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/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 /// MayReadAnyGlobal - May read global variables, but it is not known which.
53 bool MayReadAnyGlobal;
55 unsigned getInfoForGlobal(GlobalValue *GV) const {
56 unsigned Effect = MayReadAnyGlobal ? AliasAnalysis::Ref : 0;
57 std::map<GlobalValue*, unsigned>::const_iterator I = GlobalInfo.find(GV);
58 if (I != GlobalInfo.end())
63 /// FunctionEffect - Capture whether or not this function reads or writes to
64 /// ANY memory. If not, we can do a lot of aggressive analysis on it.
65 unsigned FunctionEffect;
67 FunctionRecord() : MayReadAnyGlobal (false), FunctionEffect(0) {}
70 /// GlobalsModRef - The actual analysis pass.
71 class VISIBILITY_HIDDEN GlobalsModRef
72 : public ModulePass, public AliasAnalysis {
73 /// NonAddressTakenGlobals - The globals that do not have their addresses
75 std::set<GlobalValue*> NonAddressTakenGlobals;
77 /// IndirectGlobals - The memory pointed to by this global is known to be
78 /// 'owned' by the global.
79 std::set<GlobalValue*> IndirectGlobals;
81 /// AllocsForIndirectGlobals - If an instruction allocates memory for an
82 /// indirect global, this map indicates which one.
83 std::map<Value*, GlobalValue*> AllocsForIndirectGlobals;
85 /// FunctionInfo - For each function, keep track of what globals are
87 std::map<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 Value *V1, unsigned V1Size,
110 const Value *V2, unsigned V2Size);
111 ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size);
112 ModRefResult getModRefInfo(CallSite CS1, CallSite CS2) {
113 return AliasAnalysis::getModRefInfo(CS1,CS2);
115 bool hasNoModRefInfoForCalls() const { return false; }
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 virtual ModRefBehavior getModRefBehavior(Function *F, CallSite CS,
121 std::vector<PointerAccessInfo> *Info) {
122 if (FunctionRecord *FR = getFunctionInfo(F)) {
123 if (FR->FunctionEffect == 0)
124 return DoesNotAccessMemory;
125 else if ((FR->FunctionEffect & Mod) == 0)
126 return OnlyReadsMemory;
128 return AliasAnalysis::getModRefBehavior(F, CS, Info);
131 virtual void deleteValue(Value *V);
132 virtual void copyValue(Value *From, Value *To);
135 /// getFunctionInfo - Return the function info for the function, or null if
136 /// we don't have anything useful to say about it.
137 FunctionRecord *getFunctionInfo(Function *F) {
138 std::map<Function*, FunctionRecord>::iterator I = FunctionInfo.find(F);
139 if (I != FunctionInfo.end())
144 void AnalyzeGlobals(Module &M);
145 void AnalyzeCallGraph(CallGraph &CG, Module &M);
146 bool AnalyzeUsesOfPointer(Value *V, std::vector<Function*> &Readers,
147 std::vector<Function*> &Writers,
148 GlobalValue *OkayStoreDest = 0);
149 bool AnalyzeIndirectGlobalMemory(GlobalValue *GV);
153 char GlobalsModRef::ID = 0;
154 static RegisterPass<GlobalsModRef>
155 X("globalsmodref-aa", "Simple mod/ref analysis for globals", false, true);
156 static RegisterAnalysisGroup<AliasAnalysis> Y(X);
158 Pass *llvm::createGlobalsModRefPass() { return new GlobalsModRef(); }
160 /// AnalyzeGlobals - Scan through the users of all of the internal
161 /// GlobalValue's in the program. If none of them have their "address taken"
162 /// (really, their address passed to something nontrivial), record this fact,
163 /// and record the functions that they are used directly in.
164 void GlobalsModRef::AnalyzeGlobals(Module &M) {
165 std::vector<Function*> Readers, Writers;
166 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
167 if (I->hasInternalLinkage()) {
168 if (!AnalyzeUsesOfPointer(I, Readers, Writers)) {
169 // Remember that we are tracking this global.
170 NonAddressTakenGlobals.insert(I);
171 ++NumNonAddrTakenFunctions;
173 Readers.clear(); Writers.clear();
176 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
178 if (I->hasInternalLinkage()) {
179 if (!AnalyzeUsesOfPointer(I, Readers, Writers)) {
180 // Remember that we are tracking this global, and the mod/ref fns
181 NonAddressTakenGlobals.insert(I);
183 for (unsigned i = 0, e = Readers.size(); i != e; ++i)
184 FunctionInfo[Readers[i]].GlobalInfo[I] |= Ref;
186 if (!I->isConstant()) // No need to keep track of writers to constants
187 for (unsigned i = 0, e = Writers.size(); i != e; ++i)
188 FunctionInfo[Writers[i]].GlobalInfo[I] |= Mod;
189 ++NumNonAddrTakenGlobalVars;
191 // If this global holds a pointer type, see if it is an indirect global.
192 if (isa<PointerType>(I->getType()->getElementType()) &&
193 AnalyzeIndirectGlobalMemory(I))
194 ++NumIndirectGlobalVars;
196 Readers.clear(); Writers.clear();
200 /// AnalyzeUsesOfPointer - Look at all of the users of the specified pointer.
201 /// If this is used by anything complex (i.e., the address escapes), return
202 /// true. Also, while we are at it, keep track of those functions that read and
203 /// write to the value.
205 /// If OkayStoreDest is non-null, stores into this global are allowed.
206 bool GlobalsModRef::AnalyzeUsesOfPointer(Value *V,
207 std::vector<Function*> &Readers,
208 std::vector<Function*> &Writers,
209 GlobalValue *OkayStoreDest) {
210 if (!isa<PointerType>(V->getType())) return true;
212 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
213 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
214 Readers.push_back(LI->getParent()->getParent());
215 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
216 if (V == SI->getOperand(1)) {
217 Writers.push_back(SI->getParent()->getParent());
218 } else if (SI->getOperand(1) != OkayStoreDest) {
219 return true; // Storing the pointer
221 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) {
222 if (AnalyzeUsesOfPointer(GEP, Readers, Writers)) return true;
223 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
224 // Make sure that this is just the function being called, not that it is
225 // passing into the function.
226 for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i)
227 if (CI->getOperand(i) == V) return true;
228 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
229 // Make sure that this is just the function being called, not that it is
230 // passing into the function.
231 for (unsigned i = 3, e = II->getNumOperands(); i != e; ++i)
232 if (II->getOperand(i) == V) return true;
233 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
234 if (CE->getOpcode() == Instruction::GetElementPtr ||
235 CE->getOpcode() == Instruction::BitCast) {
236 if (AnalyzeUsesOfPointer(CE, Readers, Writers))
241 } else if (ICmpInst *ICI = dyn_cast<ICmpInst>(*UI)) {
242 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
243 return true; // Allow comparison against null.
244 } else if (FreeInst *F = dyn_cast<FreeInst>(*UI)) {
245 Writers.push_back(F->getParent()->getParent());
252 /// AnalyzeIndirectGlobalMemory - We found an non-address-taken global variable
253 /// which holds a pointer type. See if the global always points to non-aliased
254 /// heap memory: that is, all initializers of the globals are allocations, and
255 /// those allocations have no use other than initialization of the global.
256 /// Further, all loads out of GV must directly use the memory, not store the
257 /// pointer somewhere. If this is true, we consider the memory pointed to by
258 /// GV to be owned by GV and can disambiguate other pointers from it.
259 bool GlobalsModRef::AnalyzeIndirectGlobalMemory(GlobalValue *GV) {
260 // Keep track of values related to the allocation of the memory, f.e. the
261 // value produced by the malloc call and any casts.
262 std::vector<Value*> AllocRelatedValues;
264 // Walk the user list of the global. If we find anything other than a direct
265 // load or store, bail out.
266 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I){
267 if (LoadInst *LI = dyn_cast<LoadInst>(*I)) {
268 // The pointer loaded from the global can only be used in simple ways:
269 // we allow addressing of it and loading storing to it. We do *not* allow
270 // storing the loaded pointer somewhere else or passing to a function.
271 std::vector<Function*> ReadersWriters;
272 if (AnalyzeUsesOfPointer(LI, ReadersWriters, ReadersWriters))
273 return false; // Loaded pointer escapes.
274 // TODO: Could try some IP mod/ref of the loaded pointer.
275 } else if (StoreInst *SI = dyn_cast<StoreInst>(*I)) {
276 // Storing the global itself.
277 if (SI->getOperand(0) == GV) return false;
279 // If storing the null pointer, ignore it.
280 if (isa<ConstantPointerNull>(SI->getOperand(0)))
283 // Check the value being stored.
284 Value *Ptr = SI->getOperand(0)->getUnderlyingObject();
286 if (isa<MallocInst>(Ptr)) {
288 } else if (CallInst *CI = dyn_cast<CallInst>(Ptr)) {
289 Function *F = CI->getCalledFunction();
290 if (!F || !F->isDeclaration()) return false; // Too hard to analyze.
291 if (F->getName() != "calloc") return false; // Not calloc.
293 return false; // Too hard to analyze.
296 // Analyze all uses of the allocation. If any of them are used in a
297 // non-simple way (e.g. stored to another global) bail out.
298 std::vector<Function*> ReadersWriters;
299 if (AnalyzeUsesOfPointer(Ptr, ReadersWriters, ReadersWriters, GV))
300 return false; // Loaded pointer escapes.
302 // Remember that this allocation is related to the indirect global.
303 AllocRelatedValues.push_back(Ptr);
305 // Something complex, bail out.
310 // Okay, this is an indirect global. Remember all of the allocations for
311 // this global in AllocsForIndirectGlobals.
312 while (!AllocRelatedValues.empty()) {
313 AllocsForIndirectGlobals[AllocRelatedValues.back()] = GV;
314 AllocRelatedValues.pop_back();
316 IndirectGlobals.insert(GV);
320 /// AnalyzeCallGraph - At this point, we know the functions where globals are
321 /// immediately stored to and read from. Propagate this information up the call
322 /// graph to all callers and compute the mod/ref info for all memory for each
324 void GlobalsModRef::AnalyzeCallGraph(CallGraph &CG, Module &M) {
325 // We do a bottom-up SCC traversal of the call graph. In other words, we
326 // visit all callees before callers (leaf-first).
327 for (scc_iterator<CallGraph*> I = scc_begin(&CG), E = scc_end(&CG); I != E;
329 std::vector<CallGraphNode *> &SCC = *I;
330 assert(!SCC.empty() && "SCC with no functions?");
332 if (!SCC[0]->getFunction()) {
333 // Calls externally - can't say anything useful. Remove any existing
334 // function records (may have been created when scanning globals).
335 for (unsigned i = 0, e = SCC.size(); i != e; ++i)
336 FunctionInfo.erase(SCC[i]->getFunction());
340 FunctionRecord &FR = FunctionInfo[SCC[0]->getFunction()];
342 bool KnowNothing = false;
343 unsigned FunctionEffect = 0;
345 // Collect the mod/ref properties due to called functions. We only compute
347 for (unsigned i = 0, e = SCC.size(); i != e && !KnowNothing; ++i) {
348 Function *F = SCC[i]->getFunction();
354 if (F->isDeclaration()) {
355 // Try to get mod/ref behaviour from function attributes.
356 if (F->doesNotAccessMemory()) {
357 // Can't do better than that!
358 } else if (F->onlyReadsMemory()) {
359 FunctionEffect |= Ref;
360 if (!F->isIntrinsic())
361 // This function might call back into the module and read a global -
362 // consider every global as possibly being read by this function.
363 FR.MayReadAnyGlobal = true;
365 FunctionEffect |= ModRef;
366 // Can't say anything useful unless it's an intrinsic - they don't
367 // read or write global variables of the kind considered here.
368 KnowNothing = !F->isIntrinsic();
373 for (CallGraphNode::iterator CI = SCC[i]->begin(), E = SCC[i]->end();
374 CI != E && !KnowNothing; ++CI)
375 if (Function *Callee = CI->second->getFunction()) {
376 if (FunctionRecord *CalleeFR = getFunctionInfo(Callee)) {
377 // Propagate function effect up.
378 FunctionEffect |= CalleeFR->FunctionEffect;
380 // Incorporate callee's effects on globals into our info.
381 for (std::map<GlobalValue*, unsigned>::iterator GI =
382 CalleeFR->GlobalInfo.begin(), E = CalleeFR->GlobalInfo.end();
384 FR.GlobalInfo[GI->first] |= GI->second;
385 FR.MayReadAnyGlobal |= CalleeFR->MayReadAnyGlobal;
387 // Can't say anything about it. However, if it is inside our SCC,
388 // then nothing needs to be done.
389 CallGraphNode *CalleeNode = CG[Callee];
390 if (std::find(SCC.begin(), SCC.end(), CalleeNode) == SCC.end())
398 // If we can't say anything useful about this SCC, remove all SCC functions
399 // from the FunctionInfo map.
401 for (unsigned i = 0, e = SCC.size(); i != e; ++i)
402 FunctionInfo.erase(SCC[i]->getFunction());
406 // Scan the function bodies for explicit loads or stores.
407 for (unsigned i = 0, e = SCC.size(); i != e && FunctionEffect != ModRef;++i)
408 for (inst_iterator II = inst_begin(SCC[i]->getFunction()),
409 E = inst_end(SCC[i]->getFunction());
410 II != E && FunctionEffect != ModRef; ++II)
411 if (isa<LoadInst>(*II)) {
412 FunctionEffect |= Ref;
413 if (cast<LoadInst>(*II).isVolatile())
414 // Volatile loads may have side-effects, so mark them as writing
415 // memory (for example, a flag inside the processor).
416 FunctionEffect |= Mod;
417 } else if (isa<StoreInst>(*II)) {
418 FunctionEffect |= Mod;
419 if (cast<StoreInst>(*II).isVolatile())
420 // Treat volatile stores as reading memory somewhere.
421 FunctionEffect |= Ref;
422 } else if (isa<MallocInst>(*II) || isa<FreeInst>(*II)) {
423 FunctionEffect |= ModRef;
426 if ((FunctionEffect & Mod) == 0)
427 ++NumReadMemFunctions;
428 if (FunctionEffect == 0)
430 FR.FunctionEffect = FunctionEffect;
432 // Finally, now that we know the full effect on this SCC, clone the
433 // information to each function in the SCC.
434 for (unsigned i = 1, e = SCC.size(); i != e; ++i)
435 FunctionInfo[SCC[i]->getFunction()] = FR;
441 /// alias - If one of the pointers is to a global that we are tracking, and the
442 /// other is some random pointer, we know there cannot be an alias, because the
443 /// address of the global isn't taken.
444 AliasAnalysis::AliasResult
445 GlobalsModRef::alias(const Value *V1, unsigned V1Size,
446 const Value *V2, unsigned V2Size) {
447 // Get the base object these pointers point to.
448 Value *UV1 = const_cast<Value*>(V1->getUnderlyingObject());
449 Value *UV2 = const_cast<Value*>(V2->getUnderlyingObject());
451 // If either of the underlying values is a global, they may be non-addr-taken
452 // globals, which we can answer queries about.
453 GlobalValue *GV1 = dyn_cast<GlobalValue>(UV1);
454 GlobalValue *GV2 = dyn_cast<GlobalValue>(UV2);
456 // If the global's address is taken, pretend we don't know it's a pointer to
458 if (GV1 && !NonAddressTakenGlobals.count(GV1)) GV1 = 0;
459 if (GV2 && !NonAddressTakenGlobals.count(GV2)) GV2 = 0;
461 // If the the two pointers are derived from two different non-addr-taken
462 // globals, or if one is and the other isn't, we know these can't alias.
463 if ((GV1 || GV2) && GV1 != GV2)
466 // Otherwise if they are both derived from the same addr-taken global, we
467 // can't know the two accesses don't overlap.
470 // These pointers may be based on the memory owned by an indirect global. If
471 // so, we may be able to handle this. First check to see if the base pointer
472 // is a direct load from an indirect global.
474 if (LoadInst *LI = dyn_cast<LoadInst>(UV1))
475 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getOperand(0)))
476 if (IndirectGlobals.count(GV))
478 if (LoadInst *LI = dyn_cast<LoadInst>(UV2))
479 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getOperand(0)))
480 if (IndirectGlobals.count(GV))
483 // These pointers may also be from an allocation for the indirect global. If
484 // so, also handle them.
485 if (AllocsForIndirectGlobals.count(UV1))
486 GV1 = AllocsForIndirectGlobals[UV1];
487 if (AllocsForIndirectGlobals.count(UV2))
488 GV2 = AllocsForIndirectGlobals[UV2];
490 // Now that we know whether the two pointers are related to indirect globals,
491 // use this to disambiguate the pointers. If either pointer is based on an
492 // indirect global and if they are not both based on the same indirect global,
493 // they cannot alias.
494 if ((GV1 || GV2) && GV1 != GV2)
497 return AliasAnalysis::alias(V1, V1Size, V2, V2Size);
500 AliasAnalysis::ModRefResult
501 GlobalsModRef::getModRefInfo(CallSite CS, Value *P, unsigned Size) {
502 unsigned Known = ModRef;
504 // If we are asking for mod/ref info of a direct call with a pointer to a
505 // global we are tracking, return information if we have it.
506 if (GlobalValue *GV = dyn_cast<GlobalValue>(P->getUnderlyingObject()))
507 if (GV->hasInternalLinkage())
508 if (Function *F = CS.getCalledFunction())
509 if (NonAddressTakenGlobals.count(GV))
510 if (FunctionRecord *FR = getFunctionInfo(F))
511 Known = FR->getInfoForGlobal(GV);
513 if (Known == NoModRef)
514 return NoModRef; // No need to query other mod/ref analyses
515 return ModRefResult(Known & AliasAnalysis::getModRefInfo(CS, P, Size));
519 //===----------------------------------------------------------------------===//
520 // Methods to update the analysis as a result of the client transformation.
522 void GlobalsModRef::deleteValue(Value *V) {
523 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
524 if (NonAddressTakenGlobals.erase(GV)) {
525 // This global might be an indirect global. If so, remove it and remove
526 // any AllocRelatedValues for it.
527 if (IndirectGlobals.erase(GV)) {
528 // Remove any entries in AllocsForIndirectGlobals for this global.
529 for (std::map<Value*, GlobalValue*>::iterator
530 I = AllocsForIndirectGlobals.begin(),
531 E = AllocsForIndirectGlobals.end(); I != E; ) {
532 if (I->second == GV) {
533 AllocsForIndirectGlobals.erase(I++);
542 // Otherwise, if this is an allocation related to an indirect global, remove
544 AllocsForIndirectGlobals.erase(V);
546 AliasAnalysis::deleteValue(V);
549 void GlobalsModRef::copyValue(Value *From, Value *To) {
550 AliasAnalysis::copyValue(From, To);