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.
PHIArgMergedDebugLoc(Instruction * Inst,PHINode & PN)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 //
foldIntegerTypedPHI(PHINode & PN)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 // Remove RoundTrip IntToPtr/PtrToInt Cast on PHI-Operand and
303 // fold Phi-operand to bitcast.
foldPHIArgIntToPtrToPHI(PHINode & PN)304 Instruction *InstCombinerImpl::foldPHIArgIntToPtrToPHI(PHINode &PN) {
305   // convert ptr2int ( phi[ int2ptr(ptr2int(x))] ) --> ptr2int ( phi [ x ] )
306   // Make sure all uses of phi are ptr2int.
307   if (!all_of(PN.users(), [](User *U) { return isa<PtrToIntInst>(U); }))
308     return nullptr;
309 
310   // Iterating over all operands to check presence of target pointers for
311   // optimization.
312   bool OperandWithRoundTripCast = false;
313   for (unsigned OpNum = 0; OpNum != PN.getNumIncomingValues(); ++OpNum) {
314     if (auto *NewOp =
315             simplifyIntToPtrRoundTripCast(PN.getIncomingValue(OpNum))) {
316       PN.setIncomingValue(OpNum, NewOp);
317       OperandWithRoundTripCast = true;
318     }
319   }
320   if (!OperandWithRoundTripCast)
321     return nullptr;
322   return &PN;
323 }
324 
325 /// If we have something like phi [insertvalue(a,b,0), insertvalue(c,d,0)],
326 /// turn this into a phi[a,c] and phi[b,d] and a single insertvalue.
327 Instruction *
foldPHIArgInsertValueInstructionIntoPHI(PHINode & PN)328 InstCombinerImpl::foldPHIArgInsertValueInstructionIntoPHI(PHINode &PN) {
329   auto *FirstIVI = cast<InsertValueInst>(PN.getIncomingValue(0));
330 
331   // Scan to see if all operands are `insertvalue`'s with the same indicies,
332   // and all have a single use.
333   for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
334     auto *I = dyn_cast<InsertValueInst>(PN.getIncomingValue(i));
335     if (!I || !I->hasOneUser() || I->getIndices() != FirstIVI->getIndices())
336       return nullptr;
337   }
338 
339   // For each operand of an `insertvalue`
340   std::array<PHINode *, 2> NewOperands;
341   for (int OpIdx : {0, 1}) {
342     auto *&NewOperand = NewOperands[OpIdx];
343     // Create a new PHI node to receive the values the operand has in each
344     // incoming basic block.
345     NewOperand = PHINode::Create(
346         FirstIVI->getOperand(OpIdx)->getType(), PN.getNumIncomingValues(),
347         FirstIVI->getOperand(OpIdx)->getName() + ".pn");
348     // And populate each operand's PHI with said values.
349     for (auto Incoming : zip(PN.blocks(), PN.incoming_values()))
350       NewOperand->addIncoming(
351           cast<InsertValueInst>(std::get<1>(Incoming))->getOperand(OpIdx),
352           std::get<0>(Incoming));
353     InsertNewInstBefore(NewOperand, PN);
354   }
355 
356   // And finally, create `insertvalue` over the newly-formed PHI nodes.
357   auto *NewIVI = InsertValueInst::Create(NewOperands[0], NewOperands[1],
358                                          FirstIVI->getIndices(), PN.getName());
359 
360   PHIArgMergedDebugLoc(NewIVI, PN);
361   ++NumPHIsOfInsertValues;
362   return NewIVI;
363 }
364 
365 /// If we have something like phi [extractvalue(a,0), extractvalue(b,0)],
366 /// turn this into a phi[a,b] and a single extractvalue.
367 Instruction *
foldPHIArgExtractValueInstructionIntoPHI(PHINode & PN)368 InstCombinerImpl::foldPHIArgExtractValueInstructionIntoPHI(PHINode &PN) {
369   auto *FirstEVI = cast<ExtractValueInst>(PN.getIncomingValue(0));
370 
371   // Scan to see if all operands are `extractvalue`'s with the same indicies,
372   // and all have a single use.
373   for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
374     auto *I = dyn_cast<ExtractValueInst>(PN.getIncomingValue(i));
375     if (!I || !I->hasOneUser() || I->getIndices() != FirstEVI->getIndices() ||
376         I->getAggregateOperand()->getType() !=
377             FirstEVI->getAggregateOperand()->getType())
378       return nullptr;
379   }
380 
381   // Create a new PHI node to receive the values the aggregate operand has
382   // in each incoming basic block.
383   auto *NewAggregateOperand = PHINode::Create(
384       FirstEVI->getAggregateOperand()->getType(), PN.getNumIncomingValues(),
385       FirstEVI->getAggregateOperand()->getName() + ".pn");
386   // And populate the PHI with said values.
387   for (auto Incoming : zip(PN.blocks(), PN.incoming_values()))
388     NewAggregateOperand->addIncoming(
389         cast<ExtractValueInst>(std::get<1>(Incoming))->getAggregateOperand(),
390         std::get<0>(Incoming));
391   InsertNewInstBefore(NewAggregateOperand, PN);
392 
393   // And finally, create `extractvalue` over the newly-formed PHI nodes.
394   auto *NewEVI = ExtractValueInst::Create(NewAggregateOperand,
395                                           FirstEVI->getIndices(), PN.getName());
396 
397   PHIArgMergedDebugLoc(NewEVI, PN);
398   ++NumPHIsOfExtractValues;
399   return NewEVI;
400 }
401 
402 /// If we have something like phi [add (a,b), add(a,c)] and if a/b/c and the
403 /// adds all have a single user, turn this into a phi and a single binop.
foldPHIArgBinOpIntoPHI(PHINode & PN)404 Instruction *InstCombinerImpl::foldPHIArgBinOpIntoPHI(PHINode &PN) {
405   Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
406   assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
407   unsigned Opc = FirstInst->getOpcode();
408   Value *LHSVal = FirstInst->getOperand(0);
409   Value *RHSVal = FirstInst->getOperand(1);
410 
411   Type *LHSType = LHSVal->getType();
412   Type *RHSType = RHSVal->getType();
413 
414   // Scan to see if all operands are the same opcode, and all have one user.
415   for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
416     Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
417     if (!I || I->getOpcode() != Opc || !I->hasOneUser() ||
418         // Verify type of the LHS matches so we don't fold cmp's of different
419         // types.
420         I->getOperand(0)->getType() != LHSType ||
421         I->getOperand(1)->getType() != RHSType)
422       return nullptr;
423 
424     // If they are CmpInst instructions, check their predicates
425     if (CmpInst *CI = dyn_cast<CmpInst>(I))
426       if (CI->getPredicate() != cast<CmpInst>(FirstInst)->getPredicate())
427         return nullptr;
428 
429     // Keep track of which operand needs a phi node.
430     if (I->getOperand(0) != LHSVal) LHSVal = nullptr;
431     if (I->getOperand(1) != RHSVal) RHSVal = nullptr;
432   }
433 
434   // If both LHS and RHS would need a PHI, don't do this transformation,
435   // because it would increase the number of PHIs entering the block,
436   // which leads to higher register pressure. This is especially
437   // bad when the PHIs are in the header of a loop.
438   if (!LHSVal && !RHSVal)
439     return nullptr;
440 
441   // Otherwise, this is safe to transform!
442 
443   Value *InLHS = FirstInst->getOperand(0);
444   Value *InRHS = FirstInst->getOperand(1);
445   PHINode *NewLHS = nullptr, *NewRHS = nullptr;
446   if (!LHSVal) {
447     NewLHS = PHINode::Create(LHSType, PN.getNumIncomingValues(),
448                              FirstInst->getOperand(0)->getName() + ".pn");
449     NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
450     InsertNewInstBefore(NewLHS, PN);
451     LHSVal = NewLHS;
452   }
453 
454   if (!RHSVal) {
455     NewRHS = PHINode::Create(RHSType, PN.getNumIncomingValues(),
456                              FirstInst->getOperand(1)->getName() + ".pn");
457     NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
458     InsertNewInstBefore(NewRHS, PN);
459     RHSVal = NewRHS;
460   }
461 
462   // Add all operands to the new PHIs.
463   if (NewLHS || NewRHS) {
464     for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
465       Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
466       if (NewLHS) {
467         Value *NewInLHS = InInst->getOperand(0);
468         NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
469       }
470       if (NewRHS) {
471         Value *NewInRHS = InInst->getOperand(1);
472         NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
473       }
474     }
475   }
476 
477   if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst)) {
478     CmpInst *NewCI = CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
479                                      LHSVal, RHSVal);
480     PHIArgMergedDebugLoc(NewCI, PN);
481     return NewCI;
482   }
483 
484   BinaryOperator *BinOp = cast<BinaryOperator>(FirstInst);
485   BinaryOperator *NewBinOp =
486     BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
487 
488   NewBinOp->copyIRFlags(PN.getIncomingValue(0));
489 
490   for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i)
491     NewBinOp->andIRFlags(PN.getIncomingValue(i));
492 
493   PHIArgMergedDebugLoc(NewBinOp, PN);
494   return NewBinOp;
495 }
496 
foldPHIArgGEPIntoPHI(PHINode & PN)497 Instruction *InstCombinerImpl::foldPHIArgGEPIntoPHI(PHINode &PN) {
498   GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
499 
500   SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
501                                         FirstInst->op_end());
502   // This is true if all GEP bases are allocas and if all indices into them are
503   // constants.
504   bool AllBasePointersAreAllocas = true;
505 
506   // We don't want to replace this phi if the replacement would require
507   // more than one phi, which leads to higher register pressure. This is
508   // especially bad when the PHIs are in the header of a loop.
509   bool NeededPhi = false;
510 
511   bool AllInBounds = true;
512 
513   // Scan to see if all operands are the same opcode, and all have one user.
514   for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
515     GetElementPtrInst *GEP =
516         dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
517     if (!GEP || !GEP->hasOneUser() || GEP->getType() != FirstInst->getType() ||
518         GEP->getNumOperands() != FirstInst->getNumOperands())
519       return nullptr;
520 
521     AllInBounds &= GEP->isInBounds();
522 
523     // Keep track of whether or not all GEPs are of alloca pointers.
524     if (AllBasePointersAreAllocas &&
525         (!isa<AllocaInst>(GEP->getOperand(0)) ||
526          !GEP->hasAllConstantIndices()))
527       AllBasePointersAreAllocas = false;
528 
529     // Compare the operand lists.
530     for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
531       if (FirstInst->getOperand(op) == GEP->getOperand(op))
532         continue;
533 
534       // Don't merge two GEPs when two operands differ (introducing phi nodes)
535       // if one of the PHIs has a constant for the index.  The index may be
536       // substantially cheaper to compute for the constants, so making it a
537       // variable index could pessimize the path.  This also handles the case
538       // for struct indices, which must always be constant.
539       if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
540           isa<ConstantInt>(GEP->getOperand(op)))
541         return nullptr;
542 
543       if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
544         return nullptr;
545 
546       // If we already needed a PHI for an earlier operand, and another operand
547       // also requires a PHI, we'd be introducing more PHIs than we're
548       // eliminating, which increases register pressure on entry to the PHI's
549       // block.
550       if (NeededPhi)
551         return nullptr;
552 
553       FixedOperands[op] = nullptr;  // Needs a PHI.
554       NeededPhi = true;
555     }
556   }
557 
558   // If all of the base pointers of the PHI'd GEPs are from allocas, don't
559   // bother doing this transformation.  At best, this will just save a bit of
560   // offset calculation, but all the predecessors will have to materialize the
561   // stack address into a register anyway.  We'd actually rather *clone* the
562   // load up into the predecessors so that we have a load of a gep of an alloca,
563   // which can usually all be folded into the load.
564   if (AllBasePointersAreAllocas)
565     return nullptr;
566 
567   // Otherwise, this is safe to transform.  Insert PHI nodes for each operand
568   // that is variable.
569   SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
570 
571   bool HasAnyPHIs = false;
572   for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
573     if (FixedOperands[i]) continue;  // operand doesn't need a phi.
574     Value *FirstOp = FirstInst->getOperand(i);
575     PHINode *NewPN = PHINode::Create(FirstOp->getType(), e,
576                                      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 
586   // Add all operands to the new PHIs.
587   if (HasAnyPHIs) {
588     for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
589       GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
590       BasicBlock *InBB = PN.getIncomingBlock(i);
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.
isSafeAndProfitableToSinkLoad(LoadInst * L)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 
foldPHIArgLoadIntoPHI(PHINode & PN)658 Instruction *InstCombinerImpl::foldPHIArgLoadIntoPHI(PHINode &PN) {
659   LoadInst *FirstLI = cast<LoadInst>(PN.getIncomingValue(0));
660 
661   // FIXME: This is overconservative; this transform is allowed in some cases
662   // for atomic operations.
663   if (FirstLI->isAtomic())
664     return nullptr;
665 
666   // When processing loads, we need to propagate two bits of information to the
667   // sunk load: whether it is volatile, and what its alignment is.  We currently
668   // don't sink loads when some have their alignment specified and some don't.
669   // visitLoadInst will propagate an alignment onto the load when TD is around,
670   // and if TD isn't around, we can't handle the mixed case.
671   bool isVolatile = FirstLI->isVolatile();
672   Align LoadAlignment = FirstLI->getAlign();
673   unsigned LoadAddrSpace = FirstLI->getPointerAddressSpace();
674 
675   // We can't sink the load if the loaded value could be modified between the
676   // load and the PHI.
677   if (FirstLI->getParent() != PN.getIncomingBlock(0) ||
678       !isSafeAndProfitableToSinkLoad(FirstLI))
679     return nullptr;
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       FirstLI->getParent()->getTerminator()->getNumSuccessors() != 1)
686     return nullptr;
687 
688   // Check to see if all arguments are the same operation.
689   for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
690     LoadInst *LI = dyn_cast<LoadInst>(PN.getIncomingValue(i));
691     if (!LI || !LI->hasOneUser())
692       return nullptr;
693 
694     // We can't sink the load if the loaded value could be modified between
695     // the load and the PHI.
696     if (LI->isVolatile() != isVolatile ||
697         LI->getParent() != PN.getIncomingBlock(i) ||
698         LI->getPointerAddressSpace() != LoadAddrSpace ||
699         !isSafeAndProfitableToSinkLoad(LI))
700       return nullptr;
701 
702     LoadAlignment = std::min(LoadAlignment, Align(LI->getAlign()));
703 
704     // If the PHI is of volatile loads and the load block has multiple
705     // successors, sinking it would remove a load of the volatile value from
706     // the path through the other successor.
707     if (isVolatile &&
708         LI->getParent()->getTerminator()->getNumSuccessors() != 1)
709       return nullptr;
710   }
711 
712   // Okay, they are all the same operation.  Create a new PHI node of the
713   // correct type, and PHI together all of the LHS's of the instructions.
714   PHINode *NewPN = PHINode::Create(FirstLI->getOperand(0)->getType(),
715                                    PN.getNumIncomingValues(),
716                                    PN.getName()+".in");
717 
718   Value *InVal = FirstLI->getOperand(0);
719   NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
720   LoadInst *NewLI =
721       new LoadInst(FirstLI->getType(), NewPN, "", isVolatile, LoadAlignment);
722 
723   unsigned KnownIDs[] = {
724     LLVMContext::MD_tbaa,
725     LLVMContext::MD_range,
726     LLVMContext::MD_invariant_load,
727     LLVMContext::MD_alias_scope,
728     LLVMContext::MD_noalias,
729     LLVMContext::MD_nonnull,
730     LLVMContext::MD_align,
731     LLVMContext::MD_dereferenceable,
732     LLVMContext::MD_dereferenceable_or_null,
733     LLVMContext::MD_access_group,
734   };
735 
736   for (unsigned ID : KnownIDs)
737     NewLI->setMetadata(ID, FirstLI->getMetadata(ID));
738 
739   // Add all operands to the new PHI and combine TBAA metadata.
740   for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
741     LoadInst *LI = cast<LoadInst>(PN.getIncomingValue(i));
742     combineMetadata(NewLI, LI, KnownIDs, true);
743     Value *NewInVal = LI->getOperand(0);
744     if (NewInVal != InVal)
745       InVal = nullptr;
746     NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
747   }
748 
749   if (InVal) {
750     // The new PHI unions all of the same values together.  This is really
751     // common, so we handle it intelligently here for compile-time speed.
752     NewLI->setOperand(0, InVal);
753     delete NewPN;
754   } else {
755     InsertNewInstBefore(NewPN, PN);
756   }
757 
758   // If this was a volatile load that we are merging, make sure to loop through
759   // and mark all the input loads as non-volatile.  If we don't do this, we will
760   // insert a new volatile load and the old ones will not be deletable.
761   if (isVolatile)
762     for (Value *IncValue : PN.incoming_values())
763       cast<LoadInst>(IncValue)->setVolatile(false);
764 
765   PHIArgMergedDebugLoc(NewLI, PN);
766   return NewLI;
767 }
768 
769 /// TODO: This function could handle other cast types, but then it might
770 /// require special-casing a cast from the 'i1' type. See the comment in
771 /// FoldPHIArgOpIntoPHI() about pessimizing illegal integer types.
foldPHIArgZextsIntoPHI(PHINode & Phi)772 Instruction *InstCombinerImpl::foldPHIArgZextsIntoPHI(PHINode &Phi) {
773   // We cannot create a new instruction after the PHI if the terminator is an
774   // EHPad because there is no valid insertion point.
775   if (Instruction *TI = Phi.getParent()->getTerminator())
776     if (TI->isEHPad())
777       return nullptr;
778 
779   // Early exit for the common case of a phi with two operands. These are
780   // handled elsewhere. See the comment below where we check the count of zexts
781   // and constants for more details.
782   unsigned NumIncomingValues = Phi.getNumIncomingValues();
783   if (NumIncomingValues < 3)
784     return nullptr;
785 
786   // Find the narrower type specified by the first zext.
787   Type *NarrowType = nullptr;
788   for (Value *V : Phi.incoming_values()) {
789     if (auto *Zext = dyn_cast<ZExtInst>(V)) {
790       NarrowType = Zext->getSrcTy();
791       break;
792     }
793   }
794   if (!NarrowType)
795     return nullptr;
796 
797   // Walk the phi operands checking that we only have zexts or constants that
798   // we can shrink for free. Store the new operands for the new phi.
799   SmallVector<Value *, 4> NewIncoming;
800   unsigned NumZexts = 0;
801   unsigned NumConsts = 0;
802   for (Value *V : Phi.incoming_values()) {
803     if (auto *Zext = dyn_cast<ZExtInst>(V)) {
804       // All zexts must be identical and have one user.
805       if (Zext->getSrcTy() != NarrowType || !Zext->hasOneUser())
806         return nullptr;
807       NewIncoming.push_back(Zext->getOperand(0));
808       NumZexts++;
809     } else if (auto *C = dyn_cast<Constant>(V)) {
810       // Make sure that constants can fit in the new type.
811       Constant *Trunc = ConstantExpr::getTrunc(C, NarrowType);
812       if (ConstantExpr::getZExt(Trunc, C->getType()) != C)
813         return nullptr;
814       NewIncoming.push_back(Trunc);
815       NumConsts++;
816     } else {
817       // If it's not a cast or a constant, bail out.
818       return nullptr;
819     }
820   }
821 
822   // The more common cases of a phi with no constant operands or just one
823   // variable operand are handled by FoldPHIArgOpIntoPHI() and foldOpIntoPhi()
824   // respectively. foldOpIntoPhi() wants to do the opposite transform that is
825   // performed here. It tries to replicate a cast in the phi operand's basic
826   // block to expose other folding opportunities. Thus, InstCombine will
827   // infinite loop without this check.
828   if (NumConsts == 0 || NumZexts < 2)
829     return nullptr;
830 
831   // All incoming values are zexts or constants that are safe to truncate.
832   // Create a new phi node of the narrow type, phi together all of the new
833   // operands, and zext the result back to the original type.
834   PHINode *NewPhi = PHINode::Create(NarrowType, NumIncomingValues,
835                                     Phi.getName() + ".shrunk");
836   for (unsigned i = 0; i != NumIncomingValues; ++i)
837     NewPhi->addIncoming(NewIncoming[i], Phi.getIncomingBlock(i));
838 
839   InsertNewInstBefore(NewPhi, Phi);
840   return CastInst::CreateZExtOrBitCast(NewPhi, Phi.getType());
841 }
842 
843 /// If all operands to a PHI node are the same "unary" operator and they all are
844 /// only used by the PHI, PHI together their inputs, and do the operation once,
845 /// to the result of the PHI.
foldPHIArgOpIntoPHI(PHINode & PN)846 Instruction *InstCombinerImpl::foldPHIArgOpIntoPHI(PHINode &PN) {
847   // We cannot create a new instruction after the PHI if the terminator is an
848   // EHPad because there is no valid insertion point.
849   if (Instruction *TI = PN.getParent()->getTerminator())
850     if (TI->isEHPad())
851       return nullptr;
852 
853   Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
854 
855   if (isa<GetElementPtrInst>(FirstInst))
856     return foldPHIArgGEPIntoPHI(PN);
857   if (isa<LoadInst>(FirstInst))
858     return foldPHIArgLoadIntoPHI(PN);
859   if (isa<InsertValueInst>(FirstInst))
860     return foldPHIArgInsertValueInstructionIntoPHI(PN);
861   if (isa<ExtractValueInst>(FirstInst))
862     return foldPHIArgExtractValueInstructionIntoPHI(PN);
863 
864   // Scan the instruction, looking for input operations that can be folded away.
865   // If all input operands to the phi are the same instruction (e.g. a cast from
866   // the same type or "+42") we can pull the operation through the PHI, reducing
867   // code size and simplifying code.
868   Constant *ConstantOp = nullptr;
869   Type *CastSrcTy = nullptr;
870 
871   if (isa<CastInst>(FirstInst)) {
872     CastSrcTy = FirstInst->getOperand(0)->getType();
873 
874     // Be careful about transforming integer PHIs.  We don't want to pessimize
875     // the code by turning an i32 into an i1293.
876     if (PN.getType()->isIntegerTy() && CastSrcTy->isIntegerTy()) {
877       if (!shouldChangeType(PN.getType(), CastSrcTy))
878         return nullptr;
879     }
880   } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
881     // Can fold binop, compare or shift here if the RHS is a constant,
882     // otherwise call FoldPHIArgBinOpIntoPHI.
883     ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
884     if (!ConstantOp)
885       return foldPHIArgBinOpIntoPHI(PN);
886   } else {
887     return nullptr;  // Cannot fold this operation.
888   }
889 
890   // Check to see if all arguments are the same operation.
891   for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
892     Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
893     if (!I || !I->hasOneUser() || !I->isSameOperationAs(FirstInst))
894       return nullptr;
895     if (CastSrcTy) {
896       if (I->getOperand(0)->getType() != CastSrcTy)
897         return nullptr;  // Cast operation must match.
898     } else if (I->getOperand(1) != ConstantOp) {
899       return nullptr;
900     }
901   }
902 
903   // Okay, they are all the same operation.  Create a new PHI node of the
904   // correct type, and PHI together all of the LHS's of the instructions.
905   PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
906                                    PN.getNumIncomingValues(),
907                                    PN.getName()+".in");
908 
909   Value *InVal = FirstInst->getOperand(0);
910   NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
911 
912   // Add all operands to the new PHI.
913   for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
914     Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
915     if (NewInVal != InVal)
916       InVal = nullptr;
917     NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
918   }
919 
920   Value *PhiVal;
921   if (InVal) {
922     // The new PHI unions all of the same values together.  This is really
923     // common, so we handle it intelligently here for compile-time speed.
924     PhiVal = InVal;
925     delete NewPN;
926   } else {
927     InsertNewInstBefore(NewPN, PN);
928     PhiVal = NewPN;
929   }
930 
931   // Insert and return the new operation.
932   if (CastInst *FirstCI = dyn_cast<CastInst>(FirstInst)) {
933     CastInst *NewCI = CastInst::Create(FirstCI->getOpcode(), PhiVal,
934                                        PN.getType());
935     PHIArgMergedDebugLoc(NewCI, PN);
936     return NewCI;
937   }
938 
939   if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst)) {
940     BinOp = BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
941     BinOp->copyIRFlags(PN.getIncomingValue(0));
942 
943     for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i)
944       BinOp->andIRFlags(PN.getIncomingValue(i));
945 
946     PHIArgMergedDebugLoc(BinOp, PN);
947     return BinOp;
948   }
949 
950   CmpInst *CIOp = cast<CmpInst>(FirstInst);
951   CmpInst *NewCI = CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
952                                    PhiVal, ConstantOp);
953   PHIArgMergedDebugLoc(NewCI, PN);
954   return NewCI;
955 }
956 
957 /// Return true if this PHI node is only used by a PHI node cycle that is dead.
DeadPHICycle(PHINode * PN,SmallPtrSetImpl<PHINode * > & PotentiallyDeadPHIs)958 static bool DeadPHICycle(PHINode *PN,
959                          SmallPtrSetImpl<PHINode*> &PotentiallyDeadPHIs) {
960   if (PN->use_empty()) return true;
961   if (!PN->hasOneUse()) return false;
962 
963   // Remember this node, and if we find the cycle, return.
964   if (!PotentiallyDeadPHIs.insert(PN).second)
965     return true;
966 
967   // Don't scan crazily complex things.
968   if (PotentiallyDeadPHIs.size() == 16)
969     return false;
970 
971   if (PHINode *PU = dyn_cast<PHINode>(PN->user_back()))
972     return DeadPHICycle(PU, PotentiallyDeadPHIs);
973 
974   return false;
975 }
976 
977 /// Return true if this phi node is always equal to NonPhiInVal.
978 /// This happens with mutually cyclic phi nodes like:
979 ///   z = some value; x = phi (y, z); y = phi (x, z)
PHIsEqualValue(PHINode * PN,Value * NonPhiInVal,SmallPtrSetImpl<PHINode * > & ValueEqualPHIs)980 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
981                            SmallPtrSetImpl<PHINode*> &ValueEqualPHIs) {
982   // See if we already saw this PHI node.
983   if (!ValueEqualPHIs.insert(PN).second)
984     return true;
985 
986   // Don't scan crazily complex things.
987   if (ValueEqualPHIs.size() == 16)
988     return false;
989 
990   // Scan the operands to see if they are either phi nodes or are equal to
991   // the value.
992   for (Value *Op : PN->incoming_values()) {
993     if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
994       if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
995         return false;
996     } else if (Op != NonPhiInVal)
997       return false;
998   }
999 
1000   return true;
1001 }
1002 
1003 /// Return an existing non-zero constant if this phi node has one, otherwise
1004 /// return constant 1.
GetAnyNonZeroConstInt(PHINode & PN)1005 static ConstantInt *GetAnyNonZeroConstInt(PHINode &PN) {
1006   assert(isa<IntegerType>(PN.getType()) && "Expect only integer type phi");
1007   for (Value *V : PN.operands())
1008     if (auto *ConstVA = dyn_cast<ConstantInt>(V))
1009       if (!ConstVA->isZero())
1010         return ConstVA;
1011   return ConstantInt::get(cast<IntegerType>(PN.getType()), 1);
1012 }
1013 
1014 namespace {
1015 struct PHIUsageRecord {
1016   unsigned PHIId;     // The ID # of the PHI (something determinstic to sort on)
1017   unsigned Shift;     // The amount shifted.
1018   Instruction *Inst;  // The trunc instruction.
1019 
PHIUsageRecord__anona2082b2d0511::PHIUsageRecord1020   PHIUsageRecord(unsigned pn, unsigned Sh, Instruction *User)
1021     : PHIId(pn), Shift(Sh), Inst(User) {}
1022 
operator <__anona2082b2d0511::PHIUsageRecord1023   bool operator<(const PHIUsageRecord &RHS) const {
1024     if (PHIId < RHS.PHIId) return true;
1025     if (PHIId > RHS.PHIId) return false;
1026     if (Shift < RHS.Shift) return true;
1027     if (Shift > RHS.Shift) return false;
1028     return Inst->getType()->getPrimitiveSizeInBits() <
1029            RHS.Inst->getType()->getPrimitiveSizeInBits();
1030   }
1031 };
1032 
1033 struct LoweredPHIRecord {
1034   PHINode *PN;        // The PHI that was lowered.
1035   unsigned Shift;     // The amount shifted.
1036   unsigned Width;     // The width extracted.
1037 
LoweredPHIRecord__anona2082b2d0511::LoweredPHIRecord1038   LoweredPHIRecord(PHINode *pn, unsigned Sh, Type *Ty)
1039     : PN(pn), Shift(Sh), Width(Ty->getPrimitiveSizeInBits()) {}
1040 
1041   // Ctor form used by DenseMap.
LoweredPHIRecord__anona2082b2d0511::LoweredPHIRecord1042   LoweredPHIRecord(PHINode *pn, unsigned Sh)
1043     : PN(pn), Shift(Sh), Width(0) {}
1044 };
1045 } // namespace
1046 
1047 namespace llvm {
1048   template<>
1049   struct DenseMapInfo<LoweredPHIRecord> {
getEmptyKeyllvm::DenseMapInfo1050     static inline LoweredPHIRecord getEmptyKey() {
1051       return LoweredPHIRecord(nullptr, 0);
1052     }
getTombstoneKeyllvm::DenseMapInfo1053     static inline LoweredPHIRecord getTombstoneKey() {
1054       return LoweredPHIRecord(nullptr, 1);
1055     }
getHashValuellvm::DenseMapInfo1056     static unsigned getHashValue(const LoweredPHIRecord &Val) {
1057       return DenseMapInfo<PHINode*>::getHashValue(Val.PN) ^ (Val.Shift>>3) ^
1058              (Val.Width>>3);
1059     }
isEqualllvm::DenseMapInfo1060     static bool isEqual(const LoweredPHIRecord &LHS,
1061                         const LoweredPHIRecord &RHS) {
1062       return LHS.PN == RHS.PN && LHS.Shift == RHS.Shift &&
1063              LHS.Width == RHS.Width;
1064     }
1065   };
1066 } // namespace llvm
1067 
1068 
1069 /// This is an integer PHI and we know that it has an illegal type: see if it is
1070 /// only used by trunc or trunc(lshr) operations. If so, we split the PHI into
1071 /// the various pieces being extracted. This sort of thing is introduced when
1072 /// SROA promotes an aggregate to large integer values.
1073 ///
1074 /// TODO: The user of the trunc may be an bitcast to float/double/vector or an
1075 /// inttoptr.  We should produce new PHIs in the right type.
1076 ///
SliceUpIllegalIntegerPHI(PHINode & FirstPhi)1077 Instruction *InstCombinerImpl::SliceUpIllegalIntegerPHI(PHINode &FirstPhi) {
1078   // PHIUsers - Keep track of all of the truncated values extracted from a set
1079   // of PHIs, along with their offset.  These are the things we want to rewrite.
1080   SmallVector<PHIUsageRecord, 16> PHIUsers;
1081 
1082   // PHIs are often mutually cyclic, so we keep track of a whole set of PHI
1083   // nodes which are extracted from. PHIsToSlice is a set we use to avoid
1084   // revisiting PHIs, PHIsInspected is a ordered list of PHIs that we need to
1085   // check the uses of (to ensure they are all extracts).
1086   SmallVector<PHINode*, 8> PHIsToSlice;
1087   SmallPtrSet<PHINode*, 8> PHIsInspected;
1088 
1089   PHIsToSlice.push_back(&FirstPhi);
1090   PHIsInspected.insert(&FirstPhi);
1091 
1092   for (unsigned PHIId = 0; PHIId != PHIsToSlice.size(); ++PHIId) {
1093     PHINode *PN = PHIsToSlice[PHIId];
1094 
1095     // Scan the input list of the PHI.  If any input is an invoke, and if the
1096     // input is defined in the predecessor, then we won't be split the critical
1097     // edge which is required to insert a truncate.  Because of this, we have to
1098     // bail out.
1099     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1100       InvokeInst *II = dyn_cast<InvokeInst>(PN->getIncomingValue(i));
1101       if (!II) continue;
1102       if (II->getParent() != PN->getIncomingBlock(i))
1103         continue;
1104 
1105       // If we have a phi, and if it's directly in the predecessor, then we have
1106       // a critical edge where we need to put the truncate.  Since we can't
1107       // split the edge in instcombine, we have to bail out.
1108       return nullptr;
1109     }
1110 
1111     for (User *U : PN->users()) {
1112       Instruction *UserI = cast<Instruction>(U);
1113 
1114       // If the user is a PHI, inspect its uses recursively.
1115       if (PHINode *UserPN = dyn_cast<PHINode>(UserI)) {
1116         if (PHIsInspected.insert(UserPN).second)
1117           PHIsToSlice.push_back(UserPN);
1118         continue;
1119       }
1120 
1121       // Truncates are always ok.
1122       if (isa<TruncInst>(UserI)) {
1123         PHIUsers.push_back(PHIUsageRecord(PHIId, 0, UserI));
1124         continue;
1125       }
1126 
1127       // Otherwise it must be a lshr which can only be used by one trunc.
1128       if (UserI->getOpcode() != Instruction::LShr ||
1129           !UserI->hasOneUse() || !isa<TruncInst>(UserI->user_back()) ||
1130           !isa<ConstantInt>(UserI->getOperand(1)))
1131         return nullptr;
1132 
1133       // Bail on out of range shifts.
1134       unsigned SizeInBits = UserI->getType()->getScalarSizeInBits();
1135       if (cast<ConstantInt>(UserI->getOperand(1))->getValue().uge(SizeInBits))
1136         return nullptr;
1137 
1138       unsigned Shift = cast<ConstantInt>(UserI->getOperand(1))->getZExtValue();
1139       PHIUsers.push_back(PHIUsageRecord(PHIId, Shift, UserI->user_back()));
1140     }
1141   }
1142 
1143   // If we have no users, they must be all self uses, just nuke the PHI.
1144   if (PHIUsers.empty())
1145     return replaceInstUsesWith(FirstPhi, PoisonValue::get(FirstPhi.getType()));
1146 
1147   // If this phi node is transformable, create new PHIs for all the pieces
1148   // extracted out of it.  First, sort the users by their offset and size.
1149   array_pod_sort(PHIUsers.begin(), PHIUsers.end());
1150 
1151   LLVM_DEBUG(dbgs() << "SLICING UP PHI: " << FirstPhi << '\n';
1152              for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i) dbgs()
1153              << "AND USER PHI #" << i << ": " << *PHIsToSlice[i] << '\n';);
1154 
1155   // PredValues - This is a temporary used when rewriting PHI nodes.  It is
1156   // hoisted out here to avoid construction/destruction thrashing.
1157   DenseMap<BasicBlock*, Value*> PredValues;
1158 
1159   // ExtractedVals - Each new PHI we introduce is saved here so we don't
1160   // introduce redundant PHIs.
1161   DenseMap<LoweredPHIRecord, PHINode*> ExtractedVals;
1162 
1163   for (unsigned UserI = 0, UserE = PHIUsers.size(); UserI != UserE; ++UserI) {
1164     unsigned PHIId = PHIUsers[UserI].PHIId;
1165     PHINode *PN = PHIsToSlice[PHIId];
1166     unsigned Offset = PHIUsers[UserI].Shift;
1167     Type *Ty = PHIUsers[UserI].Inst->getType();
1168 
1169     PHINode *EltPHI;
1170 
1171     // If we've already lowered a user like this, reuse the previously lowered
1172     // value.
1173     if ((EltPHI = ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)]) == nullptr) {
1174 
1175       // Otherwise, Create the new PHI node for this user.
1176       EltPHI = PHINode::Create(Ty, PN->getNumIncomingValues(),
1177                                PN->getName()+".off"+Twine(Offset), PN);
1178       assert(EltPHI->getType() != PN->getType() &&
1179              "Truncate didn't shrink phi?");
1180 
1181       for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1182         BasicBlock *Pred = PN->getIncomingBlock(i);
1183         Value *&PredVal = PredValues[Pred];
1184 
1185         // If we already have a value for this predecessor, reuse it.
1186         if (PredVal) {
1187           EltPHI->addIncoming(PredVal, Pred);
1188           continue;
1189         }
1190 
1191         // Handle the PHI self-reuse case.
1192         Value *InVal = PN->getIncomingValue(i);
1193         if (InVal == PN) {
1194           PredVal = EltPHI;
1195           EltPHI->addIncoming(PredVal, Pred);
1196           continue;
1197         }
1198 
1199         if (PHINode *InPHI = dyn_cast<PHINode>(PN)) {
1200           // If the incoming value was a PHI, and if it was one of the PHIs we
1201           // already rewrote it, just use the lowered value.
1202           if (Value *Res = ExtractedVals[LoweredPHIRecord(InPHI, Offset, Ty)]) {
1203             PredVal = Res;
1204             EltPHI->addIncoming(PredVal, Pred);
1205             continue;
1206           }
1207         }
1208 
1209         // Otherwise, do an extract in the predecessor.
1210         Builder.SetInsertPoint(Pred->getTerminator());
1211         Value *Res = InVal;
1212         if (Offset)
1213           Res = Builder.CreateLShr(Res, ConstantInt::get(InVal->getType(),
1214                                                           Offset), "extract");
1215         Res = Builder.CreateTrunc(Res, Ty, "extract.t");
1216         PredVal = Res;
1217         EltPHI->addIncoming(Res, Pred);
1218 
1219         // If the incoming value was a PHI, and if it was one of the PHIs we are
1220         // rewriting, we will ultimately delete the code we inserted.  This
1221         // means we need to revisit that PHI to make sure we extract out the
1222         // needed piece.
1223         if (PHINode *OldInVal = dyn_cast<PHINode>(PN->getIncomingValue(i)))
1224           if (PHIsInspected.count(OldInVal)) {
1225             unsigned RefPHIId =
1226                 find(PHIsToSlice, OldInVal) - PHIsToSlice.begin();
1227             PHIUsers.push_back(PHIUsageRecord(RefPHIId, Offset,
1228                                               cast<Instruction>(Res)));
1229             ++UserE;
1230           }
1231       }
1232       PredValues.clear();
1233 
1234       LLVM_DEBUG(dbgs() << "  Made element PHI for offset " << Offset << ": "
1235                         << *EltPHI << '\n');
1236       ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)] = EltPHI;
1237     }
1238 
1239     // Replace the use of this piece with the PHI node.
1240     replaceInstUsesWith(*PHIUsers[UserI].Inst, EltPHI);
1241   }
1242 
1243   // Replace all the remaining uses of the PHI nodes (self uses and the lshrs)
1244   // with poison.
1245   Value *Poison = PoisonValue::get(FirstPhi.getType());
1246   for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i)
1247     replaceInstUsesWith(*PHIsToSlice[i], Poison);
1248   return replaceInstUsesWith(FirstPhi, Poison);
1249 }
1250 
SimplifyUsingControlFlow(InstCombiner & Self,PHINode & PN,const DominatorTree & DT)1251 static Value *SimplifyUsingControlFlow(InstCombiner &Self, PHINode &PN,
1252                                        const DominatorTree &DT) {
1253   // Simplify the following patterns:
1254   //       if (cond)
1255   //       /       \
1256   //      ...      ...
1257   //       \       /
1258   //    phi [true] [false]
1259   if (!PN.getType()->isIntegerTy(1))
1260     return nullptr;
1261 
1262   if (PN.getNumOperands() != 2)
1263     return nullptr;
1264 
1265   // Make sure all inputs are constants.
1266   if (!all_of(PN.operands(), [](Value *V) { return isa<ConstantInt>(V); }))
1267     return nullptr;
1268 
1269   BasicBlock *BB = PN.getParent();
1270   // Do not bother with unreachable instructions.
1271   if (!DT.isReachableFromEntry(BB))
1272     return nullptr;
1273 
1274   // Same inputs.
1275   if (PN.getOperand(0) == PN.getOperand(1))
1276     return PN.getOperand(0);
1277 
1278   BasicBlock *TruePred = nullptr, *FalsePred = nullptr;
1279   for (auto *Pred : predecessors(BB)) {
1280     auto *Input = cast<ConstantInt>(PN.getIncomingValueForBlock(Pred));
1281     if (Input->isAllOnesValue())
1282       TruePred = Pred;
1283     else
1284       FalsePred = Pred;
1285   }
1286   assert(TruePred && FalsePred && "Must be!");
1287 
1288   // Check which edge of the dominator dominates the true input. If it is the
1289   // false edge, we should invert the condition.
1290   auto *IDom = DT.getNode(BB)->getIDom()->getBlock();
1291   auto *BI = dyn_cast<BranchInst>(IDom->getTerminator());
1292   if (!BI || BI->isUnconditional())
1293     return nullptr;
1294 
1295   // Check that edges outgoing from the idom's terminators dominate respective
1296   // inputs of the Phi.
1297   BasicBlockEdge TrueOutEdge(IDom, BI->getSuccessor(0));
1298   BasicBlockEdge FalseOutEdge(IDom, BI->getSuccessor(1));
1299 
1300   BasicBlockEdge TrueIncEdge(TruePred, BB);
1301   BasicBlockEdge FalseIncEdge(FalsePred, BB);
1302 
1303   auto *Cond = BI->getCondition();
1304   if (DT.dominates(TrueOutEdge, TrueIncEdge) &&
1305       DT.dominates(FalseOutEdge, FalseIncEdge))
1306     // This Phi is actually equivalent to branching condition of IDom.
1307     return Cond;
1308   else if (DT.dominates(TrueOutEdge, FalseIncEdge) &&
1309            DT.dominates(FalseOutEdge, TrueIncEdge)) {
1310     // This Phi is actually opposite to branching condition of IDom. We invert
1311     // the condition that will potentially open up some opportunities for
1312     // sinking.
1313     auto InsertPt = BB->getFirstInsertionPt();
1314     if (InsertPt != BB->end()) {
1315       Self.Builder.SetInsertPoint(&*InsertPt);
1316       return Self.Builder.CreateNot(Cond);
1317     }
1318   }
1319 
1320   return nullptr;
1321 }
1322 
1323 // PHINode simplification
1324 //
visitPHINode(PHINode & PN)1325 Instruction *InstCombinerImpl::visitPHINode(PHINode &PN) {
1326   if (Value *V = SimplifyInstruction(&PN, SQ.getWithInstruction(&PN)))
1327     return replaceInstUsesWith(PN, V);
1328 
1329   if (Instruction *Result = foldPHIArgZextsIntoPHI(PN))
1330     return Result;
1331 
1332   if (Instruction *Result = foldPHIArgIntToPtrToPHI(PN))
1333     return Result;
1334 
1335   // If all PHI operands are the same operation, pull them through the PHI,
1336   // reducing code size.
1337   if (isa<Instruction>(PN.getIncomingValue(0)) &&
1338       isa<Instruction>(PN.getIncomingValue(1)) &&
1339       cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
1340           cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
1341       PN.getIncomingValue(0)->hasOneUser())
1342     if (Instruction *Result = foldPHIArgOpIntoPHI(PN))
1343       return Result;
1344 
1345   // If the incoming values are pointer casts of the same original value,
1346   // replace the phi with a single cast iff we can insert a non-PHI instruction.
1347   if (PN.getType()->isPointerTy() &&
1348       PN.getParent()->getFirstInsertionPt() != PN.getParent()->end()) {
1349     Value *IV0 = PN.getIncomingValue(0);
1350     Value *IV0Stripped = IV0->stripPointerCasts();
1351     // Set to keep track of values known to be equal to IV0Stripped after
1352     // stripping pointer casts.
1353     SmallPtrSet<Value *, 4> CheckedIVs;
1354     CheckedIVs.insert(IV0);
1355     if (IV0 != IV0Stripped &&
1356         all_of(PN.incoming_values(), [&CheckedIVs, IV0Stripped](Value *IV) {
1357           return !CheckedIVs.insert(IV).second ||
1358                  IV0Stripped == IV->stripPointerCasts();
1359         })) {
1360       return CastInst::CreatePointerCast(IV0Stripped, PN.getType());
1361     }
1362   }
1363 
1364   // If this is a trivial cycle in the PHI node graph, remove it.  Basically, if
1365   // this PHI only has a single use (a PHI), and if that PHI only has one use (a
1366   // PHI)... break the cycle.
1367   if (PN.hasOneUse()) {
1368     if (Instruction *Result = foldIntegerTypedPHI(PN))
1369       return Result;
1370 
1371     Instruction *PHIUser = cast<Instruction>(PN.user_back());
1372     if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
1373       SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
1374       PotentiallyDeadPHIs.insert(&PN);
1375       if (DeadPHICycle(PU, PotentiallyDeadPHIs))
1376         return replaceInstUsesWith(PN, PoisonValue::get(PN.getType()));
1377     }
1378 
1379     // If this phi has a single use, and if that use just computes a value for
1380     // the next iteration of a loop, delete the phi.  This occurs with unused
1381     // induction variables, e.g. "for (int j = 0; ; ++j);".  Detecting this
1382     // common case here is good because the only other things that catch this
1383     // are induction variable analysis (sometimes) and ADCE, which is only run
1384     // late.
1385     if (PHIUser->hasOneUse() &&
1386         (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
1387         PHIUser->user_back() == &PN) {
1388       return replaceInstUsesWith(PN, PoisonValue::get(PN.getType()));
1389     }
1390     // When a PHI is used only to be compared with zero, it is safe to replace
1391     // an incoming value proved as known nonzero with any non-zero constant.
1392     // For example, in the code below, the incoming value %v can be replaced
1393     // with any non-zero constant based on the fact that the PHI is only used to
1394     // be compared with zero and %v is a known non-zero value:
1395     // %v = select %cond, 1, 2
1396     // %p = phi [%v, BB] ...
1397     //      icmp eq, %p, 0
1398     auto *CmpInst = dyn_cast<ICmpInst>(PHIUser);
1399     // FIXME: To be simple, handle only integer type for now.
1400     if (CmpInst && isa<IntegerType>(PN.getType()) && CmpInst->isEquality() &&
1401         match(CmpInst->getOperand(1), m_Zero())) {
1402       ConstantInt *NonZeroConst = nullptr;
1403       bool MadeChange = false;
1404       for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
1405         Instruction *CtxI = PN.getIncomingBlock(i)->getTerminator();
1406         Value *VA = PN.getIncomingValue(i);
1407         if (isKnownNonZero(VA, DL, 0, &AC, CtxI, &DT)) {
1408           if (!NonZeroConst)
1409             NonZeroConst = GetAnyNonZeroConstInt(PN);
1410 
1411           if (NonZeroConst != VA) {
1412             replaceOperand(PN, i, NonZeroConst);
1413             MadeChange = true;
1414           }
1415         }
1416       }
1417       if (MadeChange)
1418         return &PN;
1419     }
1420   }
1421 
1422   // We sometimes end up with phi cycles that non-obviously end up being the
1423   // same value, for example:
1424   //   z = some value; x = phi (y, z); y = phi (x, z)
1425   // where the phi nodes don't necessarily need to be in the same block.  Do a
1426   // quick check to see if the PHI node only contains a single non-phi value, if
1427   // so, scan to see if the phi cycle is actually equal to that value.
1428   {
1429     unsigned InValNo = 0, NumIncomingVals = PN.getNumIncomingValues();
1430     // Scan for the first non-phi operand.
1431     while (InValNo != NumIncomingVals &&
1432            isa<PHINode>(PN.getIncomingValue(InValNo)))
1433       ++InValNo;
1434 
1435     if (InValNo != NumIncomingVals) {
1436       Value *NonPhiInVal = PN.getIncomingValue(InValNo);
1437 
1438       // Scan the rest of the operands to see if there are any conflicts, if so
1439       // there is no need to recursively scan other phis.
1440       for (++InValNo; InValNo != NumIncomingVals; ++InValNo) {
1441         Value *OpVal = PN.getIncomingValue(InValNo);
1442         if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
1443           break;
1444       }
1445 
1446       // If we scanned over all operands, then we have one unique value plus
1447       // phi values.  Scan PHI nodes to see if they all merge in each other or
1448       // the value.
1449       if (InValNo == NumIncomingVals) {
1450         SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
1451         if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
1452           return replaceInstUsesWith(PN, NonPhiInVal);
1453       }
1454     }
1455   }
1456 
1457   // If there are multiple PHIs, sort their operands so that they all list
1458   // the blocks in the same order. This will help identical PHIs be eliminated
1459   // by other passes. Other passes shouldn't depend on this for correctness
1460   // however.
1461   PHINode *FirstPN = cast<PHINode>(PN.getParent()->begin());
1462   if (&PN != FirstPN)
1463     for (unsigned i = 0, e = FirstPN->getNumIncomingValues(); i != e; ++i) {
1464       BasicBlock *BBA = PN.getIncomingBlock(i);
1465       BasicBlock *BBB = FirstPN->getIncomingBlock(i);
1466       if (BBA != BBB) {
1467         Value *VA = PN.getIncomingValue(i);
1468         unsigned j = PN.getBasicBlockIndex(BBB);
1469         Value *VB = PN.getIncomingValue(j);
1470         PN.setIncomingBlock(i, BBB);
1471         PN.setIncomingValue(i, VB);
1472         PN.setIncomingBlock(j, BBA);
1473         PN.setIncomingValue(j, VA);
1474         // NOTE: Instcombine normally would want us to "return &PN" if we
1475         // modified any of the operands of an instruction.  However, since we
1476         // aren't adding or removing uses (just rearranging them) we don't do
1477         // this in this case.
1478       }
1479     }
1480 
1481   // Is there an identical PHI node in this basic block?
1482   for (PHINode &IdenticalPN : PN.getParent()->phis()) {
1483     // Ignore the PHI node itself.
1484     if (&IdenticalPN == &PN)
1485       continue;
1486     // Note that even though we've just canonicalized this PHI, due to the
1487     // worklist visitation order, there are no guarantess that *every* PHI
1488     // has been canonicalized, so we can't just compare operands ranges.
1489     if (!PN.isIdenticalToWhenDefined(&IdenticalPN))
1490       continue;
1491     // Just use that PHI instead then.
1492     ++NumPHICSEs;
1493     return replaceInstUsesWith(PN, &IdenticalPN);
1494   }
1495 
1496   // If this is an integer PHI and we know that it has an illegal type, see if
1497   // it is only used by trunc or trunc(lshr) operations.  If so, we split the
1498   // PHI into the various pieces being extracted.  This sort of thing is
1499   // introduced when SROA promotes an aggregate to a single large integer type.
1500   if (PN.getType()->isIntegerTy() &&
1501       !DL.isLegalInteger(PN.getType()->getPrimitiveSizeInBits()))
1502     if (Instruction *Res = SliceUpIllegalIntegerPHI(PN))
1503       return Res;
1504 
1505   // Ultimately, try to replace this Phi with a dominating condition.
1506   if (auto *V = SimplifyUsingControlFlow(*this, PN, DT))
1507     return replaceInstUsesWith(PN, V);
1508 
1509   return nullptr;
1510 }
1511