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/Support/Compiler.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 VISIBILITY_HIDDEN 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<GlobalValue*, unsigned> GlobalInfo;
52 unsigned getInfoForGlobal(GlobalValue *GV) const {
53 std::map<GlobalValue*, unsigned>::const_iterator I = GlobalInfo.find(GV);
54 if (I != GlobalInfo.end())
59 /// FunctionEffect - Capture whether or not this function reads or writes to
60 /// ANY memory. If not, we can do a lot of aggressive analysis on it.
61 unsigned FunctionEffect;
63 FunctionRecord() : FunctionEffect(0) {}
66 /// GlobalsModRef - The actual analysis pass.
67 class VISIBILITY_HIDDEN GlobalsModRef
68 : public ModulePass, public AliasAnalysis {
69 /// NonAddressTakenGlobals - The globals that do not have their addresses
71 std::set<GlobalValue*> NonAddressTakenGlobals;
73 /// ReadGlobals - The globals without addresses taken that are read by
75 std::set<GlobalValue*> ReadGlobals;
77 /// IndirectGlobals - The memory pointed to by this global is known to be
78 /// 'owned' by the global.
79 std::set<GlobalValue*> IndirectGlobals;
81 /// AllocsForIndirectGlobals - If an instruction allocates memory for an
82 /// indirect global, this map indicates which one.
83 std::map<Value*, GlobalValue*> AllocsForIndirectGlobals;
85 /// FunctionInfo - For each function, keep track of what globals are
87 std::map<Function*, FunctionRecord> FunctionInfo;
91 GlobalsModRef() : ModulePass((intptr_t)&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(CallSite CS, Value *P, unsigned Size);
112 ModRefResult getModRefInfo(CallSite CS1, CallSite CS2) {
113 return AliasAnalysis::getModRefInfo(CS1,CS2);
115 bool hasNoModRefInfoForCalls() const { return false; }
117 /// getModRefBehavior - Return the behavior of the specified function if
118 /// called from the specified call site. The call site may be null in which
119 /// case the most generic behavior of this function should be returned.
120 virtual ModRefBehavior getModRefBehavior(Function *F, CallSite CS,
121 std::vector<PointerAccessInfo> *Info) {
122 if (FunctionRecord *FR = getFunctionInfo(F)) {
123 if (FR->FunctionEffect == 0)
124 return DoesNotAccessMemory;
125 else if ((FR->FunctionEffect & Mod) == 0)
126 return OnlyReadsMemory;
128 return AliasAnalysis::getModRefBehavior(F, CS, Info);
131 virtual void deleteValue(Value *V);
132 virtual void copyValue(Value *From, Value *To);
135 /// getFunctionInfo - Return the function info for the function, or null if
136 /// we don't have anything useful to say about it.
137 FunctionRecord *getFunctionInfo(Function *F) {
138 std::map<Function*, FunctionRecord>::iterator I = FunctionInfo.find(F);
139 if (I != FunctionInfo.end())
144 void AnalyzeGlobals(Module &M);
145 void AnalyzeCallGraph(CallGraph &CG, Module &M);
146 bool AnalyzeUsesOfPointer(Value *V, std::vector<Function*> &Readers,
147 std::vector<Function*> &Writers,
148 GlobalValue *OkayStoreDest = 0);
149 bool AnalyzeIndirectGlobalMemory(GlobalValue *GV);
153 char GlobalsModRef::ID = 0;
154 static RegisterPass<GlobalsModRef>
155 X("globalsmodref-aa", "Simple mod/ref analysis for globals", false, true);
156 static RegisterAnalysisGroup<AliasAnalysis> Y(X);
158 Pass *llvm::createGlobalsModRefPass() { return new GlobalsModRef(); }
160 /// getUnderlyingObject - This traverses the use chain to figure out what object
161 /// the specified value points to. If the value points to, or is derived from,
162 /// a global object, return it.
163 static Value *getUnderlyingObject(Value *V) {
164 if (!isa<PointerType>(V->getType())) return V;
166 // If we are at some type of object... return it.
167 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) return GV;
169 // Traverse through different addressing mechanisms.
170 if (Instruction *I = dyn_cast<Instruction>(V)) {
171 if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I))
172 return getUnderlyingObject(I->getOperand(0));
173 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
174 if (CE->getOpcode() == Instruction::BitCast ||
175 CE->getOpcode() == Instruction::GetElementPtr)
176 return getUnderlyingObject(CE->getOperand(0));
179 // Otherwise, we don't know what this is, return it as the base pointer.
183 /// AnalyzeGlobals - Scan through the users of all of the internal
184 /// GlobalValue's in the program. If none of them have their "address taken"
185 /// (really, their address passed to something nontrivial), record this fact,
186 /// and record the functions that they are used directly in.
187 void GlobalsModRef::AnalyzeGlobals(Module &M) {
188 std::vector<Function*> Readers, Writers;
189 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
190 if (I->hasInternalLinkage()) {
191 if (!AnalyzeUsesOfPointer(I, Readers, Writers)) {
192 // Remember that we are tracking this global.
193 NonAddressTakenGlobals.insert(I);
194 ++NumNonAddrTakenFunctions;
196 Readers.clear(); Writers.clear();
199 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
201 if (I->hasInternalLinkage()) {
202 if (!AnalyzeUsesOfPointer(I, Readers, Writers)) {
203 // Remember that we are tracking this global, and the mod/ref fns
204 NonAddressTakenGlobals.insert(I);
206 if (!Readers.empty())
207 // Some function read this global - remember that.
208 ReadGlobals.insert(I);
210 for (unsigned i = 0, e = Readers.size(); i != e; ++i)
211 FunctionInfo[Readers[i]].GlobalInfo[I] |= Ref;
213 if (!I->isConstant()) // No need to keep track of writers to constants
214 for (unsigned i = 0, e = Writers.size(); i != e; ++i)
215 FunctionInfo[Writers[i]].GlobalInfo[I] |= Mod;
216 ++NumNonAddrTakenGlobalVars;
218 // If this global holds a pointer type, see if it is an indirect global.
219 if (isa<PointerType>(I->getType()->getElementType()) &&
220 AnalyzeIndirectGlobalMemory(I))
221 ++NumIndirectGlobalVars;
223 Readers.clear(); Writers.clear();
227 /// AnalyzeUsesOfPointer - Look at all of the users of the specified pointer.
228 /// If this is used by anything complex (i.e., the address escapes), return
229 /// true. Also, while we are at it, keep track of those functions that read and
230 /// write to the value.
232 /// If OkayStoreDest is non-null, stores into this global are allowed.
233 bool GlobalsModRef::AnalyzeUsesOfPointer(Value *V,
234 std::vector<Function*> &Readers,
235 std::vector<Function*> &Writers,
236 GlobalValue *OkayStoreDest) {
237 if (!isa<PointerType>(V->getType())) return true;
239 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
240 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
241 Readers.push_back(LI->getParent()->getParent());
242 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
243 if (V == SI->getOperand(1)) {
244 Writers.push_back(SI->getParent()->getParent());
245 } else if (SI->getOperand(1) != OkayStoreDest) {
246 return true; // Storing the pointer
248 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) {
249 if (AnalyzeUsesOfPointer(GEP, Readers, Writers)) return true;
250 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
251 // Make sure that this is just the function being called, not that it is
252 // passing into the function.
253 for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i)
254 if (CI->getOperand(i) == V) return true;
255 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
256 // Make sure that this is just the function being called, not that it is
257 // passing into the function.
258 for (unsigned i = 3, e = II->getNumOperands(); i != e; ++i)
259 if (II->getOperand(i) == V) return true;
260 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
261 if (CE->getOpcode() == Instruction::GetElementPtr ||
262 CE->getOpcode() == Instruction::BitCast) {
263 if (AnalyzeUsesOfPointer(CE, Readers, Writers))
268 } else if (ICmpInst *ICI = dyn_cast<ICmpInst>(*UI)) {
269 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
270 return true; // Allow comparison against null.
271 } else if (FreeInst *F = dyn_cast<FreeInst>(*UI)) {
272 Writers.push_back(F->getParent()->getParent());
279 /// AnalyzeIndirectGlobalMemory - We found an non-address-taken global variable
280 /// which holds a pointer type. See if the global always points to non-aliased
281 /// heap memory: that is, all initializers of the globals are allocations, and
282 /// those allocations have no use other than initialization of the global.
283 /// Further, all loads out of GV must directly use the memory, not store the
284 /// pointer somewhere. If this is true, we consider the memory pointed to by
285 /// GV to be owned by GV and can disambiguate other pointers from it.
286 bool GlobalsModRef::AnalyzeIndirectGlobalMemory(GlobalValue *GV) {
287 // Keep track of values related to the allocation of the memory, f.e. the
288 // value produced by the malloc call and any casts.
289 std::vector<Value*> AllocRelatedValues;
291 // Walk the user list of the global. If we find anything other than a direct
292 // load or store, bail out.
293 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I){
294 if (LoadInst *LI = dyn_cast<LoadInst>(*I)) {
295 // The pointer loaded from the global can only be used in simple ways:
296 // we allow addressing of it and loading storing to it. We do *not* allow
297 // storing the loaded pointer somewhere else or passing to a function.
298 std::vector<Function*> ReadersWriters;
299 if (AnalyzeUsesOfPointer(LI, ReadersWriters, ReadersWriters))
300 return false; // Loaded pointer escapes.
301 // TODO: Could try some IP mod/ref of the loaded pointer.
302 } else if (StoreInst *SI = dyn_cast<StoreInst>(*I)) {
303 // Storing the global itself.
304 if (SI->getOperand(0) == GV) return false;
306 // If storing the null pointer, ignore it.
307 if (isa<ConstantPointerNull>(SI->getOperand(0)))
310 // Check the value being stored.
311 Value *Ptr = getUnderlyingObject(SI->getOperand(0));
313 if (isa<MallocInst>(Ptr)) {
315 } else if (CallInst *CI = dyn_cast<CallInst>(Ptr)) {
316 Function *F = CI->getCalledFunction();
317 if (!F || !F->isDeclaration()) return false; // Too hard to analyze.
318 if (F->getName() != "calloc") return false; // Not calloc.
320 return false; // Too hard to analyze.
323 // Analyze all uses of the allocation. If any of them are used in a
324 // non-simple way (e.g. stored to another global) bail out.
325 std::vector<Function*> ReadersWriters;
326 if (AnalyzeUsesOfPointer(Ptr, ReadersWriters, ReadersWriters, GV))
327 return false; // Loaded pointer escapes.
329 // Remember that this allocation is related to the indirect global.
330 AllocRelatedValues.push_back(Ptr);
332 // Something complex, bail out.
337 // Okay, this is an indirect global. Remember all of the allocations for
338 // this global in AllocsForIndirectGlobals.
339 while (!AllocRelatedValues.empty()) {
340 AllocsForIndirectGlobals[AllocRelatedValues.back()] = GV;
341 AllocRelatedValues.pop_back();
343 IndirectGlobals.insert(GV);
347 /// AnalyzeCallGraph - At this point, we know the functions where globals are
348 /// immediately stored to and read from. Propagate this information up the call
349 /// graph to all callers and compute the mod/ref info for all memory for each
351 void GlobalsModRef::AnalyzeCallGraph(CallGraph &CG, Module &M) {
352 // We do a bottom-up SCC traversal of the call graph. In other words, we
353 // visit all callees before callers (leaf-first).
354 for (scc_iterator<CallGraph*> I = scc_begin(&CG), E = scc_end(&CG); I != E;
356 std::vector<CallGraphNode *> &SCC = *I;
357 assert(!SCC.empty() && "SCC with no functions?");
359 if (!SCC[0]->getFunction())
360 // Do not process the external node, assume the worst.
363 FunctionRecord &FR = FunctionInfo[SCC[0]->getFunction()];
365 bool KnowNothing = false;
366 unsigned FunctionEffect = 0;
368 // Collect the mod/ref properties due to called functions. We only compute
370 for (unsigned i = 0, e = SCC.size(); i != e && !KnowNothing; ++i) {
371 Function *F = SCC[i]->getFunction();
377 if (F->isDeclaration()) {
378 // Try to get mod/ref behaviour from function attributes.
379 if (F->doesNotAccessMemory()) {
380 // Can't do better than that!
381 } else if (F->onlyReadsMemory()) {
382 FunctionEffect |= Ref;
383 // This function might call back into the module and read a global, so
384 // mark all globals read somewhere as being read by this function.
385 for (std::set<GlobalValue*>::iterator GI = ReadGlobals.begin(),
386 E = ReadGlobals.end(); GI != E; ++GI)
387 FR.GlobalInfo[*GI] |= Ref;
389 // Can't say anything useful.
395 for (CallGraphNode::iterator CI = SCC[i]->begin(), E = SCC[i]->end();
397 if (Function *Callee = CI->second->getFunction()) {
398 if (FunctionRecord *CalleeFR = getFunctionInfo(Callee)) {
399 // Propagate function effect up.
400 FunctionEffect |= CalleeFR->FunctionEffect;
402 // Incorporate callee's effects on globals into our info.
403 for (std::map<GlobalValue*, unsigned>::iterator GI =
404 CalleeFR->GlobalInfo.begin(), E = CalleeFR->GlobalInfo.end();
406 FR.GlobalInfo[GI->first] |= GI->second;
408 // Can't say anything about it. However, if it is inside our SCC,
409 // then nothing needs to be done.
410 CallGraphNode *CalleeNode = CG[Callee];
411 if (std::find(SCC.begin(), SCC.end(), CalleeNode) == SCC.end())
419 // If we can't say anything useful about this SCC, remove all SCC functions
420 // from the FunctionInfo map.
422 for (unsigned i = 0, e = SCC.size(); i != e; ++i)
423 FunctionInfo.erase(SCC[i]->getFunction());
427 // Scan the function bodies for explicit loads or stores.
428 for (unsigned i = 0, e = SCC.size(); i != e && FunctionEffect != ModRef;++i)
429 for (inst_iterator II = inst_begin(SCC[i]->getFunction()),
430 E = inst_end(SCC[i]->getFunction());
431 II != E && FunctionEffect != ModRef; ++II)
432 if (isa<LoadInst>(*II))
433 FunctionEffect |= Ref;
434 else if (isa<StoreInst>(*II))
435 FunctionEffect |= Mod;
436 else if (isa<MallocInst>(*II) || isa<FreeInst>(*II))
437 FunctionEffect |= ModRef;
439 if ((FunctionEffect & Mod) == 0)
440 ++NumReadMemFunctions;
441 if (FunctionEffect == 0)
443 FR.FunctionEffect = FunctionEffect;
445 // Finally, now that we know the full effect on this SCC, clone the
446 // information to each function in the SCC.
447 for (unsigned i = 1, e = SCC.size(); i != e; ++i)
448 FunctionInfo[SCC[i]->getFunction()] = FR;
454 /// alias - If one of the pointers is to a global that we are tracking, and the
455 /// other is some random pointer, we know there cannot be an alias, because the
456 /// address of the global isn't taken.
457 AliasAnalysis::AliasResult
458 GlobalsModRef::alias(const Value *V1, unsigned V1Size,
459 const Value *V2, unsigned V2Size) {
460 // Get the base object these pointers point to.
461 Value *UV1 = getUnderlyingObject(const_cast<Value*>(V1));
462 Value *UV2 = getUnderlyingObject(const_cast<Value*>(V2));
464 // If either of the underlying values is a global, they may be non-addr-taken
465 // globals, which we can answer queries about.
466 GlobalValue *GV1 = dyn_cast<GlobalValue>(UV1);
467 GlobalValue *GV2 = dyn_cast<GlobalValue>(UV2);
469 // If the global's address is taken, pretend we don't know it's a pointer to
471 if (GV1 && !NonAddressTakenGlobals.count(GV1)) GV1 = 0;
472 if (GV2 && !NonAddressTakenGlobals.count(GV2)) GV2 = 0;
474 // If the the two pointers are derived from two different non-addr-taken
475 // globals, or if one is and the other isn't, we know these can't alias.
476 if ((GV1 || GV2) && GV1 != GV2)
479 // Otherwise if they are both derived from the same addr-taken global, we
480 // can't know the two accesses don't overlap.
483 // These pointers may be based on the memory owned by an indirect global. If
484 // so, we may be able to handle this. First check to see if the base pointer
485 // is a direct load from an indirect global.
487 if (LoadInst *LI = dyn_cast<LoadInst>(UV1))
488 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getOperand(0)))
489 if (IndirectGlobals.count(GV))
491 if (LoadInst *LI = dyn_cast<LoadInst>(UV2))
492 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getOperand(0)))
493 if (IndirectGlobals.count(GV))
496 // These pointers may also be from an allocation for the indirect global. If
497 // so, also handle them.
498 if (AllocsForIndirectGlobals.count(UV1))
499 GV1 = AllocsForIndirectGlobals[UV1];
500 if (AllocsForIndirectGlobals.count(UV2))
501 GV2 = AllocsForIndirectGlobals[UV2];
503 // Now that we know whether the two pointers are related to indirect globals,
504 // use this to disambiguate the pointers. If either pointer is based on an
505 // indirect global and if they are not both based on the same indirect global,
506 // they cannot alias.
507 if ((GV1 || GV2) && GV1 != GV2)
510 return AliasAnalysis::alias(V1, V1Size, V2, V2Size);
513 AliasAnalysis::ModRefResult
514 GlobalsModRef::getModRefInfo(CallSite CS, Value *P, unsigned Size) {
515 unsigned Known = ModRef;
517 // If we are asking for mod/ref info of a direct call with a pointer to a
518 // global we are tracking, return information if we have it.
519 if (GlobalValue *GV = dyn_cast<GlobalValue>(getUnderlyingObject(P)))
520 if (GV->hasInternalLinkage())
521 if (Function *F = CS.getCalledFunction())
522 if (NonAddressTakenGlobals.count(GV))
523 if (FunctionRecord *FR = getFunctionInfo(F))
524 Known = FR->getInfoForGlobal(GV);
526 if (Known == NoModRef)
527 return NoModRef; // No need to query other mod/ref analyses
528 return ModRefResult(Known & AliasAnalysis::getModRefInfo(CS, P, Size));
532 //===----------------------------------------------------------------------===//
533 // Methods to update the analysis as a result of the client transformation.
535 void GlobalsModRef::deleteValue(Value *V) {
536 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
537 if (NonAddressTakenGlobals.erase(GV)) {
538 // This global might be an indirect global. If so, remove it and remove
539 // any AllocRelatedValues for it.
540 if (IndirectGlobals.erase(GV)) {
541 // Remove any entries in AllocsForIndirectGlobals for this global.
542 for (std::map<Value*, GlobalValue*>::iterator
543 I = AllocsForIndirectGlobals.begin(),
544 E = AllocsForIndirectGlobals.end(); I != E; ) {
545 if (I->second == GV) {
546 AllocsForIndirectGlobals.erase(I++);
555 // Otherwise, if this is an allocation related to an indirect global, remove
557 AllocsForIndirectGlobals.erase(V);
559 AliasAnalysis::deleteValue(V);
562 void GlobalsModRef::copyValue(Value *From, Value *To) {
563 AliasAnalysis::copyValue(From, To);