1//===-- DifferenceEngine.cpp - Structural function/module comparison ------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This header defines the implementation of the LLVM difference
11// engine, which structurally compares global values within a module.
12//
13//===----------------------------------------------------------------------===//
14
15#include "DifferenceEngine.h"
16
17#include "llvm/Constants.h"
18#include "llvm/Function.h"
19#include "llvm/Instructions.h"
20#include "llvm/Module.h"
21#include "llvm/ADT/DenseMap.h"
22#include "llvm/ADT/DenseSet.h"
23#include "llvm/ADT/SmallVector.h"
24#include "llvm/ADT/StringRef.h"
25#include "llvm/ADT/StringSet.h"
26#include "llvm/Support/CallSite.h"
27#include "llvm/Support/CFG.h"
28#include "llvm/Support/ErrorHandling.h"
29#include "llvm/Support/raw_ostream.h"
30#include "llvm/Support/type_traits.h"
31
32#include <utility>
33
34using namespace llvm;
35
36namespace {
37
38/// A priority queue, implemented as a heap.
39template <class T, class Sorter, unsigned InlineCapacity>
40class PriorityQueue {
41  Sorter Precedes;
42  llvm::SmallVector<T, InlineCapacity> Storage;
43
44public:
45  PriorityQueue(const Sorter &Precedes) : Precedes(Precedes) {}
46
47  /// Checks whether the heap is empty.
48  bool empty() const { return Storage.empty(); }
49
50  /// Insert a new value on the heap.
51  void insert(const T &V) {
52    unsigned Index = Storage.size();
53    Storage.push_back(V);
54    if (Index == 0) return;
55
56    T *data = Storage.data();
57    while (true) {
58      unsigned Target = (Index + 1) / 2 - 1;
59      if (!Precedes(data[Index], data[Target])) return;
60      std::swap(data[Index], data[Target]);
61      if (Target == 0) return;
62      Index = Target;
63    }
64  }
65
66  /// Remove the minimum value in the heap.  Only valid on a non-empty heap.
67  T remove_min() {
68    assert(!empty());
69    T tmp = Storage[0];
70
71    unsigned NewSize = Storage.size() - 1;
72    if (NewSize) {
73      // Move the slot at the end to the beginning.
74      if (isPodLike<T>::value)
75        Storage[0] = Storage[NewSize];
76      else
77        std::swap(Storage[0], Storage[NewSize]);
78
79      // Bubble the root up as necessary.
80      unsigned Index = 0;
81      while (true) {
82        // With a 1-based index, the children would be Index*2 and Index*2+1.
83        unsigned R = (Index + 1) * 2;
84        unsigned L = R - 1;
85
86        // If R is out of bounds, we're done after this in any case.
87        if (R >= NewSize) {
88          // If L is also out of bounds, we're done immediately.
89          if (L >= NewSize) break;
90
91          // Otherwise, test whether we should swap L and Index.
92          if (Precedes(Storage[L], Storage[Index]))
93            std::swap(Storage[L], Storage[Index]);
94          break;
95        }
96
97        // Otherwise, we need to compare with the smaller of L and R.
98        // Prefer R because it's closer to the end of the array.
99        unsigned IndexToTest = (Precedes(Storage[L], Storage[R]) ? L : R);
100
101        // If Index is >= the min of L and R, then heap ordering is restored.
102        if (!Precedes(Storage[IndexToTest], Storage[Index]))
103          break;
104
105        // Otherwise, keep bubbling up.
106        std::swap(Storage[IndexToTest], Storage[Index]);
107        Index = IndexToTest;
108      }
109    }
110    Storage.pop_back();
111
112    return tmp;
113  }
114};
115
116/// A function-scope difference engine.
117class FunctionDifferenceEngine {
118  DifferenceEngine &Engine;
119
120  /// The current mapping from old local values to new local values.
121  DenseMap<Value*, Value*> Values;
122
123  /// The current mapping from old blocks to new blocks.
124  DenseMap<BasicBlock*, BasicBlock*> Blocks;
125
126  DenseSet<std::pair<Value*, Value*> > TentativeValues;
127
128  unsigned getUnprocPredCount(BasicBlock *Block) const {
129    unsigned Count = 0;
130    for (pred_iterator I = pred_begin(Block), E = pred_end(Block); I != E; ++I)
131      if (!Blocks.count(*I)) Count++;
132    return Count;
133  }
134
135  typedef std::pair<BasicBlock*, BasicBlock*> BlockPair;
136
137  /// A type which sorts a priority queue by the number of unprocessed
138  /// predecessor blocks it has remaining.
139  ///
140  /// This is actually really expensive to calculate.
141  struct QueueSorter {
142    const FunctionDifferenceEngine &fde;
143    explicit QueueSorter(const FunctionDifferenceEngine &fde) : fde(fde) {}
144
145    bool operator()(const BlockPair &Old, const BlockPair &New) {
146      return fde.getUnprocPredCount(Old.first)
147           < fde.getUnprocPredCount(New.first);
148    }
149  };
150
151  /// A queue of unified blocks to process.
152  PriorityQueue<BlockPair, QueueSorter, 20> Queue;
153
154  /// Try to unify the given two blocks.  Enqueues them for processing
155  /// if they haven't already been processed.
156  ///
157  /// Returns true if there was a problem unifying them.
158  bool tryUnify(BasicBlock *L, BasicBlock *R) {
159    BasicBlock *&Ref = Blocks[L];
160
161    if (Ref) {
162      if (Ref == R) return false;
163
164      Engine.logf("successor %l cannot be equivalent to %r; "
165                  "it's already equivalent to %r")
166        << L << R << Ref;
167      return true;
168    }
169
170    Ref = R;
171    Queue.insert(BlockPair(L, R));
172    return false;
173  }
174
175  /// Unifies two instructions, given that they're known not to have
176  /// structural differences.
177  void unify(Instruction *L, Instruction *R) {
178    DifferenceEngine::Context C(Engine, L, R);
179
180    bool Result = diff(L, R, true, true);
181    assert(!Result && "structural differences second time around?");
182    (void) Result;
183    if (!L->use_empty())
184      Values[L] = R;
185  }
186
187  void processQueue() {
188    while (!Queue.empty()) {
189      BlockPair Pair = Queue.remove_min();
190      diff(Pair.first, Pair.second);
191    }
192  }
193
194  void diff(BasicBlock *L, BasicBlock *R) {
195    DifferenceEngine::Context C(Engine, L, R);
196
197    BasicBlock::iterator LI = L->begin(), LE = L->end();
198    BasicBlock::iterator RI = R->begin();
199
200    llvm::SmallVector<std::pair<Instruction*,Instruction*>, 20> TentativePairs;
201
202    do {
203      assert(LI != LE && RI != R->end());
204      Instruction *LeftI = &*LI, *RightI = &*RI;
205
206      // If the instructions differ, start the more sophisticated diff
207      // algorithm at the start of the block.
208      if (diff(LeftI, RightI, false, false)) {
209        TentativeValues.clear();
210        return runBlockDiff(L->begin(), R->begin());
211      }
212
213      // Otherwise, tentatively unify them.
214      if (!LeftI->use_empty())
215        TentativeValues.insert(std::make_pair(LeftI, RightI));
216
217      ++LI, ++RI;
218    } while (LI != LE); // This is sufficient: we can't get equality of
219                        // terminators if there are residual instructions.
220
221    // Unify everything in the block, non-tentatively this time.
222    TentativeValues.clear();
223    for (LI = L->begin(), RI = R->begin(); LI != LE; ++LI, ++RI)
224      unify(&*LI, &*RI);
225  }
226
227  bool matchForBlockDiff(Instruction *L, Instruction *R);
228  void runBlockDiff(BasicBlock::iterator LI, BasicBlock::iterator RI);
229
230  bool diffCallSites(CallSite L, CallSite R, bool Complain) {
231    // FIXME: call attributes
232    if (!equivalentAsOperands(L.getCalledValue(), R.getCalledValue())) {
233      if (Complain) Engine.log("called functions differ");
234      return true;
235    }
236    if (L.arg_size() != R.arg_size()) {
237      if (Complain) Engine.log("argument counts differ");
238      return true;
239    }
240    for (unsigned I = 0, E = L.arg_size(); I != E; ++I)
241      if (!equivalentAsOperands(L.getArgument(I), R.getArgument(I))) {
242        if (Complain)
243          Engine.logf("arguments %l and %r differ")
244            << L.getArgument(I) << R.getArgument(I);
245        return true;
246      }
247    return false;
248  }
249
250  bool diff(Instruction *L, Instruction *R, bool Complain, bool TryUnify) {
251    // FIXME: metadata (if Complain is set)
252
253    // Different opcodes always imply different operations.
254    if (L->getOpcode() != R->getOpcode()) {
255      if (Complain) Engine.log("different instruction types");
256      return true;
257    }
258
259    if (isa<CmpInst>(L)) {
260      if (cast<CmpInst>(L)->getPredicate()
261            != cast<CmpInst>(R)->getPredicate()) {
262        if (Complain) Engine.log("different predicates");
263        return true;
264      }
265    } else if (isa<CallInst>(L)) {
266      return diffCallSites(CallSite(L), CallSite(R), Complain);
267    } else if (isa<PHINode>(L)) {
268      // FIXME: implement.
269
270      // This is really weird;  type uniquing is broken?
271      if (L->getType() != R->getType()) {
272        if (!L->getType()->isPointerTy() || !R->getType()->isPointerTy()) {
273          if (Complain) Engine.log("different phi types");
274          return true;
275        }
276      }
277      return false;
278
279    // Terminators.
280    } else if (isa<InvokeInst>(L)) {
281      InvokeInst *LI = cast<InvokeInst>(L);
282      InvokeInst *RI = cast<InvokeInst>(R);
283      if (diffCallSites(CallSite(LI), CallSite(RI), Complain))
284        return true;
285
286      if (TryUnify) {
287        tryUnify(LI->getNormalDest(), RI->getNormalDest());
288        tryUnify(LI->getUnwindDest(), RI->getUnwindDest());
289      }
290      return false;
291
292    } else if (isa<BranchInst>(L)) {
293      BranchInst *LI = cast<BranchInst>(L);
294      BranchInst *RI = cast<BranchInst>(R);
295      if (LI->isConditional() != RI->isConditional()) {
296        if (Complain) Engine.log("branch conditionality differs");
297        return true;
298      }
299
300      if (LI->isConditional()) {
301        if (!equivalentAsOperands(LI->getCondition(), RI->getCondition())) {
302          if (Complain) Engine.log("branch conditions differ");
303          return true;
304        }
305        if (TryUnify) tryUnify(LI->getSuccessor(1), RI->getSuccessor(1));
306      }
307      if (TryUnify) tryUnify(LI->getSuccessor(0), RI->getSuccessor(0));
308      return false;
309
310    } else if (isa<SwitchInst>(L)) {
311      SwitchInst *LI = cast<SwitchInst>(L);
312      SwitchInst *RI = cast<SwitchInst>(R);
313      if (!equivalentAsOperands(LI->getCondition(), RI->getCondition())) {
314        if (Complain) Engine.log("switch conditions differ");
315        return true;
316      }
317      if (TryUnify) tryUnify(LI->getDefaultDest(), RI->getDefaultDest());
318
319      bool Difference = false;
320
321      DenseMap<Constant*, BasicBlock*> LCases;
322
323      for (SwitchInst::CaseIt I = LI->case_begin(), E = LI->case_end();
324           I != E; ++I)
325        LCases[I.getCaseValueEx()] = I.getCaseSuccessor();
326
327      for (SwitchInst::CaseIt I = RI->case_begin(), E = RI->case_end();
328           I != E; ++I) {
329        IntegersSubset CaseValue = I.getCaseValueEx();
330        BasicBlock *LCase = LCases[CaseValue];
331        if (LCase) {
332          if (TryUnify) tryUnify(LCase, I.getCaseSuccessor());
333          LCases.erase(CaseValue);
334        } else if (Complain || !Difference) {
335          if (Complain)
336            Engine.logf("right switch has extra case %r") << CaseValue;
337          Difference = true;
338        }
339      }
340      if (!Difference)
341        for (DenseMap<Constant*, BasicBlock*>::iterator
342               I = LCases.begin(), E = LCases.end(); I != E; ++I) {
343          if (Complain)
344            Engine.logf("left switch has extra case %l") << I->first;
345          Difference = true;
346        }
347      return Difference;
348    } else if (isa<UnreachableInst>(L)) {
349      return false;
350    }
351
352    if (L->getNumOperands() != R->getNumOperands()) {
353      if (Complain) Engine.log("instructions have different operand counts");
354      return true;
355    }
356
357    for (unsigned I = 0, E = L->getNumOperands(); I != E; ++I) {
358      Value *LO = L->getOperand(I), *RO = R->getOperand(I);
359      if (!equivalentAsOperands(LO, RO)) {
360        if (Complain) Engine.logf("operands %l and %r differ") << LO << RO;
361        return true;
362      }
363    }
364
365    return false;
366  }
367
368  bool equivalentAsOperands(Constant *L, Constant *R) {
369    // Use equality as a preliminary filter.
370    if (L == R)
371      return true;
372
373    if (L->getValueID() != R->getValueID())
374      return false;
375
376    // Ask the engine about global values.
377    if (isa<GlobalValue>(L))
378      return Engine.equivalentAsOperands(cast<GlobalValue>(L),
379                                         cast<GlobalValue>(R));
380
381    // Compare constant expressions structurally.
382    if (isa<ConstantExpr>(L))
383      return equivalentAsOperands(cast<ConstantExpr>(L),
384                                  cast<ConstantExpr>(R));
385
386    // Nulls of the "same type" don't always actually have the same
387    // type; I don't know why.  Just white-list them.
388    if (isa<ConstantPointerNull>(L))
389      return true;
390
391    // Block addresses only match if we've already encountered the
392    // block.  FIXME: tentative matches?
393    if (isa<BlockAddress>(L))
394      return Blocks[cast<BlockAddress>(L)->getBasicBlock()]
395                 == cast<BlockAddress>(R)->getBasicBlock();
396
397    return false;
398  }
399
400  bool equivalentAsOperands(ConstantExpr *L, ConstantExpr *R) {
401    if (L == R)
402      return true;
403    if (L->getOpcode() != R->getOpcode())
404      return false;
405
406    switch (L->getOpcode()) {
407    case Instruction::ICmp:
408    case Instruction::FCmp:
409      if (L->getPredicate() != R->getPredicate())
410        return false;
411      break;
412
413    case Instruction::GetElementPtr:
414      // FIXME: inbounds?
415      break;
416
417    default:
418      break;
419    }
420
421    if (L->getNumOperands() != R->getNumOperands())
422      return false;
423
424    for (unsigned I = 0, E = L->getNumOperands(); I != E; ++I)
425      if (!equivalentAsOperands(L->getOperand(I), R->getOperand(I)))
426        return false;
427
428    return true;
429  }
430
431  bool equivalentAsOperands(Value *L, Value *R) {
432    // Fall out if the values have different kind.
433    // This possibly shouldn't take priority over oracles.
434    if (L->getValueID() != R->getValueID())
435      return false;
436
437    // Value subtypes:  Argument, Constant, Instruction, BasicBlock,
438    //                  InlineAsm, MDNode, MDString, PseudoSourceValue
439
440    if (isa<Constant>(L))
441      return equivalentAsOperands(cast<Constant>(L), cast<Constant>(R));
442
443    if (isa<Instruction>(L))
444      return Values[L] == R || TentativeValues.count(std::make_pair(L, R));
445
446    if (isa<Argument>(L))
447      return Values[L] == R;
448
449    if (isa<BasicBlock>(L))
450      return Blocks[cast<BasicBlock>(L)] != R;
451
452    // Pretend everything else is identical.
453    return true;
454  }
455
456  // Avoid a gcc warning about accessing 'this' in an initializer.
457  FunctionDifferenceEngine *this_() { return this; }
458
459public:
460  FunctionDifferenceEngine(DifferenceEngine &Engine) :
461    Engine(Engine), Queue(QueueSorter(*this_())) {}
462
463  void diff(Function *L, Function *R) {
464    if (L->arg_size() != R->arg_size())
465      Engine.log("different argument counts");
466
467    // Map the arguments.
468    for (Function::arg_iterator
469           LI = L->arg_begin(), LE = L->arg_end(),
470           RI = R->arg_begin(), RE = R->arg_end();
471         LI != LE && RI != RE; ++LI, ++RI)
472      Values[&*LI] = &*RI;
473
474    tryUnify(&*L->begin(), &*R->begin());
475    processQueue();
476  }
477};
478
479struct DiffEntry {
480  DiffEntry() : Cost(0) {}
481
482  unsigned Cost;
483  llvm::SmallVector<char, 8> Path; // actually of DifferenceEngine::DiffChange
484};
485
486bool FunctionDifferenceEngine::matchForBlockDiff(Instruction *L,
487                                                 Instruction *R) {
488  return !diff(L, R, false, false);
489}
490
491void FunctionDifferenceEngine::runBlockDiff(BasicBlock::iterator LStart,
492                                            BasicBlock::iterator RStart) {
493  BasicBlock::iterator LE = LStart->getParent()->end();
494  BasicBlock::iterator RE = RStart->getParent()->end();
495
496  unsigned NL = std::distance(LStart, LE);
497
498  SmallVector<DiffEntry, 20> Paths1(NL+1);
499  SmallVector<DiffEntry, 20> Paths2(NL+1);
500
501  DiffEntry *Cur = Paths1.data();
502  DiffEntry *Next = Paths2.data();
503
504  const unsigned LeftCost = 2;
505  const unsigned RightCost = 2;
506  const unsigned MatchCost = 0;
507
508  assert(TentativeValues.empty());
509
510  // Initialize the first column.
511  for (unsigned I = 0; I != NL+1; ++I) {
512    Cur[I].Cost = I * LeftCost;
513    for (unsigned J = 0; J != I; ++J)
514      Cur[I].Path.push_back(DC_left);
515  }
516
517  for (BasicBlock::iterator RI = RStart; RI != RE; ++RI) {
518    // Initialize the first row.
519    Next[0] = Cur[0];
520    Next[0].Cost += RightCost;
521    Next[0].Path.push_back(DC_right);
522
523    unsigned Index = 1;
524    for (BasicBlock::iterator LI = LStart; LI != LE; ++LI, ++Index) {
525      if (matchForBlockDiff(&*LI, &*RI)) {
526        Next[Index] = Cur[Index-1];
527        Next[Index].Cost += MatchCost;
528        Next[Index].Path.push_back(DC_match);
529        TentativeValues.insert(std::make_pair(&*LI, &*RI));
530      } else if (Next[Index-1].Cost <= Cur[Index].Cost) {
531        Next[Index] = Next[Index-1];
532        Next[Index].Cost += LeftCost;
533        Next[Index].Path.push_back(DC_left);
534      } else {
535        Next[Index] = Cur[Index];
536        Next[Index].Cost += RightCost;
537        Next[Index].Path.push_back(DC_right);
538      }
539    }
540
541    std::swap(Cur, Next);
542  }
543
544  // We don't need the tentative values anymore; everything from here
545  // on out should be non-tentative.
546  TentativeValues.clear();
547
548  SmallVectorImpl<char> &Path = Cur[NL].Path;
549  BasicBlock::iterator LI = LStart, RI = RStart;
550
551  DiffLogBuilder Diff(Engine.getConsumer());
552
553  // Drop trailing matches.
554  while (Path.back() == DC_match)
555    Path.pop_back();
556
557  // Skip leading matches.
558  SmallVectorImpl<char>::iterator
559    PI = Path.begin(), PE = Path.end();
560  while (PI != PE && *PI == DC_match) {
561    unify(&*LI, &*RI);
562    ++PI, ++LI, ++RI;
563  }
564
565  for (; PI != PE; ++PI) {
566    switch (static_cast<DiffChange>(*PI)) {
567    case DC_match:
568      assert(LI != LE && RI != RE);
569      {
570        Instruction *L = &*LI, *R = &*RI;
571        unify(L, R);
572        Diff.addMatch(L, R);
573      }
574      ++LI; ++RI;
575      break;
576
577    case DC_left:
578      assert(LI != LE);
579      Diff.addLeft(&*LI);
580      ++LI;
581      break;
582
583    case DC_right:
584      assert(RI != RE);
585      Diff.addRight(&*RI);
586      ++RI;
587      break;
588    }
589  }
590
591  // Finishing unifying and complaining about the tails of the block,
592  // which should be matches all the way through.
593  while (LI != LE) {
594    assert(RI != RE);
595    unify(&*LI, &*RI);
596    ++LI, ++RI;
597  }
598
599  // If the terminators have different kinds, but one is an invoke and the
600  // other is an unconditional branch immediately following a call, unify
601  // the results and the destinations.
602  TerminatorInst *LTerm = LStart->getParent()->getTerminator();
603  TerminatorInst *RTerm = RStart->getParent()->getTerminator();
604  if (isa<BranchInst>(LTerm) && isa<InvokeInst>(RTerm)) {
605    if (cast<BranchInst>(LTerm)->isConditional()) return;
606    BasicBlock::iterator I = LTerm;
607    if (I == LStart->getParent()->begin()) return;
608    --I;
609    if (!isa<CallInst>(*I)) return;
610    CallInst *LCall = cast<CallInst>(&*I);
611    InvokeInst *RInvoke = cast<InvokeInst>(RTerm);
612    if (!equivalentAsOperands(LCall->getCalledValue(), RInvoke->getCalledValue()))
613      return;
614    if (!LCall->use_empty())
615      Values[LCall] = RInvoke;
616    tryUnify(LTerm->getSuccessor(0), RInvoke->getNormalDest());
617  } else if (isa<InvokeInst>(LTerm) && isa<BranchInst>(RTerm)) {
618    if (cast<BranchInst>(RTerm)->isConditional()) return;
619    BasicBlock::iterator I = RTerm;
620    if (I == RStart->getParent()->begin()) return;
621    --I;
622    if (!isa<CallInst>(*I)) return;
623    CallInst *RCall = cast<CallInst>(I);
624    InvokeInst *LInvoke = cast<InvokeInst>(LTerm);
625    if (!equivalentAsOperands(LInvoke->getCalledValue(), RCall->getCalledValue()))
626      return;
627    if (!LInvoke->use_empty())
628      Values[LInvoke] = RCall;
629    tryUnify(LInvoke->getNormalDest(), RTerm->getSuccessor(0));
630  }
631}
632
633}
634
635void DifferenceEngine::Oracle::anchor() { }
636
637void DifferenceEngine::diff(Function *L, Function *R) {
638  Context C(*this, L, R);
639
640  // FIXME: types
641  // FIXME: attributes and CC
642  // FIXME: parameter attributes
643
644  // If both are declarations, we're done.
645  if (L->empty() && R->empty())
646    return;
647  else if (L->empty())
648    log("left function is declaration, right function is definition");
649  else if (R->empty())
650    log("right function is declaration, left function is definition");
651  else
652    FunctionDifferenceEngine(*this).diff(L, R);
653}
654
655void DifferenceEngine::diff(Module *L, Module *R) {
656  StringSet<> LNames;
657  SmallVector<std::pair<Function*,Function*>, 20> Queue;
658
659  for (Module::iterator I = L->begin(), E = L->end(); I != E; ++I) {
660    Function *LFn = &*I;
661    LNames.insert(LFn->getName());
662
663    if (Function *RFn = R->getFunction(LFn->getName()))
664      Queue.push_back(std::make_pair(LFn, RFn));
665    else
666      logf("function %l exists only in left module") << LFn;
667  }
668
669  for (Module::iterator I = R->begin(), E = R->end(); I != E; ++I) {
670    Function *RFn = &*I;
671    if (!LNames.count(RFn->getName()))
672      logf("function %r exists only in right module") << RFn;
673  }
674
675  for (SmallVectorImpl<std::pair<Function*,Function*> >::iterator
676         I = Queue.begin(), E = Queue.end(); I != E; ++I)
677    diff(I->first, I->second);
678}
679
680bool DifferenceEngine::equivalentAsOperands(GlobalValue *L, GlobalValue *R) {
681  if (globalValueOracle) return (*globalValueOracle)(L, R);
682  return L->getName() == R->getName();
683}
684