1 //===-- DifferenceEngine.cpp - Structural function/module comparison ------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This header defines the implementation of the LLVM difference
11 // engine, which structurally compares global values within a module.
13 //===----------------------------------------------------------------------===//
15 #include "DifferenceEngine.h"
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"
38 /// A priority queue, implemented as a heap.
39 template <class T, class Sorter, unsigned InlineCapacity>
42 llvm::SmallVector<T, InlineCapacity> Storage;
45 PriorityQueue(const Sorter &Precedes) : Precedes(Precedes) {}
47 /// Checks whether the heap is empty.
48 bool empty() const { return Storage.empty(); }
50 /// Insert a new value on the heap.
51 void insert(const T &V) {
52 unsigned Index = Storage.size();
54 if (Index == 0) return;
56 T *data = Storage.data();
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;
66 /// Remove the minimum value in the heap. Only valid on a non-empty heap.
71 unsigned NewSize = Storage.size() - 1;
73 // Move the slot at the end to the beginning.
74 if (isPodLike<T>::value)
75 Storage[0] = Storage[NewSize];
77 std::swap(Storage[0], Storage[NewSize]);
79 // Bubble the root up as necessary.
82 // With a 1-based index, the children would be Index*2 and Index*2+1.
83 unsigned R = (Index + 1) * 2;
86 // If R is out of bounds, we're done after this in any case.
88 // If L is also out of bounds, we're done immediately.
89 if (L >= NewSize) break;
91 // Otherwise, test whether we should swap L and Index.
92 if (Precedes(Storage[L], Storage[Index]))
93 std::swap(Storage[L], Storage[Index]);
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);
101 // If Index is >= the min of L and R, then heap ordering is restored.
102 if (!Precedes(Storage[IndexToTest], Storage[Index]))
105 // Otherwise, keep bubbling up.
106 std::swap(Storage[IndexToTest], Storage[Index]);
116 /// A function-scope difference engine.
117 class FunctionDifferenceEngine {
118 DifferenceEngine &Engine;
120 /// The current mapping from old local values to new local values.
121 DenseMap<Value*, Value*> Values;
123 /// The current mapping from old blocks to new blocks.
124 DenseMap<BasicBlock*, BasicBlock*> Blocks;
126 DenseSet<std::pair<Value*, Value*> > TentativeValues;
128 unsigned getUnprocPredCount(BasicBlock *Block) const {
130 for (pred_iterator I = pred_begin(Block), E = pred_end(Block); I != E; ++I)
131 if (!Blocks.count(*I)) Count++;
135 typedef std::pair<BasicBlock*, BasicBlock*> BlockPair;
137 /// A type which sorts a priority queue by the number of unprocessed
138 /// predecessor blocks it has remaining.
140 /// This is actually really expensive to calculate.
142 const FunctionDifferenceEngine &fde;
143 explicit QueueSorter(const FunctionDifferenceEngine &fde) : fde(fde) {}
145 bool operator()(const BlockPair &Old, const BlockPair &New) {
146 return fde.getUnprocPredCount(Old.first)
147 < fde.getUnprocPredCount(New.first);
151 /// A queue of unified blocks to process.
152 PriorityQueue<BlockPair, QueueSorter, 20> Queue;
154 /// Try to unify the given two blocks. Enqueues them for processing
155 /// if they haven't already been processed.
157 /// Returns true if there was a problem unifying them.
158 bool tryUnify(BasicBlock *L, BasicBlock *R) {
159 BasicBlock *&Ref = Blocks[L];
162 if (Ref == R) return false;
164 Engine.logf("successor %l cannot be equivalent to %r; "
165 "it's already equivalent to %r")
171 Queue.insert(BlockPair(L, R));
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);
180 bool Result = diff(L, R, true, true);
181 assert(!Result && "structural differences second time around?");
187 void processQueue() {
188 while (!Queue.empty()) {
189 BlockPair Pair = Queue.remove_min();
190 diff(Pair.first, Pair.second);
194 void diff(BasicBlock *L, BasicBlock *R) {
195 DifferenceEngine::Context C(Engine, L, R);
197 BasicBlock::iterator LI = L->begin(), LE = L->end();
198 BasicBlock::iterator RI = R->begin();
200 llvm::SmallVector<std::pair<Instruction*,Instruction*>, 20> TentativePairs;
203 assert(LI != LE && RI != R->end());
204 Instruction *LeftI = &*LI, *RightI = &*RI;
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());
213 // Otherwise, tentatively unify them.
214 if (!LeftI->use_empty())
215 TentativeValues.insert(std::make_pair(LeftI, RightI));
218 } while (LI != LE); // This is sufficient: we can't get equality of
219 // terminators if there are residual instructions.
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)
227 bool matchForBlockDiff(Instruction *L, Instruction *R);
228 void runBlockDiff(BasicBlock::iterator LI, BasicBlock::iterator RI);
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");
236 if (L.arg_size() != R.arg_size()) {
237 if (Complain) Engine.log("argument counts differ");
240 for (unsigned I = 0, E = L.arg_size(); I != E; ++I)
241 if (!equivalentAsOperands(L.getArgument(I), R.getArgument(I))) {
243 Engine.logf("arguments %l and %r differ")
244 << L.getArgument(I) << R.getArgument(I);
250 bool diff(Instruction *L, Instruction *R, bool Complain, bool TryUnify) {
251 // FIXME: metadata (if Complain is set)
253 // Different opcodes always imply different operations.
254 if (L->getOpcode() != R->getOpcode()) {
255 if (Complain) Engine.log("different instruction types");
259 if (isa<CmpInst>(L)) {
260 if (cast<CmpInst>(L)->getPredicate()
261 != cast<CmpInst>(R)->getPredicate()) {
262 if (Complain) Engine.log("different predicates");
265 } else if (isa<CallInst>(L)) {
266 return diffCallSites(CallSite(L), CallSite(R), Complain);
267 } else if (isa<PHINode>(L)) {
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");
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))
287 tryUnify(LI->getNormalDest(), RI->getNormalDest());
288 tryUnify(LI->getUnwindDest(), RI->getUnwindDest());
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");
300 if (LI->isConditional()) {
301 if (!equivalentAsOperands(LI->getCondition(), RI->getCondition())) {
302 if (Complain) Engine.log("branch conditions differ");
305 if (TryUnify) tryUnify(LI->getSuccessor(1), RI->getSuccessor(1));
307 if (TryUnify) tryUnify(LI->getSuccessor(0), RI->getSuccessor(0));
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");
317 if (TryUnify) tryUnify(LI->getDefaultDest(), RI->getDefaultDest());
319 bool Difference = false;
321 DenseMap<Constant*, BasicBlock*> LCases;
323 for (SwitchInst::CaseIt I = LI->case_begin(), E = LI->case_end();
325 LCases[I.getCaseValueEx()] = I.getCaseSuccessor();
327 for (SwitchInst::CaseIt I = RI->case_begin(), E = RI->case_end();
329 IntegersSubset CaseValue = I.getCaseValueEx();
330 BasicBlock *LCase = LCases[CaseValue];
332 if (TryUnify) tryUnify(LCase, I.getCaseSuccessor());
333 LCases.erase(CaseValue);
334 } else if (Complain || !Difference) {
336 Engine.logf("right switch has extra case %r") << CaseValue;
341 for (DenseMap<Constant*, BasicBlock*>::iterator
342 I = LCases.begin(), E = LCases.end(); I != E; ++I) {
344 Engine.logf("left switch has extra case %l") << I->first;
348 } else if (isa<UnreachableInst>(L)) {
352 if (L->getNumOperands() != R->getNumOperands()) {
353 if (Complain) Engine.log("instructions have different operand counts");
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;
368 bool equivalentAsOperands(Constant *L, Constant *R) {
369 // Use equality as a preliminary filter.
373 if (L->getValueID() != R->getValueID())
376 // Ask the engine about global values.
377 if (isa<GlobalValue>(L))
378 return Engine.equivalentAsOperands(cast<GlobalValue>(L),
379 cast<GlobalValue>(R));
381 // Compare constant expressions structurally.
382 if (isa<ConstantExpr>(L))
383 return equivalentAsOperands(cast<ConstantExpr>(L),
384 cast<ConstantExpr>(R));
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))
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();
400 bool equivalentAsOperands(ConstantExpr *L, ConstantExpr *R) {
403 if (L->getOpcode() != R->getOpcode())
406 switch (L->getOpcode()) {
407 case Instruction::ICmp:
408 case Instruction::FCmp:
409 if (L->getPredicate() != R->getPredicate())
413 case Instruction::GetElementPtr:
421 if (L->getNumOperands() != R->getNumOperands())
424 for (unsigned I = 0, E = L->getNumOperands(); I != E; ++I)
425 if (!equivalentAsOperands(L->getOperand(I), R->getOperand(I)))
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())
437 // Value subtypes: Argument, Constant, Instruction, BasicBlock,
438 // InlineAsm, MDNode, MDString, PseudoSourceValue
440 if (isa<Constant>(L))
441 return equivalentAsOperands(cast<Constant>(L), cast<Constant>(R));
443 if (isa<Instruction>(L))
444 return Values[L] == R || TentativeValues.count(std::make_pair(L, R));
446 if (isa<Argument>(L))
447 return Values[L] == R;
449 if (isa<BasicBlock>(L))
450 return Blocks[cast<BasicBlock>(L)] != R;
452 // Pretend everything else is identical.
456 // Avoid a gcc warning about accessing 'this' in an initializer.
457 FunctionDifferenceEngine *this_() { return this; }
460 FunctionDifferenceEngine(DifferenceEngine &Engine) :
461 Engine(Engine), Queue(QueueSorter(*this_())) {}
463 void diff(Function *L, Function *R) {
464 if (L->arg_size() != R->arg_size())
465 Engine.log("different argument counts");
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)
474 tryUnify(&*L->begin(), &*R->begin());
480 DiffEntry() : Cost(0) {}
483 llvm::SmallVector<char, 8> Path; // actually of DifferenceEngine::DiffChange
486 bool FunctionDifferenceEngine::matchForBlockDiff(Instruction *L,
488 return !diff(L, R, false, false);
491 void FunctionDifferenceEngine::runBlockDiff(BasicBlock::iterator LStart,
492 BasicBlock::iterator RStart) {
493 BasicBlock::iterator LE = LStart->getParent()->end();
494 BasicBlock::iterator RE = RStart->getParent()->end();
496 unsigned NL = std::distance(LStart, LE);
498 SmallVector<DiffEntry, 20> Paths1(NL+1);
499 SmallVector<DiffEntry, 20> Paths2(NL+1);
501 DiffEntry *Cur = Paths1.data();
502 DiffEntry *Next = Paths2.data();
504 const unsigned LeftCost = 2;
505 const unsigned RightCost = 2;
506 const unsigned MatchCost = 0;
508 assert(TentativeValues.empty());
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);
517 for (BasicBlock::iterator RI = RStart; RI != RE; ++RI) {
518 // Initialize the first row.
520 Next[0].Cost += RightCost;
521 Next[0].Path.push_back(DC_right);
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);
535 Next[Index] = Cur[Index];
536 Next[Index].Cost += RightCost;
537 Next[Index].Path.push_back(DC_right);
541 std::swap(Cur, Next);
544 // We don't need the tentative values anymore; everything from here
545 // on out should be non-tentative.
546 TentativeValues.clear();
548 SmallVectorImpl<char> &Path = Cur[NL].Path;
549 BasicBlock::iterator LI = LStart, RI = RStart;
551 DiffLogBuilder Diff(Engine.getConsumer());
553 // Drop trailing matches.
554 while (Path.back() == DC_match)
557 // Skip leading matches.
558 SmallVectorImpl<char>::iterator
559 PI = Path.begin(), PE = Path.end();
560 while (PI != PE && *PI == DC_match) {
565 for (; PI != PE; ++PI) {
566 switch (static_cast<DiffChange>(*PI)) {
568 assert(LI != LE && RI != RE);
570 Instruction *L = &*LI, *R = &*RI;
591 // Finishing unifying and complaining about the tails of the block,
592 // which should be matches all the way through.
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;
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()))
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;
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()))
627 if (!LInvoke->use_empty())
628 Values[LInvoke] = RCall;
629 tryUnify(LInvoke->getNormalDest(), RTerm->getSuccessor(0));
635 void DifferenceEngine::Oracle::anchor() { }
637 void DifferenceEngine::diff(Function *L, Function *R) {
638 Context C(*this, L, R);
641 // FIXME: attributes and CC
642 // FIXME: parameter attributes
644 // If both are declarations, we're done.
645 if (L->empty() && R->empty())
648 log("left function is declaration, right function is definition");
650 log("right function is declaration, left function is definition");
652 FunctionDifferenceEngine(*this).diff(L, R);
655 void DifferenceEngine::diff(Module *L, Module *R) {
657 SmallVector<std::pair<Function*,Function*>, 20> Queue;
659 for (Module::iterator I = L->begin(), E = L->end(); I != E; ++I) {
661 LNames.insert(LFn->getName());
663 if (Function *RFn = R->getFunction(LFn->getName()))
664 Queue.push_back(std::make_pair(LFn, RFn));
666 logf("function %l exists only in left module") << LFn;
669 for (Module::iterator I = R->begin(), E = R->end(); I != E; ++I) {
671 if (!LNames.count(RFn->getName()))
672 logf("function %r exists only in right module") << RFn;
675 for (SmallVectorImpl<std::pair<Function*,Function*> >::iterator
676 I = Queue.begin(), E = Queue.end(); I != E; ++I)
677 diff(I->first, I->second);
680 bool DifferenceEngine::equivalentAsOperands(GlobalValue *L, GlobalValue *R) {
681 if (globalValueOracle) return (*globalValueOracle)(L, R);
682 return L->getName() == R->getName();