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 /// getNonLocalDependency - Perform a full dependency query for the
100 /// specified instruction, returning the set of blocks that the value is
101 /// potentially live across. The returned set of results will include a
102 /// "NonLocal" result for all blocks where the value is live across.
104 /// This method assumes the instruction returns a "nonlocal" dependency
105 /// within its own block.
107 void MemoryDependenceAnalysis::
108 getNonLocalDependency(Instruction *QueryInst,
109 SmallVectorImpl<std::pair<BasicBlock*,
110 MemDepResult> > &Result) {
111 assert(getDependency(QueryInst).isNonLocal() &&
112 "getNonLocalDependency should only be used on insts with non-local deps!");
113 DenseMap<BasicBlock*, DepResultTy> &Cache = NonLocalDeps[QueryInst];
115 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
116 /// the cached case, this can happen due to instructions being deleted etc. In
117 /// the uncached case, this starts out as the set of predecessors we care
119 SmallVector<BasicBlock*, 32> DirtyBlocks;
121 if (!Cache.empty()) {
122 // If we already have a partially computed set of results, scan them to
123 // determine what is dirty, seeding our initial DirtyBlocks worklist.
124 // FIXME: In the "don't need to be updated" case, this is expensive, why not
125 // have a per-"cache" flag saying it is undirty?
126 for (DenseMap<BasicBlock*, DepResultTy>::iterator I = Cache.begin(),
127 E = Cache.end(); I != E; ++I)
128 if (I->second.getInt() == Dirty)
129 DirtyBlocks.push_back(I->first);
133 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
134 // << Cache.size() << " cached: " << *QueryInst;
136 // Seed DirtyBlocks with each of the preds of QueryInst's block.
137 BasicBlock *QueryBB = QueryInst->getParent();
138 DirtyBlocks.append(pred_begin(QueryBB), pred_end(QueryBB));
139 NumUncacheNonLocal++;
142 // Iterate while we still have blocks to update.
143 while (!DirtyBlocks.empty()) {
144 BasicBlock *DirtyBB = DirtyBlocks.back();
145 DirtyBlocks.pop_back();
147 // Get the entry for this block. Note that this relies on DepResultTy
148 // default initializing to Dirty.
149 DepResultTy &DirtyBBEntry = Cache[DirtyBB];
151 // If DirtyBBEntry isn't dirty, it ended up on the worklist multiple times.
152 if (DirtyBBEntry.getInt() != Dirty) continue;
154 // Find out if this block has a local dependency for QueryInst.
155 // FIXME: If the dirty entry has an instruction pointer, scan from it!
156 // FIXME: Don't convert back and forth for MemDepResult <-> DepResultTy.
158 // If the dirty entry has a pointer, start scanning from it so we don't have
159 // to rescan the entire block.
160 BasicBlock::iterator ScanPos = DirtyBB->end();
161 if (Instruction *Inst = DirtyBBEntry.getPointer())
164 DirtyBBEntry = ConvFromResult(getDependencyFrom(QueryInst, ScanPos,
167 // If the block has a dependency (i.e. it isn't completely transparent to
168 // the value), remember it!
169 if (DirtyBBEntry.getInt() != NonLocal) {
170 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
171 // update this when we remove instructions.
172 if (Instruction *Inst = DirtyBBEntry.getPointer())
173 ReverseNonLocalDeps[Inst].insert(QueryInst);
177 // If the block *is* completely transparent to the load, we need to check
178 // the predecessors of this block. Add them to our worklist.
179 for (pred_iterator I = pred_begin(DirtyBB), E = pred_end(DirtyBB);
181 DirtyBlocks.push_back(*I);
184 // Copy the result into the output set.
185 for (DenseMap<BasicBlock*, DepResultTy>::iterator I = Cache.begin(),
186 E = Cache.end(); I != E; ++I)
187 Result.push_back(std::make_pair(I->first, ConvToResult(I->second)));
190 /// getDependency - Return the instruction on which a memory operation
191 /// depends. The local parameter indicates if the query should only
192 /// evaluate dependencies within the same basic block.
193 MemDepResult MemoryDependenceAnalysis::
194 getDependencyFrom(Instruction *QueryInst, BasicBlock::iterator ScanIt,
196 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
197 TargetData &TD = getAnalysis<TargetData>();
199 // Get the pointer value for which dependence will be determined
201 uint64_t MemSize = 0;
202 bool MemVolatile = false;
204 if (StoreInst* S = dyn_cast<StoreInst>(QueryInst)) {
205 MemPtr = S->getPointerOperand();
206 MemSize = TD.getTypeStoreSize(S->getOperand(0)->getType());
207 MemVolatile = S->isVolatile();
208 } else if (LoadInst* L = dyn_cast<LoadInst>(QueryInst)) {
209 MemPtr = L->getPointerOperand();
210 MemSize = TD.getTypeStoreSize(L->getType());
211 MemVolatile = L->isVolatile();
212 } else if (VAArgInst* V = dyn_cast<VAArgInst>(QueryInst)) {
213 MemPtr = V->getOperand(0);
214 MemSize = TD.getTypeStoreSize(V->getType());
215 } else if (FreeInst* F = dyn_cast<FreeInst>(QueryInst)) {
216 MemPtr = F->getPointerOperand();
217 // FreeInsts erase the entire structure, not just a field.
219 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst))
220 return getCallSiteDependency(CallSite::get(QueryInst), ScanIt, BB);
221 else // Non-memory instructions depend on nothing.
222 return MemDepResult::getNone();
224 // Walk backwards through the basic block, looking for dependencies
225 while (ScanIt != BB->begin()) {
226 Instruction *Inst = --ScanIt;
228 // If the access is volatile and this is a volatile load/store, return a
231 ((isa<LoadInst>(Inst) && cast<LoadInst>(Inst)->isVolatile()) ||
232 (isa<StoreInst>(Inst) && cast<StoreInst>(Inst)->isVolatile())))
233 return MemDepResult::get(Inst);
235 // MemDep is broken w.r.t. loads: it says that two loads of the same pointer
236 // depend on each other. :(
237 // FIXME: ELIMINATE THIS!
238 if (LoadInst *L = dyn_cast<LoadInst>(Inst)) {
239 Value *Pointer = L->getPointerOperand();
240 uint64_t PointerSize = TD.getTypeStoreSize(L->getType());
242 // If we found a pointer, check if it could be the same as our pointer
243 AliasAnalysis::AliasResult R =
244 AA.alias(Pointer, PointerSize, MemPtr, MemSize);
246 if (R == AliasAnalysis::NoAlias)
249 // May-alias loads don't depend on each other without a dependence.
250 if (isa<LoadInst>(QueryInst) && R == AliasAnalysis::MayAlias)
252 return MemDepResult::get(Inst);
255 // FIXME: This claims that an access depends on the allocation. This may
256 // make sense, but is dubious at best. It would be better to fix GVN to
257 // handle a 'None' Query.
258 if (AllocationInst *AI = dyn_cast<AllocationInst>(Inst)) {
260 uint64_t PointerSize;
261 if (ConstantInt *C = dyn_cast<ConstantInt>(AI->getArraySize()))
262 // Use ABI size (size between elements), not store size (size of one
263 // element without padding).
264 PointerSize = C->getZExtValue() *
265 TD.getABITypeSize(AI->getAllocatedType());
269 AliasAnalysis::AliasResult R =
270 AA.alias(Pointer, PointerSize, MemPtr, MemSize);
272 if (R == AliasAnalysis::NoAlias)
274 return MemDepResult::get(Inst);
278 // See if this instruction mod/ref's the pointer.
279 AliasAnalysis::ModRefResult MRR = AA.getModRefInfo(Inst, MemPtr, MemSize);
281 if (MRR == AliasAnalysis::NoModRef)
284 // Loads don't depend on read-only instructions.
285 if (isa<LoadInst>(QueryInst) && MRR == AliasAnalysis::Ref)
288 // Otherwise, there is a dependence.
289 return MemDepResult::get(Inst);
292 // If we found nothing, return the non-local flag.
293 return MemDepResult::getNonLocal();
296 /// getDependency - Return the instruction on which a memory operation
298 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
299 Instruction *ScanPos = QueryInst;
301 // Check for a cached result
302 DepResultTy &LocalCache = LocalDeps[QueryInst];
304 // If the cached entry is non-dirty, just return it. Note that this depends
305 // on DepResultTy's default constructing to 'dirty'.
306 if (LocalCache.getInt() != Dirty)
307 return ConvToResult(LocalCache);
309 // Otherwise, if we have a dirty entry, we know we can start the scan at that
310 // instruction, which may save us some work.
311 if (Instruction *Inst = LocalCache.getPointer())
316 getDependencyFrom(QueryInst, ScanPos, QueryInst->getParent());
318 // Remember the result!
319 // FIXME: Don't convert back and forth! Make a shared helper function.
320 LocalCache = ConvFromResult(Res);
321 if (Instruction *I = Res.getInst())
322 ReverseLocalDeps[I].insert(QueryInst);
328 /// dropInstruction - Remove an instruction from the analysis, making
329 /// absolutely conservative assumptions when updating the cache. This is
330 /// useful, for example when an instruction is changed rather than removed.
331 void MemoryDependenceAnalysis::dropInstruction(Instruction* drop) {
332 LocalDepMapType::iterator depGraphEntry = LocalDeps.find(drop);
333 if (depGraphEntry != LocalDeps.end())
334 if (Instruction *Inst = depGraphEntry->second.getPointer())
335 ReverseLocalDeps[Inst].erase(drop);
337 // Drop dependency information for things that depended on this instr
338 SmallPtrSet<Instruction*, 4>& set = ReverseLocalDeps[drop];
339 for (SmallPtrSet<Instruction*, 4>::iterator I = set.begin(), E = set.end();
343 LocalDeps.erase(drop);
344 ReverseLocalDeps.erase(drop);
346 for (DenseMap<BasicBlock*, DepResultTy>::iterator DI =
347 NonLocalDeps[drop].begin(), DE = NonLocalDeps[drop].end();
349 if (Instruction *Inst = DI->second.getPointer())
350 ReverseNonLocalDeps[Inst].erase(drop);
352 if (ReverseNonLocalDeps.count(drop)) {
353 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
355 SmallPtrSet<Instruction*, 4>& set =
356 ReverseNonLocalDeps[drop];
357 for (SmallPtrSet<Instruction*, 4>::iterator I = set.begin(), E = set.end();
359 for (DenseMap<BasicBlock*, DepResultTy>::iterator DI =
360 NonLocalDeps[*I].begin(), DE = NonLocalDeps[*I].end();
362 if (DI->second.getPointer() == drop) {
363 // Convert to a dirty entry for the subsequent instruction.
364 DI->second.setInt(Dirty);
365 if (drop->isTerminator())
366 DI->second.setPointer(0);
368 Instruction *NextI = next(BasicBlock::iterator(drop));
369 DI->second.setPointer(NextI);
370 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I));
374 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
375 while (!ReverseDepsToAdd.empty()) {
376 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
377 .insert(ReverseDepsToAdd.back().second);
378 ReverseDepsToAdd.pop_back();
382 ReverseNonLocalDeps.erase(drop);
383 NonLocalDeps.erase(drop);
386 /// removeInstruction - Remove an instruction from the dependence analysis,
387 /// updating the dependence of instructions that previously depended on it.
388 /// This method attempts to keep the cache coherent using the reverse map.
389 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
390 // Walk through the Non-local dependencies, removing this one as the value
391 // for any cached queries.
392 for (DenseMap<BasicBlock*, DepResultTy>::iterator DI =
393 NonLocalDeps[RemInst].begin(), DE = NonLocalDeps[RemInst].end();
395 if (Instruction *Inst = DI->second.getPointer())
396 ReverseNonLocalDeps[Inst].erase(RemInst);
398 // Shortly after this, we will look for things that depend on RemInst. In
399 // order to update these, we'll need a new dependency to base them on. We
400 // could completely delete any entries that depend on this, but it is better
401 // to make a more accurate approximation where possible. Compute that better
402 // approximation if we can.
403 DepResultTy NewDependency;
405 // If we have a cached local dependence query for this instruction, remove it.
407 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
408 if (LocalDepEntry != LocalDeps.end()) {
409 DepResultTy LocalDep = LocalDepEntry->second;
411 // Remove this local dependency info.
412 LocalDeps.erase(LocalDepEntry);
414 // Remove us from DepInst's reverse set now that the local dep info is gone.
415 if (Instruction *Inst = LocalDep.getPointer())
416 ReverseLocalDeps[Inst].erase(RemInst);
418 // If we have unconfirmed info, don't trust it.
419 if (LocalDep.getInt() != Dirty) {
420 // If we have a confirmed non-local flag, use it.
421 if (LocalDep.getInt() == NonLocal || LocalDep.getInt() == None) {
422 // The only time this dependency is confirmed is if it is non-local.
423 NewDependency = LocalDep;
425 // If we have dep info for RemInst, set them to it.
426 Instruction *NDI = next(BasicBlock::iterator(LocalDep.getPointer()));
427 if (NDI != RemInst) // Don't use RemInst for the new dependency!
428 NewDependency = DepResultTy(NDI, Dirty);
433 // If we don't already have a local dependency answer for this instruction,
434 // use the immediate successor of RemInst. We use the successor because
435 // getDependence starts by checking the immediate predecessor of what is in
437 if (NewDependency == DepResultTy(0, Dirty))
438 NewDependency = DepResultTy(next(BasicBlock::iterator(RemInst)), Dirty);
440 // Loop over all of the things that depend on the instruction we're removing.
442 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
443 if (ReverseDepIt != ReverseLocalDeps.end()) {
444 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second;
445 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(),
446 E = ReverseDeps.end(); I != E; ++I) {
447 Instruction *InstDependingOnRemInst = *I;
449 // If we thought the instruction depended on itself (possible for
450 // unconfirmed dependencies) ignore the update.
451 if (InstDependingOnRemInst == RemInst) continue;
453 // Insert the new dependencies.
454 LocalDeps[InstDependingOnRemInst] = NewDependency;
456 // If our NewDependency is an instruction, make sure to remember that new
457 // things depend on it.
458 if (Instruction *Inst = NewDependency.getPointer())
459 ReverseLocalDeps[Inst].insert(InstDependingOnRemInst);
461 ReverseLocalDeps.erase(RemInst);
464 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
465 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
466 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
468 SmallPtrSet<Instruction*, 4>& set = ReverseDepIt->second;
469 for (SmallPtrSet<Instruction*, 4>::iterator I = set.begin(), E = set.end();
471 for (DenseMap<BasicBlock*, DepResultTy>::iterator
472 DI = NonLocalDeps[*I].begin(), DE = NonLocalDeps[*I].end();
474 if (DI->second.getPointer() == RemInst) {
475 // Convert to a dirty entry for the subsequent instruction.
476 DI->second.setInt(Dirty);
477 if (RemInst->isTerminator())
478 DI->second.setPointer(0);
480 Instruction *NextI = next(BasicBlock::iterator(RemInst));
481 DI->second.setPointer(NextI);
482 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I));
486 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
487 while (!ReverseDepsToAdd.empty()) {
488 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
489 .insert(ReverseDepsToAdd.back().second);
490 ReverseDepsToAdd.pop_back();
493 ReverseNonLocalDeps.erase(ReverseDepIt);
496 NonLocalDeps.erase(RemInst);
498 getAnalysis<AliasAnalysis>().deleteValue(RemInst);
500 DEBUG(verifyRemoved(RemInst));
503 /// verifyRemoved - Verify that the specified instruction does not occur
504 /// in our internal data structures.
505 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
506 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
507 E = LocalDeps.end(); I != E; ++I) {
508 assert(I->first != D && "Inst occurs in data structures");
509 assert(I->second.getPointer() != D &&
510 "Inst occurs in data structures");
513 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
514 E = NonLocalDeps.end(); I != E; ++I) {
515 assert(I->first != D && "Inst occurs in data structures");
516 for (DenseMap<BasicBlock*, DepResultTy>::iterator II = I->second.begin(),
517 EE = I->second.end(); II != EE; ++II)
518 assert(II->second.getPointer() != D && "Inst occurs in data structures");
521 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
522 E = ReverseLocalDeps.end(); I != E; ++I)
523 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
524 EE = I->second.end(); II != EE; ++II)
525 assert(*II != D && "Inst occurs in data structures");
527 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
528 E = ReverseNonLocalDeps.end();
530 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
531 EE = I->second.end(); II != EE; ++II)
532 assert(*II != D && "Inst occurs in data structures");