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 ModRefBehavior Min = UnknownModRefBehavior;
125 if (FunctionRecord *FR = getFunctionInfo(F)) {
126 if (FR->FunctionEffect == 0)
127 Min = DoesNotAccessMemory;
128 else if ((FR->FunctionEffect & Mod) == 0)
129 Min = OnlyReadsMemory;
132 return ModRefBehavior(AliasAnalysis::getModRefBehavior(F) & Min);
135 /// getModRefBehavior - Return the behavior of the specified function if
136 /// called from the specified call site. The call site may be null in which
137 /// case the most generic behavior of this function should be returned.
138 ModRefBehavior getModRefBehavior(ImmutableCallSite CS) {
139 ModRefBehavior Min = UnknownModRefBehavior;
141 if (const Function* F = CS.getCalledFunction())
142 if (FunctionRecord *FR = getFunctionInfo(F)) {
143 if (FR->FunctionEffect == 0)
144 Min = DoesNotAccessMemory;
145 else if ((FR->FunctionEffect & Mod) == 0)
146 Min = OnlyReadsMemory;
149 return ModRefBehavior(AliasAnalysis::getModRefBehavior(CS) & Min);
152 virtual void deleteValue(Value *V);
153 virtual void copyValue(Value *From, Value *To);
155 /// getAdjustedAnalysisPointer - This method is used when a pass implements
156 /// an analysis interface through multiple inheritance. If needed, it
157 /// should override this to adjust the this pointer as needed for the
158 /// specified pass info.
159 virtual void *getAdjustedAnalysisPointer(AnalysisID PI) {
160 if (PI == &AliasAnalysis::ID)
161 return (AliasAnalysis*)this;
166 /// getFunctionInfo - Return the function info for the function, or null if
167 /// we don't have anything useful to say about it.
168 FunctionRecord *getFunctionInfo(const Function *F) {
169 std::map<const Function*, FunctionRecord>::iterator I =
170 FunctionInfo.find(F);
171 if (I != FunctionInfo.end())
176 void AnalyzeGlobals(Module &M);
177 void AnalyzeCallGraph(CallGraph &CG, Module &M);
178 bool AnalyzeUsesOfPointer(Value *V, std::vector<Function*> &Readers,
179 std::vector<Function*> &Writers,
180 GlobalValue *OkayStoreDest = 0);
181 bool AnalyzeIndirectGlobalMemory(GlobalValue *GV);
185 char GlobalsModRef::ID = 0;
186 INITIALIZE_AG_PASS_BEGIN(GlobalsModRef, AliasAnalysis,
187 "globalsmodref-aa", "Simple mod/ref analysis for globals",
189 INITIALIZE_AG_DEPENDENCY(CallGraph)
190 INITIALIZE_AG_PASS_END(GlobalsModRef, AliasAnalysis,
191 "globalsmodref-aa", "Simple mod/ref analysis for globals",
194 Pass *llvm::createGlobalsModRefPass() { return new GlobalsModRef(); }
196 /// AnalyzeGlobals - Scan through the users of all of the internal
197 /// GlobalValue's in the program. If none of them have their "address taken"
198 /// (really, their address passed to something nontrivial), record this fact,
199 /// and record the functions that they are used directly in.
200 void GlobalsModRef::AnalyzeGlobals(Module &M) {
201 std::vector<Function*> Readers, Writers;
202 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
203 if (I->hasLocalLinkage()) {
204 if (!AnalyzeUsesOfPointer(I, Readers, Writers)) {
205 // Remember that we are tracking this global.
206 NonAddressTakenGlobals.insert(I);
207 ++NumNonAddrTakenFunctions;
209 Readers.clear(); Writers.clear();
212 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
214 if (I->hasLocalLinkage()) {
215 if (!AnalyzeUsesOfPointer(I, Readers, Writers)) {
216 // Remember that we are tracking this global, and the mod/ref fns
217 NonAddressTakenGlobals.insert(I);
219 for (unsigned i = 0, e = Readers.size(); i != e; ++i)
220 FunctionInfo[Readers[i]].GlobalInfo[I] |= Ref;
222 if (!I->isConstant()) // No need to keep track of writers to constants
223 for (unsigned i = 0, e = Writers.size(); i != e; ++i)
224 FunctionInfo[Writers[i]].GlobalInfo[I] |= Mod;
225 ++NumNonAddrTakenGlobalVars;
227 // If this global holds a pointer type, see if it is an indirect global.
228 if (I->getType()->getElementType()->isPointerTy() &&
229 AnalyzeIndirectGlobalMemory(I))
230 ++NumIndirectGlobalVars;
232 Readers.clear(); Writers.clear();
236 /// AnalyzeUsesOfPointer - Look at all of the users of the specified pointer.
237 /// If this is used by anything complex (i.e., the address escapes), return
238 /// true. Also, while we are at it, keep track of those functions that read and
239 /// write to the value.
241 /// If OkayStoreDest is non-null, stores into this global are allowed.
242 bool GlobalsModRef::AnalyzeUsesOfPointer(Value *V,
243 std::vector<Function*> &Readers,
244 std::vector<Function*> &Writers,
245 GlobalValue *OkayStoreDest) {
246 if (!V->getType()->isPointerTy()) return true;
248 for (Value::use_iterator UI = V->use_begin(), E=V->use_end(); UI != E; ++UI) {
250 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
251 Readers.push_back(LI->getParent()->getParent());
252 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
253 if (V == SI->getOperand(1)) {
254 Writers.push_back(SI->getParent()->getParent());
255 } else if (SI->getOperand(1) != OkayStoreDest) {
256 return true; // Storing the pointer
258 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
259 if (AnalyzeUsesOfPointer(GEP, Readers, Writers)) return true;
260 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
261 if (AnalyzeUsesOfPointer(BCI, Readers, Writers, OkayStoreDest))
263 } else if (isFreeCall(U)) {
264 Writers.push_back(cast<Instruction>(U)->getParent()->getParent());
265 } else if (CallInst *CI = dyn_cast<CallInst>(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 = CI->getNumArgOperands(); i != e; ++i)
269 if (CI->getArgOperand(i) == V) return true;
270 } else if (InvokeInst *II = dyn_cast<InvokeInst>(U)) {
271 // Make sure that this is just the function being called, not that it is
272 // passing into the function.
273 for (unsigned i = 0, e = II->getNumArgOperands(); i != e; ++i)
274 if (II->getArgOperand(i) == V) return true;
275 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
276 if (CE->getOpcode() == Instruction::GetElementPtr ||
277 CE->getOpcode() == Instruction::BitCast) {
278 if (AnalyzeUsesOfPointer(CE, Readers, Writers))
283 } else if (ICmpInst *ICI = dyn_cast<ICmpInst>(U)) {
284 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
285 return true; // Allow comparison against null.
294 /// AnalyzeIndirectGlobalMemory - We found an non-address-taken global variable
295 /// which holds a pointer type. See if the global always points to non-aliased
296 /// heap memory: that is, all initializers of the globals are allocations, and
297 /// those allocations have no use other than initialization of the global.
298 /// Further, all loads out of GV must directly use the memory, not store the
299 /// pointer somewhere. If this is true, we consider the memory pointed to by
300 /// GV to be owned by GV and can disambiguate other pointers from it.
301 bool GlobalsModRef::AnalyzeIndirectGlobalMemory(GlobalValue *GV) {
302 // Keep track of values related to the allocation of the memory, f.e. the
303 // value produced by the malloc call and any casts.
304 std::vector<Value*> AllocRelatedValues;
306 // Walk the user list of the global. If we find anything other than a direct
307 // load or store, bail out.
308 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I){
310 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
311 // The pointer loaded from the global can only be used in simple ways:
312 // we allow addressing of it and loading storing to it. We do *not* allow
313 // storing the loaded pointer somewhere else or passing to a function.
314 std::vector<Function*> ReadersWriters;
315 if (AnalyzeUsesOfPointer(LI, ReadersWriters, ReadersWriters))
316 return false; // Loaded pointer escapes.
317 // TODO: Could try some IP mod/ref of the loaded pointer.
318 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
319 // Storing the global itself.
320 if (SI->getOperand(0) == GV) return false;
322 // If storing the null pointer, ignore it.
323 if (isa<ConstantPointerNull>(SI->getOperand(0)))
326 // Check the value being stored.
327 Value *Ptr = SI->getOperand(0)->getUnderlyingObject();
331 } else if (CallInst *CI = dyn_cast<CallInst>(Ptr)) {
332 Function *F = CI->getCalledFunction();
333 if (!F || !F->isDeclaration()) return false; // Too hard to analyze.
334 if (F->getName() != "calloc") return false; // Not calloc.
336 return false; // Too hard to analyze.
339 // Analyze all uses of the allocation. If any of them are used in a
340 // non-simple way (e.g. stored to another global) bail out.
341 std::vector<Function*> ReadersWriters;
342 if (AnalyzeUsesOfPointer(Ptr, ReadersWriters, ReadersWriters, GV))
343 return false; // Loaded pointer escapes.
345 // Remember that this allocation is related to the indirect global.
346 AllocRelatedValues.push_back(Ptr);
348 // Something complex, bail out.
353 // Okay, this is an indirect global. Remember all of the allocations for
354 // this global in AllocsForIndirectGlobals.
355 while (!AllocRelatedValues.empty()) {
356 AllocsForIndirectGlobals[AllocRelatedValues.back()] = GV;
357 AllocRelatedValues.pop_back();
359 IndirectGlobals.insert(GV);
363 /// AnalyzeCallGraph - At this point, we know the functions where globals are
364 /// immediately stored to and read from. Propagate this information up the call
365 /// graph to all callers and compute the mod/ref info for all memory for each
367 void GlobalsModRef::AnalyzeCallGraph(CallGraph &CG, Module &M) {
368 // We do a bottom-up SCC traversal of the call graph. In other words, we
369 // visit all callees before callers (leaf-first).
370 for (scc_iterator<CallGraph*> I = scc_begin(&CG), E = scc_end(&CG); I != E;
372 std::vector<CallGraphNode *> &SCC = *I;
373 assert(!SCC.empty() && "SCC with no functions?");
375 if (!SCC[0]->getFunction()) {
376 // Calls externally - can't say anything useful. Remove any existing
377 // function records (may have been created when scanning globals).
378 for (unsigned i = 0, e = SCC.size(); i != e; ++i)
379 FunctionInfo.erase(SCC[i]->getFunction());
383 FunctionRecord &FR = FunctionInfo[SCC[0]->getFunction()];
385 bool KnowNothing = false;
386 unsigned FunctionEffect = 0;
388 // Collect the mod/ref properties due to called functions. We only compute
390 for (unsigned i = 0, e = SCC.size(); i != e && !KnowNothing; ++i) {
391 Function *F = SCC[i]->getFunction();
397 if (F->isDeclaration()) {
398 // Try to get mod/ref behaviour from function attributes.
399 if (F->doesNotAccessMemory()) {
400 // Can't do better than that!
401 } else if (F->onlyReadsMemory()) {
402 FunctionEffect |= Ref;
403 if (!F->isIntrinsic())
404 // This function might call back into the module and read a global -
405 // consider every global as possibly being read by this function.
406 FR.MayReadAnyGlobal = true;
408 FunctionEffect |= ModRef;
409 // Can't say anything useful unless it's an intrinsic - they don't
410 // read or write global variables of the kind considered here.
411 KnowNothing = !F->isIntrinsic();
416 for (CallGraphNode::iterator CI = SCC[i]->begin(), E = SCC[i]->end();
417 CI != E && !KnowNothing; ++CI)
418 if (Function *Callee = CI->second->getFunction()) {
419 if (FunctionRecord *CalleeFR = getFunctionInfo(Callee)) {
420 // Propagate function effect up.
421 FunctionEffect |= CalleeFR->FunctionEffect;
423 // Incorporate callee's effects on globals into our info.
424 for (std::map<const GlobalValue*, unsigned>::iterator GI =
425 CalleeFR->GlobalInfo.begin(), E = CalleeFR->GlobalInfo.end();
427 FR.GlobalInfo[GI->first] |= GI->second;
428 FR.MayReadAnyGlobal |= CalleeFR->MayReadAnyGlobal;
430 // Can't say anything about it. However, if it is inside our SCC,
431 // then nothing needs to be done.
432 CallGraphNode *CalleeNode = CG[Callee];
433 if (std::find(SCC.begin(), SCC.end(), CalleeNode) == SCC.end())
441 // If we can't say anything useful about this SCC, remove all SCC functions
442 // from the FunctionInfo map.
444 for (unsigned i = 0, e = SCC.size(); i != e; ++i)
445 FunctionInfo.erase(SCC[i]->getFunction());
449 // Scan the function bodies for explicit loads or stores.
450 for (unsigned i = 0, e = SCC.size(); i != e && FunctionEffect != ModRef;++i)
451 for (inst_iterator II = inst_begin(SCC[i]->getFunction()),
452 E = inst_end(SCC[i]->getFunction());
453 II != E && FunctionEffect != ModRef; ++II)
454 if (isa<LoadInst>(*II)) {
455 FunctionEffect |= Ref;
456 if (cast<LoadInst>(*II).isVolatile())
457 // Volatile loads may have side-effects, so mark them as writing
458 // memory (for example, a flag inside the processor).
459 FunctionEffect |= Mod;
460 } else if (isa<StoreInst>(*II)) {
461 FunctionEffect |= Mod;
462 if (cast<StoreInst>(*II).isVolatile())
463 // Treat volatile stores as reading memory somewhere.
464 FunctionEffect |= Ref;
465 } else if (isMalloc(&cast<Instruction>(*II)) ||
466 isFreeCall(&cast<Instruction>(*II))) {
467 FunctionEffect |= ModRef;
470 if ((FunctionEffect & Mod) == 0)
471 ++NumReadMemFunctions;
472 if (FunctionEffect == 0)
474 FR.FunctionEffect = FunctionEffect;
476 // Finally, now that we know the full effect on this SCC, clone the
477 // information to each function in the SCC.
478 for (unsigned i = 1, e = SCC.size(); i != e; ++i)
479 FunctionInfo[SCC[i]->getFunction()] = FR;
485 /// alias - If one of the pointers is to a global that we are tracking, and the
486 /// other is some random pointer, we know there cannot be an alias, because the
487 /// address of the global isn't taken.
488 AliasAnalysis::AliasResult
489 GlobalsModRef::alias(const Location &LocA,
490 const Location &LocB) {
491 // Get the base object these pointers point to.
492 const Value *UV1 = LocA.Ptr->getUnderlyingObject();
493 const Value *UV2 = LocB.Ptr->getUnderlyingObject();
495 // If either of the underlying values is a global, they may be non-addr-taken
496 // globals, which we can answer queries about.
497 const GlobalValue *GV1 = dyn_cast<GlobalValue>(UV1);
498 const GlobalValue *GV2 = dyn_cast<GlobalValue>(UV2);
500 // If the global's address is taken, pretend we don't know it's a pointer to
502 if (GV1 && !NonAddressTakenGlobals.count(GV1)) GV1 = 0;
503 if (GV2 && !NonAddressTakenGlobals.count(GV2)) GV2 = 0;
505 // If the two pointers are derived from two different non-addr-taken
506 // globals, or if one is and the other isn't, we know these can't alias.
507 if ((GV1 || GV2) && GV1 != GV2)
510 // Otherwise if they are both derived from the same addr-taken global, we
511 // can't know the two accesses don't overlap.
514 // These pointers may be based on the memory owned by an indirect global. If
515 // so, we may be able to handle this. First check to see if the base pointer
516 // is a direct load from an indirect global.
518 if (const LoadInst *LI = dyn_cast<LoadInst>(UV1))
519 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getOperand(0)))
520 if (IndirectGlobals.count(GV))
522 if (const LoadInst *LI = dyn_cast<LoadInst>(UV2))
523 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getOperand(0)))
524 if (IndirectGlobals.count(GV))
527 // These pointers may also be from an allocation for the indirect global. If
528 // so, also handle them.
529 if (AllocsForIndirectGlobals.count(UV1))
530 GV1 = AllocsForIndirectGlobals[UV1];
531 if (AllocsForIndirectGlobals.count(UV2))
532 GV2 = AllocsForIndirectGlobals[UV2];
534 // Now that we know whether the two pointers are related to indirect globals,
535 // use this to disambiguate the pointers. If either pointer is based on an
536 // indirect global and if they are not both based on the same indirect global,
537 // they cannot alias.
538 if ((GV1 || GV2) && GV1 != GV2)
541 return AliasAnalysis::alias(LocA, LocB);
544 AliasAnalysis::ModRefResult
545 GlobalsModRef::getModRefInfo(ImmutableCallSite CS,
546 const Location &Loc) {
547 unsigned Known = ModRef;
549 // If we are asking for mod/ref info of a direct call with a pointer to a
550 // global we are tracking, return information if we have it.
551 if (const GlobalValue *GV =
552 dyn_cast<GlobalValue>(Loc.Ptr->getUnderlyingObject()))
553 if (GV->hasLocalLinkage())
554 if (const Function *F = CS.getCalledFunction())
555 if (NonAddressTakenGlobals.count(GV))
556 if (const FunctionRecord *FR = getFunctionInfo(F))
557 Known = FR->getInfoForGlobal(GV);
559 if (Known == NoModRef)
560 return NoModRef; // No need to query other mod/ref analyses
561 return ModRefResult(Known & AliasAnalysis::getModRefInfo(CS, Loc));
565 //===----------------------------------------------------------------------===//
566 // Methods to update the analysis as a result of the client transformation.
568 void GlobalsModRef::deleteValue(Value *V) {
569 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
570 if (NonAddressTakenGlobals.erase(GV)) {
571 // This global might be an indirect global. If so, remove it and remove
572 // any AllocRelatedValues for it.
573 if (IndirectGlobals.erase(GV)) {
574 // Remove any entries in AllocsForIndirectGlobals for this global.
575 for (std::map<const Value*, const GlobalValue*>::iterator
576 I = AllocsForIndirectGlobals.begin(),
577 E = AllocsForIndirectGlobals.end(); I != E; ) {
578 if (I->second == GV) {
579 AllocsForIndirectGlobals.erase(I++);
588 // Otherwise, if this is an allocation related to an indirect global, remove
590 AllocsForIndirectGlobals.erase(V);
592 AliasAnalysis::deleteValue(V);
595 void GlobalsModRef::copyValue(Value *From, Value *To) {
596 AliasAnalysis::copyValue(From, To);