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 Value *V1, unsigned V1Size,
110
const Value *V2, unsigned V2Size);
111
ModRefResult getModRefInfo(ImmutableCallSite CS,
112
const Value *P, unsigned Size);
113
ModRefResult getModRefInfo(ImmutableCallSite CS1,
114
ImmutableCallSite CS2) {
115
return AliasAnalysis::getModRefInfo(CS1, CS2);
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(const Function *F) {
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);
131
/// getModRefBehavior - Return the behavior of the specified function if
132
/// called from the specified call site. The call site may be null in which
133
/// case the most generic behavior of this function should be returned.
134
ModRefBehavior getModRefBehavior(ImmutableCallSite CS) {
135
const Function* F = CS.getCalledFunction();
136
if (!F) return AliasAnalysis::getModRefBehavior(CS);
137
if (FunctionRecord *FR = getFunctionInfo(F)) {
138
if (FR->FunctionEffect == 0)
139
return DoesNotAccessMemory;
140
else if ((FR->FunctionEffect & Mod) == 0)
141
return OnlyReadsMemory;
143
return AliasAnalysis::getModRefBehavior(CS);
146
virtual void deleteValue(Value *V);
147
virtual void copyValue(Value *From, Value *To);
149
/// getAdjustedAnalysisPointer - This method is used when a pass implements
150
/// an analysis interface through multiple inheritance. If needed, it
151
/// should override this to adjust the this pointer as needed for the
152
/// specified pass info.
153
virtual void *getAdjustedAnalysisPointer(AnalysisID PI) {
154
if (PI == &AliasAnalysis::ID)
155
return (AliasAnalysis*)this;
160
/// getFunctionInfo - Return the function info for the function, or null if
161
/// we don't have anything useful to say about it.
162
FunctionRecord *getFunctionInfo(const Function *F) {
163
std::map<const Function*, FunctionRecord>::iterator I =
164
FunctionInfo.find(F);
165
if (I != FunctionInfo.end())
170
void AnalyzeGlobals(Module &M);
171
void AnalyzeCallGraph(CallGraph &CG, Module &M);
172
bool AnalyzeUsesOfPointer(Value *V, std::vector<Function*> &Readers,
173
std::vector<Function*> &Writers,
174
GlobalValue *OkayStoreDest = 0);
175
bool AnalyzeIndirectGlobalMemory(GlobalValue *GV);
179
char GlobalsModRef::ID = 0;
180
INITIALIZE_AG_PASS(GlobalsModRef, AliasAnalysis,
181
"globalsmodref-aa", "Simple mod/ref analysis for globals",
184
Pass *llvm::createGlobalsModRefPass() { return new GlobalsModRef(); }
186
/// AnalyzeGlobals - Scan through the users of all of the internal
187
/// GlobalValue's in the program. If none of them have their "address taken"
188
/// (really, their address passed to something nontrivial), record this fact,
189
/// and record the functions that they are used directly in.
190
void GlobalsModRef::AnalyzeGlobals(Module &M) {
191
std::vector<Function*> Readers, Writers;
192
for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
193
if (I->hasLocalLinkage()) {
194
if (!AnalyzeUsesOfPointer(I, Readers, Writers)) {
195
// Remember that we are tracking this global.
196
NonAddressTakenGlobals.insert(I);
197
++NumNonAddrTakenFunctions;
199
Readers.clear(); Writers.clear();
202
for (Module::global_iterator I = M.global_begin(), E = M.global_end();
204
if (I->hasLocalLinkage()) {
205
if (!AnalyzeUsesOfPointer(I, Readers, Writers)) {
206
// Remember that we are tracking this global, and the mod/ref fns
207
NonAddressTakenGlobals.insert(I);
209
for (unsigned i = 0, e = Readers.size(); i != e; ++i)
210
FunctionInfo[Readers[i]].GlobalInfo[I] |= Ref;
212
if (!I->isConstant()) // No need to keep track of writers to constants
213
for (unsigned i = 0, e = Writers.size(); i != e; ++i)
214
FunctionInfo[Writers[i]].GlobalInfo[I] |= Mod;
215
++NumNonAddrTakenGlobalVars;
217
// If this global holds a pointer type, see if it is an indirect global.
218
if (I->getType()->getElementType()->isPointerTy() &&
219
AnalyzeIndirectGlobalMemory(I))
220
++NumIndirectGlobalVars;
222
Readers.clear(); Writers.clear();
226
/// AnalyzeUsesOfPointer - Look at all of the users of the specified pointer.
227
/// If this is used by anything complex (i.e., the address escapes), return
228
/// true. Also, while we are at it, keep track of those functions that read and
229
/// write to the value.
231
/// If OkayStoreDest is non-null, stores into this global are allowed.
232
bool GlobalsModRef::AnalyzeUsesOfPointer(Value *V,
233
std::vector<Function*> &Readers,
234
std::vector<Function*> &Writers,
235
GlobalValue *OkayStoreDest) {
236
if (!V->getType()->isPointerTy()) return true;
238
for (Value::use_iterator UI = V->use_begin(), E=V->use_end(); UI != E; ++UI) {
240
if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
241
Readers.push_back(LI->getParent()->getParent());
242
} else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
243
if (V == SI->getOperand(1)) {
244
Writers.push_back(SI->getParent()->getParent());
245
} else if (SI->getOperand(1) != OkayStoreDest) {
246
return true; // Storing the pointer
248
} else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
249
if (AnalyzeUsesOfPointer(GEP, Readers, Writers)) return true;
250
} else if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
251
if (AnalyzeUsesOfPointer(BCI, Readers, Writers, OkayStoreDest))
253
} else if (isFreeCall(U)) {
254
Writers.push_back(cast<Instruction>(U)->getParent()->getParent());
255
} else if (CallInst *CI = dyn_cast<CallInst>(U)) {
256
// Make sure that this is just the function being called, not that it is
257
// passing into the function.
258
for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i)
259
if (CI->getArgOperand(i) == V) return true;
260
} else if (InvokeInst *II = dyn_cast<InvokeInst>(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 = II->getNumArgOperands(); i != e; ++i)
264
if (II->getArgOperand(i) == V) return true;
265
} else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
266
if (CE->getOpcode() == Instruction::GetElementPtr ||
267
CE->getOpcode() == Instruction::BitCast) {
268
if (AnalyzeUsesOfPointer(CE, Readers, Writers))
273
} else if (ICmpInst *ICI = dyn_cast<ICmpInst>(U)) {
274
if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
275
return true; // Allow comparison against null.
284
/// AnalyzeIndirectGlobalMemory - We found an non-address-taken global variable
285
/// which holds a pointer type. See if the global always points to non-aliased
286
/// heap memory: that is, all initializers of the globals are allocations, and
287
/// those allocations have no use other than initialization of the global.
288
/// Further, all loads out of GV must directly use the memory, not store the
289
/// pointer somewhere. If this is true, we consider the memory pointed to by
290
/// GV to be owned by GV and can disambiguate other pointers from it.
291
bool GlobalsModRef::AnalyzeIndirectGlobalMemory(GlobalValue *GV) {
292
// Keep track of values related to the allocation of the memory, f.e. the
293
// value produced by the malloc call and any casts.
294
std::vector<Value*> AllocRelatedValues;
296
// Walk the user list of the global. If we find anything other than a direct
297
// load or store, bail out.
298
for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I){
300
if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
301
// The pointer loaded from the global can only be used in simple ways:
302
// we allow addressing of it and loading storing to it. We do *not* allow
303
// storing the loaded pointer somewhere else or passing to a function.
304
std::vector<Function*> ReadersWriters;
305
if (AnalyzeUsesOfPointer(LI, ReadersWriters, ReadersWriters))
306
return false; // Loaded pointer escapes.
307
// TODO: Could try some IP mod/ref of the loaded pointer.
308
} else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
309
// Storing the global itself.
310
if (SI->getOperand(0) == GV) return false;
312
// If storing the null pointer, ignore it.
313
if (isa<ConstantPointerNull>(SI->getOperand(0)))
316
// Check the value being stored.
317
Value *Ptr = SI->getOperand(0)->getUnderlyingObject();
321
} else if (CallInst *CI = dyn_cast<CallInst>(Ptr)) {
322
Function *F = CI->getCalledFunction();
323
if (!F || !F->isDeclaration()) return false; // Too hard to analyze.
324
if (F->getName() != "calloc") return false; // Not calloc.
326
return false; // Too hard to analyze.
329
// Analyze all uses of the allocation. If any of them are used in a
330
// non-simple way (e.g. stored to another global) bail out.
331
std::vector<Function*> ReadersWriters;
332
if (AnalyzeUsesOfPointer(Ptr, ReadersWriters, ReadersWriters, GV))
333
return false; // Loaded pointer escapes.
335
// Remember that this allocation is related to the indirect global.
336
AllocRelatedValues.push_back(Ptr);
338
// Something complex, bail out.
343
// Okay, this is an indirect global. Remember all of the allocations for
344
// this global in AllocsForIndirectGlobals.
345
while (!AllocRelatedValues.empty()) {
346
AllocsForIndirectGlobals[AllocRelatedValues.back()] = GV;
347
AllocRelatedValues.pop_back();
349
IndirectGlobals.insert(GV);
353
/// AnalyzeCallGraph - At this point, we know the functions where globals are
354
/// immediately stored to and read from. Propagate this information up the call
355
/// graph to all callers and compute the mod/ref info for all memory for each
357
void GlobalsModRef::AnalyzeCallGraph(CallGraph &CG, Module &M) {
358
// We do a bottom-up SCC traversal of the call graph. In other words, we
359
// visit all callees before callers (leaf-first).
360
for (scc_iterator<CallGraph*> I = scc_begin(&CG), E = scc_end(&CG); I != E;
362
std::vector<CallGraphNode *> &SCC = *I;
363
assert(!SCC.empty() && "SCC with no functions?");
365
if (!SCC[0]->getFunction()) {
366
// Calls externally - can't say anything useful. Remove any existing
367
// function records (may have been created when scanning globals).
368
for (unsigned i = 0, e = SCC.size(); i != e; ++i)
369
FunctionInfo.erase(SCC[i]->getFunction());
373
FunctionRecord &FR = FunctionInfo[SCC[0]->getFunction()];
375
bool KnowNothing = false;
376
unsigned FunctionEffect = 0;
378
// Collect the mod/ref properties due to called functions. We only compute
380
for (unsigned i = 0, e = SCC.size(); i != e && !KnowNothing; ++i) {
381
Function *F = SCC[i]->getFunction();
387
if (F->isDeclaration()) {
388
// Try to get mod/ref behaviour from function attributes.
389
if (F->doesNotAccessMemory()) {
390
// Can't do better than that!
391
} else if (F->onlyReadsMemory()) {
392
FunctionEffect |= Ref;
393
if (!F->isIntrinsic())
394
// This function might call back into the module and read a global -
395
// consider every global as possibly being read by this function.
396
FR.MayReadAnyGlobal = true;
398
FunctionEffect |= ModRef;
399
// Can't say anything useful unless it's an intrinsic - they don't
400
// read or write global variables of the kind considered here.
401
KnowNothing = !F->isIntrinsic();
406
for (CallGraphNode::iterator CI = SCC[i]->begin(), E = SCC[i]->end();
407
CI != E && !KnowNothing; ++CI)
408
if (Function *Callee = CI->second->getFunction()) {
409
if (FunctionRecord *CalleeFR = getFunctionInfo(Callee)) {
410
// Propagate function effect up.
411
FunctionEffect |= CalleeFR->FunctionEffect;
413
// Incorporate callee's effects on globals into our info.
414
for (std::map<const GlobalValue*, unsigned>::iterator GI =
415
CalleeFR->GlobalInfo.begin(), E = CalleeFR->GlobalInfo.end();
417
FR.GlobalInfo[GI->first] |= GI->second;
418
FR.MayReadAnyGlobal |= CalleeFR->MayReadAnyGlobal;
420
// Can't say anything about it. However, if it is inside our SCC,
421
// then nothing needs to be done.
422
CallGraphNode *CalleeNode = CG[Callee];
423
if (std::find(SCC.begin(), SCC.end(), CalleeNode) == SCC.end())
431
// If we can't say anything useful about this SCC, remove all SCC functions
432
// from the FunctionInfo map.
434
for (unsigned i = 0, e = SCC.size(); i != e; ++i)
435
FunctionInfo.erase(SCC[i]->getFunction());
439
// Scan the function bodies for explicit loads or stores.
440
for (unsigned i = 0, e = SCC.size(); i != e && FunctionEffect != ModRef;++i)
441
for (inst_iterator II = inst_begin(SCC[i]->getFunction()),
442
E = inst_end(SCC[i]->getFunction());
443
II != E && FunctionEffect != ModRef; ++II)
444
if (isa<LoadInst>(*II)) {
445
FunctionEffect |= Ref;
446
if (cast<LoadInst>(*II).isVolatile())
447
// Volatile loads may have side-effects, so mark them as writing
448
// memory (for example, a flag inside the processor).
449
FunctionEffect |= Mod;
450
} else if (isa<StoreInst>(*II)) {
451
FunctionEffect |= Mod;
452
if (cast<StoreInst>(*II).isVolatile())
453
// Treat volatile stores as reading memory somewhere.
454
FunctionEffect |= Ref;
455
} else if (isMalloc(&cast<Instruction>(*II)) ||
456
isFreeCall(&cast<Instruction>(*II))) {
457
FunctionEffect |= ModRef;
460
if ((FunctionEffect & Mod) == 0)
461
++NumReadMemFunctions;
462
if (FunctionEffect == 0)
464
FR.FunctionEffect = FunctionEffect;
466
// Finally, now that we know the full effect on this SCC, clone the
467
// information to each function in the SCC.
468
for (unsigned i = 1, e = SCC.size(); i != e; ++i)
469
FunctionInfo[SCC[i]->getFunction()] = FR;
475
/// alias - If one of the pointers is to a global that we are tracking, and the
476
/// other is some random pointer, we know there cannot be an alias, because the
477
/// address of the global isn't taken.
478
AliasAnalysis::AliasResult
479
GlobalsModRef::alias(const Value *V1, unsigned V1Size,
480
const Value *V2, unsigned V2Size) {
481
// Get the base object these pointers point to.
482
const Value *UV1 = V1->getUnderlyingObject();
483
const Value *UV2 = V2->getUnderlyingObject();
485
// If either of the underlying values is a global, they may be non-addr-taken
486
// globals, which we can answer queries about.
487
const GlobalValue *GV1 = dyn_cast<GlobalValue>(UV1);
488
const GlobalValue *GV2 = dyn_cast<GlobalValue>(UV2);
490
// If the global's address is taken, pretend we don't know it's a pointer to
492
if (GV1 && !NonAddressTakenGlobals.count(GV1)) GV1 = 0;
493
if (GV2 && !NonAddressTakenGlobals.count(GV2)) GV2 = 0;
495
// If the two pointers are derived from two different non-addr-taken
496
// globals, or if one is and the other isn't, we know these can't alias.
497
if ((GV1 || GV2) && GV1 != GV2)
500
// Otherwise if they are both derived from the same addr-taken global, we
501
// can't know the two accesses don't overlap.
504
// These pointers may be based on the memory owned by an indirect global. If
505
// so, we may be able to handle this. First check to see if the base pointer
506
// is a direct load from an indirect global.
508
if (const LoadInst *LI = dyn_cast<LoadInst>(UV1))
509
if (GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getOperand(0)))
510
if (IndirectGlobals.count(GV))
512
if (const LoadInst *LI = dyn_cast<LoadInst>(UV2))
513
if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getOperand(0)))
514
if (IndirectGlobals.count(GV))
517
// These pointers may also be from an allocation for the indirect global. If
518
// so, also handle them.
519
if (AllocsForIndirectGlobals.count(UV1))
520
GV1 = AllocsForIndirectGlobals[UV1];
521
if (AllocsForIndirectGlobals.count(UV2))
522
GV2 = AllocsForIndirectGlobals[UV2];
524
// Now that we know whether the two pointers are related to indirect globals,
525
// use this to disambiguate the pointers. If either pointer is based on an
526
// indirect global and if they are not both based on the same indirect global,
527
// they cannot alias.
528
if ((GV1 || GV2) && GV1 != GV2)
531
return AliasAnalysis::alias(V1, V1Size, V2, V2Size);
534
AliasAnalysis::ModRefResult
535
GlobalsModRef::getModRefInfo(ImmutableCallSite CS,
536
const Value *P, unsigned Size) {
537
unsigned Known = ModRef;
539
// If we are asking for mod/ref info of a direct call with a pointer to a
540
// global we are tracking, return information if we have it.
541
if (const GlobalValue *GV = dyn_cast<GlobalValue>(P->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, P, Size));
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);