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) {
92 initializeGlobalsModRefPass(*PassRegistry::getPassRegistry());
95 bool runOnModule(Module &M) {
96 InitializeAliasAnalysis(this); // set up super class
97 AnalyzeGlobals(M); // find non-addr taken globals
98 AnalyzeCallGraph(getAnalysis<CallGraph>(), M); // Propagate on CG
102 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
103 AliasAnalysis::getAnalysisUsage(AU);
104 AU.addRequired<CallGraph>();
105 AU.setPreservesAll(); // Does not transform code
108 //------------------------------------------------
109 // Implement the AliasAnalysis API
111 AliasResult alias(const Location &LocA, const Location &LocB);
112 ModRefResult getModRefInfo(ImmutableCallSite CS,
113 const Location &Loc);
114 ModRefResult getModRefInfo(ImmutableCallSite CS1,
115 ImmutableCallSite CS2) {
116 return AliasAnalysis::getModRefInfo(CS1, CS2);
119 /// getModRefBehavior - Return the behavior of the specified function if
120 /// called from the specified call site. The call site may be null in which
121 /// case the most generic behavior of this function should be returned.
122 ModRefBehavior getModRefBehavior(const Function *F) {
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);
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 const Function* F = CS.getCalledFunction();
137 if (!F) return AliasAnalysis::getModRefBehavior(CS);
138 if (FunctionRecord *FR = getFunctionInfo(F)) {
139 if (FR->FunctionEffect == 0)
140 return DoesNotAccessMemory;
141 else if ((FR->FunctionEffect & Mod) == 0)
142 return OnlyReadsMemory;
144 return AliasAnalysis::getModRefBehavior(CS);
147 virtual void deleteValue(Value *V);
148 virtual void copyValue(Value *From, Value *To);
150 /// getAdjustedAnalysisPointer - This method is used when a pass implements
151 /// an analysis interface through multiple inheritance. If needed, it
152 /// should override this to adjust the this pointer as needed for the
153 /// specified pass info.
154 virtual void *getAdjustedAnalysisPointer(AnalysisID PI) {
155 if (PI == &AliasAnalysis::ID)
156 return (AliasAnalysis*)this;
161 /// getFunctionInfo - Return the function info for the function, or null if
162 /// we don't have anything useful to say about it.
163 FunctionRecord *getFunctionInfo(const Function *F) {
164 std::map<const Function*, FunctionRecord>::iterator I =
165 FunctionInfo.find(F);
166 if (I != FunctionInfo.end())
171 void AnalyzeGlobals(Module &M);
172 void AnalyzeCallGraph(CallGraph &CG, Module &M);
173 bool AnalyzeUsesOfPointer(Value *V, std::vector<Function*> &Readers,
174 std::vector<Function*> &Writers,
175 GlobalValue *OkayStoreDest = 0);
176 bool AnalyzeIndirectGlobalMemory(GlobalValue *GV);
180 char GlobalsModRef::ID = 0;
181 INITIALIZE_AG_PASS_BEGIN(GlobalsModRef, AliasAnalysis,
182 "globalsmodref-aa", "Simple mod/ref analysis for globals",
184 INITIALIZE_AG_DEPENDENCY(CallGraph)
185 INITIALIZE_AG_PASS_END(GlobalsModRef, AliasAnalysis,
186 "globalsmodref-aa", "Simple mod/ref analysis for globals",
189 Pass *llvm::createGlobalsModRefPass() { return new GlobalsModRef(); }
191 /// AnalyzeGlobals - Scan through the users of all of the internal
192 /// GlobalValue's in the program. If none of them have their "address taken"
193 /// (really, their address passed to something nontrivial), record this fact,
194 /// and record the functions that they are used directly in.
195 void GlobalsModRef::AnalyzeGlobals(Module &M) {
196 std::vector<Function*> Readers, Writers;
197 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
198 if (I->hasLocalLinkage()) {
199 if (!AnalyzeUsesOfPointer(I, Readers, Writers)) {
200 // Remember that we are tracking this global.
201 NonAddressTakenGlobals.insert(I);
202 ++NumNonAddrTakenFunctions;
204 Readers.clear(); Writers.clear();
207 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
209 if (I->hasLocalLinkage()) {
210 if (!AnalyzeUsesOfPointer(I, Readers, Writers)) {
211 // Remember that we are tracking this global, and the mod/ref fns
212 NonAddressTakenGlobals.insert(I);
214 for (unsigned i = 0, e = Readers.size(); i != e; ++i)
215 FunctionInfo[Readers[i]].GlobalInfo[I] |= Ref;
217 if (!I->isConstant()) // No need to keep track of writers to constants
218 for (unsigned i = 0, e = Writers.size(); i != e; ++i)
219 FunctionInfo[Writers[i]].GlobalInfo[I] |= Mod;
220 ++NumNonAddrTakenGlobalVars;
222 // If this global holds a pointer type, see if it is an indirect global.
223 if (I->getType()->getElementType()->isPointerTy() &&
224 AnalyzeIndirectGlobalMemory(I))
225 ++NumIndirectGlobalVars;
227 Readers.clear(); Writers.clear();
231 /// AnalyzeUsesOfPointer - Look at all of the users of the specified pointer.
232 /// If this is used by anything complex (i.e., the address escapes), return
233 /// true. Also, while we are at it, keep track of those functions that read and
234 /// write to the value.
236 /// If OkayStoreDest is non-null, stores into this global are allowed.
237 bool GlobalsModRef::AnalyzeUsesOfPointer(Value *V,
238 std::vector<Function*> &Readers,
239 std::vector<Function*> &Writers,
240 GlobalValue *OkayStoreDest) {
241 if (!V->getType()->isPointerTy()) return true;
243 for (Value::use_iterator UI = V->use_begin(), E=V->use_end(); UI != E; ++UI) {
245 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
246 Readers.push_back(LI->getParent()->getParent());
247 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
248 if (V == SI->getOperand(1)) {
249 Writers.push_back(SI->getParent()->getParent());
250 } else if (SI->getOperand(1) != OkayStoreDest) {
251 return true; // Storing the pointer
253 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
254 if (AnalyzeUsesOfPointer(GEP, Readers, Writers)) return true;
255 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
256 if (AnalyzeUsesOfPointer(BCI, Readers, Writers, OkayStoreDest))
258 } else if (isFreeCall(U)) {
259 Writers.push_back(cast<Instruction>(U)->getParent()->getParent());
260 } else if (CallInst *CI = dyn_cast<CallInst>(U)) {
261 // Make sure that this is just the function being called, not that it is
262 // passing into the function.
263 for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i)
264 if (CI->getArgOperand(i) == V) return true;
265 } else if (InvokeInst *II = dyn_cast<InvokeInst>(U)) {
266 // Make sure that this is just the function being called, not that it is
267 // passing into the function.
268 for (unsigned i = 0, e = II->getNumArgOperands(); i != e; ++i)
269 if (II->getArgOperand(i) == V) return true;
270 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
271 if (CE->getOpcode() == Instruction::GetElementPtr ||
272 CE->getOpcode() == Instruction::BitCast) {
273 if (AnalyzeUsesOfPointer(CE, Readers, Writers))
278 } else if (ICmpInst *ICI = dyn_cast<ICmpInst>(U)) {
279 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
280 return true; // Allow comparison against null.
289 /// AnalyzeIndirectGlobalMemory - We found an non-address-taken global variable
290 /// which holds a pointer type. See if the global always points to non-aliased
291 /// heap memory: that is, all initializers of the globals are allocations, and
292 /// those allocations have no use other than initialization of the global.
293 /// Further, all loads out of GV must directly use the memory, not store the
294 /// pointer somewhere. If this is true, we consider the memory pointed to by
295 /// GV to be owned by GV and can disambiguate other pointers from it.
296 bool GlobalsModRef::AnalyzeIndirectGlobalMemory(GlobalValue *GV) {
297 // Keep track of values related to the allocation of the memory, f.e. the
298 // value produced by the malloc call and any casts.
299 std::vector<Value*> AllocRelatedValues;
301 // Walk the user list of the global. If we find anything other than a direct
302 // load or store, bail out.
303 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I){
305 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
306 // The pointer loaded from the global can only be used in simple ways:
307 // we allow addressing of it and loading storing to it. We do *not* allow
308 // storing the loaded pointer somewhere else or passing to a function.
309 std::vector<Function*> ReadersWriters;
310 if (AnalyzeUsesOfPointer(LI, ReadersWriters, ReadersWriters))
311 return false; // Loaded pointer escapes.
312 // TODO: Could try some IP mod/ref of the loaded pointer.
313 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
314 // Storing the global itself.
315 if (SI->getOperand(0) == GV) return false;
317 // If storing the null pointer, ignore it.
318 if (isa<ConstantPointerNull>(SI->getOperand(0)))
321 // Check the value being stored.
322 Value *Ptr = SI->getOperand(0)->getUnderlyingObject();
326 } else if (CallInst *CI = dyn_cast<CallInst>(Ptr)) {
327 Function *F = CI->getCalledFunction();
328 if (!F || !F->isDeclaration()) return false; // Too hard to analyze.
329 if (F->getName() != "calloc") return false; // Not calloc.
331 return false; // Too hard to analyze.
334 // Analyze all uses of the allocation. If any of them are used in a
335 // non-simple way (e.g. stored to another global) bail out.
336 std::vector<Function*> ReadersWriters;
337 if (AnalyzeUsesOfPointer(Ptr, ReadersWriters, ReadersWriters, GV))
338 return false; // Loaded pointer escapes.
340 // Remember that this allocation is related to the indirect global.
341 AllocRelatedValues.push_back(Ptr);
343 // Something complex, bail out.
348 // Okay, this is an indirect global. Remember all of the allocations for
349 // this global in AllocsForIndirectGlobals.
350 while (!AllocRelatedValues.empty()) {
351 AllocsForIndirectGlobals[AllocRelatedValues.back()] = GV;
352 AllocRelatedValues.pop_back();
354 IndirectGlobals.insert(GV);
358 /// AnalyzeCallGraph - At this point, we know the functions where globals are
359 /// immediately stored to and read from. Propagate this information up the call
360 /// graph to all callers and compute the mod/ref info for all memory for each
362 void GlobalsModRef::AnalyzeCallGraph(CallGraph &CG, Module &M) {
363 // We do a bottom-up SCC traversal of the call graph. In other words, we
364 // visit all callees before callers (leaf-first).
365 for (scc_iterator<CallGraph*> I = scc_begin(&CG), E = scc_end(&CG); I != E;
367 std::vector<CallGraphNode *> &SCC = *I;
368 assert(!SCC.empty() && "SCC with no functions?");
370 if (!SCC[0]->getFunction()) {
371 // Calls externally - can't say anything useful. Remove any existing
372 // function records (may have been created when scanning globals).
373 for (unsigned i = 0, e = SCC.size(); i != e; ++i)
374 FunctionInfo.erase(SCC[i]->getFunction());
378 FunctionRecord &FR = FunctionInfo[SCC[0]->getFunction()];
380 bool KnowNothing = false;
381 unsigned FunctionEffect = 0;
383 // Collect the mod/ref properties due to called functions. We only compute
385 for (unsigned i = 0, e = SCC.size(); i != e && !KnowNothing; ++i) {
386 Function *F = SCC[i]->getFunction();
392 if (F->isDeclaration()) {
393 // Try to get mod/ref behaviour from function attributes.
394 if (F->doesNotAccessMemory()) {
395 // Can't do better than that!
396 } else if (F->onlyReadsMemory()) {
397 FunctionEffect |= Ref;
398 if (!F->isIntrinsic())
399 // This function might call back into the module and read a global -
400 // consider every global as possibly being read by this function.
401 FR.MayReadAnyGlobal = true;
403 FunctionEffect |= ModRef;
404 // Can't say anything useful unless it's an intrinsic - they don't
405 // read or write global variables of the kind considered here.
406 KnowNothing = !F->isIntrinsic();
411 for (CallGraphNode::iterator CI = SCC[i]->begin(), E = SCC[i]->end();
412 CI != E && !KnowNothing; ++CI)
413 if (Function *Callee = CI->second->getFunction()) {
414 if (FunctionRecord *CalleeFR = getFunctionInfo(Callee)) {
415 // Propagate function effect up.
416 FunctionEffect |= CalleeFR->FunctionEffect;
418 // Incorporate callee's effects on globals into our info.
419 for (std::map<const GlobalValue*, unsigned>::iterator GI =
420 CalleeFR->GlobalInfo.begin(), E = CalleeFR->GlobalInfo.end();
422 FR.GlobalInfo[GI->first] |= GI->second;
423 FR.MayReadAnyGlobal |= CalleeFR->MayReadAnyGlobal;
425 // Can't say anything about it. However, if it is inside our SCC,
426 // then nothing needs to be done.
427 CallGraphNode *CalleeNode = CG[Callee];
428 if (std::find(SCC.begin(), SCC.end(), CalleeNode) == SCC.end())
436 // If we can't say anything useful about this SCC, remove all SCC functions
437 // from the FunctionInfo map.
439 for (unsigned i = 0, e = SCC.size(); i != e; ++i)
440 FunctionInfo.erase(SCC[i]->getFunction());
444 // Scan the function bodies for explicit loads or stores.
445 for (unsigned i = 0, e = SCC.size(); i != e && FunctionEffect != ModRef;++i)
446 for (inst_iterator II = inst_begin(SCC[i]->getFunction()),
447 E = inst_end(SCC[i]->getFunction());
448 II != E && FunctionEffect != ModRef; ++II)
449 if (isa<LoadInst>(*II)) {
450 FunctionEffect |= Ref;
451 if (cast<LoadInst>(*II).isVolatile())
452 // Volatile loads may have side-effects, so mark them as writing
453 // memory (for example, a flag inside the processor).
454 FunctionEffect |= Mod;
455 } else if (isa<StoreInst>(*II)) {
456 FunctionEffect |= Mod;
457 if (cast<StoreInst>(*II).isVolatile())
458 // Treat volatile stores as reading memory somewhere.
459 FunctionEffect |= Ref;
460 } else if (isMalloc(&cast<Instruction>(*II)) ||
461 isFreeCall(&cast<Instruction>(*II))) {
462 FunctionEffect |= ModRef;
465 if ((FunctionEffect & Mod) == 0)
466 ++NumReadMemFunctions;
467 if (FunctionEffect == 0)
469 FR.FunctionEffect = FunctionEffect;
471 // Finally, now that we know the full effect on this SCC, clone the
472 // information to each function in the SCC.
473 for (unsigned i = 1, e = SCC.size(); i != e; ++i)
474 FunctionInfo[SCC[i]->getFunction()] = FR;
480 /// alias - If one of the pointers is to a global that we are tracking, and the
481 /// other is some random pointer, we know there cannot be an alias, because the
482 /// address of the global isn't taken.
483 AliasAnalysis::AliasResult
484 GlobalsModRef::alias(const Location &LocA,
485 const Location &LocB) {
486 // Get the base object these pointers point to.
487 const Value *UV1 = LocA.Ptr->getUnderlyingObject();
488 const Value *UV2 = LocB.Ptr->getUnderlyingObject();
490 // If either of the underlying values is a global, they may be non-addr-taken
491 // globals, which we can answer queries about.
492 const GlobalValue *GV1 = dyn_cast<GlobalValue>(UV1);
493 const GlobalValue *GV2 = dyn_cast<GlobalValue>(UV2);
495 // If the global's address is taken, pretend we don't know it's a pointer to
497 if (GV1 && !NonAddressTakenGlobals.count(GV1)) GV1 = 0;
498 if (GV2 && !NonAddressTakenGlobals.count(GV2)) GV2 = 0;
500 // If the two pointers are derived from two different non-addr-taken
501 // globals, or if one is and the other isn't, we know these can't alias.
502 if ((GV1 || GV2) && GV1 != GV2)
505 // Otherwise if they are both derived from the same addr-taken global, we
506 // can't know the two accesses don't overlap.
509 // These pointers may be based on the memory owned by an indirect global. If
510 // so, we may be able to handle this. First check to see if the base pointer
511 // is a direct load from an indirect global.
513 if (const LoadInst *LI = dyn_cast<LoadInst>(UV1))
514 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getOperand(0)))
515 if (IndirectGlobals.count(GV))
517 if (const LoadInst *LI = dyn_cast<LoadInst>(UV2))
518 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getOperand(0)))
519 if (IndirectGlobals.count(GV))
522 // These pointers may also be from an allocation for the indirect global. If
523 // so, also handle them.
524 if (AllocsForIndirectGlobals.count(UV1))
525 GV1 = AllocsForIndirectGlobals[UV1];
526 if (AllocsForIndirectGlobals.count(UV2))
527 GV2 = AllocsForIndirectGlobals[UV2];
529 // Now that we know whether the two pointers are related to indirect globals,
530 // use this to disambiguate the pointers. If either pointer is based on an
531 // indirect global and if they are not both based on the same indirect global,
532 // they cannot alias.
533 if ((GV1 || GV2) && GV1 != GV2)
536 return AliasAnalysis::alias(LocA, LocB);
539 AliasAnalysis::ModRefResult
540 GlobalsModRef::getModRefInfo(ImmutableCallSite CS,
541 const Location &Loc) {
542 unsigned Known = ModRef;
544 // If we are asking for mod/ref info of a direct call with a pointer to a
545 // global we are tracking, return information if we have it.
546 if (const GlobalValue *GV =
547 dyn_cast<GlobalValue>(Loc.Ptr->getUnderlyingObject()))
548 if (GV->hasLocalLinkage())
549 if (const Function *F = CS.getCalledFunction())
550 if (NonAddressTakenGlobals.count(GV))
551 if (const FunctionRecord *FR = getFunctionInfo(F))
552 Known = FR->getInfoForGlobal(GV);
554 if (Known == NoModRef)
555 return NoModRef; // No need to query other mod/ref analyses
556 return ModRefResult(Known & AliasAnalysis::getModRefInfo(CS, Loc));
560 //===----------------------------------------------------------------------===//
561 // Methods to update the analysis as a result of the client transformation.
563 void GlobalsModRef::deleteValue(Value *V) {
564 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
565 if (NonAddressTakenGlobals.erase(GV)) {
566 // This global might be an indirect global. If so, remove it and remove
567 // any AllocRelatedValues for it.
568 if (IndirectGlobals.erase(GV)) {
569 // Remove any entries in AllocsForIndirectGlobals for this global.
570 for (std::map<const Value*, const GlobalValue*>::iterator
571 I = AllocsForIndirectGlobals.begin(),
572 E = AllocsForIndirectGlobals.end(); I != E; ) {
573 if (I->second == GV) {
574 AllocsForIndirectGlobals.erase(I++);
583 // Otherwise, if this is an allocation related to an indirect global, remove
585 AllocsForIndirectGlobals.erase(V);
587 AliasAnalysis::deleteValue(V);
590 void GlobalsModRef::copyValue(Value *From, Value *To) {
591 AliasAnalysis::copyValue(From, To);