1 //===- GlobalsModRef.cpp - Simple Mod/Ref Analysis for Globals ------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source 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 /// IndirectGlobals - The memory pointed to by this global is known to be
74 /// 'owned' by the global.
75 std::set<GlobalValue*> IndirectGlobals;
77 /// AllocsForIndirectGlobals - If an instruction allocates memory for an
78 /// indirect global, this map indicates which one.
79 std::map<Value*, GlobalValue*> AllocsForIndirectGlobals;
81 /// FunctionInfo - For each function, keep track of what globals are
83 std::map<Function*, FunctionRecord> FunctionInfo;
87 GlobalsModRef() : ModulePass((intptr_t)&ID) {}
89 bool runOnModule(Module &M) {
90 InitializeAliasAnalysis(this); // set up super class
91 AnalyzeGlobals(M); // find non-addr taken globals
92 AnalyzeCallGraph(getAnalysis<CallGraph>(), M); // Propagate on CG
96 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
97 AliasAnalysis::getAnalysisUsage(AU);
98 AU.addRequired<CallGraph>();
99 AU.setPreservesAll(); // Does not transform code
102 //------------------------------------------------
103 // Implement the AliasAnalysis API
105 AliasResult alias(const Value *V1, unsigned V1Size,
106 const Value *V2, unsigned V2Size);
107 ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size);
108 ModRefResult getModRefInfo(CallSite CS1, CallSite CS2) {
109 return AliasAnalysis::getModRefInfo(CS1,CS2);
111 bool hasNoModRefInfoForCalls() const { return false; }
113 /// getModRefBehavior - Return the behavior of the specified function if
114 /// called from the specified call site. The call site may be null in which
115 /// case the most generic behavior of this function should be returned.
116 virtual ModRefBehavior getModRefBehavior(Function *F, CallSite CS,
117 std::vector<PointerAccessInfo> *Info) {
118 if (FunctionRecord *FR = getFunctionInfo(F))
119 if (FR->FunctionEffect == 0)
120 return DoesNotAccessMemory;
121 else if ((FR->FunctionEffect & Mod) == 0)
122 return OnlyReadsMemory;
123 return AliasAnalysis::getModRefBehavior(F, CS, Info);
126 virtual void deleteValue(Value *V);
127 virtual void copyValue(Value *From, Value *To);
130 /// getFunctionInfo - Return the function info for the function, or null if
131 /// the function calls an external function (in which case we don't have
132 /// anything useful to say about it).
133 FunctionRecord *getFunctionInfo(Function *F) {
134 std::map<Function*, FunctionRecord>::iterator I = FunctionInfo.find(F);
135 if (I != FunctionInfo.end())
140 void AnalyzeGlobals(Module &M);
141 void AnalyzeCallGraph(CallGraph &CG, Module &M);
142 void AnalyzeSCC(std::vector<CallGraphNode *> &SCC);
143 bool AnalyzeUsesOfPointer(Value *V, std::vector<Function*> &Readers,
144 std::vector<Function*> &Writers,
145 GlobalValue *OkayStoreDest = 0);
146 bool AnalyzeIndirectGlobalMemory(GlobalValue *GV);
149 const char GlobalsModRef::ID = 0;
150 RegisterPass<GlobalsModRef> X("globalsmodref-aa",
151 "Simple mod/ref analysis for globals");
152 RegisterAnalysisGroup<AliasAnalysis> Y(X);
155 Pass *llvm::createGlobalsModRefPass() { return new GlobalsModRef(); }
157 /// getUnderlyingObject - This traverses the use chain to figure out what object
158 /// the specified value points to. If the value points to, or is derived from,
159 /// a global object, return it.
160 static Value *getUnderlyingObject(Value *V) {
161 if (!isa<PointerType>(V->getType())) return V;
163 // If we are at some type of object... return it.
164 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) return GV;
166 // Traverse through different addressing mechanisms.
167 if (Instruction *I = dyn_cast<Instruction>(V)) {
168 if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I))
169 return getUnderlyingObject(I->getOperand(0));
170 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
171 if (CE->getOpcode() == Instruction::BitCast ||
172 CE->getOpcode() == Instruction::GetElementPtr)
173 return getUnderlyingObject(CE->getOperand(0));
176 // Othewise, we don't know what this is, return it as the base pointer.
180 /// AnalyzeGlobals - Scan through the users of all of the internal
181 /// GlobalValue's in the program. If none of them have their "Address taken"
182 /// (really, their address passed to something nontrivial), record this fact,
183 /// and record the functions that they are used directly in.
184 void GlobalsModRef::AnalyzeGlobals(Module &M) {
185 std::vector<Function*> Readers, Writers;
186 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
187 if (I->hasInternalLinkage()) {
188 if (!AnalyzeUsesOfPointer(I, Readers, Writers)) {
189 // Remember that we are tracking this global.
190 NonAddressTakenGlobals.insert(I);
191 ++NumNonAddrTakenFunctions;
193 Readers.clear(); Writers.clear();
196 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
198 if (I->hasInternalLinkage()) {
199 if (!AnalyzeUsesOfPointer(I, Readers, Writers)) {
200 // Remember that we are tracking this global, and the mod/ref fns
201 NonAddressTakenGlobals.insert(I);
202 for (unsigned i = 0, e = Readers.size(); i != e; ++i)
203 FunctionInfo[Readers[i]].GlobalInfo[I] |= Ref;
205 if (!I->isConstant()) // No need to keep track of writers to constants
206 for (unsigned i = 0, e = Writers.size(); i != e; ++i)
207 FunctionInfo[Writers[i]].GlobalInfo[I] |= Mod;
208 ++NumNonAddrTakenGlobalVars;
210 // If this global holds a pointer type, see if it is an indirect global.
211 if (isa<PointerType>(I->getType()->getElementType()) &&
212 AnalyzeIndirectGlobalMemory(I))
213 ++NumIndirectGlobalVars;
215 Readers.clear(); Writers.clear();
219 /// AnalyzeUsesOfPointer - Look at all of the users of the specified pointer.
220 /// If this is used by anything complex (i.e., the address escapes), return
221 /// true. Also, while we are at it, keep track of those functions that read and
222 /// write to the value.
224 /// If OkayStoreDest is non-null, stores into this global are allowed.
225 bool GlobalsModRef::AnalyzeUsesOfPointer(Value *V,
226 std::vector<Function*> &Readers,
227 std::vector<Function*> &Writers,
228 GlobalValue *OkayStoreDest) {
229 if (!isa<PointerType>(V->getType())) return true;
231 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
232 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
233 Readers.push_back(LI->getParent()->getParent());
234 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
235 if (V == SI->getOperand(1)) {
236 Writers.push_back(SI->getParent()->getParent());
237 } else if (SI->getOperand(1) != OkayStoreDest) {
238 return true; // Storing the pointer
240 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) {
241 if (AnalyzeUsesOfPointer(GEP, Readers, Writers)) return true;
242 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
243 // Make sure that this is just the function being called, not that it is
244 // passing into the function.
245 for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i)
246 if (CI->getOperand(i) == V) return true;
247 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
248 // Make sure that this is just the function being called, not that it is
249 // passing into the function.
250 for (unsigned i = 3, e = II->getNumOperands(); i != e; ++i)
251 if (II->getOperand(i) == V) return true;
252 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
253 if (CE->getOpcode() == Instruction::GetElementPtr ||
254 CE->getOpcode() == Instruction::BitCast) {
255 if (AnalyzeUsesOfPointer(CE, Readers, Writers))
260 } else if (ICmpInst *ICI = dyn_cast<ICmpInst>(*UI)) {
261 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
262 return true; // Allow comparison against null.
263 } else if (FreeInst *F = dyn_cast<FreeInst>(*UI)) {
264 Writers.push_back(F->getParent()->getParent());
271 /// AnalyzeIndirectGlobalMemory - We found an non-address-taken global variable
272 /// which holds a pointer type. See if the global always points to non-aliased
273 /// heap memory: that is, all initializers of the globals are allocations, and
274 /// those allocations have no use other than initialization of the global.
275 /// Further, all loads out of GV must directly use the memory, not store the
276 /// pointer somewhere. If this is true, we consider the memory pointed to by
277 /// GV to be owned by GV and can disambiguate other pointers from it.
278 bool GlobalsModRef::AnalyzeIndirectGlobalMemory(GlobalValue *GV) {
279 // Keep track of values related to the allocation of the memory, f.e. the
280 // value produced by the malloc call and any casts.
281 std::vector<Value*> AllocRelatedValues;
283 // Walk the user list of the global. If we find anything other than a direct
284 // load or store, bail out.
285 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I){
286 if (LoadInst *LI = dyn_cast<LoadInst>(*I)) {
287 // The pointer loaded from the global can only be used in simple ways:
288 // we allow addressing of it and loading storing to it. We do *not* allow
289 // storing the loaded pointer somewhere else or passing to a function.
290 std::vector<Function*> ReadersWriters;
291 if (AnalyzeUsesOfPointer(LI, ReadersWriters, ReadersWriters))
292 return false; // Loaded pointer escapes.
293 // TODO: Could try some IP mod/ref of the loaded pointer.
294 } else if (StoreInst *SI = dyn_cast<StoreInst>(*I)) {
295 // Storing the global itself.
296 if (SI->getOperand(0) == GV) return false;
298 // If storing the null pointer, ignore it.
299 if (isa<ConstantPointerNull>(SI->getOperand(0)))
302 // Check the value being stored.
303 Value *Ptr = getUnderlyingObject(SI->getOperand(0));
305 if (isa<MallocInst>(Ptr)) {
307 } else if (CallInst *CI = dyn_cast<CallInst>(Ptr)) {
308 Function *F = CI->getCalledFunction();
309 if (!F || !F->isDeclaration()) return false; // Too hard to analyze.
310 if (F->getName() != "calloc") return false; // Not calloc.
312 return false; // Too hard to analyze.
315 // Analyze all uses of the allocation. If any of them are used in a
316 // non-simple way (e.g. stored to another global) bail out.
317 std::vector<Function*> ReadersWriters;
318 if (AnalyzeUsesOfPointer(Ptr, ReadersWriters, ReadersWriters, GV))
319 return false; // Loaded pointer escapes.
321 // Remember that this allocation is related to the indirect global.
322 AllocRelatedValues.push_back(Ptr);
324 // Something complex, bail out.
329 // Okay, this is an indirect global. Remember all of the allocations for
330 // this global in AllocsForIndirectGlobals.
331 while (!AllocRelatedValues.empty()) {
332 AllocsForIndirectGlobals[AllocRelatedValues.back()] = GV;
333 AllocRelatedValues.pop_back();
335 IndirectGlobals.insert(GV);
339 /// AnalyzeCallGraph - At this point, we know the functions where globals are
340 /// immediately stored to and read from. Propagate this information up the call
341 /// graph to all callers and compute the mod/ref info for all memory for each
343 void GlobalsModRef::AnalyzeCallGraph(CallGraph &CG, Module &M) {
344 // We do a bottom-up SCC traversal of the call graph. In other words, we
345 // visit all callees before callers (leaf-first).
346 for (scc_iterator<CallGraph*> I = scc_begin(&CG), E = scc_end(&CG); I!=E; ++I)
347 if ((*I).size() != 1) {
349 } else if (Function *F = (*I)[0]->getFunction()) {
350 if (!F->isDeclaration()) {
351 // Nonexternal function.
354 // Otherwise external function. Handle intrinsics and other special
356 if (getAnalysis<AliasAnalysis>().doesNotAccessMemory(F))
357 // If it does not access memory, process the function, causing us to
358 // realize it doesn't do anything (the body is empty).
361 // Otherwise, don't process it. This will cause us to conservatively
366 // Do not process the external node, assume the worst.
370 void GlobalsModRef::AnalyzeSCC(std::vector<CallGraphNode *> &SCC) {
371 assert(!SCC.empty() && "SCC with no functions?");
372 FunctionRecord &FR = FunctionInfo[SCC[0]->getFunction()];
374 bool CallsExternal = false;
375 unsigned FunctionEffect = 0;
377 // Collect the mod/ref properties due to called functions. We only compute
379 for (unsigned i = 0, e = SCC.size(); i != e && !CallsExternal; ++i)
380 for (CallGraphNode::iterator CI = SCC[i]->begin(), E = SCC[i]->end();
382 if (Function *Callee = CI->second->getFunction()) {
383 if (FunctionRecord *CalleeFR = getFunctionInfo(Callee)) {
384 // Propagate function effect up.
385 FunctionEffect |= CalleeFR->FunctionEffect;
387 // Incorporate callee's effects on globals into our info.
388 for (std::map<GlobalValue*, unsigned>::iterator GI =
389 CalleeFR->GlobalInfo.begin(), E = CalleeFR->GlobalInfo.end();
391 FR.GlobalInfo[GI->first] |= GI->second;
394 // Okay, if we can't say anything about it, maybe some other alias
397 AliasAnalysis::getModRefBehavior(Callee, CallSite());
398 if (MRB != DoesNotAccessMemory) {
399 // FIXME: could make this more aggressive for functions that just
400 // read memory. We should just say they read all globals.
401 CallsExternal = true;
406 CallsExternal = true;
410 // If this SCC calls an external function, we can't say anything about it, so
411 // remove all SCC functions from the FunctionInfo map.
413 for (unsigned i = 0, e = SCC.size(); i != e; ++i)
414 FunctionInfo.erase(SCC[i]->getFunction());
418 // Otherwise, unless we already know that this function mod/refs memory, scan
419 // the function bodies to see if there are any explicit loads or stores.
420 if (FunctionEffect != ModRef) {
421 for (unsigned i = 0, e = SCC.size(); i != e && FunctionEffect != ModRef;++i)
422 for (inst_iterator II = inst_begin(SCC[i]->getFunction()),
423 E = inst_end(SCC[i]->getFunction());
424 II != E && FunctionEffect != ModRef; ++II)
425 if (isa<LoadInst>(*II))
426 FunctionEffect |= Ref;
427 else if (isa<StoreInst>(*II))
428 FunctionEffect |= Mod;
429 else if (isa<MallocInst>(*II) || isa<FreeInst>(*II))
430 FunctionEffect |= ModRef;
433 if ((FunctionEffect & Mod) == 0)
434 ++NumReadMemFunctions;
435 if (FunctionEffect == 0)
437 FR.FunctionEffect = FunctionEffect;
439 // Finally, now that we know the full effect on this SCC, clone the
440 // information to each function in the SCC.
441 for (unsigned i = 1, e = SCC.size(); i != e; ++i)
442 FunctionInfo[SCC[i]->getFunction()] = FR;
447 /// alias - If one of the pointers is to a global that we are tracking, and the
448 /// other is some random pointer, we know there cannot be an alias, because the
449 /// address of the global isn't taken.
450 AliasAnalysis::AliasResult
451 GlobalsModRef::alias(const Value *V1, unsigned V1Size,
452 const Value *V2, unsigned V2Size) {
453 // Get the base object these pointers point to.
454 Value *UV1 = getUnderlyingObject(const_cast<Value*>(V1));
455 Value *UV2 = getUnderlyingObject(const_cast<Value*>(V2));
457 // If either of the underlying values is a global, they may be non-addr-taken
458 // globals, which we can answer queries about.
459 GlobalValue *GV1 = dyn_cast<GlobalValue>(UV1);
460 GlobalValue *GV2 = dyn_cast<GlobalValue>(UV2);
462 // If the global's address is taken, pretend we don't know it's a pointer to
464 if (GV1 && !NonAddressTakenGlobals.count(GV1)) GV1 = 0;
465 if (GV2 && !NonAddressTakenGlobals.count(GV2)) GV2 = 0;
467 // If the the two pointers are derived from two different non-addr-taken
468 // globals, or if one is and the other isn't, we know these can't alias.
469 if ((GV1 || GV2) && GV1 != GV2)
472 // Otherwise if they are both derived from the same addr-taken global, we
473 // can't know the two accesses don't overlap.
476 // These pointers may be based on the memory owned by an indirect global. If
477 // so, we may be able to handle this. First check to see if the base pointer
478 // is a direct load from an indirect global.
480 if (LoadInst *LI = dyn_cast<LoadInst>(UV1))
481 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getOperand(0)))
482 if (IndirectGlobals.count(GV))
484 if (LoadInst *LI = dyn_cast<LoadInst>(UV2))
485 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getOperand(0)))
486 if (IndirectGlobals.count(GV))
489 // These pointers may also be from an allocation for the indirect global. If
490 // so, also handle them.
491 if (AllocsForIndirectGlobals.count(UV1))
492 GV1 = AllocsForIndirectGlobals[UV1];
493 if (AllocsForIndirectGlobals.count(UV2))
494 GV2 = AllocsForIndirectGlobals[UV2];
496 // Now that we know whether the two pointers are related to indirect globals,
497 // use this to disambiguate the pointers. If either pointer is based on an
498 // indirect global and if they are not both based on the same indirect global,
499 // they cannot alias.
500 if ((GV1 || GV2) && GV1 != GV2)
503 return AliasAnalysis::alias(V1, V1Size, V2, V2Size);
506 AliasAnalysis::ModRefResult
507 GlobalsModRef::getModRefInfo(CallSite CS, Value *P, unsigned Size) {
508 unsigned Known = ModRef;
510 // If we are asking for mod/ref info of a direct call with a pointer to a
511 // global we are tracking, return information if we have it.
512 if (GlobalValue *GV = dyn_cast<GlobalValue>(getUnderlyingObject(P)))
513 if (GV->hasInternalLinkage())
514 if (Function *F = CS.getCalledFunction())
515 if (NonAddressTakenGlobals.count(GV))
516 if (FunctionRecord *FR = getFunctionInfo(F))
517 Known = FR->getInfoForGlobal(GV);
519 if (Known == NoModRef)
520 return NoModRef; // No need to query other mod/ref analyses
521 return ModRefResult(Known & AliasAnalysis::getModRefInfo(CS, P, Size));
525 //===----------------------------------------------------------------------===//
526 // Methods to update the analysis as a result of the client transformation.
528 void GlobalsModRef::deleteValue(Value *V) {
529 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
530 if (NonAddressTakenGlobals.erase(GV)) {
531 // This global might be an indirect global. If so, remove it and remove
532 // any AllocRelatedValues for it.
533 if (IndirectGlobals.erase(GV)) {
534 // Remove any entries in AllocsForIndirectGlobals for this global.
535 for (std::map<Value*, GlobalValue*>::iterator
536 I = AllocsForIndirectGlobals.begin(),
537 E = AllocsForIndirectGlobals.end(); I != E; ) {
538 if (I->second == GV) {
539 AllocsForIndirectGlobals.erase(I++);
548 // Otherwise, if this is an allocation related to an indirect global, remove
550 AllocsForIndirectGlobals.erase(V);
553 void GlobalsModRef::copyValue(Value *From, Value *To) {