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 Location &LocA, const Location &LocB);
110 ModRefResult getModRefInfo(ImmutableCallSite CS,
111 const Location &Loc);
112 ModRefResult getModRefInfo(ImmutableCallSite CS1,
113 ImmutableCallSite CS2) {
114 return AliasAnalysis::getModRefInfo(CS1, CS2);
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 ModRefBehavior getModRefBehavior(const Function *F) {
121 if (FunctionRecord *FR = getFunctionInfo(F)) {
122 if (FR->FunctionEffect == 0)
123 return DoesNotAccessMemory;
124 else if ((FR->FunctionEffect & Mod) == 0)
125 return OnlyReadsMemory;
127 return AliasAnalysis::getModRefBehavior(F);
130 /// getModRefBehavior - Return the behavior of the specified function if
131 /// called from the specified call site. The call site may be null in which
132 /// case the most generic behavior of this function should be returned.
133 ModRefBehavior getModRefBehavior(ImmutableCallSite CS) {
134 const Function* F = CS.getCalledFunction();
135 if (!F) return AliasAnalysis::getModRefBehavior(CS);
136 if (FunctionRecord *FR = getFunctionInfo(F)) {
137 if (FR->FunctionEffect == 0)
138 return DoesNotAccessMemory;
139 else if ((FR->FunctionEffect & Mod) == 0)
140 return OnlyReadsMemory;
142 return AliasAnalysis::getModRefBehavior(CS);
145 virtual void deleteValue(Value *V);
146 virtual void copyValue(Value *From, Value *To);
148 /// getAdjustedAnalysisPointer - This method is used when a pass implements
149 /// an analysis interface through multiple inheritance. If needed, it
150 /// should override this to adjust the this pointer as needed for the
151 /// specified pass info.
152 virtual void *getAdjustedAnalysisPointer(AnalysisID PI) {
153 if (PI == &AliasAnalysis::ID)
154 return (AliasAnalysis*)this;
159 /// getFunctionInfo - Return the function info for the function, or null if
160 /// we don't have anything useful to say about it.
161 FunctionRecord *getFunctionInfo(const Function *F) {
162 std::map<const Function*, FunctionRecord>::iterator I =
163 FunctionInfo.find(F);
164 if (I != FunctionInfo.end())
169 void AnalyzeGlobals(Module &M);
170 void AnalyzeCallGraph(CallGraph &CG, Module &M);
171 bool AnalyzeUsesOfPointer(Value *V, std::vector<Function*> &Readers,
172 std::vector<Function*> &Writers,
173 GlobalValue *OkayStoreDest = 0);
174 bool AnalyzeIndirectGlobalMemory(GlobalValue *GV);
178 char GlobalsModRef::ID = 0;
179 INITIALIZE_AG_PASS(GlobalsModRef, AliasAnalysis,
180 "globalsmodref-aa", "Simple mod/ref analysis for globals",
183 Pass *llvm::createGlobalsModRefPass() { return new GlobalsModRef(); }
185 /// AnalyzeGlobals - Scan through the users of all of the internal
186 /// GlobalValue's in the program. If none of them have their "address taken"
187 /// (really, their address passed to something nontrivial), record this fact,
188 /// and record the functions that they are used directly in.
189 void GlobalsModRef::AnalyzeGlobals(Module &M) {
190 std::vector<Function*> Readers, Writers;
191 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
192 if (I->hasLocalLinkage()) {
193 if (!AnalyzeUsesOfPointer(I, Readers, Writers)) {
194 // Remember that we are tracking this global.
195 NonAddressTakenGlobals.insert(I);
196 ++NumNonAddrTakenFunctions;
198 Readers.clear(); Writers.clear();
201 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
203 if (I->hasLocalLinkage()) {
204 if (!AnalyzeUsesOfPointer(I, Readers, Writers)) {
205 // Remember that we are tracking this global, and the mod/ref fns
206 NonAddressTakenGlobals.insert(I);
208 for (unsigned i = 0, e = Readers.size(); i != e; ++i)
209 FunctionInfo[Readers[i]].GlobalInfo[I] |= Ref;
211 if (!I->isConstant()) // No need to keep track of writers to constants
212 for (unsigned i = 0, e = Writers.size(); i != e; ++i)
213 FunctionInfo[Writers[i]].GlobalInfo[I] |= Mod;
214 ++NumNonAddrTakenGlobalVars;
216 // If this global holds a pointer type, see if it is an indirect global.
217 if (I->getType()->getElementType()->isPointerTy() &&
218 AnalyzeIndirectGlobalMemory(I))
219 ++NumIndirectGlobalVars;
221 Readers.clear(); Writers.clear();
225 /// AnalyzeUsesOfPointer - Look at all of the users of the specified pointer.
226 /// If this is used by anything complex (i.e., the address escapes), return
227 /// true. Also, while we are at it, keep track of those functions that read and
228 /// write to the value.
230 /// If OkayStoreDest is non-null, stores into this global are allowed.
231 bool GlobalsModRef::AnalyzeUsesOfPointer(Value *V,
232 std::vector<Function*> &Readers,
233 std::vector<Function*> &Writers,
234 GlobalValue *OkayStoreDest) {
235 if (!V->getType()->isPointerTy()) return true;
237 for (Value::use_iterator UI = V->use_begin(), E=V->use_end(); UI != E; ++UI) {
239 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
240 Readers.push_back(LI->getParent()->getParent());
241 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
242 if (V == SI->getOperand(1)) {
243 Writers.push_back(SI->getParent()->getParent());
244 } else if (SI->getOperand(1) != OkayStoreDest) {
245 return true; // Storing the pointer
247 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
248 if (AnalyzeUsesOfPointer(GEP, Readers, Writers)) return true;
249 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
250 if (AnalyzeUsesOfPointer(BCI, Readers, Writers, OkayStoreDest))
252 } else if (isFreeCall(U)) {
253 Writers.push_back(cast<Instruction>(U)->getParent()->getParent());
254 } else if (CallInst *CI = dyn_cast<CallInst>(U)) {
255 // Make sure that this is just the function being called, not that it is
256 // passing into the function.
257 for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i)
258 if (CI->getArgOperand(i) == V) return true;
259 } else if (InvokeInst *II = dyn_cast<InvokeInst>(U)) {
260 // Make sure that this is just the function being called, not that it is
261 // passing into the function.
262 for (unsigned i = 0, e = II->getNumArgOperands(); i != e; ++i)
263 if (II->getArgOperand(i) == V) return true;
264 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
265 if (CE->getOpcode() == Instruction::GetElementPtr ||
266 CE->getOpcode() == Instruction::BitCast) {
267 if (AnalyzeUsesOfPointer(CE, Readers, Writers))
272 } else if (ICmpInst *ICI = dyn_cast<ICmpInst>(U)) {
273 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
274 return true; // Allow comparison against null.
283 /// AnalyzeIndirectGlobalMemory - We found an non-address-taken global variable
284 /// which holds a pointer type. See if the global always points to non-aliased
285 /// heap memory: that is, all initializers of the globals are allocations, and
286 /// those allocations have no use other than initialization of the global.
287 /// Further, all loads out of GV must directly use the memory, not store the
288 /// pointer somewhere. If this is true, we consider the memory pointed to by
289 /// GV to be owned by GV and can disambiguate other pointers from it.
290 bool GlobalsModRef::AnalyzeIndirectGlobalMemory(GlobalValue *GV) {
291 // Keep track of values related to the allocation of the memory, f.e. the
292 // value produced by the malloc call and any casts.
293 std::vector<Value*> AllocRelatedValues;
295 // Walk the user list of the global. If we find anything other than a direct
296 // load or store, bail out.
297 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I){
299 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
300 // The pointer loaded from the global can only be used in simple ways:
301 // we allow addressing of it and loading storing to it. We do *not* allow
302 // storing the loaded pointer somewhere else or passing to a function.
303 std::vector<Function*> ReadersWriters;
304 if (AnalyzeUsesOfPointer(LI, ReadersWriters, ReadersWriters))
305 return false; // Loaded pointer escapes.
306 // TODO: Could try some IP mod/ref of the loaded pointer.
307 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
308 // Storing the global itself.
309 if (SI->getOperand(0) == GV) return false;
311 // If storing the null pointer, ignore it.
312 if (isa<ConstantPointerNull>(SI->getOperand(0)))
315 // Check the value being stored.
316 Value *Ptr = SI->getOperand(0)->getUnderlyingObject();
320 } else if (CallInst *CI = dyn_cast<CallInst>(Ptr)) {
321 Function *F = CI->getCalledFunction();
322 if (!F || !F->isDeclaration()) return false; // Too hard to analyze.
323 if (F->getName() != "calloc") return false; // Not calloc.
325 return false; // Too hard to analyze.
328 // Analyze all uses of the allocation. If any of them are used in a
329 // non-simple way (e.g. stored to another global) bail out.
330 std::vector<Function*> ReadersWriters;
331 if (AnalyzeUsesOfPointer(Ptr, ReadersWriters, ReadersWriters, GV))
332 return false; // Loaded pointer escapes.
334 // Remember that this allocation is related to the indirect global.
335 AllocRelatedValues.push_back(Ptr);
337 // Something complex, bail out.
342 // Okay, this is an indirect global. Remember all of the allocations for
343 // this global in AllocsForIndirectGlobals.
344 while (!AllocRelatedValues.empty()) {
345 AllocsForIndirectGlobals[AllocRelatedValues.back()] = GV;
346 AllocRelatedValues.pop_back();
348 IndirectGlobals.insert(GV);
352 /// AnalyzeCallGraph - At this point, we know the functions where globals are
353 /// immediately stored to and read from. Propagate this information up the call
354 /// graph to all callers and compute the mod/ref info for all memory for each
356 void GlobalsModRef::AnalyzeCallGraph(CallGraph &CG, Module &M) {
357 // We do a bottom-up SCC traversal of the call graph. In other words, we
358 // visit all callees before callers (leaf-first).
359 for (scc_iterator<CallGraph*> I = scc_begin(&CG), E = scc_end(&CG); I != E;
361 std::vector<CallGraphNode *> &SCC = *I;
362 assert(!SCC.empty() && "SCC with no functions?");
364 if (!SCC[0]->getFunction()) {
365 // Calls externally - can't say anything useful. Remove any existing
366 // function records (may have been created when scanning globals).
367 for (unsigned i = 0, e = SCC.size(); i != e; ++i)
368 FunctionInfo.erase(SCC[i]->getFunction());
372 FunctionRecord &FR = FunctionInfo[SCC[0]->getFunction()];
374 bool KnowNothing = 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 && !KnowNothing; ++i) {
380 Function *F = SCC[i]->getFunction();
386 if (F->isDeclaration()) {
387 // Try to get mod/ref behaviour from function attributes.
388 if (F->doesNotAccessMemory()) {
389 // Can't do better than that!
390 } else if (F->onlyReadsMemory()) {
391 FunctionEffect |= Ref;
392 if (!F->isIntrinsic())
393 // This function might call back into the module and read a global -
394 // consider every global as possibly being read by this function.
395 FR.MayReadAnyGlobal = true;
397 FunctionEffect |= ModRef;
398 // Can't say anything useful unless it's an intrinsic - they don't
399 // read or write global variables of the kind considered here.
400 KnowNothing = !F->isIntrinsic();
405 for (CallGraphNode::iterator CI = SCC[i]->begin(), E = SCC[i]->end();
406 CI != E && !KnowNothing; ++CI)
407 if (Function *Callee = CI->second->getFunction()) {
408 if (FunctionRecord *CalleeFR = getFunctionInfo(Callee)) {
409 // Propagate function effect up.
410 FunctionEffect |= CalleeFR->FunctionEffect;
412 // Incorporate callee's effects on globals into our info.
413 for (std::map<const GlobalValue*, unsigned>::iterator GI =
414 CalleeFR->GlobalInfo.begin(), E = CalleeFR->GlobalInfo.end();
416 FR.GlobalInfo[GI->first] |= GI->second;
417 FR.MayReadAnyGlobal |= CalleeFR->MayReadAnyGlobal;
419 // Can't say anything about it. However, if it is inside our SCC,
420 // then nothing needs to be done.
421 CallGraphNode *CalleeNode = CG[Callee];
422 if (std::find(SCC.begin(), SCC.end(), CalleeNode) == SCC.end())
430 // If we can't say anything useful about this SCC, remove all SCC functions
431 // from the FunctionInfo map.
433 for (unsigned i = 0, e = SCC.size(); i != e; ++i)
434 FunctionInfo.erase(SCC[i]->getFunction());
438 // Scan the function bodies for explicit loads or stores.
439 for (unsigned i = 0, e = SCC.size(); i != e && FunctionEffect != ModRef;++i)
440 for (inst_iterator II = inst_begin(SCC[i]->getFunction()),
441 E = inst_end(SCC[i]->getFunction());
442 II != E && FunctionEffect != ModRef; ++II)
443 if (isa<LoadInst>(*II)) {
444 FunctionEffect |= Ref;
445 if (cast<LoadInst>(*II).isVolatile())
446 // Volatile loads may have side-effects, so mark them as writing
447 // memory (for example, a flag inside the processor).
448 FunctionEffect |= Mod;
449 } else if (isa<StoreInst>(*II)) {
450 FunctionEffect |= Mod;
451 if (cast<StoreInst>(*II).isVolatile())
452 // Treat volatile stores as reading memory somewhere.
453 FunctionEffect |= Ref;
454 } else if (isMalloc(&cast<Instruction>(*II)) ||
455 isFreeCall(&cast<Instruction>(*II))) {
456 FunctionEffect |= ModRef;
459 if ((FunctionEffect & Mod) == 0)
460 ++NumReadMemFunctions;
461 if (FunctionEffect == 0)
463 FR.FunctionEffect = FunctionEffect;
465 // Finally, now that we know the full effect on this SCC, clone the
466 // information to each function in the SCC.
467 for (unsigned i = 1, e = SCC.size(); i != e; ++i)
468 FunctionInfo[SCC[i]->getFunction()] = FR;
474 /// alias - If one of the pointers is to a global that we are tracking, and the
475 /// other is some random pointer, we know there cannot be an alias, because the
476 /// address of the global isn't taken.
477 AliasAnalysis::AliasResult
478 GlobalsModRef::alias(const Location &LocA,
479 const Location &LocB) {
480 // Get the base object these pointers point to.
481 const Value *UV1 = LocA.Ptr->getUnderlyingObject();
482 const Value *UV2 = LocB.Ptr->getUnderlyingObject();
484 // If either of the underlying values is a global, they may be non-addr-taken
485 // globals, which we can answer queries about.
486 const GlobalValue *GV1 = dyn_cast<GlobalValue>(UV1);
487 const GlobalValue *GV2 = dyn_cast<GlobalValue>(UV2);
489 // If the global's address is taken, pretend we don't know it's a pointer to
491 if (GV1 && !NonAddressTakenGlobals.count(GV1)) GV1 = 0;
492 if (GV2 && !NonAddressTakenGlobals.count(GV2)) GV2 = 0;
494 // If the two pointers are derived from two different non-addr-taken
495 // globals, or if one is and the other isn't, we know these can't alias.
496 if ((GV1 || GV2) && GV1 != GV2)
499 // Otherwise if they are both derived from the same addr-taken global, we
500 // can't know the two accesses don't overlap.
503 // These pointers may be based on the memory owned by an indirect global. If
504 // so, we may be able to handle this. First check to see if the base pointer
505 // is a direct load from an indirect global.
507 if (const LoadInst *LI = dyn_cast<LoadInst>(UV1))
508 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getOperand(0)))
509 if (IndirectGlobals.count(GV))
511 if (const LoadInst *LI = dyn_cast<LoadInst>(UV2))
512 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getOperand(0)))
513 if (IndirectGlobals.count(GV))
516 // These pointers may also be from an allocation for the indirect global. If
517 // so, also handle them.
518 if (AllocsForIndirectGlobals.count(UV1))
519 GV1 = AllocsForIndirectGlobals[UV1];
520 if (AllocsForIndirectGlobals.count(UV2))
521 GV2 = AllocsForIndirectGlobals[UV2];
523 // Now that we know whether the two pointers are related to indirect globals,
524 // use this to disambiguate the pointers. If either pointer is based on an
525 // indirect global and if they are not both based on the same indirect global,
526 // they cannot alias.
527 if ((GV1 || GV2) && GV1 != GV2)
530 return AliasAnalysis::alias(LocA, LocB);
533 AliasAnalysis::ModRefResult
534 GlobalsModRef::getModRefInfo(ImmutableCallSite CS,
535 const Location &Loc) {
536 unsigned Known = ModRef;
538 // If we are asking for mod/ref info of a direct call with a pointer to a
539 // global we are tracking, return information if we have it.
540 if (const GlobalValue *GV =
541 dyn_cast<GlobalValue>(Loc.Ptr->getUnderlyingObject()))
542 if (GV->hasLocalLinkage())
543 if (const Function *F = CS.getCalledFunction())
544 if (NonAddressTakenGlobals.count(GV))
545 if (const FunctionRecord *FR = getFunctionInfo(F))
546 Known = FR->getInfoForGlobal(GV);
548 if (Known == NoModRef)
549 return NoModRef; // No need to query other mod/ref analyses
550 return ModRefResult(Known & AliasAnalysis::getModRefInfo(CS, Loc));
554 //===----------------------------------------------------------------------===//
555 // Methods to update the analysis as a result of the client transformation.
557 void GlobalsModRef::deleteValue(Value *V) {
558 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
559 if (NonAddressTakenGlobals.erase(GV)) {
560 // This global might be an indirect global. If so, remove it and remove
561 // any AllocRelatedValues for it.
562 if (IndirectGlobals.erase(GV)) {
563 // Remove any entries in AllocsForIndirectGlobals for this global.
564 for (std::map<const Value*, const GlobalValue*>::iterator
565 I = AllocsForIndirectGlobals.begin(),
566 E = AllocsForIndirectGlobals.end(); I != E; ) {
567 if (I->second == GV) {
568 AllocsForIndirectGlobals.erase(I++);
577 // Otherwise, if this is an allocation related to an indirect global, remove
579 AllocsForIndirectGlobals.erase(V);
581 AliasAnalysis::deleteValue(V);
584 void GlobalsModRef::copyValue(Value *From, Value *To) {
585 AliasAnalysis::copyValue(From, To);