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/Module.h"
20 #include "llvm/Pass.h"
21 #include "llvm/Instructions.h"
22 #include "llvm/Constants.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/Analysis/AliasAnalysis.h"
25 #include "llvm/Analysis/CallGraph.h"
26 #include "llvm/Analysis/MemoryBuiltins.h"
27 #include "llvm/Support/CommandLine.h"
28 #include "llvm/Support/InstIterator.h"
29 #include "llvm/ADT/Statistic.h"
30 #include "llvm/ADT/SCCIterator.h"
34 STATISTIC(NumNonAddrTakenGlobalVars,
35 "Number of global vars without address taken");
36 STATISTIC(NumNonAddrTakenFunctions,"Number of functions without address taken");
37 STATISTIC(NumNoMemFunctions, "Number of functions that do not access memory");
38 STATISTIC(NumReadMemFunctions, "Number of functions that only read memory");
39 STATISTIC(NumIndirectGlobalVars, "Number of indirect global objects");
42 /// FunctionRecord - One instance of this structure is stored for every
43 /// function in the program. Later, the entries for these functions are
44 /// removed if the function is found to call an external function (in which
45 /// case we know nothing about it.
46 struct FunctionRecord {
47 /// GlobalInfo - Maintain mod/ref info for all of the globals without
48 /// addresses taken that are read or written (transitively) by this
50 std::map<const GlobalValue*, unsigned> GlobalInfo;
52 /// MayReadAnyGlobal - May read global variables, but it is not known which.
53 bool MayReadAnyGlobal;
55 unsigned getInfoForGlobal(const GlobalValue *GV) const {
56 unsigned Effect = MayReadAnyGlobal ? AliasAnalysis::Ref : 0;
57 std::map<const GlobalValue*, unsigned>::const_iterator I =
59 if (I != GlobalInfo.end())
64 /// FunctionEffect - Capture whether or not this function reads or writes to
65 /// ANY memory. If not, we can do a lot of aggressive analysis on it.
66 unsigned FunctionEffect;
68 FunctionRecord() : MayReadAnyGlobal (false), FunctionEffect(0) {}
71 /// GlobalsModRef - The actual analysis pass.
72 class GlobalsModRef : public ModulePass, public AliasAnalysis {
73 /// NonAddressTakenGlobals - The globals that do not have their addresses
75 std::set<const GlobalValue*> NonAddressTakenGlobals;
77 /// IndirectGlobals - The memory pointed to by this global is known to be
78 /// 'owned' by the global.
79 std::set<const GlobalValue*> IndirectGlobals;
81 /// AllocsForIndirectGlobals - If an instruction allocates memory for an
82 /// indirect global, this map indicates which one.
83 std::map<const Value*, const GlobalValue*> AllocsForIndirectGlobals;
85 /// FunctionInfo - For each function, keep track of what globals are
87 std::map<const Function*, FunctionRecord> FunctionInfo;
91 GlobalsModRef() : ModulePass(&ID) {}
93 bool runOnModule(Module &M) {
94 InitializeAliasAnalysis(this); // set up super class
95 AnalyzeGlobals(M); // find non-addr taken globals
96 AnalyzeCallGraph(getAnalysis<CallGraph>(), M); // Propagate on CG
100 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
101 AliasAnalysis::getAnalysisUsage(AU);
102 AU.addRequired<CallGraph>();
103 AU.setPreservesAll(); // Does not transform code
106 //------------------------------------------------
107 // Implement the AliasAnalysis API
109 AliasResult alias(const Value *V1, unsigned V1Size,
110 const Value *V2, unsigned V2Size);
111 ModRefResult getModRefInfo(ImmutableCallSite CS,
112 const Value *P, unsigned Size);
113 ModRefResult getModRefInfo(ImmutableCallSite CS1,
114 ImmutableCallSite CS2) {
115 return AliasAnalysis::getModRefInfo(CS1, CS2);
118 /// getModRefBehavior - Return the behavior of the specified function if
119 /// called from the specified call site. The call site may be null in which
120 /// case the most generic behavior of this function should be returned.
121 ModRefBehavior getModRefBehavior(const Function *F,
122 std::vector<PointerAccessInfo> *Info) {
123 if (FunctionRecord *FR = getFunctionInfo(F)) {
124 if (FR->FunctionEffect == 0)
125 return DoesNotAccessMemory;
126 else if ((FR->FunctionEffect & Mod) == 0)
127 return OnlyReadsMemory;
129 return AliasAnalysis::getModRefBehavior(F, Info);
132 /// getModRefBehavior - Return the behavior of the specified function if
133 /// called from the specified call site. The call site may be null in which
134 /// case the most generic behavior of this function should be returned.
135 ModRefBehavior getModRefBehavior(ImmutableCallSite CS,
136 std::vector<PointerAccessInfo> *Info) {
137 const Function* F = CS.getCalledFunction();
138 if (!F) return AliasAnalysis::getModRefBehavior(CS, Info);
139 if (FunctionRecord *FR = getFunctionInfo(F)) {
140 if (FR->FunctionEffect == 0)
141 return DoesNotAccessMemory;
142 else if ((FR->FunctionEffect & Mod) == 0)
143 return OnlyReadsMemory;
145 return AliasAnalysis::getModRefBehavior(CS, Info);
148 virtual void deleteValue(Value *V);
149 virtual void copyValue(Value *From, Value *To);
151 /// getAdjustedAnalysisPointer - This method is used when a pass implements
152 /// an analysis interface through multiple inheritance. If needed, it
153 /// should override this to adjust the this pointer as needed for the
154 /// specified pass info.
155 virtual void *getAdjustedAnalysisPointer(const PassInfo *PI) {
156 if (PI->isPassID(&AliasAnalysis::ID))
157 return (AliasAnalysis*)this;
162 /// getFunctionInfo - Return the function info for the function, or null if
163 /// we don't have anything useful to say about it.
164 FunctionRecord *getFunctionInfo(const Function *F) {
165 std::map<const Function*, FunctionRecord>::iterator I =
166 FunctionInfo.find(F);
167 if (I != FunctionInfo.end())
172 void AnalyzeGlobals(Module &M);
173 void AnalyzeCallGraph(CallGraph &CG, Module &M);
174 bool AnalyzeUsesOfPointer(Value *V, std::vector<Function*> &Readers,
175 std::vector<Function*> &Writers,
176 GlobalValue *OkayStoreDest = 0);
177 bool AnalyzeIndirectGlobalMemory(GlobalValue *GV);
181 char GlobalsModRef::ID = 0;
182 static RegisterPass<GlobalsModRef>
183 X("globalsmodref-aa", "Simple mod/ref analysis for globals", false, true);
184 static RegisterAnalysisGroup<AliasAnalysis> Y(X);
186 Pass *llvm::createGlobalsModRefPass() { return new GlobalsModRef(); }
188 /// AnalyzeGlobals - Scan through the users of all of the internal
189 /// GlobalValue's in the program. If none of them have their "address taken"
190 /// (really, their address passed to something nontrivial), record this fact,
191 /// and record the functions that they are used directly in.
192 void GlobalsModRef::AnalyzeGlobals(Module &M) {
193 std::vector<Function*> Readers, Writers;
194 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
195 if (I->hasLocalLinkage()) {
196 if (!AnalyzeUsesOfPointer(I, Readers, Writers)) {
197 // Remember that we are tracking this global.
198 NonAddressTakenGlobals.insert(I);
199 ++NumNonAddrTakenFunctions;
201 Readers.clear(); Writers.clear();
204 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
206 if (I->hasLocalLinkage()) {
207 if (!AnalyzeUsesOfPointer(I, Readers, Writers)) {
208 // Remember that we are tracking this global, and the mod/ref fns
209 NonAddressTakenGlobals.insert(I);
211 for (unsigned i = 0, e = Readers.size(); i != e; ++i)
212 FunctionInfo[Readers[i]].GlobalInfo[I] |= Ref;
214 if (!I->isConstant()) // No need to keep track of writers to constants
215 for (unsigned i = 0, e = Writers.size(); i != e; ++i)
216 FunctionInfo[Writers[i]].GlobalInfo[I] |= Mod;
217 ++NumNonAddrTakenGlobalVars;
219 // If this global holds a pointer type, see if it is an indirect global.
220 if (I->getType()->getElementType()->isPointerTy() &&
221 AnalyzeIndirectGlobalMemory(I))
222 ++NumIndirectGlobalVars;
224 Readers.clear(); Writers.clear();
228 /// AnalyzeUsesOfPointer - Look at all of the users of the specified pointer.
229 /// If this is used by anything complex (i.e., the address escapes), return
230 /// true. Also, while we are at it, keep track of those functions that read and
231 /// write to the value.
233 /// If OkayStoreDest is non-null, stores into this global are allowed.
234 bool GlobalsModRef::AnalyzeUsesOfPointer(Value *V,
235 std::vector<Function*> &Readers,
236 std::vector<Function*> &Writers,
237 GlobalValue *OkayStoreDest) {
238 if (!V->getType()->isPointerTy()) return true;
240 for (Value::use_iterator UI = V->use_begin(), E=V->use_end(); UI != E; ++UI) {
242 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
243 Readers.push_back(LI->getParent()->getParent());
244 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
245 if (V == SI->getOperand(1)) {
246 Writers.push_back(SI->getParent()->getParent());
247 } else if (SI->getOperand(1) != OkayStoreDest) {
248 return true; // Storing the pointer
250 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
251 if (AnalyzeUsesOfPointer(GEP, Readers, Writers)) return true;
252 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
253 if (AnalyzeUsesOfPointer(BCI, Readers, Writers, OkayStoreDest))
255 } else if (isFreeCall(U)) {
256 Writers.push_back(cast<Instruction>(U)->getParent()->getParent());
257 } else if (CallInst *CI = dyn_cast<CallInst>(U)) {
258 // Make sure that this is just the function being called, not that it is
259 // passing into the function.
260 for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i)
261 if (CI->getArgOperand(i) == V) return true;
262 } else if (InvokeInst *II = dyn_cast<InvokeInst>(U)) {
263 // Make sure that this is just the function being called, not that it is
264 // passing into the function.
265 for (unsigned i = 0, e = II->getNumArgOperands(); i != e; ++i)
266 if (II->getArgOperand(i) == V) return true;
267 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
268 if (CE->getOpcode() == Instruction::GetElementPtr ||
269 CE->getOpcode() == Instruction::BitCast) {
270 if (AnalyzeUsesOfPointer(CE, Readers, Writers))
275 } else if (ICmpInst *ICI = dyn_cast<ICmpInst>(U)) {
276 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
277 return true; // Allow comparison against null.
286 /// AnalyzeIndirectGlobalMemory - We found an non-address-taken global variable
287 /// which holds a pointer type. See if the global always points to non-aliased
288 /// heap memory: that is, all initializers of the globals are allocations, and
289 /// those allocations have no use other than initialization of the global.
290 /// Further, all loads out of GV must directly use the memory, not store the
291 /// pointer somewhere. If this is true, we consider the memory pointed to by
292 /// GV to be owned by GV and can disambiguate other pointers from it.
293 bool GlobalsModRef::AnalyzeIndirectGlobalMemory(GlobalValue *GV) {
294 // Keep track of values related to the allocation of the memory, f.e. the
295 // value produced by the malloc call and any casts.
296 std::vector<Value*> AllocRelatedValues;
298 // Walk the user list of the global. If we find anything other than a direct
299 // load or store, bail out.
300 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I){
302 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
303 // The pointer loaded from the global can only be used in simple ways:
304 // we allow addressing of it and loading storing to it. We do *not* allow
305 // storing the loaded pointer somewhere else or passing to a function.
306 std::vector<Function*> ReadersWriters;
307 if (AnalyzeUsesOfPointer(LI, ReadersWriters, ReadersWriters))
308 return false; // Loaded pointer escapes.
309 // TODO: Could try some IP mod/ref of the loaded pointer.
310 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
311 // Storing the global itself.
312 if (SI->getOperand(0) == GV) return false;
314 // If storing the null pointer, ignore it.
315 if (isa<ConstantPointerNull>(SI->getOperand(0)))
318 // Check the value being stored.
319 Value *Ptr = SI->getOperand(0)->getUnderlyingObject();
323 } else if (CallInst *CI = dyn_cast<CallInst>(Ptr)) {
324 Function *F = CI->getCalledFunction();
325 if (!F || !F->isDeclaration()) return false; // Too hard to analyze.
326 if (F->getName() != "calloc") return false; // Not calloc.
328 return false; // Too hard to analyze.
331 // Analyze all uses of the allocation. If any of them are used in a
332 // non-simple way (e.g. stored to another global) bail out.
333 std::vector<Function*> ReadersWriters;
334 if (AnalyzeUsesOfPointer(Ptr, ReadersWriters, ReadersWriters, GV))
335 return false; // Loaded pointer escapes.
337 // Remember that this allocation is related to the indirect global.
338 AllocRelatedValues.push_back(Ptr);
340 // Something complex, bail out.
345 // Okay, this is an indirect global. Remember all of the allocations for
346 // this global in AllocsForIndirectGlobals.
347 while (!AllocRelatedValues.empty()) {
348 AllocsForIndirectGlobals[AllocRelatedValues.back()] = GV;
349 AllocRelatedValues.pop_back();
351 IndirectGlobals.insert(GV);
355 /// AnalyzeCallGraph - At this point, we know the functions where globals are
356 /// immediately stored to and read from. Propagate this information up the call
357 /// graph to all callers and compute the mod/ref info for all memory for each
359 void GlobalsModRef::AnalyzeCallGraph(CallGraph &CG, Module &M) {
360 // We do a bottom-up SCC traversal of the call graph. In other words, we
361 // visit all callees before callers (leaf-first).
362 for (scc_iterator<CallGraph*> I = scc_begin(&CG), E = scc_end(&CG); I != E;
364 std::vector<CallGraphNode *> &SCC = *I;
365 assert(!SCC.empty() && "SCC with no functions?");
367 if (!SCC[0]->getFunction()) {
368 // Calls externally - can't say anything useful. Remove any existing
369 // function records (may have been created when scanning globals).
370 for (unsigned i = 0, e = SCC.size(); i != e; ++i)
371 FunctionInfo.erase(SCC[i]->getFunction());
375 FunctionRecord &FR = FunctionInfo[SCC[0]->getFunction()];
377 bool KnowNothing = false;
378 unsigned FunctionEffect = 0;
380 // Collect the mod/ref properties due to called functions. We only compute
382 for (unsigned i = 0, e = SCC.size(); i != e && !KnowNothing; ++i) {
383 Function *F = SCC[i]->getFunction();
389 if (F->isDeclaration()) {
390 // Try to get mod/ref behaviour from function attributes.
391 if (F->doesNotAccessMemory()) {
392 // Can't do better than that!
393 } else if (F->onlyReadsMemory()) {
394 FunctionEffect |= Ref;
395 if (!F->isIntrinsic())
396 // This function might call back into the module and read a global -
397 // consider every global as possibly being read by this function.
398 FR.MayReadAnyGlobal = true;
400 FunctionEffect |= ModRef;
401 // Can't say anything useful unless it's an intrinsic - they don't
402 // read or write global variables of the kind considered here.
403 KnowNothing = !F->isIntrinsic();
408 for (CallGraphNode::iterator CI = SCC[i]->begin(), E = SCC[i]->end();
409 CI != E && !KnowNothing; ++CI)
410 if (Function *Callee = CI->second->getFunction()) {
411 if (FunctionRecord *CalleeFR = getFunctionInfo(Callee)) {
412 // Propagate function effect up.
413 FunctionEffect |= CalleeFR->FunctionEffect;
415 // Incorporate callee's effects on globals into our info.
416 for (std::map<const GlobalValue*, unsigned>::iterator GI =
417 CalleeFR->GlobalInfo.begin(), E = CalleeFR->GlobalInfo.end();
419 FR.GlobalInfo[GI->first] |= GI->second;
420 FR.MayReadAnyGlobal |= CalleeFR->MayReadAnyGlobal;
422 // Can't say anything about it. However, if it is inside our SCC,
423 // then nothing needs to be done.
424 CallGraphNode *CalleeNode = CG[Callee];
425 if (std::find(SCC.begin(), SCC.end(), CalleeNode) == SCC.end())
433 // If we can't say anything useful about this SCC, remove all SCC functions
434 // from the FunctionInfo map.
436 for (unsigned i = 0, e = SCC.size(); i != e; ++i)
437 FunctionInfo.erase(SCC[i]->getFunction());
441 // Scan the function bodies for explicit loads or stores.
442 for (unsigned i = 0, e = SCC.size(); i != e && FunctionEffect != ModRef;++i)
443 for (inst_iterator II = inst_begin(SCC[i]->getFunction()),
444 E = inst_end(SCC[i]->getFunction());
445 II != E && FunctionEffect != ModRef; ++II)
446 if (isa<LoadInst>(*II)) {
447 FunctionEffect |= Ref;
448 if (cast<LoadInst>(*II).isVolatile())
449 // Volatile loads may have side-effects, so mark them as writing
450 // memory (for example, a flag inside the processor).
451 FunctionEffect |= Mod;
452 } else if (isa<StoreInst>(*II)) {
453 FunctionEffect |= Mod;
454 if (cast<StoreInst>(*II).isVolatile())
455 // Treat volatile stores as reading memory somewhere.
456 FunctionEffect |= Ref;
457 } else if (isMalloc(&cast<Instruction>(*II)) ||
458 isFreeCall(&cast<Instruction>(*II))) {
459 FunctionEffect |= ModRef;
462 if ((FunctionEffect & Mod) == 0)
463 ++NumReadMemFunctions;
464 if (FunctionEffect == 0)
466 FR.FunctionEffect = FunctionEffect;
468 // Finally, now that we know the full effect on this SCC, clone the
469 // information to each function in the SCC.
470 for (unsigned i = 1, e = SCC.size(); i != e; ++i)
471 FunctionInfo[SCC[i]->getFunction()] = FR;
477 /// alias - If one of the pointers is to a global that we are tracking, and the
478 /// other is some random pointer, we know there cannot be an alias, because the
479 /// address of the global isn't taken.
480 AliasAnalysis::AliasResult
481 GlobalsModRef::alias(const Value *V1, unsigned V1Size,
482 const Value *V2, unsigned V2Size) {
483 // Get the base object these pointers point to.
484 const Value *UV1 = V1->getUnderlyingObject();
485 const Value *UV2 = V2->getUnderlyingObject();
487 // If either of the underlying values is a global, they may be non-addr-taken
488 // globals, which we can answer queries about.
489 const GlobalValue *GV1 = dyn_cast<GlobalValue>(UV1);
490 const GlobalValue *GV2 = dyn_cast<GlobalValue>(UV2);
492 // If the global's address is taken, pretend we don't know it's a pointer to
494 if (GV1 && !NonAddressTakenGlobals.count(GV1)) GV1 = 0;
495 if (GV2 && !NonAddressTakenGlobals.count(GV2)) GV2 = 0;
497 // If the two pointers are derived from two different non-addr-taken
498 // globals, or if one is and the other isn't, we know these can't alias.
499 if ((GV1 || GV2) && GV1 != GV2)
502 // Otherwise if they are both derived from the same addr-taken global, we
503 // can't know the two accesses don't overlap.
506 // These pointers may be based on the memory owned by an indirect global. If
507 // so, we may be able to handle this. First check to see if the base pointer
508 // is a direct load from an indirect global.
510 if (const LoadInst *LI = dyn_cast<LoadInst>(UV1))
511 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getOperand(0)))
512 if (IndirectGlobals.count(GV))
514 if (const LoadInst *LI = dyn_cast<LoadInst>(UV2))
515 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getOperand(0)))
516 if (IndirectGlobals.count(GV))
519 // These pointers may also be from an allocation for the indirect global. If
520 // so, also handle them.
521 if (AllocsForIndirectGlobals.count(UV1))
522 GV1 = AllocsForIndirectGlobals[UV1];
523 if (AllocsForIndirectGlobals.count(UV2))
524 GV2 = AllocsForIndirectGlobals[UV2];
526 // Now that we know whether the two pointers are related to indirect globals,
527 // use this to disambiguate the pointers. If either pointer is based on an
528 // indirect global and if they are not both based on the same indirect global,
529 // they cannot alias.
530 if ((GV1 || GV2) && GV1 != GV2)
533 return AliasAnalysis::alias(V1, V1Size, V2, V2Size);
536 AliasAnalysis::ModRefResult
537 GlobalsModRef::getModRefInfo(ImmutableCallSite CS,
538 const Value *P, unsigned Size) {
539 unsigned Known = ModRef;
541 // If we are asking for mod/ref info of a direct call with a pointer to a
542 // global we are tracking, return information if we have it.
543 if (const GlobalValue *GV = dyn_cast<GlobalValue>(P->getUnderlyingObject()))
544 if (GV->hasLocalLinkage())
545 if (const Function *F = CS.getCalledFunction())
546 if (NonAddressTakenGlobals.count(GV))
547 if (const FunctionRecord *FR = getFunctionInfo(F))
548 Known = FR->getInfoForGlobal(GV);
550 if (Known == NoModRef)
551 return NoModRef; // No need to query other mod/ref analyses
552 return ModRefResult(Known & AliasAnalysis::getModRefInfo(CS, P, Size));
556 //===----------------------------------------------------------------------===//
557 // Methods to update the analysis as a result of the client transformation.
559 void GlobalsModRef::deleteValue(Value *V) {
560 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
561 if (NonAddressTakenGlobals.erase(GV)) {
562 // This global might be an indirect global. If so, remove it and remove
563 // any AllocRelatedValues for it.
564 if (IndirectGlobals.erase(GV)) {
565 // Remove any entries in AllocsForIndirectGlobals for this global.
566 for (std::map<const Value*, const GlobalValue*>::iterator
567 I = AllocsForIndirectGlobals.begin(),
568 E = AllocsForIndirectGlobals.end(); I != E; ) {
569 if (I->second == GV) {
570 AllocsForIndirectGlobals.erase(I++);
579 // Otherwise, if this is an allocation related to an indirect global, remove
581 AllocsForIndirectGlobals.erase(V);
583 AliasAnalysis::deleteValue(V);
586 void GlobalsModRef::copyValue(Value *From, Value *To) {
587 AliasAnalysis::copyValue(From, To);