1 //===- GlobalsModRef.cpp - Simple Mod/Ref Analysis for Globals ------------===//
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 simple pass provides alias and mod/ref information for global values
11 // that do not have their address taken, and keeps track of whether functions
12 // read or write memory (are "pure"). For this simple (but very common) case,
13 // we can provide pretty accurate and useful information.
15 //===----------------------------------------------------------------------===//
17 #define DEBUG_TYPE "globalsmodref-aa"
18 #include "llvm/Analysis/Passes.h"
19 #include "llvm/ADT/SCCIterator.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/AliasAnalysis.h"
22 #include "llvm/Analysis/CallGraph.h"
23 #include "llvm/Analysis/MemoryBuiltins.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/DerivedTypes.h"
27 #include "llvm/IR/Instructions.h"
28 #include "llvm/IR/IntrinsicInst.h"
29 #include "llvm/IR/Module.h"
30 #include "llvm/Pass.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/InstIterator.h"
36 STATISTIC(NumNonAddrTakenGlobalVars,
37 "Number of global vars without address taken");
38 STATISTIC(NumNonAddrTakenFunctions,"Number of functions without address taken");
39 STATISTIC(NumNoMemFunctions, "Number of functions that do not access memory");
40 STATISTIC(NumReadMemFunctions, "Number of functions that only read memory");
41 STATISTIC(NumIndirectGlobalVars, "Number of indirect global objects");
44 /// FunctionRecord - One instance of this structure is stored for every
45 /// function in the program. Later, the entries for these functions are
46 /// removed if the function is found to call an external function (in which
47 /// case we know nothing about it.
48 struct FunctionRecord {
49 /// GlobalInfo - Maintain mod/ref info for all of the globals without
50 /// addresses taken that are read or written (transitively) by this
52 std::map<const GlobalValue*, unsigned> GlobalInfo;
54 /// MayReadAnyGlobal - May read global variables, but it is not known which.
55 bool MayReadAnyGlobal;
57 unsigned getInfoForGlobal(const GlobalValue *GV) const {
58 unsigned Effect = MayReadAnyGlobal ? AliasAnalysis::Ref : 0;
59 std::map<const GlobalValue*, unsigned>::const_iterator I =
61 if (I != GlobalInfo.end())
66 /// FunctionEffect - Capture whether or not this function reads or writes to
67 /// ANY memory. If not, we can do a lot of aggressive analysis on it.
68 unsigned FunctionEffect;
70 FunctionRecord() : MayReadAnyGlobal (false), FunctionEffect(0) {}
73 /// GlobalsModRef - The actual analysis pass.
74 class GlobalsModRef : public ModulePass, public AliasAnalysis {
75 /// NonAddressTakenGlobals - The globals that do not have their addresses
77 std::set<const GlobalValue*> NonAddressTakenGlobals;
79 /// IndirectGlobals - The memory pointed to by this global is known to be
80 /// 'owned' by the global.
81 std::set<const GlobalValue*> IndirectGlobals;
83 /// AllocsForIndirectGlobals - If an instruction allocates memory for an
84 /// indirect global, this map indicates which one.
85 std::map<const Value*, const GlobalValue*> AllocsForIndirectGlobals;
87 /// FunctionInfo - For each function, keep track of what globals are
89 std::map<const Function*, FunctionRecord> FunctionInfo;
93 GlobalsModRef() : ModulePass(ID) {
94 initializeGlobalsModRefPass(*PassRegistry::getPassRegistry());
97 bool runOnModule(Module &M) {
98 InitializeAliasAnalysis(this);
100 // Find non-addr taken globals.
104 AnalyzeCallGraph(getAnalysis<CallGraphWrapperPass>().getCallGraph(), M);
108 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
109 AliasAnalysis::getAnalysisUsage(AU);
110 AU.addRequired<CallGraphWrapperPass>();
111 AU.setPreservesAll(); // Does not transform code
114 //------------------------------------------------
115 // Implement the AliasAnalysis API
117 AliasResult alias(const Location &LocA, const Location &LocB);
118 ModRefResult getModRefInfo(ImmutableCallSite CS,
119 const Location &Loc);
120 ModRefResult getModRefInfo(ImmutableCallSite CS1,
121 ImmutableCallSite CS2) {
122 return AliasAnalysis::getModRefInfo(CS1, CS2);
125 /// getModRefBehavior - Return the behavior of the specified function if
126 /// called from the specified call site. The call site may be null in which
127 /// case the most generic behavior of this function should be returned.
128 ModRefBehavior getModRefBehavior(const Function *F) {
129 ModRefBehavior Min = UnknownModRefBehavior;
131 if (FunctionRecord *FR = getFunctionInfo(F)) {
132 if (FR->FunctionEffect == 0)
133 Min = DoesNotAccessMemory;
134 else if ((FR->FunctionEffect & Mod) == 0)
135 Min = OnlyReadsMemory;
138 return ModRefBehavior(AliasAnalysis::getModRefBehavior(F) & Min);
141 /// getModRefBehavior - Return the behavior of the specified function if
142 /// called from the specified call site. The call site may be null in which
143 /// case the most generic behavior of this function should be returned.
144 ModRefBehavior getModRefBehavior(ImmutableCallSite CS) {
145 ModRefBehavior Min = UnknownModRefBehavior;
147 if (const Function* F = CS.getCalledFunction())
148 if (FunctionRecord *FR = getFunctionInfo(F)) {
149 if (FR->FunctionEffect == 0)
150 Min = DoesNotAccessMemory;
151 else if ((FR->FunctionEffect & Mod) == 0)
152 Min = OnlyReadsMemory;
155 return ModRefBehavior(AliasAnalysis::getModRefBehavior(CS) & Min);
158 virtual void deleteValue(Value *V);
159 virtual void copyValue(Value *From, Value *To);
160 virtual void addEscapingUse(Use &U);
162 /// getAdjustedAnalysisPointer - This method is used when a pass implements
163 /// an analysis interface through multiple inheritance. If needed, it
164 /// should override this to adjust the this pointer as needed for the
165 /// specified pass info.
166 virtual void *getAdjustedAnalysisPointer(AnalysisID PI) {
167 if (PI == &AliasAnalysis::ID)
168 return (AliasAnalysis*)this;
173 /// getFunctionInfo - Return the function info for the function, or null if
174 /// we don't have anything useful to say about it.
175 FunctionRecord *getFunctionInfo(const Function *F) {
176 std::map<const Function*, FunctionRecord>::iterator I =
177 FunctionInfo.find(F);
178 if (I != FunctionInfo.end())
183 void AnalyzeGlobals(Module &M);
184 void AnalyzeCallGraph(CallGraph &CG, Module &M);
185 bool AnalyzeUsesOfPointer(Value *V, std::vector<Function*> &Readers,
186 std::vector<Function*> &Writers,
187 GlobalValue *OkayStoreDest = 0);
188 bool AnalyzeIndirectGlobalMemory(GlobalValue *GV);
192 char GlobalsModRef::ID = 0;
193 INITIALIZE_AG_PASS_BEGIN(GlobalsModRef, AliasAnalysis,
194 "globalsmodref-aa", "Simple mod/ref analysis for globals",
196 INITIALIZE_PASS_DEPENDENCY(CallGraphWrapperPass)
197 INITIALIZE_AG_PASS_END(GlobalsModRef, AliasAnalysis,
198 "globalsmodref-aa", "Simple mod/ref analysis for globals",
201 Pass *llvm::createGlobalsModRefPass() { return new GlobalsModRef(); }
203 /// AnalyzeGlobals - Scan through the users of all of the internal
204 /// GlobalValue's in the program. If none of them have their "address taken"
205 /// (really, their address passed to something nontrivial), record this fact,
206 /// and record the functions that they are used directly in.
207 void GlobalsModRef::AnalyzeGlobals(Module &M) {
208 std::vector<Function*> Readers, Writers;
209 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
210 if (I->hasLocalLinkage()) {
211 if (!AnalyzeUsesOfPointer(I, Readers, Writers)) {
212 // Remember that we are tracking this global.
213 NonAddressTakenGlobals.insert(I);
214 ++NumNonAddrTakenFunctions;
216 Readers.clear(); Writers.clear();
219 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
221 if (I->hasLocalLinkage()) {
222 if (!AnalyzeUsesOfPointer(I, Readers, Writers)) {
223 // Remember that we are tracking this global, and the mod/ref fns
224 NonAddressTakenGlobals.insert(I);
226 for (unsigned i = 0, e = Readers.size(); i != e; ++i)
227 FunctionInfo[Readers[i]].GlobalInfo[I] |= Ref;
229 if (!I->isConstant()) // No need to keep track of writers to constants
230 for (unsigned i = 0, e = Writers.size(); i != e; ++i)
231 FunctionInfo[Writers[i]].GlobalInfo[I] |= Mod;
232 ++NumNonAddrTakenGlobalVars;
234 // If this global holds a pointer type, see if it is an indirect global.
235 if (I->getType()->getElementType()->isPointerTy() &&
236 AnalyzeIndirectGlobalMemory(I))
237 ++NumIndirectGlobalVars;
239 Readers.clear(); Writers.clear();
243 /// AnalyzeUsesOfPointer - Look at all of the users of the specified pointer.
244 /// If this is used by anything complex (i.e., the address escapes), return
245 /// true. Also, while we are at it, keep track of those functions that read and
246 /// write to the value.
248 /// If OkayStoreDest is non-null, stores into this global are allowed.
249 bool GlobalsModRef::AnalyzeUsesOfPointer(Value *V,
250 std::vector<Function*> &Readers,
251 std::vector<Function*> &Writers,
252 GlobalValue *OkayStoreDest) {
253 if (!V->getType()->isPointerTy()) return true;
255 for (Value::use_iterator UI = V->use_begin(), E=V->use_end(); UI != E; ++UI) {
257 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
258 Readers.push_back(LI->getParent()->getParent());
259 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
260 if (V == SI->getOperand(1)) {
261 Writers.push_back(SI->getParent()->getParent());
262 } else if (SI->getOperand(1) != OkayStoreDest) {
263 return true; // Storing the pointer
265 } else if (Operator::getOpcode(U) == Instruction::GetElementPtr) {
266 if (AnalyzeUsesOfPointer(U, Readers, Writers))
268 } else if (Operator::getOpcode(U) == Instruction::BitCast) {
269 if (AnalyzeUsesOfPointer(U, Readers, Writers, OkayStoreDest))
271 } else if (CallSite CS = U) {
272 // Make sure that this is just the function being called, not that it is
273 // passing into the function.
274 if (!CS.isCallee(UI)) {
275 // Detect calls to free.
276 if (isFreeCall(U, TLI))
277 Writers.push_back(CS->getParent()->getParent());
279 return true; // Argument of an unknown call.
281 } else if (ICmpInst *ICI = dyn_cast<ICmpInst>(U)) {
282 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
283 return true; // Allow comparison against null.
292 /// AnalyzeIndirectGlobalMemory - We found an non-address-taken global variable
293 /// which holds a pointer type. See if the global always points to non-aliased
294 /// heap memory: that is, all initializers of the globals are allocations, and
295 /// those allocations have no use other than initialization of the global.
296 /// Further, all loads out of GV must directly use the memory, not store the
297 /// pointer somewhere. If this is true, we consider the memory pointed to by
298 /// GV to be owned by GV and can disambiguate other pointers from it.
299 bool GlobalsModRef::AnalyzeIndirectGlobalMemory(GlobalValue *GV) {
300 // Keep track of values related to the allocation of the memory, f.e. the
301 // value produced by the malloc call and any casts.
302 std::vector<Value*> AllocRelatedValues;
304 // Walk the user list of the global. If we find anything other than a direct
305 // load or store, bail out.
306 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I){
308 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
309 // The pointer loaded from the global can only be used in simple ways:
310 // we allow addressing of it and loading storing to it. We do *not* allow
311 // storing the loaded pointer somewhere else or passing to a function.
312 std::vector<Function*> ReadersWriters;
313 if (AnalyzeUsesOfPointer(LI, ReadersWriters, ReadersWriters))
314 return false; // Loaded pointer escapes.
315 // TODO: Could try some IP mod/ref of the loaded pointer.
316 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
317 // Storing the global itself.
318 if (SI->getOperand(0) == GV) return false;
320 // If storing the null pointer, ignore it.
321 if (isa<ConstantPointerNull>(SI->getOperand(0)))
324 // Check the value being stored.
325 Value *Ptr = GetUnderlyingObject(SI->getOperand(0));
327 if (!isAllocLikeFn(Ptr, TLI))
328 return false; // Too hard to analyze.
330 // Analyze all uses of the allocation. If any of them are used in a
331 // non-simple way (e.g. stored to another global) bail out.
332 std::vector<Function*> ReadersWriters;
333 if (AnalyzeUsesOfPointer(Ptr, ReadersWriters, ReadersWriters, GV))
334 return false; // Loaded pointer escapes.
336 // Remember that this allocation is related to the indirect global.
337 AllocRelatedValues.push_back(Ptr);
339 // Something complex, bail out.
344 // Okay, this is an indirect global. Remember all of the allocations for
345 // this global in AllocsForIndirectGlobals.
346 while (!AllocRelatedValues.empty()) {
347 AllocsForIndirectGlobals[AllocRelatedValues.back()] = GV;
348 AllocRelatedValues.pop_back();
350 IndirectGlobals.insert(GV);
354 /// AnalyzeCallGraph - At this point, we know the functions where globals are
355 /// immediately stored to and read from. Propagate this information up the call
356 /// graph to all callers and compute the mod/ref info for all memory for each
358 void GlobalsModRef::AnalyzeCallGraph(CallGraph &CG, Module &M) {
359 // We do a bottom-up SCC traversal of the call graph. In other words, we
360 // visit all callees before callers (leaf-first).
361 for (scc_iterator<CallGraph*> I = scc_begin(&CG); !I.isAtEnd(); ++I) {
362 std::vector<CallGraphNode *> &SCC = *I;
363 assert(!SCC.empty() && "SCC with no functions?");
365 if (!SCC[0]->getFunction()) {
366 // Calls externally - can't say anything useful. Remove any existing
367 // function records (may have been created when scanning globals).
368 for (unsigned i = 0, e = SCC.size(); i != e; ++i)
369 FunctionInfo.erase(SCC[i]->getFunction());
373 FunctionRecord &FR = FunctionInfo[SCC[0]->getFunction()];
375 bool KnowNothing = false;
376 unsigned FunctionEffect = 0;
378 // Collect the mod/ref properties due to called functions. We only compute
380 for (unsigned i = 0, e = SCC.size(); i != e && !KnowNothing; ++i) {
381 Function *F = SCC[i]->getFunction();
387 if (F->isDeclaration()) {
388 // Try to get mod/ref behaviour from function attributes.
389 if (F->doesNotAccessMemory()) {
390 // Can't do better than that!
391 } else if (F->onlyReadsMemory()) {
392 FunctionEffect |= Ref;
393 if (!F->isIntrinsic())
394 // This function might call back into the module and read a global -
395 // consider every global as possibly being read by this function.
396 FR.MayReadAnyGlobal = true;
398 FunctionEffect |= ModRef;
399 // Can't say anything useful unless it's an intrinsic - they don't
400 // read or write global variables of the kind considered here.
401 KnowNothing = !F->isIntrinsic();
406 for (CallGraphNode::iterator CI = SCC[i]->begin(), E = SCC[i]->end();
407 CI != E && !KnowNothing; ++CI)
408 if (Function *Callee = CI->second->getFunction()) {
409 if (FunctionRecord *CalleeFR = getFunctionInfo(Callee)) {
410 // Propagate function effect up.
411 FunctionEffect |= CalleeFR->FunctionEffect;
413 // Incorporate callee's effects on globals into our info.
414 for (std::map<const GlobalValue*, unsigned>::iterator GI =
415 CalleeFR->GlobalInfo.begin(), E = CalleeFR->GlobalInfo.end();
417 FR.GlobalInfo[GI->first] |= GI->second;
418 FR.MayReadAnyGlobal |= CalleeFR->MayReadAnyGlobal;
420 // Can't say anything about it. However, if it is inside our SCC,
421 // then nothing needs to be done.
422 CallGraphNode *CalleeNode = CG[Callee];
423 if (std::find(SCC.begin(), SCC.end(), CalleeNode) == SCC.end())
431 // If we can't say anything useful about this SCC, remove all SCC functions
432 // from the FunctionInfo map.
434 for (unsigned i = 0, e = SCC.size(); i != e; ++i)
435 FunctionInfo.erase(SCC[i]->getFunction());
439 // Scan the function bodies for explicit loads or stores.
440 for (unsigned i = 0, e = SCC.size(); i != e && FunctionEffect != ModRef;++i)
441 for (inst_iterator II = inst_begin(SCC[i]->getFunction()),
442 E = inst_end(SCC[i]->getFunction());
443 II != E && FunctionEffect != ModRef; ++II)
444 if (LoadInst *LI = dyn_cast<LoadInst>(&*II)) {
445 FunctionEffect |= Ref;
446 if (LI->isVolatile())
447 // Volatile loads may have side-effects, so mark them as writing
448 // memory (for example, a flag inside the processor).
449 FunctionEffect |= Mod;
450 } else if (StoreInst *SI = dyn_cast<StoreInst>(&*II)) {
451 FunctionEffect |= Mod;
452 if (SI->isVolatile())
453 // Treat volatile stores as reading memory somewhere.
454 FunctionEffect |= Ref;
455 } else if (isAllocationFn(&*II, TLI) || isFreeCall(&*II, TLI)) {
456 FunctionEffect |= ModRef;
457 } else if (IntrinsicInst *Intrinsic = dyn_cast<IntrinsicInst>(&*II)) {
458 // The callgraph doesn't include intrinsic calls.
459 Function *Callee = Intrinsic->getCalledFunction();
460 ModRefBehavior Behaviour = AliasAnalysis::getModRefBehavior(Callee);
461 FunctionEffect |= (Behaviour & ModRef);
464 if ((FunctionEffect & Mod) == 0)
465 ++NumReadMemFunctions;
466 if (FunctionEffect == 0)
468 FR.FunctionEffect = FunctionEffect;
470 // Finally, now that we know the full effect on this SCC, clone the
471 // information to each function in the SCC.
472 for (unsigned i = 1, e = SCC.size(); i != e; ++i)
473 FunctionInfo[SCC[i]->getFunction()] = FR;
479 /// alias - If one of the pointers is to a global that we are tracking, and the
480 /// other is some random pointer, we know there cannot be an alias, because the
481 /// address of the global isn't taken.
482 AliasAnalysis::AliasResult
483 GlobalsModRef::alias(const Location &LocA,
484 const Location &LocB) {
485 // Get the base object these pointers point to.
486 const Value *UV1 = GetUnderlyingObject(LocA.Ptr);
487 const Value *UV2 = GetUnderlyingObject(LocB.Ptr);
489 // If either of the underlying values is a global, they may be non-addr-taken
490 // globals, which we can answer queries about.
491 const GlobalValue *GV1 = dyn_cast<GlobalValue>(UV1);
492 const GlobalValue *GV2 = dyn_cast<GlobalValue>(UV2);
494 // If the global's address is taken, pretend we don't know it's a pointer to
496 if (GV1 && !NonAddressTakenGlobals.count(GV1)) GV1 = 0;
497 if (GV2 && !NonAddressTakenGlobals.count(GV2)) GV2 = 0;
499 // If the two pointers are derived from two different non-addr-taken
500 // globals, or if one is and the other isn't, we know these can't alias.
501 if ((GV1 || GV2) && GV1 != GV2)
504 // Otherwise if they are both derived from the same addr-taken global, we
505 // can't know the two accesses don't overlap.
508 // These pointers may be based on the memory owned by an indirect global. If
509 // so, we may be able to handle this. First check to see if the base pointer
510 // is a direct load from an indirect global.
512 if (const LoadInst *LI = dyn_cast<LoadInst>(UV1))
513 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getOperand(0)))
514 if (IndirectGlobals.count(GV))
516 if (const LoadInst *LI = dyn_cast<LoadInst>(UV2))
517 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getOperand(0)))
518 if (IndirectGlobals.count(GV))
521 // These pointers may also be from an allocation for the indirect global. If
522 // so, also handle them.
523 if (AllocsForIndirectGlobals.count(UV1))
524 GV1 = AllocsForIndirectGlobals[UV1];
525 if (AllocsForIndirectGlobals.count(UV2))
526 GV2 = AllocsForIndirectGlobals[UV2];
528 // Now that we know whether the two pointers are related to indirect globals,
529 // use this to disambiguate the pointers. If either pointer is based on an
530 // indirect global and if they are not both based on the same indirect global,
531 // they cannot alias.
532 if ((GV1 || GV2) && GV1 != GV2)
535 return AliasAnalysis::alias(LocA, LocB);
538 AliasAnalysis::ModRefResult
539 GlobalsModRef::getModRefInfo(ImmutableCallSite CS,
540 const Location &Loc) {
541 unsigned Known = ModRef;
543 // If we are asking for mod/ref info of a direct call with a pointer to a
544 // global we are tracking, return information if we have it.
545 if (const GlobalValue *GV =
546 dyn_cast<GlobalValue>(GetUnderlyingObject(Loc.Ptr)))
547 if (GV->hasLocalLinkage())
548 if (const Function *F = CS.getCalledFunction())
549 if (NonAddressTakenGlobals.count(GV))
550 if (const FunctionRecord *FR = getFunctionInfo(F))
551 Known = FR->getInfoForGlobal(GV);
553 if (Known == NoModRef)
554 return NoModRef; // No need to query other mod/ref analyses
555 return ModRefResult(Known & AliasAnalysis::getModRefInfo(CS, Loc));
559 //===----------------------------------------------------------------------===//
560 // Methods to update the analysis as a result of the client transformation.
562 void GlobalsModRef::deleteValue(Value *V) {
563 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
564 if (NonAddressTakenGlobals.erase(GV)) {
565 // This global might be an indirect global. If so, remove it and remove
566 // any AllocRelatedValues for it.
567 if (IndirectGlobals.erase(GV)) {
568 // Remove any entries in AllocsForIndirectGlobals for this global.
569 for (std::map<const Value*, const GlobalValue*>::iterator
570 I = AllocsForIndirectGlobals.begin(),
571 E = AllocsForIndirectGlobals.end(); I != E; ) {
572 if (I->second == GV) {
573 AllocsForIndirectGlobals.erase(I++);
582 // Otherwise, if this is an allocation related to an indirect global, remove
584 AllocsForIndirectGlobals.erase(V);
586 AliasAnalysis::deleteValue(V);
589 void GlobalsModRef::copyValue(Value *From, Value *To) {
590 AliasAnalysis::copyValue(From, To);
593 void GlobalsModRef::addEscapingUse(Use &U) {
594 // For the purposes of this analysis, it is conservatively correct to treat
595 // a newly escaping value equivalently to a deleted one. We could perhaps
596 // be more precise by processing the new use and attempting to update our
597 // saved analysis results to accommodate it.
600 AliasAnalysis::addEscapingUse(U);