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 (Value *V : drop_begin(PN.incoming_values())) {
50     auto *I = cast<Instruction>(V);
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 (auto Incoming : zip(PN.blocks(), PN.incoming_values())) {
142     BasicBlock *BB = std::get<0>(Incoming);
143     Value *Arg = std::get<1>(Incoming);
144 
145     // First look backward:
146     if (auto *PI = dyn_cast<PtrToIntInst>(Arg)) {
147       AvailablePtrVals.emplace_back(PI->getOperand(0));
148       continue;
149     }
150 
151     // Next look forward:
152     Value *ArgIntToPtr = nullptr;
153     for (User *U : Arg->users()) {
154       if (isa<IntToPtrInst>(U) && U->getType() == IntToPtr->getType() &&
155           (DT.dominates(cast<Instruction>(U), BB) ||
156            cast<Instruction>(U)->getParent() == BB)) {
157         ArgIntToPtr = U;
158         break;
159       }
160     }
161 
162     if (ArgIntToPtr) {
163       AvailablePtrVals.emplace_back(ArgIntToPtr);
164       continue;
165     }
166 
167     // If Arg is defined by a PHI, allow it. This will also create
168     // more opportunities iteratively.
169     if (isa<PHINode>(Arg)) {
170       AvailablePtrVals.emplace_back(Arg);
171       continue;
172     }
173 
174     // For a single use integer load:
175     auto *LoadI = dyn_cast<LoadInst>(Arg);
176     if (!LoadI)
177       return nullptr;
178 
179     if (!LoadI->hasOneUse())
180       return nullptr;
181 
182     // Push the integer typed Load instruction into the available
183     // value set, and fix it up later when the pointer typed PHI
184     // is synthesized.
185     AvailablePtrVals.emplace_back(LoadI);
186   }
187 
188   // Now search for a matching PHI
189   auto *BB = PN.getParent();
190   assert(AvailablePtrVals.size() == PN.getNumIncomingValues() &&
191          "Not enough available ptr typed incoming values");
192   PHINode *MatchingPtrPHI = nullptr;
193   unsigned NumPhis = 0;
194   for (PHINode &PtrPHI : BB->phis()) {
195     // FIXME: consider handling this in AggressiveInstCombine
196     if (NumPhis++ > MaxNumPhis)
197       return nullptr;
198     if (&PtrPHI == &PN || PtrPHI.getType() != IntToPtr->getType())
199       continue;
200     if (any_of(zip(PN.blocks(), AvailablePtrVals),
201                [&](const auto &BlockAndValue) {
202                  BasicBlock *BB = std::get<0>(BlockAndValue);
203                  Value *V = std::get<1>(BlockAndValue);
204                  return PtrPHI.getIncomingValueForBlock(BB) != V;
205                }))
206       continue;
207     MatchingPtrPHI = &PtrPHI;
208     break;
209   }
210 
211   if (MatchingPtrPHI) {
212     assert(MatchingPtrPHI->getType() == IntToPtr->getType() &&
213            "Phi's Type does not match with IntToPtr");
214     // The PtrToCast + IntToPtr will be simplified later
215     return CastInst::CreateBitOrPointerCast(MatchingPtrPHI,
216                                             IntToPtr->getOperand(0)->getType());
217   }
218 
219   // If it requires a conversion for every PHI operand, do not do it.
220   if (all_of(AvailablePtrVals, [&](Value *V) {
221         return (V->getType() != IntToPtr->getType()) || isa<IntToPtrInst>(V);
222       }))
223     return nullptr;
224 
225   // If any of the operand that requires casting is a terminator
226   // instruction, do not do it. Similarly, do not do the transform if the value
227   // is PHI in a block with no insertion point, for example, a catchswitch
228   // block, since we will not be able to insert a cast after the PHI.
229   if (any_of(AvailablePtrVals, [&](Value *V) {
230         if (V->getType() == IntToPtr->getType())
231           return false;
232         auto *Inst = dyn_cast<Instruction>(V);
233         if (!Inst)
234           return false;
235         if (Inst->isTerminator())
236           return true;
237         auto *BB = Inst->getParent();
238         if (isa<PHINode>(Inst) && BB->getFirstInsertionPt() == BB->end())
239           return true;
240         return false;
241       }))
242     return nullptr;
243 
244   PHINode *NewPtrPHI = PHINode::Create(
245       IntToPtr->getType(), PN.getNumIncomingValues(), PN.getName() + ".ptr");
246 
247   InsertNewInstBefore(NewPtrPHI, PN);
248   SmallDenseMap<Value *, Instruction *> Casts;
249   for (auto Incoming : zip(PN.blocks(), AvailablePtrVals)) {
250     auto *IncomingBB = std::get<0>(Incoming);
251     auto *IncomingVal = std::get<1>(Incoming);
252 
253     if (IncomingVal->getType() == IntToPtr->getType()) {
254       NewPtrPHI->addIncoming(IncomingVal, IncomingBB);
255       continue;
256     }
257 
258 #ifndef NDEBUG
259     LoadInst *LoadI = dyn_cast<LoadInst>(IncomingVal);
260     assert((isa<PHINode>(IncomingVal) ||
261             IncomingVal->getType()->isPointerTy() ||
262             (LoadI && LoadI->hasOneUse())) &&
263            "Can not replace LoadInst with multiple uses");
264 #endif
265     // Need to insert a BitCast.
266     // For an integer Load instruction with a single use, the load + IntToPtr
267     // cast will be simplified into a pointer load:
268     // %v = load i64, i64* %a.ip, align 8
269     // %v.cast = inttoptr i64 %v to float **
270     // ==>
271     // %v.ptrp = bitcast i64 * %a.ip to float **
272     // %v.cast = load float *, float ** %v.ptrp, align 8
273     Instruction *&CI = Casts[IncomingVal];
274     if (!CI) {
275       CI = CastInst::CreateBitOrPointerCast(IncomingVal, IntToPtr->getType(),
276                                             IncomingVal->getName() + ".ptr");
277       if (auto *IncomingI = dyn_cast<Instruction>(IncomingVal)) {
278         BasicBlock::iterator InsertPos(IncomingI);
279         InsertPos++;
280         BasicBlock *BB = IncomingI->getParent();
281         if (isa<PHINode>(IncomingI))
282           InsertPos = BB->getFirstInsertionPt();
283         assert(InsertPos != BB->end() && "should have checked above");
284         InsertNewInstBefore(CI, *InsertPos);
285       } else {
286         auto *InsertBB = &IncomingBB->getParent()->getEntryBlock();
287         InsertNewInstBefore(CI, *InsertBB->getFirstInsertionPt());
288       }
289     }
290     NewPtrPHI->addIncoming(CI, IncomingBB);
291   }
292 
293   // The PtrToCast + IntToPtr will be simplified later
294   return CastInst::CreateBitOrPointerCast(NewPtrPHI,
295                                           IntToPtr->getOperand(0)->getType());
296 }
297 
298 // Remove RoundTrip IntToPtr/PtrToInt Cast on PHI-Operand and
299 // fold Phi-operand to bitcast.
300 Instruction *InstCombinerImpl::foldPHIArgIntToPtrToPHI(PHINode &PN) {
301   // convert ptr2int ( phi[ int2ptr(ptr2int(x))] ) --> ptr2int ( phi [ x ] )
302   // Make sure all uses of phi are ptr2int.
303   if (!all_of(PN.users(), [](User *U) { return isa<PtrToIntInst>(U); }))
304     return nullptr;
305 
306   // Iterating over all operands to check presence of target pointers for
307   // optimization.
308   bool OperandWithRoundTripCast = false;
309   for (unsigned OpNum = 0; OpNum != PN.getNumIncomingValues(); ++OpNum) {
310     if (auto *NewOp =
311             simplifyIntToPtrRoundTripCast(PN.getIncomingValue(OpNum))) {
312       PN.setIncomingValue(OpNum, NewOp);
313       OperandWithRoundTripCast = true;
314     }
315   }
316   if (!OperandWithRoundTripCast)
317     return nullptr;
318   return &PN;
319 }
320 
321 /// If we have something like phi [insertvalue(a,b,0), insertvalue(c,d,0)],
322 /// turn this into a phi[a,c] and phi[b,d] and a single insertvalue.
323 Instruction *
324 InstCombinerImpl::foldPHIArgInsertValueInstructionIntoPHI(PHINode &PN) {
325   auto *FirstIVI = cast<InsertValueInst>(PN.getIncomingValue(0));
326 
327   // Scan to see if all operands are `insertvalue`'s with the same indicies,
328   // and all have a single use.
329   for (Value *V : drop_begin(PN.incoming_values())) {
330     auto *I = dyn_cast<InsertValueInst>(V);
331     if (!I || !I->hasOneUser() || I->getIndices() != FirstIVI->getIndices())
332       return nullptr;
333   }
334 
335   // For each operand of an `insertvalue`
336   std::array<PHINode *, 2> NewOperands;
337   for (int OpIdx : {0, 1}) {
338     auto *&NewOperand = NewOperands[OpIdx];
339     // Create a new PHI node to receive the values the operand has in each
340     // incoming basic block.
341     NewOperand = PHINode::Create(
342         FirstIVI->getOperand(OpIdx)->getType(), PN.getNumIncomingValues(),
343         FirstIVI->getOperand(OpIdx)->getName() + ".pn");
344     // And populate each operand's PHI with said values.
345     for (auto Incoming : zip(PN.blocks(), PN.incoming_values()))
346       NewOperand->addIncoming(
347           cast<InsertValueInst>(std::get<1>(Incoming))->getOperand(OpIdx),
348           std::get<0>(Incoming));
349     InsertNewInstBefore(NewOperand, PN);
350   }
351 
352   // And finally, create `insertvalue` over the newly-formed PHI nodes.
353   auto *NewIVI = InsertValueInst::Create(NewOperands[0], NewOperands[1],
354                                          FirstIVI->getIndices(), PN.getName());
355 
356   PHIArgMergedDebugLoc(NewIVI, PN);
357   ++NumPHIsOfInsertValues;
358   return NewIVI;
359 }
360 
361 /// If we have something like phi [extractvalue(a,0), extractvalue(b,0)],
362 /// turn this into a phi[a,b] and a single extractvalue.
363 Instruction *
364 InstCombinerImpl::foldPHIArgExtractValueInstructionIntoPHI(PHINode &PN) {
365   auto *FirstEVI = cast<ExtractValueInst>(PN.getIncomingValue(0));
366 
367   // Scan to see if all operands are `extractvalue`'s with the same indicies,
368   // and all have a single use.
369   for (Value *V : drop_begin(PN.incoming_values())) {
370     auto *I = dyn_cast<ExtractValueInst>(V);
371     if (!I || !I->hasOneUser() || I->getIndices() != FirstEVI->getIndices() ||
372         I->getAggregateOperand()->getType() !=
373             FirstEVI->getAggregateOperand()->getType())
374       return nullptr;
375   }
376 
377   // Create a new PHI node to receive the values the aggregate operand has
378   // in each incoming basic block.
379   auto *NewAggregateOperand = PHINode::Create(
380       FirstEVI->getAggregateOperand()->getType(), PN.getNumIncomingValues(),
381       FirstEVI->getAggregateOperand()->getName() + ".pn");
382   // And populate the PHI with said values.
383   for (auto Incoming : zip(PN.blocks(), PN.incoming_values()))
384     NewAggregateOperand->addIncoming(
385         cast<ExtractValueInst>(std::get<1>(Incoming))->getAggregateOperand(),
386         std::get<0>(Incoming));
387   InsertNewInstBefore(NewAggregateOperand, PN);
388 
389   // And finally, create `extractvalue` over the newly-formed PHI nodes.
390   auto *NewEVI = ExtractValueInst::Create(NewAggregateOperand,
391                                           FirstEVI->getIndices(), PN.getName());
392 
393   PHIArgMergedDebugLoc(NewEVI, PN);
394   ++NumPHIsOfExtractValues;
395   return NewEVI;
396 }
397 
398 /// If we have something like phi [add (a,b), add(a,c)] and if a/b/c and the
399 /// adds all have a single user, turn this into a phi and a single binop.
400 Instruction *InstCombinerImpl::foldPHIArgBinOpIntoPHI(PHINode &PN) {
401   Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
402   assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
403   unsigned Opc = FirstInst->getOpcode();
404   Value *LHSVal = FirstInst->getOperand(0);
405   Value *RHSVal = FirstInst->getOperand(1);
406 
407   Type *LHSType = LHSVal->getType();
408   Type *RHSType = RHSVal->getType();
409 
410   // Scan to see if all operands are the same opcode, and all have one user.
411   for (Value *V : drop_begin(PN.incoming_values())) {
412     Instruction *I = dyn_cast<Instruction>(V);
413     if (!I || I->getOpcode() != Opc || !I->hasOneUser() ||
414         // Verify type of the LHS matches so we don't fold cmp's of different
415         // types.
416         I->getOperand(0)->getType() != LHSType ||
417         I->getOperand(1)->getType() != RHSType)
418       return nullptr;
419 
420     // If they are CmpInst instructions, check their predicates
421     if (CmpInst *CI = dyn_cast<CmpInst>(I))
422       if (CI->getPredicate() != cast<CmpInst>(FirstInst)->getPredicate())
423         return nullptr;
424 
425     // Keep track of which operand needs a phi node.
426     if (I->getOperand(0) != LHSVal) LHSVal = nullptr;
427     if (I->getOperand(1) != RHSVal) RHSVal = nullptr;
428   }
429 
430   // If both LHS and RHS would need a PHI, don't do this transformation,
431   // because it would increase the number of PHIs entering the block,
432   // which leads to higher register pressure. This is especially
433   // bad when the PHIs are in the header of a loop.
434   if (!LHSVal && !RHSVal)
435     return nullptr;
436 
437   // Otherwise, this is safe to transform!
438 
439   Value *InLHS = FirstInst->getOperand(0);
440   Value *InRHS = FirstInst->getOperand(1);
441   PHINode *NewLHS = nullptr, *NewRHS = nullptr;
442   if (!LHSVal) {
443     NewLHS = PHINode::Create(LHSType, PN.getNumIncomingValues(),
444                              FirstInst->getOperand(0)->getName() + ".pn");
445     NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
446     InsertNewInstBefore(NewLHS, PN);
447     LHSVal = NewLHS;
448   }
449 
450   if (!RHSVal) {
451     NewRHS = PHINode::Create(RHSType, PN.getNumIncomingValues(),
452                              FirstInst->getOperand(1)->getName() + ".pn");
453     NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
454     InsertNewInstBefore(NewRHS, PN);
455     RHSVal = NewRHS;
456   }
457 
458   // Add all operands to the new PHIs.
459   if (NewLHS || NewRHS) {
460     for (auto Incoming : drop_begin(zip(PN.blocks(), PN.incoming_values()))) {
461       BasicBlock *InBB = std::get<0>(Incoming);
462       Value *InVal = std::get<1>(Incoming);
463       Instruction *InInst = cast<Instruction>(InVal);
464       if (NewLHS) {
465         Value *NewInLHS = InInst->getOperand(0);
466         NewLHS->addIncoming(NewInLHS, InBB);
467       }
468       if (NewRHS) {
469         Value *NewInRHS = InInst->getOperand(1);
470         NewRHS->addIncoming(NewInRHS, InBB);
471       }
472     }
473   }
474 
475   if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst)) {
476     CmpInst *NewCI = CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
477                                      LHSVal, RHSVal);
478     PHIArgMergedDebugLoc(NewCI, PN);
479     return NewCI;
480   }
481 
482   BinaryOperator *BinOp = cast<BinaryOperator>(FirstInst);
483   BinaryOperator *NewBinOp =
484     BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
485 
486   NewBinOp->copyIRFlags(PN.getIncomingValue(0));
487 
488   for (Value *V : drop_begin(PN.incoming_values()))
489     NewBinOp->andIRFlags(V);
490 
491   PHIArgMergedDebugLoc(NewBinOp, PN);
492   return NewBinOp;
493 }
494 
495 Instruction *InstCombinerImpl::foldPHIArgGEPIntoPHI(PHINode &PN) {
496   GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
497 
498   SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
499                                         FirstInst->op_end());
500   // This is true if all GEP bases are allocas and if all indices into them are
501   // constants.
502   bool AllBasePointersAreAllocas = true;
503 
504   // We don't want to replace this phi if the replacement would require
505   // more than one phi, which leads to higher register pressure. This is
506   // especially bad when the PHIs are in the header of a loop.
507   bool NeededPhi = false;
508 
509   bool AllInBounds = true;
510 
511   // Scan to see if all operands are the same opcode, and all have one user.
512   for (Value *V : drop_begin(PN.incoming_values())) {
513     GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V);
514     if (!GEP || !GEP->hasOneUser() ||
515         GEP->getSourceElementType() != FirstInst->getSourceElementType() ||
516         GEP->getNumOperands() != FirstInst->getNumOperands())
517       return nullptr;
518 
519     AllInBounds &= GEP->isInBounds();
520 
521     // Keep track of whether or not all GEPs are of alloca pointers.
522     if (AllBasePointersAreAllocas &&
523         (!isa<AllocaInst>(GEP->getOperand(0)) ||
524          !GEP->hasAllConstantIndices()))
525       AllBasePointersAreAllocas = false;
526 
527     // Compare the operand lists.
528     for (unsigned Op = 0, E = FirstInst->getNumOperands(); Op != E; ++Op) {
529       if (FirstInst->getOperand(Op) == GEP->getOperand(Op))
530         continue;
531 
532       // Don't merge two GEPs when two operands differ (introducing phi nodes)
533       // if one of the PHIs has a constant for the index.  The index may be
534       // substantially cheaper to compute for the constants, so making it a
535       // variable index could pessimize the path.  This also handles the case
536       // for struct indices, which must always be constant.
537       if (isa<ConstantInt>(FirstInst->getOperand(Op)) ||
538           isa<ConstantInt>(GEP->getOperand(Op)))
539         return nullptr;
540 
541       if (FirstInst->getOperand(Op)->getType() !=
542           GEP->getOperand(Op)->getType())
543         return nullptr;
544 
545       // If we already needed a PHI for an earlier operand, and another operand
546       // also requires a PHI, we'd be introducing more PHIs than we're
547       // eliminating, which increases register pressure on entry to the PHI's
548       // block.
549       if (NeededPhi)
550         return nullptr;
551 
552       FixedOperands[Op] = nullptr; // Needs a PHI.
553       NeededPhi = true;
554     }
555   }
556 
557   // If all of the base pointers of the PHI'd GEPs are from allocas, don't
558   // bother doing this transformation.  At best, this will just save a bit of
559   // offset calculation, but all the predecessors will have to materialize the
560   // stack address into a register anyway.  We'd actually rather *clone* the
561   // load up into the predecessors so that we have a load of a gep of an alloca,
562   // which can usually all be folded into the load.
563   if (AllBasePointersAreAllocas)
564     return nullptr;
565 
566   // Otherwise, this is safe to transform.  Insert PHI nodes for each operand
567   // that is variable.
568   SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
569 
570   bool HasAnyPHIs = false;
571   for (unsigned I = 0, E = FixedOperands.size(); I != E; ++I) {
572     if (FixedOperands[I])
573       continue; // operand doesn't need a phi.
574     Value *FirstOp = FirstInst->getOperand(I);
575     PHINode *NewPN =
576         PHINode::Create(FirstOp->getType(), E, FirstOp->getName() + ".pn");
577     InsertNewInstBefore(NewPN, PN);
578 
579     NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
580     OperandPhis[I] = NewPN;
581     FixedOperands[I] = NewPN;
582     HasAnyPHIs = true;
583   }
584 
585   // Add all operands to the new PHIs.
586   if (HasAnyPHIs) {
587     for (auto Incoming : drop_begin(zip(PN.blocks(), PN.incoming_values()))) {
588       BasicBlock *InBB = std::get<0>(Incoming);
589       Value *InVal = std::get<1>(Incoming);
590       GetElementPtrInst *InGEP = cast<GetElementPtrInst>(InVal);
591 
592       for (unsigned Op = 0, E = OperandPhis.size(); Op != E; ++Op)
593         if (PHINode *OpPhi = OperandPhis[Op])
594           OpPhi->addIncoming(InGEP->getOperand(Op), InBB);
595     }
596   }
597 
598   Value *Base = FixedOperands[0];
599   GetElementPtrInst *NewGEP =
600       GetElementPtrInst::Create(FirstInst->getSourceElementType(), Base,
601                                 makeArrayRef(FixedOperands).slice(1));
602   if (AllInBounds) NewGEP->setIsInBounds();
603   PHIArgMergedDebugLoc(NewGEP, PN);
604   return NewGEP;
605 }
606 
607 /// Return true if we know that it is safe to sink the load out of the block
608 /// that defines it. This means that it must be obvious the value of the load is
609 /// not changed from the point of the load to the end of the block it is in.
610 ///
611 /// Finally, it is safe, but not profitable, to sink a load targeting a
612 /// non-address-taken alloca.  Doing so will cause us to not promote the alloca
613 /// to a register.
614 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
615   BasicBlock::iterator BBI = L->getIterator(), E = L->getParent()->end();
616 
617   for (++BBI; BBI != E; ++BBI)
618     if (BBI->mayWriteToMemory()) {
619       // Calls that only access inaccessible memory do not block sinking the
620       // load.
621       if (auto *CB = dyn_cast<CallBase>(BBI))
622         if (CB->onlyAccessesInaccessibleMemory())
623           continue;
624       return false;
625     }
626 
627   // Check for non-address taken alloca.  If not address-taken already, it isn't
628   // profitable to do this xform.
629   if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
630     bool IsAddressTaken = false;
631     for (User *U : AI->users()) {
632       if (isa<LoadInst>(U)) continue;
633       if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
634         // If storing TO the alloca, then the address isn't taken.
635         if (SI->getOperand(1) == AI) continue;
636       }
637       IsAddressTaken = true;
638       break;
639     }
640 
641     if (!IsAddressTaken && AI->isStaticAlloca())
642       return false;
643   }
644 
645   // If this load is a load from a GEP with a constant offset from an alloca,
646   // then we don't want to sink it.  In its present form, it will be
647   // load [constant stack offset].  Sinking it will cause us to have to
648   // materialize the stack addresses in each predecessor in a register only to
649   // do a shared load from register in the successor.
650   if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))
651     if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))
652       if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
653         return false;
654 
655   return true;
656 }
657 
658 Instruction *InstCombinerImpl::foldPHIArgLoadIntoPHI(PHINode &PN) {
659   LoadInst *FirstLI = cast<LoadInst>(PN.getIncomingValue(0));
660 
661   // Can't forward swifterror through a phi.
662   if (FirstLI->getOperand(0)->isSwiftError())
663     return nullptr;
664 
665   // FIXME: This is overconservative; this transform is allowed in some cases
666   // for atomic operations.
667   if (FirstLI->isAtomic())
668     return nullptr;
669 
670   // When processing loads, we need to propagate two bits of information to the
671   // sunk load: whether it is volatile, and what its alignment is.
672   bool IsVolatile = FirstLI->isVolatile();
673   Align LoadAlignment = FirstLI->getAlign();
674   const unsigned LoadAddrSpace = FirstLI->getPointerAddressSpace();
675 
676   // We can't sink the load if the loaded value could be modified between the
677   // load and the PHI.
678   if (FirstLI->getParent() != PN.getIncomingBlock(0) ||
679       !isSafeAndProfitableToSinkLoad(FirstLI))
680     return nullptr;
681 
682   // If the PHI is of volatile loads and the load block has multiple
683   // successors, sinking it would remove a load of the volatile value from
684   // the path through the other successor.
685   if (IsVolatile &&
686       FirstLI->getParent()->getTerminator()->getNumSuccessors() != 1)
687     return nullptr;
688 
689   for (auto Incoming : drop_begin(zip(PN.blocks(), PN.incoming_values()))) {
690     BasicBlock *InBB = std::get<0>(Incoming);
691     Value *InVal = std::get<1>(Incoming);
692     LoadInst *LI = dyn_cast<LoadInst>(InVal);
693     if (!LI || !LI->hasOneUser() || LI->isAtomic())
694       return nullptr;
695 
696     // Make sure all arguments are the same type of operation.
697     if (LI->isVolatile() != IsVolatile ||
698         LI->getPointerAddressSpace() != LoadAddrSpace)
699       return nullptr;
700 
701     // Can't forward swifterror through a phi.
702     if (LI->getOperand(0)->isSwiftError())
703       return nullptr;
704 
705     // We can't sink the load if the loaded value could be modified between
706     // the load and the PHI.
707     if (LI->getParent() != InBB || !isSafeAndProfitableToSinkLoad(LI))
708       return nullptr;
709 
710     LoadAlignment = std::min(LoadAlignment, LI->getAlign());
711 
712     // If the PHI is of volatile loads and the load block has multiple
713     // successors, sinking it would remove a load of the volatile value from
714     // the path through the other successor.
715     if (IsVolatile && LI->getParent()->getTerminator()->getNumSuccessors() != 1)
716       return nullptr;
717   }
718 
719   // Okay, they are all the same operation.  Create a new PHI node of the
720   // correct type, and PHI together all of the LHS's of the instructions.
721   PHINode *NewPN = PHINode::Create(FirstLI->getOperand(0)->getType(),
722                                    PN.getNumIncomingValues(),
723                                    PN.getName()+".in");
724 
725   Value *InVal = FirstLI->getOperand(0);
726   NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
727   LoadInst *NewLI =
728       new LoadInst(FirstLI->getType(), NewPN, "", IsVolatile, LoadAlignment);
729 
730   unsigned KnownIDs[] = {
731     LLVMContext::MD_tbaa,
732     LLVMContext::MD_range,
733     LLVMContext::MD_invariant_load,
734     LLVMContext::MD_alias_scope,
735     LLVMContext::MD_noalias,
736     LLVMContext::MD_nonnull,
737     LLVMContext::MD_align,
738     LLVMContext::MD_dereferenceable,
739     LLVMContext::MD_dereferenceable_or_null,
740     LLVMContext::MD_access_group,
741   };
742 
743   for (unsigned ID : KnownIDs)
744     NewLI->setMetadata(ID, FirstLI->getMetadata(ID));
745 
746   // Add all operands to the new PHI and combine TBAA metadata.
747   for (auto Incoming : drop_begin(zip(PN.blocks(), PN.incoming_values()))) {
748     BasicBlock *BB = std::get<0>(Incoming);
749     Value *V = std::get<1>(Incoming);
750     LoadInst *LI = cast<LoadInst>(V);
751     combineMetadata(NewLI, LI, KnownIDs, true);
752     Value *NewInVal = LI->getOperand(0);
753     if (NewInVal != InVal)
754       InVal = nullptr;
755     NewPN->addIncoming(NewInVal, BB);
756   }
757 
758   if (InVal) {
759     // The new PHI unions all of the same values together.  This is really
760     // common, so we handle it intelligently here for compile-time speed.
761     NewLI->setOperand(0, InVal);
762     delete NewPN;
763   } else {
764     InsertNewInstBefore(NewPN, PN);
765   }
766 
767   // If this was a volatile load that we are merging, make sure to loop through
768   // and mark all the input loads as non-volatile.  If we don't do this, we will
769   // insert a new volatile load and the old ones will not be deletable.
770   if (IsVolatile)
771     for (Value *IncValue : PN.incoming_values())
772       cast<LoadInst>(IncValue)->setVolatile(false);
773 
774   PHIArgMergedDebugLoc(NewLI, PN);
775   return NewLI;
776 }
777 
778 /// TODO: This function could handle other cast types, but then it might
779 /// require special-casing a cast from the 'i1' type. See the comment in
780 /// FoldPHIArgOpIntoPHI() about pessimizing illegal integer types.
781 Instruction *InstCombinerImpl::foldPHIArgZextsIntoPHI(PHINode &Phi) {
782   // We cannot create a new instruction after the PHI if the terminator is an
783   // EHPad because there is no valid insertion point.
784   if (Instruction *TI = Phi.getParent()->getTerminator())
785     if (TI->isEHPad())
786       return nullptr;
787 
788   // Early exit for the common case of a phi with two operands. These are
789   // handled elsewhere. See the comment below where we check the count of zexts
790   // and constants for more details.
791   unsigned NumIncomingValues = Phi.getNumIncomingValues();
792   if (NumIncomingValues < 3)
793     return nullptr;
794 
795   // Find the narrower type specified by the first zext.
796   Type *NarrowType = nullptr;
797   for (Value *V : Phi.incoming_values()) {
798     if (auto *Zext = dyn_cast<ZExtInst>(V)) {
799       NarrowType = Zext->getSrcTy();
800       break;
801     }
802   }
803   if (!NarrowType)
804     return nullptr;
805 
806   // Walk the phi operands checking that we only have zexts or constants that
807   // we can shrink for free. Store the new operands for the new phi.
808   SmallVector<Value *, 4> NewIncoming;
809   unsigned NumZexts = 0;
810   unsigned NumConsts = 0;
811   for (Value *V : Phi.incoming_values()) {
812     if (auto *Zext = dyn_cast<ZExtInst>(V)) {
813       // All zexts must be identical and have one user.
814       if (Zext->getSrcTy() != NarrowType || !Zext->hasOneUser())
815         return nullptr;
816       NewIncoming.push_back(Zext->getOperand(0));
817       NumZexts++;
818     } else if (auto *C = dyn_cast<Constant>(V)) {
819       // Make sure that constants can fit in the new type.
820       Constant *Trunc = ConstantExpr::getTrunc(C, NarrowType);
821       if (ConstantExpr::getZExt(Trunc, C->getType()) != C)
822         return nullptr;
823       NewIncoming.push_back(Trunc);
824       NumConsts++;
825     } else {
826       // If it's not a cast or a constant, bail out.
827       return nullptr;
828     }
829   }
830 
831   // The more common cases of a phi with no constant operands or just one
832   // variable operand are handled by FoldPHIArgOpIntoPHI() and foldOpIntoPhi()
833   // respectively. foldOpIntoPhi() wants to do the opposite transform that is
834   // performed here. It tries to replicate a cast in the phi operand's basic
835   // block to expose other folding opportunities. Thus, InstCombine will
836   // infinite loop without this check.
837   if (NumConsts == 0 || NumZexts < 2)
838     return nullptr;
839 
840   // All incoming values are zexts or constants that are safe to truncate.
841   // Create a new phi node of the narrow type, phi together all of the new
842   // operands, and zext the result back to the original type.
843   PHINode *NewPhi = PHINode::Create(NarrowType, NumIncomingValues,
844                                     Phi.getName() + ".shrunk");
845   for (unsigned I = 0; I != NumIncomingValues; ++I)
846     NewPhi->addIncoming(NewIncoming[I], Phi.getIncomingBlock(I));
847 
848   InsertNewInstBefore(NewPhi, Phi);
849   return CastInst::CreateZExtOrBitCast(NewPhi, Phi.getType());
850 }
851 
852 /// If all operands to a PHI node are the same "unary" operator and they all are
853 /// only used by the PHI, PHI together their inputs, and do the operation once,
854 /// to the result of the PHI.
855 Instruction *InstCombinerImpl::foldPHIArgOpIntoPHI(PHINode &PN) {
856   // We cannot create a new instruction after the PHI if the terminator is an
857   // EHPad because there is no valid insertion point.
858   if (Instruction *TI = PN.getParent()->getTerminator())
859     if (TI->isEHPad())
860       return nullptr;
861 
862   Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
863 
864   if (isa<GetElementPtrInst>(FirstInst))
865     return foldPHIArgGEPIntoPHI(PN);
866   if (isa<LoadInst>(FirstInst))
867     return foldPHIArgLoadIntoPHI(PN);
868   if (isa<InsertValueInst>(FirstInst))
869     return foldPHIArgInsertValueInstructionIntoPHI(PN);
870   if (isa<ExtractValueInst>(FirstInst))
871     return foldPHIArgExtractValueInstructionIntoPHI(PN);
872 
873   // Scan the instruction, looking for input operations that can be folded away.
874   // If all input operands to the phi are the same instruction (e.g. a cast from
875   // the same type or "+42") we can pull the operation through the PHI, reducing
876   // code size and simplifying code.
877   Constant *ConstantOp = nullptr;
878   Type *CastSrcTy = nullptr;
879 
880   if (isa<CastInst>(FirstInst)) {
881     CastSrcTy = FirstInst->getOperand(0)->getType();
882 
883     // Be careful about transforming integer PHIs.  We don't want to pessimize
884     // the code by turning an i32 into an i1293.
885     if (PN.getType()->isIntegerTy() && CastSrcTy->isIntegerTy()) {
886       if (!shouldChangeType(PN.getType(), CastSrcTy))
887         return nullptr;
888     }
889   } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
890     // Can fold binop, compare or shift here if the RHS is a constant,
891     // otherwise call FoldPHIArgBinOpIntoPHI.
892     ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
893     if (!ConstantOp)
894       return foldPHIArgBinOpIntoPHI(PN);
895   } else {
896     return nullptr;  // Cannot fold this operation.
897   }
898 
899   // Check to see if all arguments are the same operation.
900   for (Value *V : drop_begin(PN.incoming_values())) {
901     Instruction *I = dyn_cast<Instruction>(V);
902     if (!I || !I->hasOneUser() || !I->isSameOperationAs(FirstInst))
903       return nullptr;
904     if (CastSrcTy) {
905       if (I->getOperand(0)->getType() != CastSrcTy)
906         return nullptr; // Cast operation must match.
907     } else if (I->getOperand(1) != ConstantOp) {
908       return nullptr;
909     }
910   }
911 
912   // Okay, they are all the same operation.  Create a new PHI node of the
913   // correct type, and PHI together all of the LHS's of the instructions.
914   PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
915                                    PN.getNumIncomingValues(),
916                                    PN.getName()+".in");
917 
918   Value *InVal = FirstInst->getOperand(0);
919   NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
920 
921   // Add all operands to the new PHI.
922   for (auto Incoming : drop_begin(zip(PN.blocks(), PN.incoming_values()))) {
923     BasicBlock *BB = std::get<0>(Incoming);
924     Value *V = std::get<1>(Incoming);
925     Value *NewInVal = cast<Instruction>(V)->getOperand(0);
926     if (NewInVal != InVal)
927       InVal = nullptr;
928     NewPN->addIncoming(NewInVal, BB);
929   }
930 
931   Value *PhiVal;
932   if (InVal) {
933     // The new PHI unions all of the same values together.  This is really
934     // common, so we handle it intelligently here for compile-time speed.
935     PhiVal = InVal;
936     delete NewPN;
937   } else {
938     InsertNewInstBefore(NewPN, PN);
939     PhiVal = NewPN;
940   }
941 
942   // Insert and return the new operation.
943   if (CastInst *FirstCI = dyn_cast<CastInst>(FirstInst)) {
944     CastInst *NewCI = CastInst::Create(FirstCI->getOpcode(), PhiVal,
945                                        PN.getType());
946     PHIArgMergedDebugLoc(NewCI, PN);
947     return NewCI;
948   }
949 
950   if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst)) {
951     BinOp = BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
952     BinOp->copyIRFlags(PN.getIncomingValue(0));
953 
954     for (Value *V : drop_begin(PN.incoming_values()))
955       BinOp->andIRFlags(V);
956 
957     PHIArgMergedDebugLoc(BinOp, PN);
958     return BinOp;
959   }
960 
961   CmpInst *CIOp = cast<CmpInst>(FirstInst);
962   CmpInst *NewCI = CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
963                                    PhiVal, ConstantOp);
964   PHIArgMergedDebugLoc(NewCI, PN);
965   return NewCI;
966 }
967 
968 /// Return true if this PHI node is only used by a PHI node cycle that is dead.
969 static bool isDeadPHICycle(PHINode *PN,
970                            SmallPtrSetImpl<PHINode *> &PotentiallyDeadPHIs) {
971   if (PN->use_empty()) return true;
972   if (!PN->hasOneUse()) return false;
973 
974   // Remember this node, and if we find the cycle, return.
975   if (!PotentiallyDeadPHIs.insert(PN).second)
976     return true;
977 
978   // Don't scan crazily complex things.
979   if (PotentiallyDeadPHIs.size() == 16)
980     return false;
981 
982   if (PHINode *PU = dyn_cast<PHINode>(PN->user_back()))
983     return isDeadPHICycle(PU, PotentiallyDeadPHIs);
984 
985   return false;
986 }
987 
988 /// Return true if this phi node is always equal to NonPhiInVal.
989 /// This happens with mutually cyclic phi nodes like:
990 ///   z = some value; x = phi (y, z); y = phi (x, z)
991 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
992                            SmallPtrSetImpl<PHINode*> &ValueEqualPHIs) {
993   // See if we already saw this PHI node.
994   if (!ValueEqualPHIs.insert(PN).second)
995     return true;
996 
997   // Don't scan crazily complex things.
998   if (ValueEqualPHIs.size() == 16)
999     return false;
1000 
1001   // Scan the operands to see if they are either phi nodes or are equal to
1002   // the value.
1003   for (Value *Op : PN->incoming_values()) {
1004     if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
1005       if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
1006         return false;
1007     } else if (Op != NonPhiInVal)
1008       return false;
1009   }
1010 
1011   return true;
1012 }
1013 
1014 /// Return an existing non-zero constant if this phi node has one, otherwise
1015 /// return constant 1.
1016 static ConstantInt *getAnyNonZeroConstInt(PHINode &PN) {
1017   assert(isa<IntegerType>(PN.getType()) && "Expect only integer type phi");
1018   for (Value *V : PN.operands())
1019     if (auto *ConstVA = dyn_cast<ConstantInt>(V))
1020       if (!ConstVA->isZero())
1021         return ConstVA;
1022   return ConstantInt::get(cast<IntegerType>(PN.getType()), 1);
1023 }
1024 
1025 namespace {
1026 struct PHIUsageRecord {
1027   unsigned PHIId;     // The ID # of the PHI (something determinstic to sort on)
1028   unsigned Shift;     // The amount shifted.
1029   Instruction *Inst;  // The trunc instruction.
1030 
1031   PHIUsageRecord(unsigned Pn, unsigned Sh, Instruction *User)
1032       : PHIId(Pn), Shift(Sh), Inst(User) {}
1033 
1034   bool operator<(const PHIUsageRecord &RHS) const {
1035     if (PHIId < RHS.PHIId) return true;
1036     if (PHIId > RHS.PHIId) return false;
1037     if (Shift < RHS.Shift) return true;
1038     if (Shift > RHS.Shift) return false;
1039     return Inst->getType()->getPrimitiveSizeInBits() <
1040            RHS.Inst->getType()->getPrimitiveSizeInBits();
1041   }
1042 };
1043 
1044 struct LoweredPHIRecord {
1045   PHINode *PN;        // The PHI that was lowered.
1046   unsigned Shift;     // The amount shifted.
1047   unsigned Width;     // The width extracted.
1048 
1049   LoweredPHIRecord(PHINode *Phi, unsigned Sh, Type *Ty)
1050       : PN(Phi), Shift(Sh), Width(Ty->getPrimitiveSizeInBits()) {}
1051 
1052   // Ctor form used by DenseMap.
1053   LoweredPHIRecord(PHINode *Phi, unsigned Sh) : PN(Phi), Shift(Sh), Width(0) {}
1054 };
1055 } // namespace
1056 
1057 namespace llvm {
1058   template<>
1059   struct DenseMapInfo<LoweredPHIRecord> {
1060     static inline LoweredPHIRecord getEmptyKey() {
1061       return LoweredPHIRecord(nullptr, 0);
1062     }
1063     static inline LoweredPHIRecord getTombstoneKey() {
1064       return LoweredPHIRecord(nullptr, 1);
1065     }
1066     static unsigned getHashValue(const LoweredPHIRecord &Val) {
1067       return DenseMapInfo<PHINode*>::getHashValue(Val.PN) ^ (Val.Shift>>3) ^
1068              (Val.Width>>3);
1069     }
1070     static bool isEqual(const LoweredPHIRecord &LHS,
1071                         const LoweredPHIRecord &RHS) {
1072       return LHS.PN == RHS.PN && LHS.Shift == RHS.Shift &&
1073              LHS.Width == RHS.Width;
1074     }
1075   };
1076 } // namespace llvm
1077 
1078 
1079 /// This is an integer PHI and we know that it has an illegal type: see if it is
1080 /// only used by trunc or trunc(lshr) operations. If so, we split the PHI into
1081 /// the various pieces being extracted. This sort of thing is introduced when
1082 /// SROA promotes an aggregate to large integer values.
1083 ///
1084 /// TODO: The user of the trunc may be an bitcast to float/double/vector or an
1085 /// inttoptr.  We should produce new PHIs in the right type.
1086 ///
1087 Instruction *InstCombinerImpl::SliceUpIllegalIntegerPHI(PHINode &FirstPhi) {
1088   // PHIUsers - Keep track of all of the truncated values extracted from a set
1089   // of PHIs, along with their offset.  These are the things we want to rewrite.
1090   SmallVector<PHIUsageRecord, 16> PHIUsers;
1091 
1092   // PHIs are often mutually cyclic, so we keep track of a whole set of PHI
1093   // nodes which are extracted from. PHIsToSlice is a set we use to avoid
1094   // revisiting PHIs, PHIsInspected is a ordered list of PHIs that we need to
1095   // check the uses of (to ensure they are all extracts).
1096   SmallVector<PHINode*, 8> PHIsToSlice;
1097   SmallPtrSet<PHINode*, 8> PHIsInspected;
1098 
1099   PHIsToSlice.push_back(&FirstPhi);
1100   PHIsInspected.insert(&FirstPhi);
1101 
1102   for (unsigned PHIId = 0; PHIId != PHIsToSlice.size(); ++PHIId) {
1103     PHINode *PN = PHIsToSlice[PHIId];
1104 
1105     // Scan the input list of the PHI.  If any input is an invoke, and if the
1106     // input is defined in the predecessor, then we won't be split the critical
1107     // edge which is required to insert a truncate.  Because of this, we have to
1108     // bail out.
1109     for (auto Incoming : zip(PN->blocks(), PN->incoming_values())) {
1110       BasicBlock *BB = std::get<0>(Incoming);
1111       Value *V = std::get<1>(Incoming);
1112       InvokeInst *II = dyn_cast<InvokeInst>(V);
1113       if (!II)
1114         continue;
1115       if (II->getParent() != BB)
1116         continue;
1117 
1118       // If we have a phi, and if it's directly in the predecessor, then we have
1119       // a critical edge where we need to put the truncate.  Since we can't
1120       // split the edge in instcombine, we have to bail out.
1121       return nullptr;
1122     }
1123 
1124     // If the incoming value is a PHI node before a catchswitch, we cannot
1125     // extract the value within that BB because we cannot insert any non-PHI
1126     // instructions in the BB.
1127     for (auto *Pred : PN->blocks())
1128       if (Pred->getFirstInsertionPt() == Pred->end())
1129         return nullptr;
1130 
1131     for (User *U : PN->users()) {
1132       Instruction *UserI = cast<Instruction>(U);
1133 
1134       // If the user is a PHI, inspect its uses recursively.
1135       if (PHINode *UserPN = dyn_cast<PHINode>(UserI)) {
1136         if (PHIsInspected.insert(UserPN).second)
1137           PHIsToSlice.push_back(UserPN);
1138         continue;
1139       }
1140 
1141       // Truncates are always ok.
1142       if (isa<TruncInst>(UserI)) {
1143         PHIUsers.push_back(PHIUsageRecord(PHIId, 0, UserI));
1144         continue;
1145       }
1146 
1147       // Otherwise it must be a lshr which can only be used by one trunc.
1148       if (UserI->getOpcode() != Instruction::LShr ||
1149           !UserI->hasOneUse() || !isa<TruncInst>(UserI->user_back()) ||
1150           !isa<ConstantInt>(UserI->getOperand(1)))
1151         return nullptr;
1152 
1153       // Bail on out of range shifts.
1154       unsigned SizeInBits = UserI->getType()->getScalarSizeInBits();
1155       if (cast<ConstantInt>(UserI->getOperand(1))->getValue().uge(SizeInBits))
1156         return nullptr;
1157 
1158       unsigned Shift = cast<ConstantInt>(UserI->getOperand(1))->getZExtValue();
1159       PHIUsers.push_back(PHIUsageRecord(PHIId, Shift, UserI->user_back()));
1160     }
1161   }
1162 
1163   // If we have no users, they must be all self uses, just nuke the PHI.
1164   if (PHIUsers.empty())
1165     return replaceInstUsesWith(FirstPhi, PoisonValue::get(FirstPhi.getType()));
1166 
1167   // If this phi node is transformable, create new PHIs for all the pieces
1168   // extracted out of it.  First, sort the users by their offset and size.
1169   array_pod_sort(PHIUsers.begin(), PHIUsers.end());
1170 
1171   LLVM_DEBUG(dbgs() << "SLICING UP PHI: " << FirstPhi << '\n';
1172              for (unsigned I = 1; I != PHIsToSlice.size(); ++I) dbgs()
1173              << "AND USER PHI #" << I << ": " << *PHIsToSlice[I] << '\n');
1174 
1175   // PredValues - This is a temporary used when rewriting PHI nodes.  It is
1176   // hoisted out here to avoid construction/destruction thrashing.
1177   DenseMap<BasicBlock*, Value*> PredValues;
1178 
1179   // ExtractedVals - Each new PHI we introduce is saved here so we don't
1180   // introduce redundant PHIs.
1181   DenseMap<LoweredPHIRecord, PHINode*> ExtractedVals;
1182 
1183   for (unsigned UserI = 0, UserE = PHIUsers.size(); UserI != UserE; ++UserI) {
1184     unsigned PHIId = PHIUsers[UserI].PHIId;
1185     PHINode *PN = PHIsToSlice[PHIId];
1186     unsigned Offset = PHIUsers[UserI].Shift;
1187     Type *Ty = PHIUsers[UserI].Inst->getType();
1188 
1189     PHINode *EltPHI;
1190 
1191     // If we've already lowered a user like this, reuse the previously lowered
1192     // value.
1193     if ((EltPHI = ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)]) == nullptr) {
1194 
1195       // Otherwise, Create the new PHI node for this user.
1196       EltPHI = PHINode::Create(Ty, PN->getNumIncomingValues(),
1197                                PN->getName()+".off"+Twine(Offset), PN);
1198       assert(EltPHI->getType() != PN->getType() &&
1199              "Truncate didn't shrink phi?");
1200 
1201       for (auto Incoming : zip(PN->blocks(), PN->incoming_values())) {
1202         BasicBlock *Pred = std::get<0>(Incoming);
1203         Value *InVal = std::get<1>(Incoming);
1204         Value *&PredVal = PredValues[Pred];
1205 
1206         // If we already have a value for this predecessor, reuse it.
1207         if (PredVal) {
1208           EltPHI->addIncoming(PredVal, Pred);
1209           continue;
1210         }
1211 
1212         // Handle the PHI self-reuse case.
1213         if (InVal == PN) {
1214           PredVal = EltPHI;
1215           EltPHI->addIncoming(PredVal, Pred);
1216           continue;
1217         }
1218 
1219         if (PHINode *InPHI = dyn_cast<PHINode>(PN)) {
1220           // If the incoming value was a PHI, and if it was one of the PHIs we
1221           // already rewrote it, just use the lowered value.
1222           if (Value *Res = ExtractedVals[LoweredPHIRecord(InPHI, Offset, Ty)]) {
1223             PredVal = Res;
1224             EltPHI->addIncoming(PredVal, Pred);
1225             continue;
1226           }
1227         }
1228 
1229         // Otherwise, do an extract in the predecessor.
1230         Builder.SetInsertPoint(Pred->getTerminator());
1231         Value *Res = InVal;
1232         if (Offset)
1233           Res = Builder.CreateLShr(
1234               Res, ConstantInt::get(InVal->getType(), Offset), "extract");
1235         Res = Builder.CreateTrunc(Res, Ty, "extract.t");
1236         PredVal = Res;
1237         EltPHI->addIncoming(Res, Pred);
1238 
1239         // If the incoming value was a PHI, and if it was one of the PHIs we are
1240         // rewriting, we will ultimately delete the code we inserted.  This
1241         // means we need to revisit that PHI to make sure we extract out the
1242         // needed piece.
1243         if (PHINode *OldInVal = dyn_cast<PHINode>(InVal))
1244           if (PHIsInspected.count(OldInVal)) {
1245             unsigned RefPHIId =
1246                 find(PHIsToSlice, OldInVal) - PHIsToSlice.begin();
1247             PHIUsers.push_back(
1248                 PHIUsageRecord(RefPHIId, Offset, cast<Instruction>(Res)));
1249             ++UserE;
1250           }
1251       }
1252       PredValues.clear();
1253 
1254       LLVM_DEBUG(dbgs() << "  Made element PHI for offset " << Offset << ": "
1255                         << *EltPHI << '\n');
1256       ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)] = EltPHI;
1257     }
1258 
1259     // Replace the use of this piece with the PHI node.
1260     replaceInstUsesWith(*PHIUsers[UserI].Inst, EltPHI);
1261   }
1262 
1263   // Replace all the remaining uses of the PHI nodes (self uses and the lshrs)
1264   // with poison.
1265   Value *Poison = PoisonValue::get(FirstPhi.getType());
1266   for (PHINode *PHI : drop_begin(PHIsToSlice))
1267     replaceInstUsesWith(*PHI, Poison);
1268   return replaceInstUsesWith(FirstPhi, Poison);
1269 }
1270 
1271 static Value *simplifyUsingControlFlow(InstCombiner &Self, PHINode &PN,
1272                                        const DominatorTree &DT) {
1273   // Simplify the following patterns:
1274   //       if (cond)
1275   //       /       \
1276   //      ...      ...
1277   //       \       /
1278   //    phi [true] [false]
1279   // and
1280   //        switch (cond)
1281   // case v1: /       \ case v2:
1282   //         ...      ...
1283   //          \       /
1284   //       phi [v1] [v2]
1285   // Make sure all inputs are constants.
1286   if (!all_of(PN.operands(), [](Value *V) { return isa<ConstantInt>(V); }))
1287     return nullptr;
1288 
1289   BasicBlock *BB = PN.getParent();
1290   // Do not bother with unreachable instructions.
1291   if (!DT.isReachableFromEntry(BB))
1292     return nullptr;
1293 
1294   // Determine which value the condition of the idom has for which successor.
1295   LLVMContext &Context = PN.getContext();
1296   auto *IDom = DT.getNode(BB)->getIDom()->getBlock();
1297   Value *Cond;
1298   SmallDenseMap<ConstantInt *, BasicBlock *, 8> SuccForValue;
1299   SmallDenseMap<BasicBlock *, unsigned, 8> SuccCount;
1300   auto AddSucc = [&](ConstantInt *C, BasicBlock *Succ) {
1301     SuccForValue[C] = Succ;
1302     ++SuccCount[Succ];
1303   };
1304   if (auto *BI = dyn_cast<BranchInst>(IDom->getTerminator())) {
1305     if (BI->isUnconditional())
1306       return nullptr;
1307 
1308     Cond = BI->getCondition();
1309     AddSucc(ConstantInt::getTrue(Context), BI->getSuccessor(0));
1310     AddSucc(ConstantInt::getFalse(Context), BI->getSuccessor(1));
1311   } else if (auto *SI = dyn_cast<SwitchInst>(IDom->getTerminator())) {
1312     Cond = SI->getCondition();
1313     ++SuccCount[SI->getDefaultDest()];
1314     for (auto Case : SI->cases())
1315       AddSucc(Case.getCaseValue(), Case.getCaseSuccessor());
1316   } else {
1317     return nullptr;
1318   }
1319 
1320   if (Cond->getType() != PN.getType())
1321     return nullptr;
1322 
1323   // Check that edges outgoing from the idom's terminators dominate respective
1324   // inputs of the Phi.
1325   Optional<bool> Invert;
1326   for (auto Pair : zip(PN.incoming_values(), PN.blocks())) {
1327     auto *Input = cast<ConstantInt>(std::get<0>(Pair));
1328     BasicBlock *Pred = std::get<1>(Pair);
1329     auto IsCorrectInput = [&](ConstantInt *Input) {
1330       // The input needs to be dominated by the corresponding edge of the idom.
1331       // This edge cannot be a multi-edge, as that would imply that multiple
1332       // different condition values follow the same edge.
1333       auto It = SuccForValue.find(Input);
1334       return It != SuccForValue.end() && SuccCount[It->second] == 1 &&
1335              DT.dominates(BasicBlockEdge(IDom, It->second),
1336                           BasicBlockEdge(Pred, BB));
1337     };
1338 
1339     // Depending on the constant, the condition may need to be inverted.
1340     bool NeedsInvert;
1341     if (IsCorrectInput(Input))
1342       NeedsInvert = false;
1343     else if (IsCorrectInput(cast<ConstantInt>(ConstantExpr::getNot(Input))))
1344       NeedsInvert = true;
1345     else
1346       return nullptr;
1347 
1348     // Make sure the inversion requirement is always the same.
1349     if (Invert && *Invert != NeedsInvert)
1350       return nullptr;
1351 
1352     Invert = NeedsInvert;
1353   }
1354 
1355   if (!*Invert)
1356     return Cond;
1357 
1358   // This Phi is actually opposite to branching condition of IDom. We invert
1359   // the condition that will potentially open up some opportunities for
1360   // sinking.
1361   auto InsertPt = BB->getFirstInsertionPt();
1362   if (InsertPt != BB->end()) {
1363     Self.Builder.SetInsertPoint(&*InsertPt);
1364     return Self.Builder.CreateNot(Cond);
1365   }
1366 
1367   return nullptr;
1368 }
1369 
1370 // PHINode simplification
1371 //
1372 Instruction *InstCombinerImpl::visitPHINode(PHINode &PN) {
1373   if (Value *V = simplifyInstruction(&PN, SQ.getWithInstruction(&PN)))
1374     return replaceInstUsesWith(PN, V);
1375 
1376   if (Instruction *Result = foldPHIArgZextsIntoPHI(PN))
1377     return Result;
1378 
1379   if (Instruction *Result = foldPHIArgIntToPtrToPHI(PN))
1380     return Result;
1381 
1382   // If all PHI operands are the same operation, pull them through the PHI,
1383   // reducing code size.
1384   if (isa<Instruction>(PN.getIncomingValue(0)) &&
1385       isa<Instruction>(PN.getIncomingValue(1)) &&
1386       cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
1387           cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
1388       PN.getIncomingValue(0)->hasOneUser())
1389     if (Instruction *Result = foldPHIArgOpIntoPHI(PN))
1390       return Result;
1391 
1392   // If the incoming values are pointer casts of the same original value,
1393   // replace the phi with a single cast iff we can insert a non-PHI instruction.
1394   if (PN.getType()->isPointerTy() &&
1395       PN.getParent()->getFirstInsertionPt() != PN.getParent()->end()) {
1396     Value *IV0 = PN.getIncomingValue(0);
1397     Value *IV0Stripped = IV0->stripPointerCasts();
1398     // Set to keep track of values known to be equal to IV0Stripped after
1399     // stripping pointer casts.
1400     SmallPtrSet<Value *, 4> CheckedIVs;
1401     CheckedIVs.insert(IV0);
1402     if (IV0 != IV0Stripped &&
1403         all_of(PN.incoming_values(), [&CheckedIVs, IV0Stripped](Value *IV) {
1404           return !CheckedIVs.insert(IV).second ||
1405                  IV0Stripped == IV->stripPointerCasts();
1406         })) {
1407       return CastInst::CreatePointerCast(IV0Stripped, PN.getType());
1408     }
1409   }
1410 
1411   // If this is a trivial cycle in the PHI node graph, remove it.  Basically, if
1412   // this PHI only has a single use (a PHI), and if that PHI only has one use (a
1413   // PHI)... break the cycle.
1414   if (PN.hasOneUse()) {
1415     if (Instruction *Result = foldIntegerTypedPHI(PN))
1416       return Result;
1417 
1418     Instruction *PHIUser = cast<Instruction>(PN.user_back());
1419     if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
1420       SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
1421       PotentiallyDeadPHIs.insert(&PN);
1422       if (isDeadPHICycle(PU, PotentiallyDeadPHIs))
1423         return replaceInstUsesWith(PN, PoisonValue::get(PN.getType()));
1424     }
1425 
1426     // If this phi has a single use, and if that use just computes a value for
1427     // the next iteration of a loop, delete the phi.  This occurs with unused
1428     // induction variables, e.g. "for (int j = 0; ; ++j);".  Detecting this
1429     // common case here is good because the only other things that catch this
1430     // are induction variable analysis (sometimes) and ADCE, which is only run
1431     // late.
1432     if (PHIUser->hasOneUse() &&
1433         (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
1434         PHIUser->user_back() == &PN) {
1435       return replaceInstUsesWith(PN, PoisonValue::get(PN.getType()));
1436     }
1437     // When a PHI is used only to be compared with zero, it is safe to replace
1438     // an incoming value proved as known nonzero with any non-zero constant.
1439     // For example, in the code below, the incoming value %v can be replaced
1440     // with any non-zero constant based on the fact that the PHI is only used to
1441     // be compared with zero and %v is a known non-zero value:
1442     // %v = select %cond, 1, 2
1443     // %p = phi [%v, BB] ...
1444     //      icmp eq, %p, 0
1445     auto *CmpInst = dyn_cast<ICmpInst>(PHIUser);
1446     // FIXME: To be simple, handle only integer type for now.
1447     if (CmpInst && isa<IntegerType>(PN.getType()) && CmpInst->isEquality() &&
1448         match(CmpInst->getOperand(1), m_Zero())) {
1449       ConstantInt *NonZeroConst = nullptr;
1450       bool MadeChange = false;
1451       for (unsigned I = 0, E = PN.getNumIncomingValues(); I != E; ++I) {
1452         Instruction *CtxI = PN.getIncomingBlock(I)->getTerminator();
1453         Value *VA = PN.getIncomingValue(I);
1454         if (isKnownNonZero(VA, DL, 0, &AC, CtxI, &DT)) {
1455           if (!NonZeroConst)
1456             NonZeroConst = getAnyNonZeroConstInt(PN);
1457 
1458           if (NonZeroConst != VA) {
1459             replaceOperand(PN, I, NonZeroConst);
1460             MadeChange = true;
1461           }
1462         }
1463       }
1464       if (MadeChange)
1465         return &PN;
1466     }
1467   }
1468 
1469   // We sometimes end up with phi cycles that non-obviously end up being the
1470   // same value, for example:
1471   //   z = some value; x = phi (y, z); y = phi (x, z)
1472   // where the phi nodes don't necessarily need to be in the same block.  Do a
1473   // quick check to see if the PHI node only contains a single non-phi value, if
1474   // so, scan to see if the phi cycle is actually equal to that value.
1475   {
1476     unsigned InValNo = 0, NumIncomingVals = PN.getNumIncomingValues();
1477     // Scan for the first non-phi operand.
1478     while (InValNo != NumIncomingVals &&
1479            isa<PHINode>(PN.getIncomingValue(InValNo)))
1480       ++InValNo;
1481 
1482     if (InValNo != NumIncomingVals) {
1483       Value *NonPhiInVal = PN.getIncomingValue(InValNo);
1484 
1485       // Scan the rest of the operands to see if there are any conflicts, if so
1486       // there is no need to recursively scan other phis.
1487       for (++InValNo; InValNo != NumIncomingVals; ++InValNo) {
1488         Value *OpVal = PN.getIncomingValue(InValNo);
1489         if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
1490           break;
1491       }
1492 
1493       // If we scanned over all operands, then we have one unique value plus
1494       // phi values.  Scan PHI nodes to see if they all merge in each other or
1495       // the value.
1496       if (InValNo == NumIncomingVals) {
1497         SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
1498         if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
1499           return replaceInstUsesWith(PN, NonPhiInVal);
1500       }
1501     }
1502   }
1503 
1504   // If there are multiple PHIs, sort their operands so that they all list
1505   // the blocks in the same order. This will help identical PHIs be eliminated
1506   // by other passes. Other passes shouldn't depend on this for correctness
1507   // however.
1508   PHINode *FirstPN = cast<PHINode>(PN.getParent()->begin());
1509   if (&PN != FirstPN)
1510     for (unsigned I = 0, E = FirstPN->getNumIncomingValues(); I != E; ++I) {
1511       BasicBlock *BBA = PN.getIncomingBlock(I);
1512       BasicBlock *BBB = FirstPN->getIncomingBlock(I);
1513       if (BBA != BBB) {
1514         Value *VA = PN.getIncomingValue(I);
1515         unsigned J = PN.getBasicBlockIndex(BBB);
1516         Value *VB = PN.getIncomingValue(J);
1517         PN.setIncomingBlock(I, BBB);
1518         PN.setIncomingValue(I, VB);
1519         PN.setIncomingBlock(J, BBA);
1520         PN.setIncomingValue(J, VA);
1521         // NOTE: Instcombine normally would want us to "return &PN" if we
1522         // modified any of the operands of an instruction.  However, since we
1523         // aren't adding or removing uses (just rearranging them) we don't do
1524         // this in this case.
1525       }
1526     }
1527 
1528   // Is there an identical PHI node in this basic block?
1529   for (PHINode &IdenticalPN : PN.getParent()->phis()) {
1530     // Ignore the PHI node itself.
1531     if (&IdenticalPN == &PN)
1532       continue;
1533     // Note that even though we've just canonicalized this PHI, due to the
1534     // worklist visitation order, there are no guarantess that *every* PHI
1535     // has been canonicalized, so we can't just compare operands ranges.
1536     if (!PN.isIdenticalToWhenDefined(&IdenticalPN))
1537       continue;
1538     // Just use that PHI instead then.
1539     ++NumPHICSEs;
1540     return replaceInstUsesWith(PN, &IdenticalPN);
1541   }
1542 
1543   // If this is an integer PHI and we know that it has an illegal type, see if
1544   // it is only used by trunc or trunc(lshr) operations.  If so, we split the
1545   // PHI into the various pieces being extracted.  This sort of thing is
1546   // introduced when SROA promotes an aggregate to a single large integer type.
1547   if (PN.getType()->isIntegerTy() &&
1548       !DL.isLegalInteger(PN.getType()->getPrimitiveSizeInBits()))
1549     if (Instruction *Res = SliceUpIllegalIntegerPHI(PN))
1550       return Res;
1551 
1552   // Ultimately, try to replace this Phi with a dominating condition.
1553   if (auto *V = simplifyUsingControlFlow(*this, PN, DT))
1554     return replaceInstUsesWith(PN, V);
1555 
1556   return nullptr;
1557 }
1558