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/MallocHelper.h"
27 #include "llvm/Support/Compiler.h"
28 #include "llvm/Support/CommandLine.h"
29 #include "llvm/Support/InstIterator.h"
30 #include "llvm/ADT/Statistic.h"
31 #include "llvm/ADT/SCCIterator.h"
35 STATISTIC(NumNonAddrTakenGlobalVars,
36 "Number of global vars without address taken");
37 STATISTIC(NumNonAddrTakenFunctions,"Number of functions without address taken");
38 STATISTIC(NumNoMemFunctions, "Number of functions that do not access memory");
39 STATISTIC(NumReadMemFunctions, "Number of functions that only read memory");
40 STATISTIC(NumIndirectGlobalVars, "Number of indirect global objects");
43 /// FunctionRecord - One instance of this structure is stored for every
44 /// function in the program. Later, the entries for these functions are
45 /// removed if the function is found to call an external function (in which
46 /// case we know nothing about it.
47 struct VISIBILITY_HIDDEN FunctionRecord {
48 /// GlobalInfo - Maintain mod/ref info for all of the globals without
49 /// addresses taken that are read or written (transitively) by this
51 std::map<GlobalValue*, unsigned> GlobalInfo;
53 /// MayReadAnyGlobal - May read global variables, but it is not known which.
54 bool MayReadAnyGlobal;
56 unsigned getInfoForGlobal(GlobalValue *GV) const {
57 unsigned Effect = MayReadAnyGlobal ? AliasAnalysis::Ref : 0;
58 std::map<GlobalValue*, unsigned>::const_iterator I = GlobalInfo.find(GV);
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 VISIBILITY_HIDDEN GlobalsModRef
73 : public ModulePass, public AliasAnalysis {
74 /// NonAddressTakenGlobals - The globals that do not have their addresses
76 std::set<GlobalValue*> NonAddressTakenGlobals;
78 /// IndirectGlobals - The memory pointed to by this global is known to be
79 /// 'owned' by the global.
80 std::set<GlobalValue*> IndirectGlobals;
82 /// AllocsForIndirectGlobals - If an instruction allocates memory for an
83 /// indirect global, this map indicates which one.
84 std::map<Value*, GlobalValue*> AllocsForIndirectGlobals;
86 /// FunctionInfo - For each function, keep track of what globals are
88 std::map<Function*, FunctionRecord> FunctionInfo;
92 GlobalsModRef() : ModulePass(&ID) {}
94 bool runOnModule(Module &M) {
95 InitializeAliasAnalysis(this); // set up super class
96 AnalyzeGlobals(M); // find non-addr taken globals
97 AnalyzeCallGraph(getAnalysis<CallGraph>(), M); // Propagate on CG
101 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
102 AliasAnalysis::getAnalysisUsage(AU);
103 AU.addRequired<CallGraph>();
104 AU.setPreservesAll(); // Does not transform code
107 //------------------------------------------------
108 // Implement the AliasAnalysis API
110 AliasResult alias(const Value *V1, unsigned V1Size,
111 const Value *V2, unsigned V2Size);
112 ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size);
113 ModRefResult getModRefInfo(CallSite CS1, CallSite CS2) {
114 return AliasAnalysis::getModRefInfo(CS1,CS2);
116 bool hasNoModRefInfoForCalls() const { return false; }
118 /// getModRefBehavior - Return the behavior of the specified function if
119 /// called from the specified call site. The call site may be null in which
120 /// case the most generic behavior of this function should be returned.
121 ModRefBehavior getModRefBehavior(Function *F,
122 std::vector<PointerAccessInfo> *Info) {
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, Info);
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(CallSite CS,
136 std::vector<PointerAccessInfo> *Info) {
137 Function* F = CS.getCalledFunction();
138 if (!F) return AliasAnalysis::getModRefBehavior(CS, Info);
139 if (FunctionRecord *FR = getFunctionInfo(F)) {
140 if (FR->FunctionEffect == 0)
141 return DoesNotAccessMemory;
142 else if ((FR->FunctionEffect & Mod) == 0)
143 return OnlyReadsMemory;
145 return AliasAnalysis::getModRefBehavior(CS, Info);
148 virtual void deleteValue(Value *V);
149 virtual void copyValue(Value *From, Value *To);
152 /// getFunctionInfo - Return the function info for the function, or null if
153 /// we don't have anything useful to say about it.
154 FunctionRecord *getFunctionInfo(Function *F) {
155 std::map<Function*, FunctionRecord>::iterator I = FunctionInfo.find(F);
156 if (I != FunctionInfo.end())
161 void AnalyzeGlobals(Module &M);
162 void AnalyzeCallGraph(CallGraph &CG, Module &M);
163 bool AnalyzeUsesOfPointer(Value *V, std::vector<Function*> &Readers,
164 std::vector<Function*> &Writers,
165 GlobalValue *OkayStoreDest = 0);
166 bool AnalyzeIndirectGlobalMemory(GlobalValue *GV);
170 char GlobalsModRef::ID = 0;
171 static RegisterPass<GlobalsModRef>
172 X("globalsmodref-aa", "Simple mod/ref analysis for globals", false, true);
173 static RegisterAnalysisGroup<AliasAnalysis> Y(X);
175 Pass *llvm::createGlobalsModRefPass() { return new GlobalsModRef(); }
177 /// AnalyzeGlobals - Scan through the users of all of the internal
178 /// GlobalValue's in the program. If none of them have their "address taken"
179 /// (really, their address passed to something nontrivial), record this fact,
180 /// and record the functions that they are used directly in.
181 void GlobalsModRef::AnalyzeGlobals(Module &M) {
182 std::vector<Function*> Readers, Writers;
183 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
184 if (I->hasLocalLinkage()) {
185 if (!AnalyzeUsesOfPointer(I, Readers, Writers)) {
186 // Remember that we are tracking this global.
187 NonAddressTakenGlobals.insert(I);
188 ++NumNonAddrTakenFunctions;
190 Readers.clear(); Writers.clear();
193 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
195 if (I->hasLocalLinkage()) {
196 if (!AnalyzeUsesOfPointer(I, Readers, Writers)) {
197 // Remember that we are tracking this global, and the mod/ref fns
198 NonAddressTakenGlobals.insert(I);
200 for (unsigned i = 0, e = Readers.size(); i != e; ++i)
201 FunctionInfo[Readers[i]].GlobalInfo[I] |= Ref;
203 if (!I->isConstant()) // No need to keep track of writers to constants
204 for (unsigned i = 0, e = Writers.size(); i != e; ++i)
205 FunctionInfo[Writers[i]].GlobalInfo[I] |= Mod;
206 ++NumNonAddrTakenGlobalVars;
208 // If this global holds a pointer type, see if it is an indirect global.
209 if (isa<PointerType>(I->getType()->getElementType()) &&
210 AnalyzeIndirectGlobalMemory(I))
211 ++NumIndirectGlobalVars;
213 Readers.clear(); Writers.clear();
217 /// AnalyzeUsesOfPointer - Look at all of the users of the specified pointer.
218 /// If this is used by anything complex (i.e., the address escapes), return
219 /// true. Also, while we are at it, keep track of those functions that read and
220 /// write to the value.
222 /// If OkayStoreDest is non-null, stores into this global are allowed.
223 bool GlobalsModRef::AnalyzeUsesOfPointer(Value *V,
224 std::vector<Function*> &Readers,
225 std::vector<Function*> &Writers,
226 GlobalValue *OkayStoreDest) {
227 if (!isa<PointerType>(V->getType())) return true;
229 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
230 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
231 Readers.push_back(LI->getParent()->getParent());
232 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
233 if (V == SI->getOperand(1)) {
234 Writers.push_back(SI->getParent()->getParent());
235 } else if (SI->getOperand(1) != OkayStoreDest) {
236 return true; // Storing the pointer
238 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) {
239 if (AnalyzeUsesOfPointer(GEP, Readers, Writers)) return true;
240 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) {
241 if (AnalyzeUsesOfPointer(BCI, Readers, Writers, OkayStoreDest))
243 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
244 // Make sure that this is just the function being called, not that it is
245 // passing into the function.
246 for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i)
247 if (CI->getOperand(i) == V) return true;
248 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
249 // Make sure that this is just the function being called, not that it is
250 // passing into the function.
251 for (unsigned i = 3, e = II->getNumOperands(); i != e; ++i)
252 if (II->getOperand(i) == V) return true;
253 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
254 if (CE->getOpcode() == Instruction::GetElementPtr ||
255 CE->getOpcode() == Instruction::BitCast) {
256 if (AnalyzeUsesOfPointer(CE, Readers, Writers))
261 } else if (ICmpInst *ICI = dyn_cast<ICmpInst>(*UI)) {
262 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
263 return true; // Allow comparison against null.
264 } else if (FreeInst *F = dyn_cast<FreeInst>(*UI)) {
265 Writers.push_back(F->getParent()->getParent());
272 /// AnalyzeIndirectGlobalMemory - We found an non-address-taken global variable
273 /// which holds a pointer type. See if the global always points to non-aliased
274 /// heap memory: that is, all initializers of the globals are allocations, and
275 /// those allocations have no use other than initialization of the global.
276 /// Further, all loads out of GV must directly use the memory, not store the
277 /// pointer somewhere. If this is true, we consider the memory pointed to by
278 /// GV to be owned by GV and can disambiguate other pointers from it.
279 bool GlobalsModRef::AnalyzeIndirectGlobalMemory(GlobalValue *GV) {
280 // Keep track of values related to the allocation of the memory, f.e. the
281 // value produced by the malloc call and any casts.
282 std::vector<Value*> AllocRelatedValues;
284 // Walk the user list of the global. If we find anything other than a direct
285 // load or store, bail out.
286 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I){
287 if (LoadInst *LI = dyn_cast<LoadInst>(*I)) {
288 // The pointer loaded from the global can only be used in simple ways:
289 // we allow addressing of it and loading storing to it. We do *not* allow
290 // storing the loaded pointer somewhere else or passing to a function.
291 std::vector<Function*> ReadersWriters;
292 if (AnalyzeUsesOfPointer(LI, ReadersWriters, ReadersWriters))
293 return false; // Loaded pointer escapes.
294 // TODO: Could try some IP mod/ref of the loaded pointer.
295 } else if (StoreInst *SI = dyn_cast<StoreInst>(*I)) {
296 // Storing the global itself.
297 if (SI->getOperand(0) == GV) return false;
299 // If storing the null pointer, ignore it.
300 if (isa<ConstantPointerNull>(SI->getOperand(0)))
303 // Check the value being stored.
304 Value *Ptr = SI->getOperand(0)->getUnderlyingObject();
306 if (isa<MallocInst>(Ptr) || isMalloc(Ptr)) {
308 } else if (CallInst *CI = dyn_cast<CallInst>(Ptr)) {
309 Function *F = CI->getCalledFunction();
310 if (!F || !F->isDeclaration()) return false; // Too hard to analyze.
311 if (F->getName() != "calloc") return false; // Not calloc.
313 return false; // Too hard to analyze.
316 // Analyze all uses of the allocation. If any of them are used in a
317 // non-simple way (e.g. stored to another global) bail out.
318 std::vector<Function*> ReadersWriters;
319 if (AnalyzeUsesOfPointer(Ptr, ReadersWriters, ReadersWriters, GV))
320 return false; // Loaded pointer escapes.
322 // Remember that this allocation is related to the indirect global.
323 AllocRelatedValues.push_back(Ptr);
325 // Something complex, bail out.
330 // Okay, this is an indirect global. Remember all of the allocations for
331 // this global in AllocsForIndirectGlobals.
332 while (!AllocRelatedValues.empty()) {
333 AllocsForIndirectGlobals[AllocRelatedValues.back()] = GV;
334 AllocRelatedValues.pop_back();
336 IndirectGlobals.insert(GV);
340 /// AnalyzeCallGraph - At this point, we know the functions where globals are
341 /// immediately stored to and read from. Propagate this information up the call
342 /// graph to all callers and compute the mod/ref info for all memory for each
344 void GlobalsModRef::AnalyzeCallGraph(CallGraph &CG, Module &M) {
345 // We do a bottom-up SCC traversal of the call graph. In other words, we
346 // visit all callees before callers (leaf-first).
347 for (scc_iterator<CallGraph*> I = scc_begin(&CG), E = scc_end(&CG); I != E;
349 std::vector<CallGraphNode *> &SCC = *I;
350 assert(!SCC.empty() && "SCC with no functions?");
352 if (!SCC[0]->getFunction()) {
353 // Calls externally - can't say anything useful. Remove any existing
354 // function records (may have been created when scanning globals).
355 for (unsigned i = 0, e = SCC.size(); i != e; ++i)
356 FunctionInfo.erase(SCC[i]->getFunction());
360 FunctionRecord &FR = FunctionInfo[SCC[0]->getFunction()];
362 bool KnowNothing = false;
363 unsigned FunctionEffect = 0;
365 // Collect the mod/ref properties due to called functions. We only compute
367 for (unsigned i = 0, e = SCC.size(); i != e && !KnowNothing; ++i) {
368 Function *F = SCC[i]->getFunction();
374 if (F->isDeclaration()) {
375 // Try to get mod/ref behaviour from function attributes.
376 if (F->doesNotAccessMemory()) {
377 // Can't do better than that!
378 } else if (F->onlyReadsMemory()) {
379 FunctionEffect |= Ref;
380 if (!F->isIntrinsic())
381 // This function might call back into the module and read a global -
382 // consider every global as possibly being read by this function.
383 FR.MayReadAnyGlobal = true;
385 FunctionEffect |= ModRef;
386 // Can't say anything useful unless it's an intrinsic - they don't
387 // read or write global variables of the kind considered here.
388 KnowNothing = !F->isIntrinsic();
393 for (CallGraphNode::iterator CI = SCC[i]->begin(), E = SCC[i]->end();
394 CI != E && !KnowNothing; ++CI)
395 if (Function *Callee = CI->second->getFunction()) {
396 if (FunctionRecord *CalleeFR = getFunctionInfo(Callee)) {
397 // Propagate function effect up.
398 FunctionEffect |= CalleeFR->FunctionEffect;
400 // Incorporate callee's effects on globals into our info.
401 for (std::map<GlobalValue*, unsigned>::iterator GI =
402 CalleeFR->GlobalInfo.begin(), E = CalleeFR->GlobalInfo.end();
404 FR.GlobalInfo[GI->first] |= GI->second;
405 FR.MayReadAnyGlobal |= CalleeFR->MayReadAnyGlobal;
407 // Can't say anything about it. However, if it is inside our SCC,
408 // then nothing needs to be done.
409 CallGraphNode *CalleeNode = CG[Callee];
410 if (std::find(SCC.begin(), SCC.end(), CalleeNode) == SCC.end())
418 // If we can't say anything useful about this SCC, remove all SCC functions
419 // from the FunctionInfo map.
421 for (unsigned i = 0, e = SCC.size(); i != e; ++i)
422 FunctionInfo.erase(SCC[i]->getFunction());
426 // Scan the function bodies for explicit loads or stores.
427 for (unsigned i = 0, e = SCC.size(); i != e && FunctionEffect != ModRef;++i)
428 for (inst_iterator II = inst_begin(SCC[i]->getFunction()),
429 E = inst_end(SCC[i]->getFunction());
430 II != E && FunctionEffect != ModRef; ++II)
431 if (isa<LoadInst>(*II)) {
432 FunctionEffect |= Ref;
433 if (cast<LoadInst>(*II).isVolatile())
434 // Volatile loads may have side-effects, so mark them as writing
435 // memory (for example, a flag inside the processor).
436 FunctionEffect |= Mod;
437 } else if (isa<StoreInst>(*II)) {
438 FunctionEffect |= Mod;
439 if (cast<StoreInst>(*II).isVolatile())
440 // Treat volatile stores as reading memory somewhere.
441 FunctionEffect |= Ref;
442 } else if (isa<MallocInst>(*II) || isa<FreeInst>(*II) ||
443 isMalloc(&cast<Instruction>(*II))) {
444 FunctionEffect |= ModRef;
447 if ((FunctionEffect & Mod) == 0)
448 ++NumReadMemFunctions;
449 if (FunctionEffect == 0)
451 FR.FunctionEffect = FunctionEffect;
453 // Finally, now that we know the full effect on this SCC, clone the
454 // information to each function in the SCC.
455 for (unsigned i = 1, e = SCC.size(); i != e; ++i)
456 FunctionInfo[SCC[i]->getFunction()] = FR;
462 /// alias - If one of the pointers is to a global that we are tracking, and the
463 /// other is some random pointer, we know there cannot be an alias, because the
464 /// address of the global isn't taken.
465 AliasAnalysis::AliasResult
466 GlobalsModRef::alias(const Value *V1, unsigned V1Size,
467 const Value *V2, unsigned V2Size) {
468 // Get the base object these pointers point to.
469 Value *UV1 = const_cast<Value*>(V1->getUnderlyingObject());
470 Value *UV2 = const_cast<Value*>(V2->getUnderlyingObject());
472 // If either of the underlying values is a global, they may be non-addr-taken
473 // globals, which we can answer queries about.
474 GlobalValue *GV1 = dyn_cast<GlobalValue>(UV1);
475 GlobalValue *GV2 = dyn_cast<GlobalValue>(UV2);
477 // If the global's address is taken, pretend we don't know it's a pointer to
479 if (GV1 && !NonAddressTakenGlobals.count(GV1)) GV1 = 0;
480 if (GV2 && !NonAddressTakenGlobals.count(GV2)) GV2 = 0;
482 // If the the two pointers are derived from two different non-addr-taken
483 // globals, or if one is and the other isn't, we know these can't alias.
484 if ((GV1 || GV2) && GV1 != GV2)
487 // Otherwise if they are both derived from the same addr-taken global, we
488 // can't know the two accesses don't overlap.
491 // These pointers may be based on the memory owned by an indirect global. If
492 // so, we may be able to handle this. First check to see if the base pointer
493 // is a direct load from an indirect global.
495 if (LoadInst *LI = dyn_cast<LoadInst>(UV1))
496 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getOperand(0)))
497 if (IndirectGlobals.count(GV))
499 if (LoadInst *LI = dyn_cast<LoadInst>(UV2))
500 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getOperand(0)))
501 if (IndirectGlobals.count(GV))
504 // These pointers may also be from an allocation for the indirect global. If
505 // so, also handle them.
506 if (AllocsForIndirectGlobals.count(UV1))
507 GV1 = AllocsForIndirectGlobals[UV1];
508 if (AllocsForIndirectGlobals.count(UV2))
509 GV2 = AllocsForIndirectGlobals[UV2];
511 // Now that we know whether the two pointers are related to indirect globals,
512 // use this to disambiguate the pointers. If either pointer is based on an
513 // indirect global and if they are not both based on the same indirect global,
514 // they cannot alias.
515 if ((GV1 || GV2) && GV1 != GV2)
518 return AliasAnalysis::alias(V1, V1Size, V2, V2Size);
521 AliasAnalysis::ModRefResult
522 GlobalsModRef::getModRefInfo(CallSite CS, Value *P, unsigned Size) {
523 unsigned Known = ModRef;
525 // If we are asking for mod/ref info of a direct call with a pointer to a
526 // global we are tracking, return information if we have it.
527 if (GlobalValue *GV = dyn_cast<GlobalValue>(P->getUnderlyingObject()))
528 if (GV->hasLocalLinkage())
529 if (Function *F = CS.getCalledFunction())
530 if (NonAddressTakenGlobals.count(GV))
531 if (FunctionRecord *FR = getFunctionInfo(F))
532 Known = FR->getInfoForGlobal(GV);
534 if (Known == NoModRef)
535 return NoModRef; // No need to query other mod/ref analyses
536 return ModRefResult(Known & AliasAnalysis::getModRefInfo(CS, P, Size));
540 //===----------------------------------------------------------------------===//
541 // Methods to update the analysis as a result of the client transformation.
543 void GlobalsModRef::deleteValue(Value *V) {
544 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
545 if (NonAddressTakenGlobals.erase(GV)) {
546 // This global might be an indirect global. If so, remove it and remove
547 // any AllocRelatedValues for it.
548 if (IndirectGlobals.erase(GV)) {
549 // Remove any entries in AllocsForIndirectGlobals for this global.
550 for (std::map<Value*, GlobalValue*>::iterator
551 I = AllocsForIndirectGlobals.begin(),
552 E = AllocsForIndirectGlobals.end(); I != E; ) {
553 if (I->second == GV) {
554 AllocsForIndirectGlobals.erase(I++);
563 // Otherwise, if this is an allocation related to an indirect global, remove
565 AllocsForIndirectGlobals.erase(V);
567 AliasAnalysis::deleteValue(V);
570 void GlobalsModRef::copyValue(Value *From, Value *To) {
571 AliasAnalysis::copyValue(From, To);