[PowerPC] Convert r+r instructions to r+i (pre and post RA)
[llvm-core.git] / tools / llvm-diff / DifferenceEngine.cpp
blob95a63d7f9c835c540088b5cde5d29f657f46bf2c
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.
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/StringSet.h"
20 #include "llvm/IR/CFG.h"
21 #include "llvm/IR/CallSite.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>
31 using namespace llvm;
33 namespace {
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;
41 public:
42 PriorityQueue(const Sorter &Precedes) : Precedes(Precedes) {}
44 /// Checks whether the heap is empty.
45 bool empty() const { return Storage.empty(); }
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;
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;
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];
68 unsigned NewSize = Storage.size() - 1;
69 if (NewSize) {
70 // Move the slot at the end to the beginning.
71 if (isPodLike<T>::value)
72 Storage[0] = Storage[NewSize];
73 else
74 std::swap(Storage[0], Storage[NewSize]);
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;
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;
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;
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);
98 // If Index is >= the min of L and R, then heap ordering is restored.
99 if (!Precedes(Storage[IndexToTest], Storage[Index]))
100 break;
102 // Otherwise, keep bubbling up.
103 std::swap(Storage[IndexToTest], Storage[Index]);
104 Index = IndexToTest;
107 Storage.pop_back();
109 return tmp;
113 /// A function-scope difference engine.
114 class FunctionDifferenceEngine {
115 DifferenceEngine &Engine;
117 /// The current mapping from old local values to new local values.
118 DenseMap<Value*, Value*> Values;
120 /// The current mapping from old blocks to new blocks.
121 DenseMap<BasicBlock*, BasicBlock*> Blocks;
123 DenseSet<std::pair<Value*, Value*> > TentativeValues;
125 unsigned getUnprocPredCount(BasicBlock *Block) const {
126 unsigned Count = 0;
127 for (pred_iterator I = pred_begin(Block), E = pred_end(Block); I != E; ++I)
128 if (!Blocks.count(*I)) Count++;
129 return Count;
132 typedef std::pair<BasicBlock*, BasicBlock*> BlockPair;
134 /// A type which sorts a priority queue by the number of unprocessed
135 /// predecessor blocks it has remaining.
137 /// This is actually really expensive to calculate.
138 struct QueueSorter {
139 const FunctionDifferenceEngine &fde;
140 explicit QueueSorter(const FunctionDifferenceEngine &fde) : fde(fde) {}
142 bool operator()(const BlockPair &Old, const BlockPair &New) {
143 return fde.getUnprocPredCount(Old.first)
144 < fde.getUnprocPredCount(New.first);
148 /// A queue of unified blocks to process.
149 PriorityQueue<BlockPair, QueueSorter, 20> Queue;
151 /// Try to unify the given two blocks. Enqueues them for processing
152 /// if they haven't already been processed.
154 /// Returns true if there was a problem unifying them.
155 bool tryUnify(BasicBlock *L, BasicBlock *R) {
156 BasicBlock *&Ref = Blocks[L];
158 if (Ref) {
159 if (Ref == R) return false;
161 Engine.logf("successor %l cannot be equivalent to %r; "
162 "it's already equivalent to %r")
163 << L << R << Ref;
164 return true;
167 Ref = R;
168 Queue.insert(BlockPair(L, R));
169 return false;
172 /// Unifies two instructions, given that they're known not to have
173 /// structural differences.
174 void unify(Instruction *L, Instruction *R) {
175 DifferenceEngine::Context C(Engine, L, R);
177 bool Result = diff(L, R, true, true);
178 assert(!Result && "structural differences second time around?");
179 (void) Result;
180 if (!L->use_empty())
181 Values[L] = R;
184 void processQueue() {
185 while (!Queue.empty()) {
186 BlockPair Pair = Queue.remove_min();
187 diff(Pair.first, Pair.second);
191 void diff(BasicBlock *L, BasicBlock *R) {
192 DifferenceEngine::Context C(Engine, L, R);
194 BasicBlock::iterator LI = L->begin(), LE = L->end();
195 BasicBlock::iterator RI = R->begin();
197 do {
198 assert(LI != LE && RI != R->end());
199 Instruction *LeftI = &*LI, *RightI = &*RI;
201 // If the instructions differ, start the more sophisticated diff
202 // algorithm at the start of the block.
203 if (diff(LeftI, RightI, false, false)) {
204 TentativeValues.clear();
205 return runBlockDiff(L->begin(), R->begin());
208 // Otherwise, tentatively unify them.
209 if (!LeftI->use_empty())
210 TentativeValues.insert(std::make_pair(LeftI, RightI));
212 ++LI;
213 ++RI;
214 } while (LI != LE); // This is sufficient: we can't get equality of
215 // terminators if there are residual instructions.
217 // Unify everything in the block, non-tentatively this time.
218 TentativeValues.clear();
219 for (LI = L->begin(), RI = R->begin(); LI != LE; ++LI, ++RI)
220 unify(&*LI, &*RI);
223 bool matchForBlockDiff(Instruction *L, Instruction *R);
224 void runBlockDiff(BasicBlock::iterator LI, BasicBlock::iterator RI);
226 bool diffCallSites(CallSite L, CallSite R, bool Complain) {
227 // FIXME: call attributes
228 if (!equivalentAsOperands(L.getCalledValue(), R.getCalledValue())) {
229 if (Complain) Engine.log("called functions differ");
230 return true;
232 if (L.arg_size() != R.arg_size()) {
233 if (Complain) Engine.log("argument counts differ");
234 return true;
236 for (unsigned I = 0, E = L.arg_size(); I != E; ++I)
237 if (!equivalentAsOperands(L.getArgument(I), R.getArgument(I))) {
238 if (Complain)
239 Engine.logf("arguments %l and %r differ")
240 << L.getArgument(I) << R.getArgument(I);
241 return true;
243 return false;
246 bool diff(Instruction *L, Instruction *R, bool Complain, bool TryUnify) {
247 // FIXME: metadata (if Complain is set)
249 // Different opcodes always imply different operations.
250 if (L->getOpcode() != R->getOpcode()) {
251 if (Complain) Engine.log("different instruction types");
252 return true;
255 if (isa<CmpInst>(L)) {
256 if (cast<CmpInst>(L)->getPredicate()
257 != cast<CmpInst>(R)->getPredicate()) {
258 if (Complain) Engine.log("different predicates");
259 return true;
261 } else if (isa<CallInst>(L)) {
262 return diffCallSites(CallSite(L), CallSite(R), Complain);
263 } else if (isa<PHINode>(L)) {
264 // FIXME: implement.
266 // This is really weird; type uniquing is broken?
267 if (L->getType() != R->getType()) {
268 if (!L->getType()->isPointerTy() || !R->getType()->isPointerTy()) {
269 if (Complain) Engine.log("different phi types");
270 return true;
273 return false;
275 // Terminators.
276 } else if (isa<InvokeInst>(L)) {
277 InvokeInst *LI = cast<InvokeInst>(L);
278 InvokeInst *RI = cast<InvokeInst>(R);
279 if (diffCallSites(CallSite(LI), CallSite(RI), Complain))
280 return true;
282 if (TryUnify) {
283 tryUnify(LI->getNormalDest(), RI->getNormalDest());
284 tryUnify(LI->getUnwindDest(), RI->getUnwindDest());
286 return false;
288 } else if (isa<BranchInst>(L)) {
289 BranchInst *LI = cast<BranchInst>(L);
290 BranchInst *RI = cast<BranchInst>(R);
291 if (LI->isConditional() != RI->isConditional()) {
292 if (Complain) Engine.log("branch conditionality differs");
293 return true;
296 if (LI->isConditional()) {
297 if (!equivalentAsOperands(LI->getCondition(), RI->getCondition())) {
298 if (Complain) Engine.log("branch conditions differ");
299 return true;
301 if (TryUnify) tryUnify(LI->getSuccessor(1), RI->getSuccessor(1));
303 if (TryUnify) tryUnify(LI->getSuccessor(0), RI->getSuccessor(0));
304 return false;
306 } else if (isa<SwitchInst>(L)) {
307 SwitchInst *LI = cast<SwitchInst>(L);
308 SwitchInst *RI = cast<SwitchInst>(R);
309 if (!equivalentAsOperands(LI->getCondition(), RI->getCondition())) {
310 if (Complain) Engine.log("switch conditions differ");
311 return true;
313 if (TryUnify) tryUnify(LI->getDefaultDest(), RI->getDefaultDest());
315 bool Difference = false;
317 DenseMap<ConstantInt*,BasicBlock*> LCases;
318 for (auto Case : LI->cases())
319 LCases[Case.getCaseValue()] = Case.getCaseSuccessor();
321 for (auto Case : RI->cases()) {
322 ConstantInt *CaseValue = Case.getCaseValue();
323 BasicBlock *LCase = LCases[CaseValue];
324 if (LCase) {
325 if (TryUnify)
326 tryUnify(LCase, Case.getCaseSuccessor());
327 LCases.erase(CaseValue);
328 } else if (Complain || !Difference) {
329 if (Complain)
330 Engine.logf("right switch has extra case %r") << CaseValue;
331 Difference = true;
334 if (!Difference)
335 for (DenseMap<ConstantInt*,BasicBlock*>::iterator
336 I = LCases.begin(), E = LCases.end(); I != E; ++I) {
337 if (Complain)
338 Engine.logf("left switch has extra case %l") << I->first;
339 Difference = true;
341 return Difference;
342 } else if (isa<UnreachableInst>(L)) {
343 return false;
346 if (L->getNumOperands() != R->getNumOperands()) {
347 if (Complain) Engine.log("instructions have different operand counts");
348 return true;
351 for (unsigned I = 0, E = L->getNumOperands(); I != E; ++I) {
352 Value *LO = L->getOperand(I), *RO = R->getOperand(I);
353 if (!equivalentAsOperands(LO, RO)) {
354 if (Complain) Engine.logf("operands %l and %r differ") << LO << RO;
355 return true;
359 return false;
362 bool equivalentAsOperands(Constant *L, Constant *R) {
363 // Use equality as a preliminary filter.
364 if (L == R)
365 return true;
367 if (L->getValueID() != R->getValueID())
368 return false;
370 // Ask the engine about global values.
371 if (isa<GlobalValue>(L))
372 return Engine.equivalentAsOperands(cast<GlobalValue>(L),
373 cast<GlobalValue>(R));
375 // Compare constant expressions structurally.
376 if (isa<ConstantExpr>(L))
377 return equivalentAsOperands(cast<ConstantExpr>(L),
378 cast<ConstantExpr>(R));
380 // Nulls of the "same type" don't always actually have the same
381 // type; I don't know why. Just white-list them.
382 if (isa<ConstantPointerNull>(L))
383 return true;
385 // Block addresses only match if we've already encountered the
386 // block. FIXME: tentative matches?
387 if (isa<BlockAddress>(L))
388 return Blocks[cast<BlockAddress>(L)->getBasicBlock()]
389 == cast<BlockAddress>(R)->getBasicBlock();
391 return false;
394 bool equivalentAsOperands(ConstantExpr *L, ConstantExpr *R) {
395 if (L == R)
396 return true;
397 if (L->getOpcode() != R->getOpcode())
398 return false;
400 switch (L->getOpcode()) {
401 case Instruction::ICmp:
402 case Instruction::FCmp:
403 if (L->getPredicate() != R->getPredicate())
404 return false;
405 break;
407 case Instruction::GetElementPtr:
408 // FIXME: inbounds?
409 break;
411 default:
412 break;
415 if (L->getNumOperands() != R->getNumOperands())
416 return false;
418 for (unsigned I = 0, E = L->getNumOperands(); I != E; ++I)
419 if (!equivalentAsOperands(L->getOperand(I), R->getOperand(I)))
420 return false;
422 return true;
425 bool equivalentAsOperands(Value *L, Value *R) {
426 // Fall out if the values have different kind.
427 // This possibly shouldn't take priority over oracles.
428 if (L->getValueID() != R->getValueID())
429 return false;
431 // Value subtypes: Argument, Constant, Instruction, BasicBlock,
432 // InlineAsm, MDNode, MDString, PseudoSourceValue
434 if (isa<Constant>(L))
435 return equivalentAsOperands(cast<Constant>(L), cast<Constant>(R));
437 if (isa<Instruction>(L))
438 return Values[L] == R || TentativeValues.count(std::make_pair(L, R));
440 if (isa<Argument>(L))
441 return Values[L] == R;
443 if (isa<BasicBlock>(L))
444 return Blocks[cast<BasicBlock>(L)] != R;
446 // Pretend everything else is identical.
447 return true;
450 // Avoid a gcc warning about accessing 'this' in an initializer.
451 FunctionDifferenceEngine *this_() { return this; }
453 public:
454 FunctionDifferenceEngine(DifferenceEngine &Engine) :
455 Engine(Engine), Queue(QueueSorter(*this_())) {}
457 void diff(Function *L, Function *R) {
458 if (L->arg_size() != R->arg_size())
459 Engine.log("different argument counts");
461 // Map the arguments.
462 for (Function::arg_iterator
463 LI = L->arg_begin(), LE = L->arg_end(),
464 RI = R->arg_begin(), RE = R->arg_end();
465 LI != LE && RI != RE; ++LI, ++RI)
466 Values[&*LI] = &*RI;
468 tryUnify(&*L->begin(), &*R->begin());
469 processQueue();
473 struct DiffEntry {
474 DiffEntry() : Cost(0) {}
476 unsigned Cost;
477 llvm::SmallVector<char, 8> Path; // actually of DifferenceEngine::DiffChange
480 bool FunctionDifferenceEngine::matchForBlockDiff(Instruction *L,
481 Instruction *R) {
482 return !diff(L, R, false, false);
485 void FunctionDifferenceEngine::runBlockDiff(BasicBlock::iterator LStart,
486 BasicBlock::iterator RStart) {
487 BasicBlock::iterator LE = LStart->getParent()->end();
488 BasicBlock::iterator RE = RStart->getParent()->end();
490 unsigned NL = std::distance(LStart, LE);
492 SmallVector<DiffEntry, 20> Paths1(NL+1);
493 SmallVector<DiffEntry, 20> Paths2(NL+1);
495 DiffEntry *Cur = Paths1.data();
496 DiffEntry *Next = Paths2.data();
498 const unsigned LeftCost = 2;
499 const unsigned RightCost = 2;
500 const unsigned MatchCost = 0;
502 assert(TentativeValues.empty());
504 // Initialize the first column.
505 for (unsigned I = 0; I != NL+1; ++I) {
506 Cur[I].Cost = I * LeftCost;
507 for (unsigned J = 0; J != I; ++J)
508 Cur[I].Path.push_back(DC_left);
511 for (BasicBlock::iterator RI = RStart; RI != RE; ++RI) {
512 // Initialize the first row.
513 Next[0] = Cur[0];
514 Next[0].Cost += RightCost;
515 Next[0].Path.push_back(DC_right);
517 unsigned Index = 1;
518 for (BasicBlock::iterator LI = LStart; LI != LE; ++LI, ++Index) {
519 if (matchForBlockDiff(&*LI, &*RI)) {
520 Next[Index] = Cur[Index-1];
521 Next[Index].Cost += MatchCost;
522 Next[Index].Path.push_back(DC_match);
523 TentativeValues.insert(std::make_pair(&*LI, &*RI));
524 } else if (Next[Index-1].Cost <= Cur[Index].Cost) {
525 Next[Index] = Next[Index-1];
526 Next[Index].Cost += LeftCost;
527 Next[Index].Path.push_back(DC_left);
528 } else {
529 Next[Index] = Cur[Index];
530 Next[Index].Cost += RightCost;
531 Next[Index].Path.push_back(DC_right);
535 std::swap(Cur, Next);
538 // We don't need the tentative values anymore; everything from here
539 // on out should be non-tentative.
540 TentativeValues.clear();
542 SmallVectorImpl<char> &Path = Cur[NL].Path;
543 BasicBlock::iterator LI = LStart, RI = RStart;
545 DiffLogBuilder Diff(Engine.getConsumer());
547 // Drop trailing matches.
548 while (Path.back() == DC_match)
549 Path.pop_back();
551 // Skip leading matches.
552 SmallVectorImpl<char>::iterator
553 PI = Path.begin(), PE = Path.end();
554 while (PI != PE && *PI == DC_match) {
555 unify(&*LI, &*RI);
556 ++PI;
557 ++LI;
558 ++RI;
561 for (; PI != PE; ++PI) {
562 switch (static_cast<DiffChange>(*PI)) {
563 case DC_match:
564 assert(LI != LE && RI != RE);
566 Instruction *L = &*LI, *R = &*RI;
567 unify(L, R);
568 Diff.addMatch(L, R);
570 ++LI; ++RI;
571 break;
573 case DC_left:
574 assert(LI != LE);
575 Diff.addLeft(&*LI);
576 ++LI;
577 break;
579 case DC_right:
580 assert(RI != RE);
581 Diff.addRight(&*RI);
582 ++RI;
583 break;
587 // Finishing unifying and complaining about the tails of the block,
588 // which should be matches all the way through.
589 while (LI != LE) {
590 assert(RI != RE);
591 unify(&*LI, &*RI);
592 ++LI;
593 ++RI;
596 // If the terminators have different kinds, but one is an invoke and the
597 // other is an unconditional branch immediately following a call, unify
598 // the results and the destinations.
599 TerminatorInst *LTerm = LStart->getParent()->getTerminator();
600 TerminatorInst *RTerm = RStart->getParent()->getTerminator();
601 if (isa<BranchInst>(LTerm) && isa<InvokeInst>(RTerm)) {
602 if (cast<BranchInst>(LTerm)->isConditional()) return;
603 BasicBlock::iterator I = LTerm->getIterator();
604 if (I == LStart->getParent()->begin()) return;
605 --I;
606 if (!isa<CallInst>(*I)) return;
607 CallInst *LCall = cast<CallInst>(&*I);
608 InvokeInst *RInvoke = cast<InvokeInst>(RTerm);
609 if (!equivalentAsOperands(LCall->getCalledValue(), RInvoke->getCalledValue()))
610 return;
611 if (!LCall->use_empty())
612 Values[LCall] = RInvoke;
613 tryUnify(LTerm->getSuccessor(0), RInvoke->getNormalDest());
614 } else if (isa<InvokeInst>(LTerm) && isa<BranchInst>(RTerm)) {
615 if (cast<BranchInst>(RTerm)->isConditional()) return;
616 BasicBlock::iterator I = RTerm->getIterator();
617 if (I == RStart->getParent()->begin()) return;
618 --I;
619 if (!isa<CallInst>(*I)) return;
620 CallInst *RCall = cast<CallInst>(I);
621 InvokeInst *LInvoke = cast<InvokeInst>(LTerm);
622 if (!equivalentAsOperands(LInvoke->getCalledValue(), RCall->getCalledValue()))
623 return;
624 if (!LInvoke->use_empty())
625 Values[LInvoke] = RCall;
626 tryUnify(LInvoke->getNormalDest(), RTerm->getSuccessor(0));
632 void DifferenceEngine::Oracle::anchor() { }
634 void DifferenceEngine::diff(Function *L, Function *R) {
635 Context C(*this, L, R);
637 // FIXME: types
638 // FIXME: attributes and CC
639 // FIXME: parameter attributes
641 // If both are declarations, we're done.
642 if (L->empty() && R->empty())
643 return;
644 else if (L->empty())
645 log("left function is declaration, right function is definition");
646 else if (R->empty())
647 log("right function is declaration, left function is definition");
648 else
649 FunctionDifferenceEngine(*this).diff(L, R);
652 void DifferenceEngine::diff(Module *L, Module *R) {
653 StringSet<> LNames;
654 SmallVector<std::pair<Function*,Function*>, 20> Queue;
656 for (Module::iterator I = L->begin(), E = L->end(); I != E; ++I) {
657 Function *LFn = &*I;
658 LNames.insert(LFn->getName());
660 if (Function *RFn = R->getFunction(LFn->getName()))
661 Queue.push_back(std::make_pair(LFn, RFn));
662 else
663 logf("function %l exists only in left module") << LFn;
666 for (Module::iterator I = R->begin(), E = R->end(); I != E; ++I) {
667 Function *RFn = &*I;
668 if (!LNames.count(RFn->getName()))
669 logf("function %r exists only in right module") << RFn;
672 for (SmallVectorImpl<std::pair<Function*,Function*> >::iterator
673 I = Queue.begin(), E = Queue.end(); I != E; ++I)
674 diff(I->first, I->second);
677 bool DifferenceEngine::equivalentAsOperands(GlobalValue *L, GlobalValue *R) {
678 if (globalValueOracle) return (*globalValueOracle)(L, R);
679 return L->getName() == R->getName();