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"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/DenseSet.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/ADT/StringRef.h"
20 #include "llvm/ADT/StringSet.h"
21 #include "llvm/IR/Constants.h"
22 #include "llvm/IR/Function.h"
23 #include "llvm/IR/Instructions.h"
24 #include "llvm/IR/Module.h"
25 #include "llvm/Support/CFG.h"
26 #include "llvm/Support/CallSite.h"
27 #include "llvm/Support/ErrorHandling.h"
28 #include "llvm/Support/raw_ostream.h"
29 #include "llvm/Support/type_traits.h"
36 /// A priority queue, implemented as a heap.
37 template <class T, class Sorter, unsigned InlineCapacity>
40 llvm::SmallVector<T, InlineCapacity> Storage;
43 PriorityQueue(const Sorter &Precedes) : Precedes(Precedes) {}
45 /// Checks whether the heap is empty.
46 bool empty() const { return Storage.empty(); }
48 /// Insert a new value on the heap.
49 void insert(const T &V) {
50 unsigned Index = Storage.size();
52 if (Index == 0) return;
54 T *data = Storage.data();
56 unsigned Target = (Index + 1) / 2 - 1;
57 if (!Precedes(data[Index], data[Target])) return;
58 std::swap(data[Index], data[Target]);
59 if (Target == 0) return;
64 /// Remove the minimum value in the heap. Only valid on a non-empty heap.
69 unsigned NewSize = Storage.size() - 1;
71 // Move the slot at the end to the beginning.
72 if (isPodLike<T>::value)
73 Storage[0] = Storage[NewSize];
75 std::swap(Storage[0], Storage[NewSize]);
77 // Bubble the root up as necessary.
80 // With a 1-based index, the children would be Index*2 and Index*2+1.
81 unsigned R = (Index + 1) * 2;
84 // If R is out of bounds, we're done after this in any case.
86 // If L is also out of bounds, we're done immediately.
87 if (L >= NewSize) break;
89 // Otherwise, test whether we should swap L and Index.
90 if (Precedes(Storage[L], Storage[Index]))
91 std::swap(Storage[L], Storage[Index]);
95 // Otherwise, we need to compare with the smaller of L and R.
96 // Prefer R because it's closer to the end of the array.
97 unsigned IndexToTest = (Precedes(Storage[L], Storage[R]) ? L : R);
99 // If Index is >= the min of L and R, then heap ordering is restored.
100 if (!Precedes(Storage[IndexToTest], Storage[Index]))
103 // Otherwise, keep bubbling up.
104 std::swap(Storage[IndexToTest], Storage[Index]);
114 /// A function-scope difference engine.
115 class FunctionDifferenceEngine {
116 DifferenceEngine &Engine;
118 /// The current mapping from old local values to new local values.
119 DenseMap<Value*, Value*> Values;
121 /// The current mapping from old blocks to new blocks.
122 DenseMap<BasicBlock*, BasicBlock*> Blocks;
124 DenseSet<std::pair<Value*, Value*> > TentativeValues;
126 unsigned getUnprocPredCount(BasicBlock *Block) const {
128 for (pred_iterator I = pred_begin(Block), E = pred_end(Block); I != E; ++I)
129 if (!Blocks.count(*I)) Count++;
133 typedef std::pair<BasicBlock*, BasicBlock*> BlockPair;
135 /// A type which sorts a priority queue by the number of unprocessed
136 /// predecessor blocks it has remaining.
138 /// This is actually really expensive to calculate.
140 const FunctionDifferenceEngine &fde;
141 explicit QueueSorter(const FunctionDifferenceEngine &fde) : fde(fde) {}
143 bool operator()(const BlockPair &Old, const BlockPair &New) {
144 return fde.getUnprocPredCount(Old.first)
145 < fde.getUnprocPredCount(New.first);
149 /// A queue of unified blocks to process.
150 PriorityQueue<BlockPair, QueueSorter, 20> Queue;
152 /// Try to unify the given two blocks. Enqueues them for processing
153 /// if they haven't already been processed.
155 /// Returns true if there was a problem unifying them.
156 bool tryUnify(BasicBlock *L, BasicBlock *R) {
157 BasicBlock *&Ref = Blocks[L];
160 if (Ref == R) return false;
162 Engine.logf("successor %l cannot be equivalent to %r; "
163 "it's already equivalent to %r")
169 Queue.insert(BlockPair(L, R));
173 /// Unifies two instructions, given that they're known not to have
174 /// structural differences.
175 void unify(Instruction *L, Instruction *R) {
176 DifferenceEngine::Context C(Engine, L, R);
178 bool Result = diff(L, R, true, true);
179 assert(!Result && "structural differences second time around?");
185 void processQueue() {
186 while (!Queue.empty()) {
187 BlockPair Pair = Queue.remove_min();
188 diff(Pair.first, Pair.second);
192 void diff(BasicBlock *L, BasicBlock *R) {
193 DifferenceEngine::Context C(Engine, L, R);
195 BasicBlock::iterator LI = L->begin(), LE = L->end();
196 BasicBlock::iterator RI = R->begin();
198 llvm::SmallVector<std::pair<Instruction*,Instruction*>, 20> TentativePairs;
201 assert(LI != LE && RI != R->end());
202 Instruction *LeftI = &*LI, *RightI = &*RI;
204 // If the instructions differ, start the more sophisticated diff
205 // algorithm at the start of the block.
206 if (diff(LeftI, RightI, false, false)) {
207 TentativeValues.clear();
208 return runBlockDiff(L->begin(), R->begin());
211 // Otherwise, tentatively unify them.
212 if (!LeftI->use_empty())
213 TentativeValues.insert(std::make_pair(LeftI, RightI));
216 } while (LI != LE); // This is sufficient: we can't get equality of
217 // terminators if there are residual instructions.
219 // Unify everything in the block, non-tentatively this time.
220 TentativeValues.clear();
221 for (LI = L->begin(), RI = R->begin(); LI != LE; ++LI, ++RI)
225 bool matchForBlockDiff(Instruction *L, Instruction *R);
226 void runBlockDiff(BasicBlock::iterator LI, BasicBlock::iterator RI);
228 bool diffCallSites(CallSite L, CallSite R, bool Complain) {
229 // FIXME: call attributes
230 if (!equivalentAsOperands(L.getCalledValue(), R.getCalledValue())) {
231 if (Complain) Engine.log("called functions differ");
234 if (L.arg_size() != R.arg_size()) {
235 if (Complain) Engine.log("argument counts differ");
238 for (unsigned I = 0, E = L.arg_size(); I != E; ++I)
239 if (!equivalentAsOperands(L.getArgument(I), R.getArgument(I))) {
241 Engine.logf("arguments %l and %r differ")
242 << L.getArgument(I) << R.getArgument(I);
248 bool diff(Instruction *L, Instruction *R, bool Complain, bool TryUnify) {
249 // FIXME: metadata (if Complain is set)
251 // Different opcodes always imply different operations.
252 if (L->getOpcode() != R->getOpcode()) {
253 if (Complain) Engine.log("different instruction types");
257 if (isa<CmpInst>(L)) {
258 if (cast<CmpInst>(L)->getPredicate()
259 != cast<CmpInst>(R)->getPredicate()) {
260 if (Complain) Engine.log("different predicates");
263 } else if (isa<CallInst>(L)) {
264 return diffCallSites(CallSite(L), CallSite(R), Complain);
265 } else if (isa<PHINode>(L)) {
268 // This is really weird; type uniquing is broken?
269 if (L->getType() != R->getType()) {
270 if (!L->getType()->isPointerTy() || !R->getType()->isPointerTy()) {
271 if (Complain) Engine.log("different phi types");
278 } else if (isa<InvokeInst>(L)) {
279 InvokeInst *LI = cast<InvokeInst>(L);
280 InvokeInst *RI = cast<InvokeInst>(R);
281 if (diffCallSites(CallSite(LI), CallSite(RI), Complain))
285 tryUnify(LI->getNormalDest(), RI->getNormalDest());
286 tryUnify(LI->getUnwindDest(), RI->getUnwindDest());
290 } else if (isa<BranchInst>(L)) {
291 BranchInst *LI = cast<BranchInst>(L);
292 BranchInst *RI = cast<BranchInst>(R);
293 if (LI->isConditional() != RI->isConditional()) {
294 if (Complain) Engine.log("branch conditionality differs");
298 if (LI->isConditional()) {
299 if (!equivalentAsOperands(LI->getCondition(), RI->getCondition())) {
300 if (Complain) Engine.log("branch conditions differ");
303 if (TryUnify) tryUnify(LI->getSuccessor(1), RI->getSuccessor(1));
305 if (TryUnify) tryUnify(LI->getSuccessor(0), RI->getSuccessor(0));
308 } else if (isa<SwitchInst>(L)) {
309 SwitchInst *LI = cast<SwitchInst>(L);
310 SwitchInst *RI = cast<SwitchInst>(R);
311 if (!equivalentAsOperands(LI->getCondition(), RI->getCondition())) {
312 if (Complain) Engine.log("switch conditions differ");
315 if (TryUnify) tryUnify(LI->getDefaultDest(), RI->getDefaultDest());
317 bool Difference = false;
319 DenseMap<ConstantInt*,BasicBlock*> LCases;
321 for (SwitchInst::CaseIt I = LI->case_begin(), E = LI->case_end();
323 LCases[I.getCaseValue()] = I.getCaseSuccessor();
325 for (SwitchInst::CaseIt I = RI->case_begin(), E = RI->case_end();
327 ConstantInt *CaseValue = I.getCaseValue();
328 BasicBlock *LCase = LCases[CaseValue];
330 if (TryUnify) tryUnify(LCase, I.getCaseSuccessor());
331 LCases.erase(CaseValue);
332 } else if (Complain || !Difference) {
334 Engine.logf("right switch has extra case %r") << CaseValue;
339 for (DenseMap<ConstantInt*,BasicBlock*>::iterator
340 I = LCases.begin(), E = LCases.end(); I != E; ++I) {
342 Engine.logf("left switch has extra case %l") << I->first;
346 } else if (isa<UnreachableInst>(L)) {
350 if (L->getNumOperands() != R->getNumOperands()) {
351 if (Complain) Engine.log("instructions have different operand counts");
355 for (unsigned I = 0, E = L->getNumOperands(); I != E; ++I) {
356 Value *LO = L->getOperand(I), *RO = R->getOperand(I);
357 if (!equivalentAsOperands(LO, RO)) {
358 if (Complain) Engine.logf("operands %l and %r differ") << LO << RO;
366 bool equivalentAsOperands(Constant *L, Constant *R) {
367 // Use equality as a preliminary filter.
371 if (L->getValueID() != R->getValueID())
374 // Ask the engine about global values.
375 if (isa<GlobalValue>(L))
376 return Engine.equivalentAsOperands(cast<GlobalValue>(L),
377 cast<GlobalValue>(R));
379 // Compare constant expressions structurally.
380 if (isa<ConstantExpr>(L))
381 return equivalentAsOperands(cast<ConstantExpr>(L),
382 cast<ConstantExpr>(R));
384 // Nulls of the "same type" don't always actually have the same
385 // type; I don't know why. Just white-list them.
386 if (isa<ConstantPointerNull>(L))
389 // Block addresses only match if we've already encountered the
390 // block. FIXME: tentative matches?
391 if (isa<BlockAddress>(L))
392 return Blocks[cast<BlockAddress>(L)->getBasicBlock()]
393 == cast<BlockAddress>(R)->getBasicBlock();
398 bool equivalentAsOperands(ConstantExpr *L, ConstantExpr *R) {
401 if (L->getOpcode() != R->getOpcode())
404 switch (L->getOpcode()) {
405 case Instruction::ICmp:
406 case Instruction::FCmp:
407 if (L->getPredicate() != R->getPredicate())
411 case Instruction::GetElementPtr:
419 if (L->getNumOperands() != R->getNumOperands())
422 for (unsigned I = 0, E = L->getNumOperands(); I != E; ++I)
423 if (!equivalentAsOperands(L->getOperand(I), R->getOperand(I)))
429 bool equivalentAsOperands(Value *L, Value *R) {
430 // Fall out if the values have different kind.
431 // This possibly shouldn't take priority over oracles.
432 if (L->getValueID() != R->getValueID())
435 // Value subtypes: Argument, Constant, Instruction, BasicBlock,
436 // InlineAsm, MDNode, MDString, PseudoSourceValue
438 if (isa<Constant>(L))
439 return equivalentAsOperands(cast<Constant>(L), cast<Constant>(R));
441 if (isa<Instruction>(L))
442 return Values[L] == R || TentativeValues.count(std::make_pair(L, R));
444 if (isa<Argument>(L))
445 return Values[L] == R;
447 if (isa<BasicBlock>(L))
448 return Blocks[cast<BasicBlock>(L)] != R;
450 // Pretend everything else is identical.
454 // Avoid a gcc warning about accessing 'this' in an initializer.
455 FunctionDifferenceEngine *this_() { return this; }
458 FunctionDifferenceEngine(DifferenceEngine &Engine) :
459 Engine(Engine), Queue(QueueSorter(*this_())) {}
461 void diff(Function *L, Function *R) {
462 if (L->arg_size() != R->arg_size())
463 Engine.log("different argument counts");
465 // Map the arguments.
466 for (Function::arg_iterator
467 LI = L->arg_begin(), LE = L->arg_end(),
468 RI = R->arg_begin(), RE = R->arg_end();
469 LI != LE && RI != RE; ++LI, ++RI)
472 tryUnify(&*L->begin(), &*R->begin());
478 DiffEntry() : Cost(0) {}
481 llvm::SmallVector<char, 8> Path; // actually of DifferenceEngine::DiffChange
484 bool FunctionDifferenceEngine::matchForBlockDiff(Instruction *L,
486 return !diff(L, R, false, false);
489 void FunctionDifferenceEngine::runBlockDiff(BasicBlock::iterator LStart,
490 BasicBlock::iterator RStart) {
491 BasicBlock::iterator LE = LStart->getParent()->end();
492 BasicBlock::iterator RE = RStart->getParent()->end();
494 unsigned NL = std::distance(LStart, LE);
496 SmallVector<DiffEntry, 20> Paths1(NL+1);
497 SmallVector<DiffEntry, 20> Paths2(NL+1);
499 DiffEntry *Cur = Paths1.data();
500 DiffEntry *Next = Paths2.data();
502 const unsigned LeftCost = 2;
503 const unsigned RightCost = 2;
504 const unsigned MatchCost = 0;
506 assert(TentativeValues.empty());
508 // Initialize the first column.
509 for (unsigned I = 0; I != NL+1; ++I) {
510 Cur[I].Cost = I * LeftCost;
511 for (unsigned J = 0; J != I; ++J)
512 Cur[I].Path.push_back(DC_left);
515 for (BasicBlock::iterator RI = RStart; RI != RE; ++RI) {
516 // Initialize the first row.
518 Next[0].Cost += RightCost;
519 Next[0].Path.push_back(DC_right);
522 for (BasicBlock::iterator LI = LStart; LI != LE; ++LI, ++Index) {
523 if (matchForBlockDiff(&*LI, &*RI)) {
524 Next[Index] = Cur[Index-1];
525 Next[Index].Cost += MatchCost;
526 Next[Index].Path.push_back(DC_match);
527 TentativeValues.insert(std::make_pair(&*LI, &*RI));
528 } else if (Next[Index-1].Cost <= Cur[Index].Cost) {
529 Next[Index] = Next[Index-1];
530 Next[Index].Cost += LeftCost;
531 Next[Index].Path.push_back(DC_left);
533 Next[Index] = Cur[Index];
534 Next[Index].Cost += RightCost;
535 Next[Index].Path.push_back(DC_right);
539 std::swap(Cur, Next);
542 // We don't need the tentative values anymore; everything from here
543 // on out should be non-tentative.
544 TentativeValues.clear();
546 SmallVectorImpl<char> &Path = Cur[NL].Path;
547 BasicBlock::iterator LI = LStart, RI = RStart;
549 DiffLogBuilder Diff(Engine.getConsumer());
551 // Drop trailing matches.
552 while (Path.back() == DC_match)
555 // Skip leading matches.
556 SmallVectorImpl<char>::iterator
557 PI = Path.begin(), PE = Path.end();
558 while (PI != PE && *PI == DC_match) {
563 for (; PI != PE; ++PI) {
564 switch (static_cast<DiffChange>(*PI)) {
566 assert(LI != LE && RI != RE);
568 Instruction *L = &*LI, *R = &*RI;
589 // Finishing unifying and complaining about the tails of the block,
590 // which should be matches all the way through.
597 // If the terminators have different kinds, but one is an invoke and the
598 // other is an unconditional branch immediately following a call, unify
599 // the results and the destinations.
600 TerminatorInst *LTerm = LStart->getParent()->getTerminator();
601 TerminatorInst *RTerm = RStart->getParent()->getTerminator();
602 if (isa<BranchInst>(LTerm) && isa<InvokeInst>(RTerm)) {
603 if (cast<BranchInst>(LTerm)->isConditional()) return;
604 BasicBlock::iterator I = LTerm;
605 if (I == LStart->getParent()->begin()) return;
607 if (!isa<CallInst>(*I)) return;
608 CallInst *LCall = cast<CallInst>(&*I);
609 InvokeInst *RInvoke = cast<InvokeInst>(RTerm);
610 if (!equivalentAsOperands(LCall->getCalledValue(), RInvoke->getCalledValue()))
612 if (!LCall->use_empty())
613 Values[LCall] = RInvoke;
614 tryUnify(LTerm->getSuccessor(0), RInvoke->getNormalDest());
615 } else if (isa<InvokeInst>(LTerm) && isa<BranchInst>(RTerm)) {
616 if (cast<BranchInst>(RTerm)->isConditional()) return;
617 BasicBlock::iterator I = RTerm;
618 if (I == RStart->getParent()->begin()) return;
620 if (!isa<CallInst>(*I)) return;
621 CallInst *RCall = cast<CallInst>(I);
622 InvokeInst *LInvoke = cast<InvokeInst>(LTerm);
623 if (!equivalentAsOperands(LInvoke->getCalledValue(), RCall->getCalledValue()))
625 if (!LInvoke->use_empty())
626 Values[LInvoke] = RCall;
627 tryUnify(LInvoke->getNormalDest(), RTerm->getSuccessor(0));
633 void DifferenceEngine::Oracle::anchor() { }
635 void DifferenceEngine::diff(Function *L, Function *R) {
636 Context C(*this, L, R);
639 // FIXME: attributes and CC
640 // FIXME: parameter attributes
642 // If both are declarations, we're done.
643 if (L->empty() && R->empty())
646 log("left function is declaration, right function is definition");
648 log("right function is declaration, left function is definition");
650 FunctionDifferenceEngine(*this).diff(L, R);
653 void DifferenceEngine::diff(Module *L, Module *R) {
655 SmallVector<std::pair<Function*,Function*>, 20> Queue;
657 for (Module::iterator I = L->begin(), E = L->end(); I != E; ++I) {
659 LNames.insert(LFn->getName());
661 if (Function *RFn = R->getFunction(LFn->getName()))
662 Queue.push_back(std::make_pair(LFn, RFn));
664 logf("function %l exists only in left module") << LFn;
667 for (Module::iterator I = R->begin(), E = R->end(); I != E; ++I) {
669 if (!LNames.count(RFn->getName()))
670 logf("function %r exists only in right module") << RFn;
673 for (SmallVectorImpl<std::pair<Function*,Function*> >::iterator
674 I = Queue.begin(), E = Queue.end(); I != E; ++I)
675 diff(I->first, I->second);
678 bool DifferenceEngine::equivalentAsOperands(GlobalValue *L, GlobalValue *R) {
679 if (globalValueOracle) return (*globalValueOracle)(L, R);
680 return L->getName() == R->getName();