1 //===- MemoryDependenceAnalysis.cpp - Mem Deps Implementation --*- C++ -*-===//
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 file implements an analysis that determines, for a given memory
11 // operation, what preceding memory operations it depends on. It builds on
12 // alias analysis information, and tries to provide a lazy, caching interface to
13 // a common kind of alias information query.
15 //===----------------------------------------------------------------------===//
17 #define DEBUG_TYPE "memdep"
18 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
19 #include "llvm/Constants.h"
20 #include "llvm/Instructions.h"
21 #include "llvm/Function.h"
22 #include "llvm/Analysis/AliasAnalysis.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/Support/CFG.h"
26 #include "llvm/Support/CommandLine.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Target/TargetData.h"
31 STATISTIC(NumCacheNonLocal, "Number of cached non-local responses");
32 STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses");
34 char MemoryDependenceAnalysis::ID = 0;
36 // Register this pass...
37 static RegisterPass<MemoryDependenceAnalysis> X("memdep",
38 "Memory Dependence Analysis", false, true);
40 /// getAnalysisUsage - Does not modify anything. It uses Alias Analysis.
42 void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
44 AU.addRequiredTransitive<AliasAnalysis>();
45 AU.addRequiredTransitive<TargetData>();
48 /// getCallSiteDependency - Private helper for finding the local dependencies
50 MemDepResult MemoryDependenceAnalysis::
51 getCallSiteDependency(CallSite C, BasicBlock::iterator ScanIt,
53 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
54 TargetData &TD = getAnalysis<TargetData>();
56 // Walk backwards through the block, looking for dependencies
57 while (ScanIt != BB->begin()) {
58 Instruction *Inst = --ScanIt;
60 // If this inst is a memory op, get the pointer it accessed
62 uint64_t PointerSize = 0;
63 if (StoreInst *S = dyn_cast<StoreInst>(Inst)) {
64 Pointer = S->getPointerOperand();
65 PointerSize = TD.getTypeStoreSize(S->getOperand(0)->getType());
66 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(Inst)) {
68 if (ConstantInt *C = dyn_cast<ConstantInt>(AI->getArraySize()))
69 // Use ABI size (size between elements), not store size (size of one
70 // element without padding).
71 PointerSize = C->getZExtValue() *
72 TD.getABITypeSize(AI->getAllocatedType());
75 } else if (VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
76 Pointer = V->getOperand(0);
77 PointerSize = TD.getTypeStoreSize(V->getType());
78 } else if (FreeInst *F = dyn_cast<FreeInst>(Inst)) {
79 Pointer = F->getPointerOperand();
81 // FreeInsts erase the entire structure
83 } else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
84 if (AA.getModRefBehavior(CallSite::get(Inst)) ==
85 AliasAnalysis::DoesNotAccessMemory)
87 return MemDepResult::get(Inst);
91 if (AA.getModRefInfo(C, Pointer, PointerSize) != AliasAnalysis::NoModRef)
92 return MemDepResult::get(Inst);
95 // No dependence found.
96 return MemDepResult::getNonLocal();
99 /// getDependency - Return the instruction on which a memory operation
100 /// depends. The local parameter indicates if the query should only
101 /// evaluate dependencies within the same basic block.
102 MemDepResult MemoryDependenceAnalysis::
103 getDependencyFrom(Instruction *QueryInst, BasicBlock::iterator ScanIt,
105 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
106 TargetData &TD = getAnalysis<TargetData>();
108 // Get the pointer value for which dependence will be determined
110 uint64_t MemSize = 0;
111 bool MemVolatile = false;
113 if (StoreInst* S = dyn_cast<StoreInst>(QueryInst)) {
114 MemPtr = S->getPointerOperand();
115 MemSize = TD.getTypeStoreSize(S->getOperand(0)->getType());
116 MemVolatile = S->isVolatile();
117 } else if (LoadInst* L = dyn_cast<LoadInst>(QueryInst)) {
118 MemPtr = L->getPointerOperand();
119 MemSize = TD.getTypeStoreSize(L->getType());
120 MemVolatile = L->isVolatile();
121 } else if (VAArgInst* V = dyn_cast<VAArgInst>(QueryInst)) {
122 MemPtr = V->getOperand(0);
123 MemSize = TD.getTypeStoreSize(V->getType());
124 } else if (FreeInst* F = dyn_cast<FreeInst>(QueryInst)) {
125 MemPtr = F->getPointerOperand();
126 // FreeInsts erase the entire structure, not just a field.
128 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst))
129 return getCallSiteDependency(CallSite::get(QueryInst), ScanIt, BB);
130 else // Non-memory instructions depend on nothing.
131 return MemDepResult::getNone();
133 // Walk backwards through the basic block, looking for dependencies
134 while (ScanIt != BB->begin()) {
135 Instruction *Inst = --ScanIt;
137 // If the access is volatile and this is a volatile load/store, return a
140 ((isa<LoadInst>(Inst) && cast<LoadInst>(Inst)->isVolatile()) ||
141 (isa<StoreInst>(Inst) && cast<StoreInst>(Inst)->isVolatile())))
142 return MemDepResult::get(Inst);
144 // MemDep is broken w.r.t. loads: it says that two loads of the same pointer
145 // depend on each other. :(
146 if (LoadInst *L = dyn_cast<LoadInst>(Inst)) {
147 Value *Pointer = L->getPointerOperand();
148 uint64_t PointerSize = TD.getTypeStoreSize(L->getType());
150 // If we found a pointer, check if it could be the same as our pointer
151 AliasAnalysis::AliasResult R =
152 AA.alias(Pointer, PointerSize, MemPtr, MemSize);
154 if (R == AliasAnalysis::NoAlias)
157 // May-alias loads don't depend on each other without a dependence.
158 if (isa<LoadInst>(QueryInst) && R == AliasAnalysis::MayAlias)
160 return MemDepResult::get(Inst);
163 // FIXME: This claims that an access depends on the allocation. This may
164 // make sense, but is dubious at best. It would be better to fix GVN to
165 // handle a 'None' Query.
166 if (AllocationInst *AI = dyn_cast<AllocationInst>(Inst)) {
168 uint64_t PointerSize;
169 if (ConstantInt *C = dyn_cast<ConstantInt>(AI->getArraySize()))
170 // Use ABI size (size between elements), not store size (size of one
171 // element without padding).
172 PointerSize = C->getZExtValue() *
173 TD.getABITypeSize(AI->getAllocatedType());
177 AliasAnalysis::AliasResult R =
178 AA.alias(Pointer, PointerSize, MemPtr, MemSize);
180 if (R == AliasAnalysis::NoAlias)
182 return MemDepResult::get(Inst);
186 // See if this instruction mod/ref's the pointer.
187 AliasAnalysis::ModRefResult MRR = AA.getModRefInfo(Inst, MemPtr, MemSize);
189 if (MRR == AliasAnalysis::NoModRef)
192 // Loads don't depend on read-only instructions.
193 if (isa<LoadInst>(QueryInst) && MRR == AliasAnalysis::Ref)
196 // Otherwise, there is a dependence.
197 return MemDepResult::get(Inst);
200 // If we found nothing, return the non-local flag.
201 return MemDepResult::getNonLocal();
204 /// getDependency - Return the instruction on which a memory operation
206 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
207 Instruction *ScanPos = QueryInst;
209 // Check for a cached result
210 DepResultTy &LocalCache = LocalDeps[QueryInst];
212 // If the cached entry is non-dirty, just return it. Note that this depends
213 // on DepResultTy's default constructing to 'dirty'.
214 if (LocalCache.getInt() != Dirty)
215 return ConvToResult(LocalCache);
217 // Otherwise, if we have a dirty entry, we know we can start the scan at that
218 // instruction, which may save us some work.
219 if (Instruction *Inst = LocalCache.getPointer())
224 getDependencyFrom(QueryInst, ScanPos, QueryInst->getParent());
226 // Remember the result!
227 // FIXME: Don't convert back and forth! Make a shared helper function.
228 LocalCache = ConvFromResult(Res);
229 if (Instruction *I = Res.getInst())
230 ReverseLocalDeps[I].insert(QueryInst);
235 /// getNonLocalDependency - Perform a full dependency query for the
236 /// specified instruction, returning the set of blocks that the value is
237 /// potentially live across. The returned set of results will include a
238 /// "NonLocal" result for all blocks where the value is live across.
240 /// This method assumes the instruction returns a "nonlocal" dependency
241 /// within its own block.
243 void MemoryDependenceAnalysis::
244 getNonLocalDependency(Instruction *QueryInst,
245 SmallVectorImpl<std::pair<BasicBlock*,
246 MemDepResult> > &Result) {
247 assert(getDependency(QueryInst).isNonLocal() &&
248 "getNonLocalDependency should only be used on insts with non-local deps!");
249 DenseMap<BasicBlock*, DepResultTy> &Cache = NonLocalDeps[QueryInst];
251 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
252 /// the cached case, this can happen due to instructions being deleted etc. In
253 /// the uncached case, this starts out as the set of predecessors we care
255 SmallVector<BasicBlock*, 32> DirtyBlocks;
257 if (!Cache.empty()) {
258 // If we already have a partially computed set of results, scan them to
259 // determine what is dirty, seeding our initial DirtyBlocks worklist.
260 // FIXME: In the "don't need to be updated" case, this is expensive, why not
261 // have a per-"cache" flag saying it is undirty?
262 for (DenseMap<BasicBlock*, DepResultTy>::iterator I = Cache.begin(),
263 E = Cache.end(); I != E; ++I)
264 if (I->second.getInt() == Dirty)
265 DirtyBlocks.push_back(I->first);
269 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
270 // << Cache.size() << " cached: " << *QueryInst;
272 // Seed DirtyBlocks with each of the preds of QueryInst's block.
273 BasicBlock *QueryBB = QueryInst->getParent();
274 DirtyBlocks.append(pred_begin(QueryBB), pred_end(QueryBB));
275 NumUncacheNonLocal++;
278 // Iterate while we still have blocks to update.
279 while (!DirtyBlocks.empty()) {
280 BasicBlock *DirtyBB = DirtyBlocks.back();
281 DirtyBlocks.pop_back();
283 // Get the entry for this block. Note that this relies on DepResultTy
284 // default initializing to Dirty.
285 DepResultTy &DirtyBBEntry = Cache[DirtyBB];
287 // If DirtyBBEntry isn't dirty, it ended up on the worklist multiple times.
288 if (DirtyBBEntry.getInt() != Dirty) continue;
290 // Find out if this block has a local dependency for QueryInst.
291 // FIXME: Don't convert back and forth for MemDepResult <-> DepResultTy.
293 // If the dirty entry has a pointer, start scanning from it so we don't have
294 // to rescan the entire block.
295 BasicBlock::iterator ScanPos = DirtyBB->end();
296 if (Instruction *Inst = DirtyBBEntry.getPointer())
299 DirtyBBEntry = ConvFromResult(getDependencyFrom(QueryInst, ScanPos,
302 // If the block has a dependency (i.e. it isn't completely transparent to
303 // the value), remember it!
304 if (DirtyBBEntry.getInt() != NonLocal) {
305 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
306 // update this when we remove instructions.
307 if (Instruction *Inst = DirtyBBEntry.getPointer())
308 ReverseNonLocalDeps[Inst].insert(QueryInst);
312 // If the block *is* completely transparent to the load, we need to check
313 // the predecessors of this block. Add them to our worklist.
314 DirtyBlocks.append(pred_begin(DirtyBB), pred_end(DirtyBB));
318 // Copy the result into the output set.
319 for (DenseMap<BasicBlock*, DepResultTy>::iterator I = Cache.begin(),
320 E = Cache.end(); I != E; ++I)
321 Result.push_back(std::make_pair(I->first, ConvToResult(I->second)));
324 /// removeInstruction - Remove an instruction from the dependence analysis,
325 /// updating the dependence of instructions that previously depended on it.
326 /// This method attempts to keep the cache coherent using the reverse map.
327 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
328 // Walk through the Non-local dependencies, removing this one as the value
329 // for any cached queries.
330 for (DenseMap<BasicBlock*, DepResultTy>::iterator DI =
331 NonLocalDeps[RemInst].begin(), DE = NonLocalDeps[RemInst].end();
333 if (Instruction *Inst = DI->second.getPointer())
334 ReverseNonLocalDeps[Inst].erase(RemInst);
336 // If we have a cached local dependence query for this instruction, remove it.
338 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
339 if (LocalDepEntry != LocalDeps.end()) {
340 // Remove us from DepInst's reverse set now that the local dep info is gone.
341 if (Instruction *Inst = LocalDepEntry->second.getPointer()) {
342 SmallPtrSet<Instruction*, 4> &RLD = ReverseLocalDeps[Inst];
345 ReverseLocalDeps.erase(Inst);
348 // Remove this local dependency info.
349 LocalDeps.erase(LocalDepEntry);
352 // Loop over all of the things that depend on the instruction we're removing.
354 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
356 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
357 if (ReverseDepIt != ReverseLocalDeps.end()) {
358 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second;
359 // RemInst can't be the terminator if it has stuff depending on it.
360 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) &&
361 "Nothing can locally depend on a terminator");
363 // Anything that was locally dependent on RemInst is now going to be
364 // dependent on the instruction after RemInst. It will have the dirty flag
365 // set so it will rescan. This saves having to scan the entire block to get
367 Instruction *NewDepInst = next(BasicBlock::iterator(RemInst));
369 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(),
370 E = ReverseDeps.end(); I != E; ++I) {
371 Instruction *InstDependingOnRemInst = *I;
373 // If we thought the instruction depended on itself (possible for
374 // unconfirmed dependencies) ignore the update.
375 if (InstDependingOnRemInst == RemInst) continue;
377 LocalDeps[InstDependingOnRemInst] = DepResultTy(NewDepInst, Dirty);
379 // Make sure to remember that new things depend on NewDepInst.
380 ReverseDepsToAdd.push_back(std::make_pair(NewDepInst,
381 InstDependingOnRemInst));
384 ReverseLocalDeps.erase(ReverseDepIt);
386 // Add new reverse deps after scanning the set, to avoid invalidating the
387 // 'ReverseDeps' reference.
388 while (!ReverseDepsToAdd.empty()) {
389 ReverseLocalDeps[ReverseDepsToAdd.back().first]
390 .insert(ReverseDepsToAdd.back().second);
391 ReverseDepsToAdd.pop_back();
395 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
396 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
397 SmallPtrSet<Instruction*, 4>& set = ReverseDepIt->second;
398 for (SmallPtrSet<Instruction*, 4>::iterator I = set.begin(), E = set.end();
400 for (DenseMap<BasicBlock*, DepResultTy>::iterator
401 DI = NonLocalDeps[*I].begin(), DE = NonLocalDeps[*I].end();
403 if (DI->second.getPointer() == RemInst) {
404 // Convert to a dirty entry for the subsequent instruction.
405 DI->second.setInt(Dirty);
406 if (RemInst->isTerminator())
407 DI->second.setPointer(0);
409 Instruction *NextI = next(BasicBlock::iterator(RemInst));
410 DI->second.setPointer(NextI);
411 assert(NextI != RemInst);
412 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I));
416 ReverseNonLocalDeps.erase(ReverseDepIt);
418 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
419 while (!ReverseDepsToAdd.empty()) {
420 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
421 .insert(ReverseDepsToAdd.back().second);
422 ReverseDepsToAdd.pop_back();
426 NonLocalDeps.erase(RemInst);
427 getAnalysis<AliasAnalysis>().deleteValue(RemInst);
428 DEBUG(verifyRemoved(RemInst));
431 /// verifyRemoved - Verify that the specified instruction does not occur
432 /// in our internal data structures.
433 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
434 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
435 E = LocalDeps.end(); I != E; ++I) {
436 assert(I->first != D && "Inst occurs in data structures");
437 assert(I->second.getPointer() != D &&
438 "Inst occurs in data structures");
441 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
442 E = NonLocalDeps.end(); I != E; ++I) {
443 assert(I->first != D && "Inst occurs in data structures");
444 for (DenseMap<BasicBlock*, DepResultTy>::iterator II = I->second.begin(),
445 EE = I->second.end(); II != EE; ++II)
446 assert(II->second.getPointer() != D && "Inst occurs in data structures");
449 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
450 E = ReverseLocalDeps.end(); I != E; ++I)
451 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
452 EE = I->second.end(); II != EE; ++II)
453 assert(*II != D && "Inst occurs in data structures");
455 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
456 E = ReverseNonLocalDeps.end();
458 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
459 EE = I->second.end(); II != EE; ++II)
460 assert(*II != D && "Inst occurs in data structures");