1 //===-- DifferenceEngine.cpp - Structural function/module comparison ------===//
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 header defines the implementation of the LLVM difference
10 // engine, which structurally compares global values within a module.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "DifferenceEngine.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/DenseSet.h"
17 #include "llvm/ADT/SmallString.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/ADT/StringSet.h"
20 #include "llvm/IR/BasicBlock.h"
21 #include "llvm/IR/CFG.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/Function.h"
24 #include "llvm/IR/Instructions.h"
25 #include "llvm/IR/Module.h"
26 #include "llvm/Support/ErrorHandling.h"
27 #include "llvm/Support/raw_ostream.h"
28 #include "llvm/Support/type_traits.h"
29 #include <utility>
30 
31 using namespace llvm;
32 
33 namespace {
34 
35 /// A priority queue, implemented as a heap.
36 template <class T, class Sorter, unsigned InlineCapacity>
37 class PriorityQueue {
38   Sorter Precedes;
39   llvm::SmallVector<T, InlineCapacity> Storage;
40 
41 public:
42   PriorityQueue(const Sorter &Precedes) : Precedes(Precedes) {}
43 
44   /// Checks whether the heap is empty.
45   bool empty() const { return Storage.empty(); }
46 
47   /// Insert a new value on the heap.
48   void insert(const T &V) {
49     unsigned Index = Storage.size();
50     Storage.push_back(V);
51     if (Index == 0) return;
52 
53     T *data = Storage.data();
54     while (true) {
55       unsigned Target = (Index + 1) / 2 - 1;
56       if (!Precedes(data[Index], data[Target])) return;
57       std::swap(data[Index], data[Target]);
58       if (Target == 0) return;
59       Index = Target;
60     }
61   }
62 
63   /// Remove the minimum value in the heap.  Only valid on a non-empty heap.
64   T remove_min() {
65     assert(!empty());
66     T tmp = Storage[0];
67 
68     unsigned NewSize = Storage.size() - 1;
69     if (NewSize) {
70       // Move the slot at the end to the beginning.
71       if (std::is_trivially_copyable<T>::value)
72         Storage[0] = Storage[NewSize];
73       else
74         std::swap(Storage[0], Storage[NewSize]);
75 
76       // Bubble the root up as necessary.
77       unsigned Index = 0;
78       while (true) {
79         // With a 1-based index, the children would be Index*2 and Index*2+1.
80         unsigned R = (Index + 1) * 2;
81         unsigned L = R - 1;
82 
83         // If R is out of bounds, we're done after this in any case.
84         if (R >= NewSize) {
85           // If L is also out of bounds, we're done immediately.
86           if (L >= NewSize) break;
87 
88           // Otherwise, test whether we should swap L and Index.
89           if (Precedes(Storage[L], Storage[Index]))
90             std::swap(Storage[L], Storage[Index]);
91           break;
92         }
93 
94         // Otherwise, we need to compare with the smaller of L and R.
95         // Prefer R because it's closer to the end of the array.
96         unsigned IndexToTest = (Precedes(Storage[L], Storage[R]) ? L : R);
97 
98         // If Index is >= the min of L and R, then heap ordering is restored.
99         if (!Precedes(Storage[IndexToTest], Storage[Index]))
100           break;
101 
102         // Otherwise, keep bubbling up.
103         std::swap(Storage[IndexToTest], Storage[Index]);
104         Index = IndexToTest;
105       }
106     }
107     Storage.pop_back();
108 
109     return tmp;
110   }
111 };
112 
113 /// A function-scope difference engine.
114 class FunctionDifferenceEngine {
115   DifferenceEngine &Engine;
116 
117   // Some initializers may reference the variable we're currently checking. This
118   // can cause an infinite loop. The Saved[LR]HS ivars can be checked to prevent
119   // recursing.
120   const Value *SavedLHS;
121   const Value *SavedRHS;
122 
123   // The current mapping from old local values to new local values.
124   DenseMap<const Value *, const Value *> Values;
125 
126   // The current mapping from old blocks to new blocks.
127   DenseMap<const BasicBlock *, const BasicBlock *> Blocks;
128 
129   // The tentative mapping from old local values while comparing a pair of
130   // basic blocks. Once the pair has been processed, the tentative mapping is
131   // committed to the Values map.
132   DenseSet<std::pair<const Value *, const Value *>> TentativeValues;
133 
134   // Equivalence Assumptions
135   //
136   // For basic blocks in loops, some values in phi nodes may depend on
137   // values from not yet processed basic blocks in the loop. When encountering
138   // such values, we optimistically asssume their equivalence and store this
139   // assumption in a BlockDiffCandidate for the pair of compared BBs.
140   //
141   // Once we have diffed all BBs, for every BlockDiffCandidate, we check all
142   // stored assumptions using the Values map that stores proven equivalences
143   // between the old and new values, and report a diff if an assumption cannot
144   // be proven to be true.
145   //
146   // Note that after having made an assumption, all further determined
147   // equivalences implicitly depend on that assumption. These will not be
148   // reverted or reported if the assumption proves to be false, because these
149   // are considered indirect diffs caused by earlier direct diffs.
150   //
151   // We aim to avoid false negatives in llvm-diff, that is, ensure that
152   // whenever no diff is reported, the functions are indeed equal. If
153   // assumptions were made, this is not entirely clear, because in principle we
154   // could end up with a circular proof where the proof of equivalence of two
155   // nodes is depending on the assumption of their equivalence.
156   //
157   // To see that assumptions do not add false negatives, note that if we do not
158   // report a diff, this means that there is an equivalence mapping between old
159   // and new values that is consistent with all assumptions made. The circular
160   // dependency that exists on an IR value level does not exist at run time,
161   // because the values selected by the phi nodes must always already have been
162   // computed. Hence, we can prove equivalence of the old and new functions by
163   // considering step-wise parallel execution, and incrementally proving
164   // equivalence of every new computed value. Another way to think about it is
165   // to imagine cloning the loop BBs for every iteration, turning the loops
166   // into (possibly infinite) DAGs, and proving equivalence by induction on the
167   // iteration, using the computed value mapping.
168 
169   // The class BlockDiffCandidate stores pairs which either have already been
170   // proven to differ, or pairs whose equivalence depends on assumptions to be
171   // verified later.
172   struct BlockDiffCandidate {
173     const BasicBlock *LBB;
174     const BasicBlock *RBB;
175     // Maps old values to assumed-to-be-equivalent new values
176     SmallDenseMap<const Value *, const Value *> EquivalenceAssumptions;
177     // If set, we already know the blocks differ.
178     bool KnownToDiffer;
179   };
180 
181   // List of block diff candidates in the order found by processing.
182   // We generate reports in this order.
183   // For every LBB, there may only be one corresponding RBB.
184   SmallVector<BlockDiffCandidate> BlockDiffCandidates;
185   // Maps LBB to the index of its BlockDiffCandidate, if existing.
186   DenseMap<const BasicBlock *, uint64_t> BlockDiffCandidateIndices;
187 
188   // Note: Every LBB must always be queried together with the same RBB.
189   // The returned reference is not permanently valid and should not be stored.
190   BlockDiffCandidate &getOrCreateBlockDiffCandidate(const BasicBlock *LBB,
191                                                     const BasicBlock *RBB) {
192     auto It = BlockDiffCandidateIndices.find(LBB);
193     // Check if LBB already has a diff candidate
194     if (It == BlockDiffCandidateIndices.end()) {
195       // Add new one
196       BlockDiffCandidateIndices[LBB] = BlockDiffCandidates.size();
197       BlockDiffCandidates.push_back(
198           {LBB, RBB, SmallDenseMap<const Value *, const Value *>(), false});
199       return BlockDiffCandidates.back();
200     }
201     // Use existing one
202     BlockDiffCandidate &Result = BlockDiffCandidates[It->second];
203     assert(Result.RBB == RBB && "Inconsistent basic block pairing!");
204     return Result;
205   }
206 
207   // Optionally passed to equivalence checker functions, so these can add
208   // assumptions in BlockDiffCandidates. Its presence controls whether
209   // assumptions are generated.
210   struct AssumptionContext {
211     // The two basic blocks that need the two compared values to be equivalent.
212     const BasicBlock *LBB;
213     const BasicBlock *RBB;
214   };
215 
216   unsigned getUnprocPredCount(const BasicBlock *Block) const {
217     unsigned Count = 0;
218     for (const_pred_iterator I = pred_begin(Block), E = pred_end(Block); I != E;
219          ++I)
220       if (!Blocks.count(*I)) Count++;
221     return Count;
222   }
223 
224   typedef std::pair<const BasicBlock *, const BasicBlock *> BlockPair;
225 
226   /// A type which sorts a priority queue by the number of unprocessed
227   /// predecessor blocks it has remaining.
228   ///
229   /// This is actually really expensive to calculate.
230   struct QueueSorter {
231     const FunctionDifferenceEngine &fde;
232     explicit QueueSorter(const FunctionDifferenceEngine &fde) : fde(fde) {}
233 
234     bool operator()(BlockPair &Old, BlockPair &New) {
235       return fde.getUnprocPredCount(Old.first)
236            < fde.getUnprocPredCount(New.first);
237     }
238   };
239 
240   /// A queue of unified blocks to process.
241   PriorityQueue<BlockPair, QueueSorter, 20> Queue;
242 
243   /// Try to unify the given two blocks.  Enqueues them for processing
244   /// if they haven't already been processed.
245   ///
246   /// Returns true if there was a problem unifying them.
247   bool tryUnify(const BasicBlock *L, const BasicBlock *R) {
248     const BasicBlock *&Ref = Blocks[L];
249 
250     if (Ref) {
251       if (Ref == R) return false;
252 
253       Engine.logf("successor %l cannot be equivalent to %r; "
254                   "it's already equivalent to %r")
255         << L << R << Ref;
256       return true;
257     }
258 
259     Ref = R;
260     Queue.insert(BlockPair(L, R));
261     return false;
262   }
263 
264   /// Unifies two instructions, given that they're known not to have
265   /// structural differences.
266   void unify(const Instruction *L, const Instruction *R) {
267     DifferenceEngine::Context C(Engine, L, R);
268 
269     bool Result = diff(L, R, true, true, true);
270     assert(!Result && "structural differences second time around?");
271     (void) Result;
272     if (!L->use_empty())
273       Values[L] = R;
274   }
275 
276   void processQueue() {
277     while (!Queue.empty()) {
278       BlockPair Pair = Queue.remove_min();
279       diff(Pair.first, Pair.second);
280     }
281   }
282 
283   void checkAndReportDiffCandidates() {
284     for (BlockDiffCandidate &BDC : BlockDiffCandidates) {
285 
286       // Check assumptions
287       for (const auto &[L, R] : BDC.EquivalenceAssumptions) {
288         auto It = Values.find(L);
289         if (It == Values.end() || It->second != R) {
290           BDC.KnownToDiffer = true;
291           break;
292         }
293       }
294 
295       // Run block diff if the BBs differ
296       if (BDC.KnownToDiffer) {
297         DifferenceEngine::Context C(Engine, BDC.LBB, BDC.RBB);
298         runBlockDiff(BDC.LBB->begin(), BDC.RBB->begin());
299       }
300     }
301   }
302 
303   void diff(const BasicBlock *L, const BasicBlock *R) {
304     DifferenceEngine::Context C(Engine, L, R);
305 
306     BasicBlock::const_iterator LI = L->begin(), LE = L->end();
307     BasicBlock::const_iterator RI = R->begin();
308 
309     do {
310       assert(LI != LE && RI != R->end());
311       const Instruction *LeftI = &*LI, *RightI = &*RI;
312 
313       // If the instructions differ, start the more sophisticated diff
314       // algorithm at the start of the block.
315       if (diff(LeftI, RightI, false, false, true)) {
316         TentativeValues.clear();
317         // Register (L, R) as diffing pair. Note that we could directly emit a
318         // block diff here, but this way we ensure all diffs are emitted in one
319         // consistent order, independent of whether the diffs were detected
320         // immediately or via invalid assumptions.
321         getOrCreateBlockDiffCandidate(L, R).KnownToDiffer = true;
322         return;
323       }
324 
325       // Otherwise, tentatively unify them.
326       if (!LeftI->use_empty())
327         TentativeValues.insert(std::make_pair(LeftI, RightI));
328 
329       ++LI;
330       ++RI;
331     } while (LI != LE); // This is sufficient: we can't get equality of
332                         // terminators if there are residual instructions.
333 
334     // Unify everything in the block, non-tentatively this time.
335     TentativeValues.clear();
336     for (LI = L->begin(), RI = R->begin(); LI != LE; ++LI, ++RI)
337       unify(&*LI, &*RI);
338   }
339 
340   bool matchForBlockDiff(const Instruction *L, const Instruction *R);
341   void runBlockDiff(BasicBlock::const_iterator LI,
342                     BasicBlock::const_iterator RI);
343 
344   bool diffCallSites(const CallBase &L, const CallBase &R, bool Complain) {
345     // FIXME: call attributes
346     AssumptionContext AC = {L.getParent(), R.getParent()};
347     if (!equivalentAsOperands(L.getCalledOperand(), R.getCalledOperand(),
348                               &AC)) {
349       if (Complain) Engine.log("called functions differ");
350       return true;
351     }
352     if (L.arg_size() != R.arg_size()) {
353       if (Complain) Engine.log("argument counts differ");
354       return true;
355     }
356     for (unsigned I = 0, E = L.arg_size(); I != E; ++I)
357       if (!equivalentAsOperands(L.getArgOperand(I), R.getArgOperand(I), &AC)) {
358         if (Complain)
359           Engine.logf("arguments %l and %r differ")
360               << L.getArgOperand(I) << R.getArgOperand(I);
361         return true;
362       }
363     return false;
364   }
365 
366   // If AllowAssumptions is enabled, whenever we encounter a pair of values
367   // that we cannot prove to be equivalent, we assume equivalence and store that
368   // assumption to be checked later in BlockDiffCandidates.
369   bool diff(const Instruction *L, const Instruction *R, bool Complain,
370             bool TryUnify, bool AllowAssumptions) {
371     // FIXME: metadata (if Complain is set)
372     AssumptionContext ACValue = {L->getParent(), R->getParent()};
373     // nullptr AssumptionContext disables assumption generation.
374     const AssumptionContext *AC = AllowAssumptions ? &ACValue : nullptr;
375 
376     // Different opcodes always imply different operations.
377     if (L->getOpcode() != R->getOpcode()) {
378       if (Complain) Engine.log("different instruction types");
379       return true;
380     }
381 
382     if (isa<CmpInst>(L)) {
383       if (cast<CmpInst>(L)->getPredicate()
384             != cast<CmpInst>(R)->getPredicate()) {
385         if (Complain) Engine.log("different predicates");
386         return true;
387       }
388     } else if (isa<CallInst>(L)) {
389       return diffCallSites(cast<CallInst>(*L), cast<CallInst>(*R), Complain);
390     } else if (isa<PHINode>(L)) {
391       const PHINode &LI = cast<PHINode>(*L);
392       const PHINode &RI = cast<PHINode>(*R);
393 
394       // This is really weird;  type uniquing is broken?
395       if (LI.getType() != RI.getType()) {
396         if (!LI.getType()->isPointerTy() || !RI.getType()->isPointerTy()) {
397           if (Complain) Engine.log("different phi types");
398           return true;
399         }
400       }
401 
402       if (LI.getNumIncomingValues() != RI.getNumIncomingValues()) {
403         if (Complain)
404           Engine.log("PHI node # of incoming values differ");
405         return true;
406       }
407 
408       for (unsigned I = 0; I < LI.getNumIncomingValues(); ++I) {
409         if (TryUnify)
410           tryUnify(LI.getIncomingBlock(I), RI.getIncomingBlock(I));
411 
412         if (!equivalentAsOperands(LI.getIncomingValue(I),
413                                   RI.getIncomingValue(I), AC)) {
414           if (Complain)
415             Engine.log("PHI node incoming values differ");
416           return true;
417         }
418       }
419 
420       return false;
421 
422     // Terminators.
423     } else if (isa<InvokeInst>(L)) {
424       const InvokeInst &LI = cast<InvokeInst>(*L);
425       const InvokeInst &RI = cast<InvokeInst>(*R);
426       if (diffCallSites(LI, RI, Complain))
427         return true;
428 
429       if (TryUnify) {
430         tryUnify(LI.getNormalDest(), RI.getNormalDest());
431         tryUnify(LI.getUnwindDest(), RI.getUnwindDest());
432       }
433       return false;
434 
435     } else if (isa<CallBrInst>(L)) {
436       const CallBrInst &LI = cast<CallBrInst>(*L);
437       const CallBrInst &RI = cast<CallBrInst>(*R);
438       if (LI.getNumIndirectDests() != RI.getNumIndirectDests()) {
439         if (Complain)
440           Engine.log("callbr # of indirect destinations differ");
441         return true;
442       }
443 
444       // Perform the "try unify" step so that we can equate the indirect
445       // destinations before checking the call site.
446       for (unsigned I = 0; I < LI.getNumIndirectDests(); I++)
447         tryUnify(LI.getIndirectDest(I), RI.getIndirectDest(I));
448 
449       if (diffCallSites(LI, RI, Complain))
450         return true;
451 
452       if (TryUnify)
453         tryUnify(LI.getDefaultDest(), RI.getDefaultDest());
454       return false;
455 
456     } else if (isa<BranchInst>(L)) {
457       const BranchInst *LI = cast<BranchInst>(L);
458       const BranchInst *RI = cast<BranchInst>(R);
459       if (LI->isConditional() != RI->isConditional()) {
460         if (Complain) Engine.log("branch conditionality differs");
461         return true;
462       }
463 
464       if (LI->isConditional()) {
465         if (!equivalentAsOperands(LI->getCondition(), RI->getCondition(), AC)) {
466           if (Complain) Engine.log("branch conditions differ");
467           return true;
468         }
469         if (TryUnify) tryUnify(LI->getSuccessor(1), RI->getSuccessor(1));
470       }
471       if (TryUnify) tryUnify(LI->getSuccessor(0), RI->getSuccessor(0));
472       return false;
473 
474     } else if (isa<IndirectBrInst>(L)) {
475       const IndirectBrInst *LI = cast<IndirectBrInst>(L);
476       const IndirectBrInst *RI = cast<IndirectBrInst>(R);
477       if (LI->getNumDestinations() != RI->getNumDestinations()) {
478         if (Complain) Engine.log("indirectbr # of destinations differ");
479         return true;
480       }
481 
482       if (!equivalentAsOperands(LI->getAddress(), RI->getAddress(), AC)) {
483         if (Complain) Engine.log("indirectbr addresses differ");
484         return true;
485       }
486 
487       if (TryUnify) {
488         for (unsigned i = 0; i < LI->getNumDestinations(); i++) {
489           tryUnify(LI->getDestination(i), RI->getDestination(i));
490         }
491       }
492       return false;
493 
494     } else if (isa<SwitchInst>(L)) {
495       const SwitchInst *LI = cast<SwitchInst>(L);
496       const SwitchInst *RI = cast<SwitchInst>(R);
497       if (!equivalentAsOperands(LI->getCondition(), RI->getCondition(), AC)) {
498         if (Complain) Engine.log("switch conditions differ");
499         return true;
500       }
501       if (TryUnify) tryUnify(LI->getDefaultDest(), RI->getDefaultDest());
502 
503       bool Difference = false;
504 
505       DenseMap<const ConstantInt *, const BasicBlock *> LCases;
506       for (auto Case : LI->cases())
507         LCases[Case.getCaseValue()] = Case.getCaseSuccessor();
508 
509       for (auto Case : RI->cases()) {
510         const ConstantInt *CaseValue = Case.getCaseValue();
511         const BasicBlock *LCase = LCases[CaseValue];
512         if (LCase) {
513           if (TryUnify)
514             tryUnify(LCase, Case.getCaseSuccessor());
515           LCases.erase(CaseValue);
516         } else if (Complain || !Difference) {
517           if (Complain)
518             Engine.logf("right switch has extra case %r") << CaseValue;
519           Difference = true;
520         }
521       }
522       if (!Difference)
523         for (DenseMap<const ConstantInt *, const BasicBlock *>::iterator
524                  I = LCases.begin(),
525                  E = LCases.end();
526              I != E; ++I) {
527           if (Complain)
528             Engine.logf("left switch has extra case %l") << I->first;
529           Difference = true;
530         }
531       return Difference;
532     } else if (isa<UnreachableInst>(L)) {
533       return false;
534     }
535 
536     if (L->getNumOperands() != R->getNumOperands()) {
537       if (Complain) Engine.log("instructions have different operand counts");
538       return true;
539     }
540 
541     for (unsigned I = 0, E = L->getNumOperands(); I != E; ++I) {
542       Value *LO = L->getOperand(I), *RO = R->getOperand(I);
543       if (!equivalentAsOperands(LO, RO, AC)) {
544         if (Complain) Engine.logf("operands %l and %r differ") << LO << RO;
545         return true;
546       }
547     }
548 
549     return false;
550   }
551 
552 public:
553   bool equivalentAsOperands(const Constant *L, const Constant *R,
554                             const AssumptionContext *AC) {
555     // Use equality as a preliminary filter.
556     if (L == R)
557       return true;
558 
559     if (L->getValueID() != R->getValueID())
560       return false;
561 
562     // Ask the engine about global values.
563     if (isa<GlobalValue>(L))
564       return Engine.equivalentAsOperands(cast<GlobalValue>(L),
565                                          cast<GlobalValue>(R));
566 
567     // Compare constant expressions structurally.
568     if (isa<ConstantExpr>(L))
569       return equivalentAsOperands(cast<ConstantExpr>(L), cast<ConstantExpr>(R),
570                                   AC);
571 
572     // Constants of the "same type" don't always actually have the same
573     // type; I don't know why.  Just white-list them.
574     if (isa<ConstantPointerNull>(L) || isa<UndefValue>(L) || isa<ConstantAggregateZero>(L))
575       return true;
576 
577     // Block addresses only match if we've already encountered the
578     // block.  FIXME: tentative matches?
579     if (isa<BlockAddress>(L))
580       return Blocks[cast<BlockAddress>(L)->getBasicBlock()]
581                  == cast<BlockAddress>(R)->getBasicBlock();
582 
583     // If L and R are ConstantVectors, compare each element
584     if (isa<ConstantVector>(L)) {
585       const ConstantVector *CVL = cast<ConstantVector>(L);
586       const ConstantVector *CVR = cast<ConstantVector>(R);
587       if (CVL->getType()->getNumElements() != CVR->getType()->getNumElements())
588         return false;
589       for (unsigned i = 0; i < CVL->getType()->getNumElements(); i++) {
590         if (!equivalentAsOperands(CVL->getOperand(i), CVR->getOperand(i), AC))
591           return false;
592       }
593       return true;
594     }
595 
596     // If L and R are ConstantArrays, compare the element count and types.
597     if (isa<ConstantArray>(L)) {
598       const ConstantArray *CAL = cast<ConstantArray>(L);
599       const ConstantArray *CAR = cast<ConstantArray>(R);
600       // Sometimes a type may be equivalent, but not uniquified---e.g. it may
601       // contain a GEP instruction. Do a deeper comparison of the types.
602       if (CAL->getType()->getNumElements() != CAR->getType()->getNumElements())
603         return false;
604 
605       for (unsigned I = 0; I < CAL->getType()->getNumElements(); ++I) {
606         if (!equivalentAsOperands(CAL->getAggregateElement(I),
607                                   CAR->getAggregateElement(I), AC))
608           return false;
609       }
610 
611       return true;
612     }
613 
614     // If L and R are ConstantStructs, compare each field and type.
615     if (isa<ConstantStruct>(L)) {
616       const ConstantStruct *CSL = cast<ConstantStruct>(L);
617       const ConstantStruct *CSR = cast<ConstantStruct>(R);
618 
619       const StructType *LTy = cast<StructType>(CSL->getType());
620       const StructType *RTy = cast<StructType>(CSR->getType());
621 
622       // The StructTypes should have the same attributes. Don't use
623       // isLayoutIdentical(), because that just checks the element pointers,
624       // which may not work here.
625       if (LTy->getNumElements() != RTy->getNumElements() ||
626           LTy->isPacked() != RTy->isPacked())
627         return false;
628 
629       for (unsigned I = 0; I < LTy->getNumElements(); I++) {
630         const Value *LAgg = CSL->getAggregateElement(I);
631         const Value *RAgg = CSR->getAggregateElement(I);
632 
633         if (LAgg == SavedLHS || RAgg == SavedRHS) {
634           if (LAgg != SavedLHS || RAgg != SavedRHS)
635             // If the left and right operands aren't both re-analyzing the
636             // variable, then the initialiers don't match, so report "false".
637             // Otherwise, we skip these operands..
638             return false;
639 
640           continue;
641         }
642 
643         if (!equivalentAsOperands(LAgg, RAgg, AC)) {
644           return false;
645         }
646       }
647 
648       return true;
649     }
650 
651     return false;
652   }
653 
654   bool equivalentAsOperands(const ConstantExpr *L, const ConstantExpr *R,
655                             const AssumptionContext *AC) {
656     if (L == R)
657       return true;
658 
659     if (L->getOpcode() != R->getOpcode())
660       return false;
661 
662     switch (L->getOpcode()) {
663     case Instruction::ICmp:
664     case Instruction::FCmp:
665       if (L->getPredicate() != R->getPredicate())
666         return false;
667       break;
668 
669     case Instruction::GetElementPtr:
670       // FIXME: inbounds?
671       break;
672 
673     default:
674       break;
675     }
676 
677     if (L->getNumOperands() != R->getNumOperands())
678       return false;
679 
680     for (unsigned I = 0, E = L->getNumOperands(); I != E; ++I) {
681       const auto *LOp = L->getOperand(I);
682       const auto *ROp = R->getOperand(I);
683 
684       if (LOp == SavedLHS || ROp == SavedRHS) {
685         if (LOp != SavedLHS || ROp != SavedRHS)
686           // If the left and right operands aren't both re-analyzing the
687           // variable, then the initialiers don't match, so report "false".
688           // Otherwise, we skip these operands..
689           return false;
690 
691         continue;
692       }
693 
694       if (!equivalentAsOperands(LOp, ROp, AC))
695         return false;
696     }
697 
698     return true;
699   }
700 
701   // There are cases where we cannot determine whether two values are
702   // equivalent, because it depends on not yet processed basic blocks -- see the
703   // documentation on assumptions.
704   //
705   // AC is the context in which we are currently performing a diff.
706   // When we encounter a pair of values for which we can neither prove
707   // equivalence nor the opposite, we do the following:
708   //  * If AC is nullptr, we treat the pair as non-equivalent.
709   //  * If AC is set, we add an assumption for the basic blocks given by AC,
710   //    and treat the pair as equivalent. The assumption is checked later.
711   bool equivalentAsOperands(const Value *L, const Value *R,
712                             const AssumptionContext *AC) {
713     // Fall out if the values have different kind.
714     // This possibly shouldn't take priority over oracles.
715     if (L->getValueID() != R->getValueID())
716       return false;
717 
718     // Value subtypes:  Argument, Constant, Instruction, BasicBlock,
719     //                  InlineAsm, MDNode, MDString, PseudoSourceValue
720 
721     if (isa<Constant>(L))
722       return equivalentAsOperands(cast<Constant>(L), cast<Constant>(R), AC);
723 
724     if (isa<Instruction>(L)) {
725       auto It = Values.find(L);
726       if (It != Values.end())
727         return It->second == R;
728 
729       if (TentativeValues.count(std::make_pair(L, R)))
730         return true;
731 
732       // L and R might be equivalent, this could depend on not yet processed
733       // basic blocks, so we cannot decide here.
734       if (AC) {
735         // Add an assumption, unless there is a conflict with an existing one
736         BlockDiffCandidate &BDC =
737             getOrCreateBlockDiffCandidate(AC->LBB, AC->RBB);
738         auto InsertionResult = BDC.EquivalenceAssumptions.insert({L, R});
739         if (!InsertionResult.second && InsertionResult.first->second != R) {
740           // We already have a conflicting equivalence assumption for L, so at
741           // least one must be wrong, and we know that there is a diff.
742           BDC.KnownToDiffer = true;
743           BDC.EquivalenceAssumptions.clear();
744           return false;
745         }
746         // Optimistically assume equivalence, and check later once all BBs
747         // have been processed.
748         return true;
749       }
750 
751       // Assumptions disabled, so pessimistically assume non-equivalence.
752       return false;
753     }
754 
755     if (isa<Argument>(L))
756       return Values[L] == R;
757 
758     if (isa<BasicBlock>(L))
759       return Blocks[cast<BasicBlock>(L)] != R;
760 
761     // Pretend everything else is identical.
762     return true;
763   }
764 
765   // Avoid a gcc warning about accessing 'this' in an initializer.
766   FunctionDifferenceEngine *this_() { return this; }
767 
768 public:
769   FunctionDifferenceEngine(DifferenceEngine &Engine,
770                            const Value *SavedLHS = nullptr,
771                            const Value *SavedRHS = nullptr)
772       : Engine(Engine), SavedLHS(SavedLHS), SavedRHS(SavedRHS),
773         Queue(QueueSorter(*this_())) {}
774 
775   void diff(const Function *L, const Function *R) {
776     assert(Values.empty() && "Multiple diffs per engine are not supported!");
777 
778     if (L->arg_size() != R->arg_size())
779       Engine.log("different argument counts");
780 
781     // Map the arguments.
782     for (Function::const_arg_iterator LI = L->arg_begin(), LE = L->arg_end(),
783                                       RI = R->arg_begin(), RE = R->arg_end();
784          LI != LE && RI != RE; ++LI, ++RI)
785       Values[&*LI] = &*RI;
786 
787     tryUnify(&*L->begin(), &*R->begin());
788     processQueue();
789     checkAndReportDiffCandidates();
790   }
791 };
792 
793 struct DiffEntry {
794   DiffEntry() : Cost(0) {}
795 
796   unsigned Cost;
797   llvm::SmallVector<char, 8> Path; // actually of DifferenceEngine::DiffChange
798 };
799 
800 bool FunctionDifferenceEngine::matchForBlockDiff(const Instruction *L,
801                                                  const Instruction *R) {
802   return !diff(L, R, false, false, false);
803 }
804 
805 void FunctionDifferenceEngine::runBlockDiff(BasicBlock::const_iterator LStart,
806                                             BasicBlock::const_iterator RStart) {
807   BasicBlock::const_iterator LE = LStart->getParent()->end();
808   BasicBlock::const_iterator RE = RStart->getParent()->end();
809 
810   unsigned NL = std::distance(LStart, LE);
811 
812   SmallVector<DiffEntry, 20> Paths1(NL+1);
813   SmallVector<DiffEntry, 20> Paths2(NL+1);
814 
815   DiffEntry *Cur = Paths1.data();
816   DiffEntry *Next = Paths2.data();
817 
818   const unsigned LeftCost = 2;
819   const unsigned RightCost = 2;
820   const unsigned MatchCost = 0;
821 
822   assert(TentativeValues.empty());
823 
824   // Initialize the first column.
825   for (unsigned I = 0; I != NL+1; ++I) {
826     Cur[I].Cost = I * LeftCost;
827     for (unsigned J = 0; J != I; ++J)
828       Cur[I].Path.push_back(DC_left);
829   }
830 
831   for (BasicBlock::const_iterator RI = RStart; RI != RE; ++RI) {
832     // Initialize the first row.
833     Next[0] = Cur[0];
834     Next[0].Cost += RightCost;
835     Next[0].Path.push_back(DC_right);
836 
837     unsigned Index = 1;
838     for (BasicBlock::const_iterator LI = LStart; LI != LE; ++LI, ++Index) {
839       if (matchForBlockDiff(&*LI, &*RI)) {
840         Next[Index] = Cur[Index-1];
841         Next[Index].Cost += MatchCost;
842         Next[Index].Path.push_back(DC_match);
843         TentativeValues.insert(std::make_pair(&*LI, &*RI));
844       } else if (Next[Index-1].Cost <= Cur[Index].Cost) {
845         Next[Index] = Next[Index-1];
846         Next[Index].Cost += LeftCost;
847         Next[Index].Path.push_back(DC_left);
848       } else {
849         Next[Index] = Cur[Index];
850         Next[Index].Cost += RightCost;
851         Next[Index].Path.push_back(DC_right);
852       }
853     }
854 
855     std::swap(Cur, Next);
856   }
857 
858   // We don't need the tentative values anymore; everything from here
859   // on out should be non-tentative.
860   TentativeValues.clear();
861 
862   SmallVectorImpl<char> &Path = Cur[NL].Path;
863   BasicBlock::const_iterator LI = LStart, RI = RStart;
864 
865   DiffLogBuilder Diff(Engine.getConsumer());
866 
867   // Drop trailing matches.
868   while (Path.size() && Path.back() == DC_match)
869     Path.pop_back();
870 
871   // Skip leading matches.
872   SmallVectorImpl<char>::iterator
873     PI = Path.begin(), PE = Path.end();
874   while (PI != PE && *PI == DC_match) {
875     unify(&*LI, &*RI);
876     ++PI;
877     ++LI;
878     ++RI;
879   }
880 
881   for (; PI != PE; ++PI) {
882     switch (static_cast<DiffChange>(*PI)) {
883     case DC_match:
884       assert(LI != LE && RI != RE);
885       {
886         const Instruction *L = &*LI, *R = &*RI;
887         unify(L, R);
888         Diff.addMatch(L, R);
889       }
890       ++LI; ++RI;
891       break;
892 
893     case DC_left:
894       assert(LI != LE);
895       Diff.addLeft(&*LI);
896       ++LI;
897       break;
898 
899     case DC_right:
900       assert(RI != RE);
901       Diff.addRight(&*RI);
902       ++RI;
903       break;
904     }
905   }
906 
907   // Finishing unifying and complaining about the tails of the block,
908   // which should be matches all the way through.
909   while (LI != LE) {
910     assert(RI != RE);
911     unify(&*LI, &*RI);
912     ++LI;
913     ++RI;
914   }
915 
916   // If the terminators have different kinds, but one is an invoke and the
917   // other is an unconditional branch immediately following a call, unify
918   // the results and the destinations.
919   const Instruction *LTerm = LStart->getParent()->getTerminator();
920   const Instruction *RTerm = RStart->getParent()->getTerminator();
921   if (isa<BranchInst>(LTerm) && isa<InvokeInst>(RTerm)) {
922     if (cast<BranchInst>(LTerm)->isConditional()) return;
923     BasicBlock::const_iterator I = LTerm->getIterator();
924     if (I == LStart->getParent()->begin()) return;
925     --I;
926     if (!isa<CallInst>(*I)) return;
927     const CallInst *LCall = cast<CallInst>(&*I);
928     const InvokeInst *RInvoke = cast<InvokeInst>(RTerm);
929     if (!equivalentAsOperands(LCall->getCalledOperand(),
930                               RInvoke->getCalledOperand(), nullptr))
931       return;
932     if (!LCall->use_empty())
933       Values[LCall] = RInvoke;
934     tryUnify(LTerm->getSuccessor(0), RInvoke->getNormalDest());
935   } else if (isa<InvokeInst>(LTerm) && isa<BranchInst>(RTerm)) {
936     if (cast<BranchInst>(RTerm)->isConditional()) return;
937     BasicBlock::const_iterator I = RTerm->getIterator();
938     if (I == RStart->getParent()->begin()) return;
939     --I;
940     if (!isa<CallInst>(*I)) return;
941     const CallInst *RCall = cast<CallInst>(I);
942     const InvokeInst *LInvoke = cast<InvokeInst>(LTerm);
943     if (!equivalentAsOperands(LInvoke->getCalledOperand(),
944                               RCall->getCalledOperand(), nullptr))
945       return;
946     if (!LInvoke->use_empty())
947       Values[LInvoke] = RCall;
948     tryUnify(LInvoke->getNormalDest(), RTerm->getSuccessor(0));
949   }
950 }
951 }
952 
953 void DifferenceEngine::Oracle::anchor() { }
954 
955 void DifferenceEngine::diff(const Function *L, const Function *R) {
956   Context C(*this, L, R);
957 
958   // FIXME: types
959   // FIXME: attributes and CC
960   // FIXME: parameter attributes
961 
962   // If both are declarations, we're done.
963   if (L->empty() && R->empty())
964     return;
965   else if (L->empty())
966     log("left function is declaration, right function is definition");
967   else if (R->empty())
968     log("right function is declaration, left function is definition");
969   else
970     FunctionDifferenceEngine(*this).diff(L, R);
971 }
972 
973 void DifferenceEngine::diff(const Module *L, const Module *R) {
974   StringSet<> LNames;
975   SmallVector<std::pair<const Function *, const Function *>, 20> Queue;
976 
977   unsigned LeftAnonCount = 0;
978   unsigned RightAnonCount = 0;
979 
980   for (Module::const_iterator I = L->begin(), E = L->end(); I != E; ++I) {
981     const Function *LFn = &*I;
982     StringRef Name = LFn->getName();
983     if (Name.empty()) {
984       ++LeftAnonCount;
985       continue;
986     }
987 
988     LNames.insert(Name);
989 
990     if (Function *RFn = R->getFunction(LFn->getName()))
991       Queue.push_back(std::make_pair(LFn, RFn));
992     else
993       logf("function %l exists only in left module") << LFn;
994   }
995 
996   for (Module::const_iterator I = R->begin(), E = R->end(); I != E; ++I) {
997     const Function *RFn = &*I;
998     StringRef Name = RFn->getName();
999     if (Name.empty()) {
1000       ++RightAnonCount;
1001       continue;
1002     }
1003 
1004     if (!LNames.count(Name))
1005       logf("function %r exists only in right module") << RFn;
1006   }
1007 
1008   if (LeftAnonCount != 0 || RightAnonCount != 0) {
1009     SmallString<32> Tmp;
1010     logf(("not comparing " + Twine(LeftAnonCount) +
1011           " anonymous functions in the left module and " +
1012           Twine(RightAnonCount) + " in the right module")
1013              .toStringRef(Tmp));
1014   }
1015 
1016   for (SmallVectorImpl<std::pair<const Function *, const Function *>>::iterator
1017            I = Queue.begin(),
1018            E = Queue.end();
1019        I != E; ++I)
1020     diff(I->first, I->second);
1021 }
1022 
1023 bool DifferenceEngine::equivalentAsOperands(const GlobalValue *L,
1024                                             const GlobalValue *R) {
1025   if (globalValueOracle) return (*globalValueOracle)(L, R);
1026 
1027   if (isa<GlobalVariable>(L) && isa<GlobalVariable>(R)) {
1028     const GlobalVariable *GVL = cast<GlobalVariable>(L);
1029     const GlobalVariable *GVR = cast<GlobalVariable>(R);
1030     if (GVL->hasLocalLinkage() && GVL->hasUniqueInitializer() &&
1031         GVR->hasLocalLinkage() && GVR->hasUniqueInitializer())
1032       return FunctionDifferenceEngine(*this, GVL, GVR)
1033           .equivalentAsOperands(GVL->getInitializer(), GVR->getInitializer(),
1034                                 nullptr);
1035   }
1036 
1037   return L->getName() == R->getName();
1038 }
1039