInstUsesWith(PN, NonPhiInVal);
1408       }
1409     }
1410   }
1411 
1412   // If there are multiple PHIs, sort their operands so that they all list
1413   // the blocks in the same order. This will help identical PHIs be eliminated
1414   // by other passes. Other passes shouldn't depend on this for correctness
1415   // however.
1416   PHINode *FirstPN = cast<PHINode>(PN.getParent()->begin());
1417   if (&PN != FirstPN)
1418     for (unsigned i = 0, e = FirstPN->getNumIncomingValues(); i != e; ++i) {
1419       BasicBlock *BBA = PN.getIncomingBlock(i);
1420       BasicBlock *BBB = FirstPN->getIncomingBlock(i);
1421       if (BBA != BBB) {
1422         Value *VA = PN.getIncomingValue(i);
1423         unsigned j = PN.getBasicBlockIndex(BBB);
1424         Value *VB = PN.getIncomingValue(j);
1425         PN.setIncomingBlock(i, BBB);
1426         PN.setIncomingValue(i, VB);
1427         PN.setIncomingBlock(j, BBA);
1428         PN.setIncomingValue(j, VA);
1429         // NOTE: Instcombine normally would want us to "return &PN" if we
1430         // modified any of the operands of an instruction.  However, since we
1431         // aren't adding or removing uses (just rearranging them) we don't do
1432         // this in this case.
1433       }
1434     }
1435 
1436   // Is there an identical PHI node in this basic block?
1437   for (PHINode &IdenticalPN : PN.getParent()->phis()) {
1438     // Ignore the PHI node itself.
1439     if (&IdenticalPN == &PN)
1440       continue;
1441     // Note that even though we've just canonicalized this PHI, due to the
1442     // worklist visitation order, there are no guarantess that *every* PHI
1443     // has been canonicalized, so we can't just compare operands ranges.
1444     if (!PN.isIdenticalToWhenDefined(&IdenticalPN))
1445       continue;
1446     // Just use that PHI instead then.
1447     ++NumPHICSEs;
1448     return replaceInstUsesWith(PN, &IdenticalPN);
1449   }
1450 
1451   // If this is an integer PHI and we know that it has an illegal type, see if
1452   // it is only used by trunc or trunc(lshr) operations.  If so, we split the
1453   // PHI into the various pieces being extracted.  This sort of thing is
1454   // introduced when SROA promotes an aggregate to a single large integer type.
1455   if (PN.getType()->isIntegerTy() &&
1456       !DL.isLegalInteger(PN.getType()->getPrimitiveSizeInBits()))
1457     if (Instruction *Res = SliceUpIllegalIntegerPHI(PN))
1458       return Res;
1459 
1460   // Ultimately, try to replace this Phi with a dominating condition.
1461   if (auto *V = SimplifyUsingControlFlow(*this, PN, DT))
1462     return replaceInstUsesWith(PN, V);
1463 
1464   return nullptr;
1465 }
1466