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 #include "llvm/Analysis/GlobalsModRef.h"
18 #include "llvm/ADT/SCCIterator.h"
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/MemoryBuiltins.h"
22 #include "llvm/Analysis/TargetLibraryInfo.h"
23 #include "llvm/Analysis/ValueTracking.h"
24 #include "llvm/IR/DerivedTypes.h"
25 #include "llvm/IR/InstIterator.h"
26 #include "llvm/IR/Instructions.h"
27 #include "llvm/IR/IntrinsicInst.h"
28 #include "llvm/IR/Module.h"
29 #include "llvm/Pass.h"
30 #include "llvm/Support/CommandLine.h"
33 #define DEBUG_TYPE "globalsmodref-aa"
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");
42 // An option to enable unsafe alias results from the GlobalsModRef analysis.
43 // When enabled, GlobalsModRef will provide no-alias results which in extremely
44 // rare cases may not be conservatively correct. In particular, in the face of
45 // transforms which cause assymetry between how effective GetUnderlyingObject
46 // is for two pointers, it may produce incorrect results.
48 // These unsafe results have been returned by GMR for many years without
49 // causing significant issues in the wild and so we provide a mechanism to
50 // re-enable them for users of LLVM that have a particular performance
51 // sensitivity and no known issues. The option also makes it easy to evaluate
52 // the performance impact of these results.
53 static cl::opt<bool> EnableUnsafeGlobalsModRefAliasResults(
54 "enable-unsafe-globalsmodref-alias-results", cl::init(false), cl::Hidden);
56 /// The mod/ref information collected for a particular function.
58 /// We collect information about mod/ref behavior of a function here, both in
59 /// general and as pertains to specific globals. We only have this detailed
60 /// information when we know *something* useful about the behavior. If we
61 /// saturate to fully general mod/ref, we remove the info for the function.
62 class GlobalsAAResult::FunctionInfo {
63 typedef SmallDenseMap<const GlobalValue *, ModRefInfo, 16> GlobalInfoMapType;
65 /// Build a wrapper struct that has 8-byte alignment. All heap allocations
66 /// should provide this much alignment at least, but this makes it clear we
67 /// specifically rely on this amount of alignment.
68 struct LLVM_ALIGNAS(8) AlignedMap {
70 AlignedMap(const AlignedMap &Arg) : Map(Arg.Map) {}
71 GlobalInfoMapType Map;
74 /// Pointer traits for our aligned map.
75 struct AlignedMapPointerTraits {
76 static inline void *getAsVoidPointer(AlignedMap *P) { return P; }
77 static inline AlignedMap *getFromVoidPointer(void *P) {
78 return (AlignedMap *)P;
80 enum { NumLowBitsAvailable = 3 };
81 static_assert(alignof(AlignedMap) >= (1 << NumLowBitsAvailable),
82 "AlignedMap insufficiently aligned to have enough low bits.");
85 /// The bit that flags that this function may read any global. This is
86 /// chosen to mix together with ModRefInfo bits.
87 /// FIXME: This assumes ModRefInfo lattice will remain 4 bits!
88 /// It overlaps with ModRefInfo::Must bit!
89 /// FunctionInfo.getModRefInfo() masks out everything except ModRef so
90 /// this remains correct, but the Must info is lost.
91 enum { MayReadAnyGlobal = 4 };
93 /// Checks to document the invariants of the bit packing here.
94 static_assert((MayReadAnyGlobal & static_cast<int>(ModRefInfo::MustModRef)) ==
96 "ModRef and the MayReadAnyGlobal flag bits overlap.");
97 static_assert(((MayReadAnyGlobal |
98 static_cast<int>(ModRefInfo::MustModRef)) >>
99 AlignedMapPointerTraits::NumLowBitsAvailable) == 0,
100 "Insufficient low bits to store our flag and ModRef info.");
103 FunctionInfo() : Info() {}
105 delete Info.getPointer();
107 // Spell out the copy ond move constructors and assignment operators to get
108 // deep copy semantics and correct move semantics in the face of the
110 FunctionInfo(const FunctionInfo &Arg)
111 : Info(nullptr, Arg.Info.getInt()) {
112 if (const auto *ArgPtr = Arg.Info.getPointer())
113 Info.setPointer(new AlignedMap(*ArgPtr));
115 FunctionInfo(FunctionInfo &&Arg)
116 : Info(Arg.Info.getPointer(), Arg.Info.getInt()) {
117 Arg.Info.setPointerAndInt(nullptr, 0);
119 FunctionInfo &operator=(const FunctionInfo &RHS) {
120 delete Info.getPointer();
121 Info.setPointerAndInt(nullptr, RHS.Info.getInt());
122 if (const auto *RHSPtr = RHS.Info.getPointer())
123 Info.setPointer(new AlignedMap(*RHSPtr));
126 FunctionInfo &operator=(FunctionInfo &&RHS) {
127 delete Info.getPointer();
128 Info.setPointerAndInt(RHS.Info.getPointer(), RHS.Info.getInt());
129 RHS.Info.setPointerAndInt(nullptr, 0);
133 /// This method clears MayReadAnyGlobal bit added by GlobalsAAResult to return
134 /// the corresponding ModRefInfo. It must align in functionality with
136 ModRefInfo globalClearMayReadAnyGlobal(int I) const {
137 return ModRefInfo((I & static_cast<int>(ModRefInfo::ModRef)) |
138 static_cast<int>(ModRefInfo::NoModRef));
141 /// Returns the \c ModRefInfo info for this function.
142 ModRefInfo getModRefInfo() const {
143 return globalClearMayReadAnyGlobal(Info.getInt());
146 /// Adds new \c ModRefInfo for this function to its state.
147 void addModRefInfo(ModRefInfo NewMRI) {
148 Info.setInt(Info.getInt() | static_cast<int>(setMust(NewMRI)));
151 /// Returns whether this function may read any global variable, and we don't
152 /// know which global.
153 bool mayReadAnyGlobal() const { return Info.getInt() & MayReadAnyGlobal; }
155 /// Sets this function as potentially reading from any global.
156 void setMayReadAnyGlobal() { Info.setInt(Info.getInt() | MayReadAnyGlobal); }
158 /// Returns the \c ModRefInfo info for this function w.r.t. a particular
159 /// global, which may be more precise than the general information above.
160 ModRefInfo getModRefInfoForGlobal(const GlobalValue &GV) const {
161 ModRefInfo GlobalMRI =
162 mayReadAnyGlobal() ? ModRefInfo::Ref : ModRefInfo::NoModRef;
163 if (AlignedMap *P = Info.getPointer()) {
164 auto I = P->Map.find(&GV);
165 if (I != P->Map.end())
166 GlobalMRI = unionModRef(GlobalMRI, I->second);
171 /// Add mod/ref info from another function into ours, saturating towards
173 void addFunctionInfo(const FunctionInfo &FI) {
174 addModRefInfo(FI.getModRefInfo());
176 if (FI.mayReadAnyGlobal())
177 setMayReadAnyGlobal();
179 if (AlignedMap *P = FI.Info.getPointer())
180 for (const auto &G : P->Map)
181 addModRefInfoForGlobal(*G.first, G.second);
184 void addModRefInfoForGlobal(const GlobalValue &GV, ModRefInfo NewMRI) {
185 AlignedMap *P = Info.getPointer();
187 P = new AlignedMap();
190 auto &GlobalMRI = P->Map[&GV];
191 GlobalMRI = unionModRef(GlobalMRI, NewMRI);
194 /// Clear a global's ModRef info. Should be used when a global is being
196 void eraseModRefInfoForGlobal(const GlobalValue &GV) {
197 if (AlignedMap *P = Info.getPointer())
202 /// All of the information is encoded into a single pointer, with a three bit
203 /// integer in the low three bits. The high bit provides a flag for when this
204 /// function may read any global. The low two bits are the ModRefInfo. And
205 /// the pointer, when non-null, points to a map from GlobalValue to
206 /// ModRefInfo specific to that GlobalValue.
207 PointerIntPair<AlignedMap *, 3, unsigned, AlignedMapPointerTraits> Info;
210 void GlobalsAAResult::DeletionCallbackHandle::deleted() {
211 Value *V = getValPtr();
212 if (auto *F = dyn_cast<Function>(V))
213 GAR->FunctionInfos.erase(F);
215 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
216 if (GAR->NonAddressTakenGlobals.erase(GV)) {
217 // This global might be an indirect global. If so, remove it and
218 // remove any AllocRelatedValues for it.
219 if (GAR->IndirectGlobals.erase(GV)) {
220 // Remove any entries in AllocsForIndirectGlobals for this global.
221 for (auto I = GAR->AllocsForIndirectGlobals.begin(),
222 E = GAR->AllocsForIndirectGlobals.end();
225 GAR->AllocsForIndirectGlobals.erase(I);
228 // Scan the function info we have collected and remove this global
230 for (auto &FIPair : GAR->FunctionInfos)
231 FIPair.second.eraseModRefInfoForGlobal(*GV);
235 // If this is an allocation related to an indirect global, remove it.
236 GAR->AllocsForIndirectGlobals.erase(V);
238 // And clear out the handle.
240 GAR->Handles.erase(I);
241 // This object is now destroyed!
244 FunctionModRefBehavior GlobalsAAResult::getModRefBehavior(const Function *F) {
245 FunctionModRefBehavior Min = FMRB_UnknownModRefBehavior;
247 if (FunctionInfo *FI = getFunctionInfo(F)) {
248 if (!isModOrRefSet(FI->getModRefInfo()))
249 Min = FMRB_DoesNotAccessMemory;
250 else if (!isModSet(FI->getModRefInfo()))
251 Min = FMRB_OnlyReadsMemory;
254 return FunctionModRefBehavior(AAResultBase::getModRefBehavior(F) & Min);
257 FunctionModRefBehavior
258 GlobalsAAResult::getModRefBehavior(ImmutableCallSite CS) {
259 FunctionModRefBehavior Min = FMRB_UnknownModRefBehavior;
261 if (!CS.hasOperandBundles())
262 if (const Function *F = CS.getCalledFunction())
263 if (FunctionInfo *FI = getFunctionInfo(F)) {
264 if (!isModOrRefSet(FI->getModRefInfo()))
265 Min = FMRB_DoesNotAccessMemory;
266 else if (!isModSet(FI->getModRefInfo()))
267 Min = FMRB_OnlyReadsMemory;
270 return FunctionModRefBehavior(AAResultBase::getModRefBehavior(CS) & Min);
273 /// Returns the function info for the function, or null if we don't have
274 /// anything useful to say about it.
275 GlobalsAAResult::FunctionInfo *
276 GlobalsAAResult::getFunctionInfo(const Function *F) {
277 auto I = FunctionInfos.find(F);
278 if (I != FunctionInfos.end())
283 /// AnalyzeGlobals - Scan through the users of all of the internal
284 /// GlobalValue's in the program. If none of them have their "address taken"
285 /// (really, their address passed to something nontrivial), record this fact,
286 /// and record the functions that they are used directly in.
287 void GlobalsAAResult::AnalyzeGlobals(Module &M) {
288 SmallPtrSet<Function *, 32> TrackedFunctions;
289 for (Function &F : M)
290 if (F.hasLocalLinkage())
291 if (!AnalyzeUsesOfPointer(&F)) {
292 // Remember that we are tracking this global.
293 NonAddressTakenGlobals.insert(&F);
294 TrackedFunctions.insert(&F);
295 Handles.emplace_front(*this, &F);
296 Handles.front().I = Handles.begin();
297 ++NumNonAddrTakenFunctions;
300 SmallPtrSet<Function *, 16> Readers, Writers;
301 for (GlobalVariable &GV : M.globals())
302 if (GV.hasLocalLinkage()) {
303 if (!AnalyzeUsesOfPointer(&GV, &Readers,
304 GV.isConstant() ? nullptr : &Writers)) {
305 // Remember that we are tracking this global, and the mod/ref fns
306 NonAddressTakenGlobals.insert(&GV);
307 Handles.emplace_front(*this, &GV);
308 Handles.front().I = Handles.begin();
310 for (Function *Reader : Readers) {
311 if (TrackedFunctions.insert(Reader).second) {
312 Handles.emplace_front(*this, Reader);
313 Handles.front().I = Handles.begin();
315 FunctionInfos[Reader].addModRefInfoForGlobal(GV, ModRefInfo::Ref);
318 if (!GV.isConstant()) // No need to keep track of writers to constants
319 for (Function *Writer : Writers) {
320 if (TrackedFunctions.insert(Writer).second) {
321 Handles.emplace_front(*this, Writer);
322 Handles.front().I = Handles.begin();
324 FunctionInfos[Writer].addModRefInfoForGlobal(GV, ModRefInfo::Mod);
326 ++NumNonAddrTakenGlobalVars;
328 // If this global holds a pointer type, see if it is an indirect global.
329 if (GV.getValueType()->isPointerTy() &&
330 AnalyzeIndirectGlobalMemory(&GV))
331 ++NumIndirectGlobalVars;
338 /// AnalyzeUsesOfPointer - Look at all of the users of the specified pointer.
339 /// If this is used by anything complex (i.e., the address escapes), return
340 /// true. Also, while we are at it, keep track of those functions that read and
341 /// write to the value.
343 /// If OkayStoreDest is non-null, stores into this global are allowed.
344 bool GlobalsAAResult::AnalyzeUsesOfPointer(Value *V,
345 SmallPtrSetImpl<Function *> *Readers,
346 SmallPtrSetImpl<Function *> *Writers,
347 GlobalValue *OkayStoreDest) {
348 if (!V->getType()->isPointerTy())
351 for (Use &U : V->uses()) {
352 User *I = U.getUser();
353 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
355 Readers->insert(LI->getParent()->getParent());
356 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
357 if (V == SI->getOperand(1)) {
359 Writers->insert(SI->getParent()->getParent());
360 } else if (SI->getOperand(1) != OkayStoreDest) {
361 return true; // Storing the pointer
363 } else if (Operator::getOpcode(I) == Instruction::GetElementPtr) {
364 if (AnalyzeUsesOfPointer(I, Readers, Writers))
366 } else if (Operator::getOpcode(I) == Instruction::BitCast) {
367 if (AnalyzeUsesOfPointer(I, Readers, Writers, OkayStoreDest))
369 } else if (auto CS = CallSite(I)) {
370 // Make sure that this is just the function being called, not that it is
371 // passing into the function.
372 if (CS.isDataOperand(&U)) {
373 // Detect calls to free.
374 if (CS.isArgOperand(&U) && isFreeCall(I, &TLI)) {
376 Writers->insert(CS->getParent()->getParent());
378 return true; // Argument of an unknown call.
381 } else if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) {
382 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
383 return true; // Allow comparison against null.
384 } else if (Constant *C = dyn_cast<Constant>(I)) {
385 // Ignore constants which don't have any live uses.
386 if (isa<GlobalValue>(C) || C->isConstantUsed())
396 /// AnalyzeIndirectGlobalMemory - We found an non-address-taken global variable
397 /// which holds a pointer type. See if the global always points to non-aliased
398 /// heap memory: that is, all initializers of the globals are allocations, and
399 /// those allocations have no use other than initialization of the global.
400 /// Further, all loads out of GV must directly use the memory, not store the
401 /// pointer somewhere. If this is true, we consider the memory pointed to by
402 /// GV to be owned by GV and can disambiguate other pointers from it.
403 bool GlobalsAAResult::AnalyzeIndirectGlobalMemory(GlobalVariable *GV) {
404 // Keep track of values related to the allocation of the memory, f.e. the
405 // value produced by the malloc call and any casts.
406 std::vector<Value *> AllocRelatedValues;
408 // If the initializer is a valid pointer, bail.
409 if (Constant *C = GV->getInitializer())
410 if (!C->isNullValue())
413 // Walk the user list of the global. If we find anything other than a direct
414 // load or store, bail out.
415 for (User *U : GV->users()) {
416 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
417 // The pointer loaded from the global can only be used in simple ways:
418 // we allow addressing of it and loading storing to it. We do *not* allow
419 // storing the loaded pointer somewhere else or passing to a function.
420 if (AnalyzeUsesOfPointer(LI))
421 return false; // Loaded pointer escapes.
422 // TODO: Could try some IP mod/ref of the loaded pointer.
423 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
424 // Storing the global itself.
425 if (SI->getOperand(0) == GV)
428 // If storing the null pointer, ignore it.
429 if (isa<ConstantPointerNull>(SI->getOperand(0)))
432 // Check the value being stored.
433 Value *Ptr = GetUnderlyingObject(SI->getOperand(0),
434 GV->getParent()->getDataLayout());
436 if (!isAllocLikeFn(Ptr, &TLI))
437 return false; // Too hard to analyze.
439 // Analyze all uses of the allocation. If any of them are used in a
440 // non-simple way (e.g. stored to another global) bail out.
441 if (AnalyzeUsesOfPointer(Ptr, /*Readers*/ nullptr, /*Writers*/ nullptr,
443 return false; // Loaded pointer escapes.
445 // Remember that this allocation is related to the indirect global.
446 AllocRelatedValues.push_back(Ptr);
448 // Something complex, bail out.
453 // Okay, this is an indirect global. Remember all of the allocations for
454 // this global in AllocsForIndirectGlobals.
455 while (!AllocRelatedValues.empty()) {
456 AllocsForIndirectGlobals[AllocRelatedValues.back()] = GV;
457 Handles.emplace_front(*this, AllocRelatedValues.back());
458 Handles.front().I = Handles.begin();
459 AllocRelatedValues.pop_back();
461 IndirectGlobals.insert(GV);
462 Handles.emplace_front(*this, GV);
463 Handles.front().I = Handles.begin();
467 void GlobalsAAResult::CollectSCCMembership(CallGraph &CG) {
468 // We do a bottom-up SCC traversal of the call graph. In other words, we
469 // visit all callees before callers (leaf-first).
471 for (scc_iterator<CallGraph *> I = scc_begin(&CG); !I.isAtEnd(); ++I) {
472 const std::vector<CallGraphNode *> &SCC = *I;
473 assert(!SCC.empty() && "SCC with no functions?");
475 for (auto *CGN : SCC)
476 if (Function *F = CGN->getFunction())
477 FunctionToSCCMap[F] = SCCID;
482 /// AnalyzeCallGraph - At this point, we know the functions where globals are
483 /// immediately stored to and read from. Propagate this information up the call
484 /// graph to all callers and compute the mod/ref info for all memory for each
486 void GlobalsAAResult::AnalyzeCallGraph(CallGraph &CG, Module &M) {
487 // We do a bottom-up SCC traversal of the call graph. In other words, we
488 // visit all callees before callers (leaf-first).
489 for (scc_iterator<CallGraph *> I = scc_begin(&CG); !I.isAtEnd(); ++I) {
490 const std::vector<CallGraphNode *> &SCC = *I;
491 assert(!SCC.empty() && "SCC with no functions?");
493 Function *F = SCC[0]->getFunction();
495 if (!F || !F->isDefinitionExact()) {
496 // Calls externally or not exact - can't say anything useful. Remove any
497 // existing function records (may have been created when scanning
499 for (auto *Node : SCC)
500 FunctionInfos.erase(Node->getFunction());
504 FunctionInfo &FI = FunctionInfos[F];
505 bool KnowNothing = false;
507 // Collect the mod/ref properties due to called functions. We only compute
509 for (unsigned i = 0, e = SCC.size(); i != e && !KnowNothing; ++i) {
515 if (F->isDeclaration() || F->hasFnAttribute(Attribute::OptimizeNone)) {
516 // Try to get mod/ref behaviour from function attributes.
517 if (F->doesNotAccessMemory()) {
518 // Can't do better than that!
519 } else if (F->onlyReadsMemory()) {
520 FI.addModRefInfo(ModRefInfo::Ref);
521 if (!F->isIntrinsic() && !F->onlyAccessesArgMemory())
522 // This function might call back into the module and read a global -
523 // consider every global as possibly being read by this function.
524 FI.setMayReadAnyGlobal();
526 FI.addModRefInfo(ModRefInfo::ModRef);
527 // Can't say anything useful unless it's an intrinsic - they don't
528 // read or write global variables of the kind considered here.
529 KnowNothing = !F->isIntrinsic();
534 for (CallGraphNode::iterator CI = SCC[i]->begin(), E = SCC[i]->end();
535 CI != E && !KnowNothing; ++CI)
536 if (Function *Callee = CI->second->getFunction()) {
537 if (FunctionInfo *CalleeFI = getFunctionInfo(Callee)) {
538 // Propagate function effect up.
539 FI.addFunctionInfo(*CalleeFI);
541 // Can't say anything about it. However, if it is inside our SCC,
542 // then nothing needs to be done.
543 CallGraphNode *CalleeNode = CG[Callee];
544 if (!is_contained(SCC, CalleeNode))
552 // If we can't say anything useful about this SCC, remove all SCC functions
553 // from the FunctionInfos map.
555 for (auto *Node : SCC)
556 FunctionInfos.erase(Node->getFunction());
560 // Scan the function bodies for explicit loads or stores.
561 for (auto *Node : SCC) {
562 if (isModAndRefSet(FI.getModRefInfo()))
563 break; // The mod/ref lattice saturates here.
565 // Don't prove any properties based on the implementation of an optnone
566 // function. Function attributes were already used as a best approximation
568 if (Node->getFunction()->hasFnAttribute(Attribute::OptimizeNone))
571 for (Instruction &I : instructions(Node->getFunction())) {
572 if (isModAndRefSet(FI.getModRefInfo()))
573 break; // The mod/ref lattice saturates here.
575 // We handle calls specially because the graph-relevant aspects are
577 if (auto CS = CallSite(&I)) {
578 if (isAllocationFn(&I, &TLI) || isFreeCall(&I, &TLI)) {
579 // FIXME: It is completely unclear why this is necessary and not
580 // handled by the above graph code.
581 FI.addModRefInfo(ModRefInfo::ModRef);
582 } else if (Function *Callee = CS.getCalledFunction()) {
583 // The callgraph doesn't include intrinsic calls.
584 if (Callee->isIntrinsic()) {
585 FunctionModRefBehavior Behaviour =
586 AAResultBase::getModRefBehavior(Callee);
587 FI.addModRefInfo(createModRefInfo(Behaviour));
593 // All non-call instructions we use the primary predicates for whether
594 // thay read or write memory.
595 if (I.mayReadFromMemory())
596 FI.addModRefInfo(ModRefInfo::Ref);
597 if (I.mayWriteToMemory())
598 FI.addModRefInfo(ModRefInfo::Mod);
602 if (!isModSet(FI.getModRefInfo()))
603 ++NumReadMemFunctions;
604 if (!isModOrRefSet(FI.getModRefInfo()))
607 // Finally, now that we know the full effect on this SCC, clone the
608 // information to each function in the SCC.
609 // FI is a reference into FunctionInfos, so copy it now so that it doesn't
610 // get invalidated if DenseMap decides to re-hash.
611 FunctionInfo CachedFI = FI;
612 for (unsigned i = 1, e = SCC.size(); i != e; ++i)
613 FunctionInfos[SCC[i]->getFunction()] = CachedFI;
617 // GV is a non-escaping global. V is a pointer address that has been loaded from.
618 // If we can prove that V must escape, we can conclude that a load from V cannot
620 static bool isNonEscapingGlobalNoAliasWithLoad(const GlobalValue *GV,
623 const DataLayout &DL) {
624 SmallPtrSet<const Value *, 8> Visited;
625 SmallVector<const Value *, 8> Inputs;
629 const Value *Input = Inputs.pop_back_val();
631 if (isa<GlobalValue>(Input) || isa<Argument>(Input) || isa<CallInst>(Input) ||
632 isa<InvokeInst>(Input))
633 // Arguments to functions or returns from functions are inherently
634 // escaping, so we can immediately classify those as not aliasing any
635 // non-addr-taken globals.
637 // (Transitive) loads from a global are also safe - if this aliased
638 // another global, its address would escape, so no alias.
641 // Recurse through a limited number of selects, loads and PHIs. This is an
642 // arbitrary depth of 4, lower numbers could be used to fix compile time
643 // issues if needed, but this is generally expected to be only be important
648 if (auto *LI = dyn_cast<LoadInst>(Input)) {
649 Inputs.push_back(GetUnderlyingObject(LI->getPointerOperand(), DL));
652 if (auto *SI = dyn_cast<SelectInst>(Input)) {
653 const Value *LHS = GetUnderlyingObject(SI->getTrueValue(), DL);
654 const Value *RHS = GetUnderlyingObject(SI->getFalseValue(), DL);
655 if (Visited.insert(LHS).second)
656 Inputs.push_back(LHS);
657 if (Visited.insert(RHS).second)
658 Inputs.push_back(RHS);
661 if (auto *PN = dyn_cast<PHINode>(Input)) {
662 for (const Value *Op : PN->incoming_values()) {
663 Op = GetUnderlyingObject(Op, DL);
664 if (Visited.insert(Op).second)
665 Inputs.push_back(Op);
671 } while (!Inputs.empty());
673 // All inputs were known to be no-alias.
677 // There are particular cases where we can conclude no-alias between
678 // a non-addr-taken global and some other underlying object. Specifically,
679 // a non-addr-taken global is known to not be escaped from any function. It is
680 // also incorrect for a transformation to introduce an escape of a global in
681 // a way that is observable when it was not there previously. One function
682 // being transformed to introduce an escape which could possibly be observed
683 // (via loading from a global or the return value for example) within another
684 // function is never safe. If the observation is made through non-atomic
685 // operations on different threads, it is a data-race and UB. If the
686 // observation is well defined, by being observed the transformation would have
687 // changed program behavior by introducing the observed escape, making it an
688 // invalid transform.
690 // This property does require that transformations which *temporarily* escape
691 // a global that was not previously escaped, prior to restoring it, cannot rely
692 // on the results of GMR::alias. This seems a reasonable restriction, although
693 // currently there is no way to enforce it. There is also no realistic
694 // optimization pass that would make this mistake. The closest example is
695 // a transformation pass which does reg2mem of SSA values but stores them into
696 // global variables temporarily before restoring the global variable's value.
697 // This could be useful to expose "benign" races for example. However, it seems
698 // reasonable to require that a pass which introduces escapes of global
699 // variables in this way to either not trust AA results while the escape is
700 // active, or to be forced to operate as a module pass that cannot co-exist
701 // with an alias analysis such as GMR.
702 bool GlobalsAAResult::isNonEscapingGlobalNoAlias(const GlobalValue *GV,
704 // In order to know that the underlying object cannot alias the
705 // non-addr-taken global, we must know that it would have to be an escape.
706 // Thus if the underlying object is a function argument, a load from
707 // a global, or the return of a function, it cannot alias. We can also
708 // recurse through PHI nodes and select nodes provided all of their inputs
709 // resolve to one of these known-escaping roots.
710 SmallPtrSet<const Value *, 8> Visited;
711 SmallVector<const Value *, 8> Inputs;
716 const Value *Input = Inputs.pop_back_val();
718 if (auto *InputGV = dyn_cast<GlobalValue>(Input)) {
719 // If one input is the very global we're querying against, then we can't
720 // conclude no-alias.
724 // Distinct GlobalVariables never alias, unless overriden or zero-sized.
725 // FIXME: The condition can be refined, but be conservative for now.
726 auto *GVar = dyn_cast<GlobalVariable>(GV);
727 auto *InputGVar = dyn_cast<GlobalVariable>(InputGV);
728 if (GVar && InputGVar &&
729 !GVar->isDeclaration() && !InputGVar->isDeclaration() &&
730 !GVar->isInterposable() && !InputGVar->isInterposable()) {
731 Type *GVType = GVar->getInitializer()->getType();
732 Type *InputGVType = InputGVar->getInitializer()->getType();
733 if (GVType->isSized() && InputGVType->isSized() &&
734 (DL.getTypeAllocSize(GVType) > 0) &&
735 (DL.getTypeAllocSize(InputGVType) > 0))
739 // Conservatively return false, even though we could be smarter
740 // (e.g. look through GlobalAliases).
744 if (isa<Argument>(Input) || isa<CallInst>(Input) ||
745 isa<InvokeInst>(Input)) {
746 // Arguments to functions or returns from functions are inherently
747 // escaping, so we can immediately classify those as not aliasing any
748 // non-addr-taken globals.
752 // Recurse through a limited number of selects, loads and PHIs. This is an
753 // arbitrary depth of 4, lower numbers could be used to fix compile time
754 // issues if needed, but this is generally expected to be only be important
759 if (auto *LI = dyn_cast<LoadInst>(Input)) {
760 // A pointer loaded from a global would have been captured, and we know
761 // that the global is non-escaping, so no alias.
762 const Value *Ptr = GetUnderlyingObject(LI->getPointerOperand(), DL);
763 if (isNonEscapingGlobalNoAliasWithLoad(GV, Ptr, Depth, DL))
764 // The load does not alias with GV.
766 // Otherwise, a load could come from anywhere, so bail.
769 if (auto *SI = dyn_cast<SelectInst>(Input)) {
770 const Value *LHS = GetUnderlyingObject(SI->getTrueValue(), DL);
771 const Value *RHS = GetUnderlyingObject(SI->getFalseValue(), DL);
772 if (Visited.insert(LHS).second)
773 Inputs.push_back(LHS);
774 if (Visited.insert(RHS).second)
775 Inputs.push_back(RHS);
778 if (auto *PN = dyn_cast<PHINode>(Input)) {
779 for (const Value *Op : PN->incoming_values()) {
780 Op = GetUnderlyingObject(Op, DL);
781 if (Visited.insert(Op).second)
782 Inputs.push_back(Op);
787 // FIXME: It would be good to handle other obvious no-alias cases here, but
788 // it isn't clear how to do so reasonbly without building a small version
789 // of BasicAA into this code. We could recurse into AAResultBase::alias
790 // here but that seems likely to go poorly as we're inside the
791 // implementation of such a query. Until then, just conservatievly retun
794 } while (!Inputs.empty());
796 // If all the inputs to V were definitively no-alias, then V is no-alias.
800 /// alias - If one of the pointers is to a global that we are tracking, and the
801 /// other is some random pointer, we know there cannot be an alias, because the
802 /// address of the global isn't taken.
803 AliasResult GlobalsAAResult::alias(const MemoryLocation &LocA,
804 const MemoryLocation &LocB) {
805 // Get the base object these pointers point to.
806 const Value *UV1 = GetUnderlyingObject(LocA.Ptr, DL);
807 const Value *UV2 = GetUnderlyingObject(LocB.Ptr, DL);
809 // If either of the underlying values is a global, they may be non-addr-taken
810 // globals, which we can answer queries about.
811 const GlobalValue *GV1 = dyn_cast<GlobalValue>(UV1);
812 const GlobalValue *GV2 = dyn_cast<GlobalValue>(UV2);
814 // If the global's address is taken, pretend we don't know it's a pointer to
816 if (GV1 && !NonAddressTakenGlobals.count(GV1))
818 if (GV2 && !NonAddressTakenGlobals.count(GV2))
821 // If the two pointers are derived from two different non-addr-taken
822 // globals we know these can't alias.
823 if (GV1 && GV2 && GV1 != GV2)
826 // If one is and the other isn't, it isn't strictly safe but we can fake
827 // this result if necessary for performance. This does not appear to be
828 // a common problem in practice.
829 if (EnableUnsafeGlobalsModRefAliasResults)
830 if ((GV1 || GV2) && GV1 != GV2)
833 // Check for a special case where a non-escaping global can be used to
834 // conclude no-alias.
835 if ((GV1 || GV2) && GV1 != GV2) {
836 const GlobalValue *GV = GV1 ? GV1 : GV2;
837 const Value *UV = GV1 ? UV2 : UV1;
838 if (isNonEscapingGlobalNoAlias(GV, UV))
842 // Otherwise if they are both derived from the same addr-taken global, we
843 // can't know the two accesses don't overlap.
846 // These pointers may be based on the memory owned by an indirect global. If
847 // so, we may be able to handle this. First check to see if the base pointer
848 // is a direct load from an indirect global.
850 if (const LoadInst *LI = dyn_cast<LoadInst>(UV1))
851 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getOperand(0)))
852 if (IndirectGlobals.count(GV))
854 if (const LoadInst *LI = dyn_cast<LoadInst>(UV2))
855 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getOperand(0)))
856 if (IndirectGlobals.count(GV))
859 // These pointers may also be from an allocation for the indirect global. If
860 // so, also handle them.
862 GV1 = AllocsForIndirectGlobals.lookup(UV1);
864 GV2 = AllocsForIndirectGlobals.lookup(UV2);
866 // Now that we know whether the two pointers are related to indirect globals,
867 // use this to disambiguate the pointers. If the pointers are based on
868 // different indirect globals they cannot alias.
869 if (GV1 && GV2 && GV1 != GV2)
872 // If one is based on an indirect global and the other isn't, it isn't
873 // strictly safe but we can fake this result if necessary for performance.
874 // This does not appear to be a common problem in practice.
875 if (EnableUnsafeGlobalsModRefAliasResults)
876 if ((GV1 || GV2) && GV1 != GV2)
879 return AAResultBase::alias(LocA, LocB);
882 ModRefInfo GlobalsAAResult::getModRefInfoForArgument(ImmutableCallSite CS,
883 const GlobalValue *GV) {
884 if (CS.doesNotAccessMemory())
885 return ModRefInfo::NoModRef;
886 ModRefInfo ConservativeResult =
887 CS.onlyReadsMemory() ? ModRefInfo::Ref : ModRefInfo::ModRef;
889 // Iterate through all the arguments to the called function. If any argument
890 // is based on GV, return the conservative result.
891 for (auto &A : CS.args()) {
892 SmallVector<Value*, 4> Objects;
893 GetUnderlyingObjects(A, Objects, DL);
895 // All objects must be identified.
896 if (!all_of(Objects, isIdentifiedObject) &&
897 // Try ::alias to see if all objects are known not to alias GV.
898 !all_of(Objects, [&](Value *V) {
899 return this->alias(MemoryLocation(V), MemoryLocation(GV)) == NoAlias;
901 return ConservativeResult;
903 if (is_contained(Objects, GV))
904 return ConservativeResult;
907 // We identified all objects in the argument list, and none of them were GV.
908 return ModRefInfo::NoModRef;
911 ModRefInfo GlobalsAAResult::getModRefInfo(ImmutableCallSite CS,
912 const MemoryLocation &Loc) {
913 ModRefInfo Known = ModRefInfo::ModRef;
915 // If we are asking for mod/ref info of a direct call with a pointer to a
916 // global we are tracking, return information if we have it.
917 if (const GlobalValue *GV =
918 dyn_cast<GlobalValue>(GetUnderlyingObject(Loc.Ptr, DL)))
919 if (GV->hasLocalLinkage())
920 if (const Function *F = CS.getCalledFunction())
921 if (NonAddressTakenGlobals.count(GV))
922 if (const FunctionInfo *FI = getFunctionInfo(F))
923 Known = unionModRef(FI->getModRefInfoForGlobal(*GV),
924 getModRefInfoForArgument(CS, GV));
926 if (!isModOrRefSet(Known))
927 return ModRefInfo::NoModRef; // No need to query other mod/ref analyses
928 return intersectModRef(Known, AAResultBase::getModRefInfo(CS, Loc));
931 GlobalsAAResult::GlobalsAAResult(const DataLayout &DL,
932 const TargetLibraryInfo &TLI)
933 : AAResultBase(), DL(DL), TLI(TLI) {}
935 GlobalsAAResult::GlobalsAAResult(GlobalsAAResult &&Arg)
936 : AAResultBase(std::move(Arg)), DL(Arg.DL), TLI(Arg.TLI),
937 NonAddressTakenGlobals(std::move(Arg.NonAddressTakenGlobals)),
938 IndirectGlobals(std::move(Arg.IndirectGlobals)),
939 AllocsForIndirectGlobals(std::move(Arg.AllocsForIndirectGlobals)),
940 FunctionInfos(std::move(Arg.FunctionInfos)),
941 Handles(std::move(Arg.Handles)) {
942 // Update the parent for each DeletionCallbackHandle.
943 for (auto &H : Handles) {
944 assert(H.GAR == &Arg);
949 GlobalsAAResult::~GlobalsAAResult() {}
951 /*static*/ GlobalsAAResult
952 GlobalsAAResult::analyzeModule(Module &M, const TargetLibraryInfo &TLI,
954 GlobalsAAResult Result(M.getDataLayout(), TLI);
956 // Discover which functions aren't recursive, to feed into AnalyzeGlobals.
957 Result.CollectSCCMembership(CG);
959 // Find non-addr taken globals.
960 Result.AnalyzeGlobals(M);
963 Result.AnalyzeCallGraph(CG, M);
968 AnalysisKey GlobalsAA::Key;
970 GlobalsAAResult GlobalsAA::run(Module &M, ModuleAnalysisManager &AM) {
971 return GlobalsAAResult::analyzeModule(M,
972 AM.getResult<TargetLibraryAnalysis>(M),
973 AM.getResult<CallGraphAnalysis>(M));
976 char GlobalsAAWrapperPass::ID = 0;
977 INITIALIZE_PASS_BEGIN(GlobalsAAWrapperPass, "globals-aa",
978 "Globals Alias Analysis", false, true)
979 INITIALIZE_PASS_DEPENDENCY(CallGraphWrapperPass)
980 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
981 INITIALIZE_PASS_END(GlobalsAAWrapperPass, "globals-aa",
982 "Globals Alias Analysis", false, true)
984 ModulePass *llvm::createGlobalsAAWrapperPass() {
985 return new GlobalsAAWrapperPass();
988 GlobalsAAWrapperPass::GlobalsAAWrapperPass() : ModulePass(ID) {
989 initializeGlobalsAAWrapperPassPass(*PassRegistry::getPassRegistry());
992 bool GlobalsAAWrapperPass::runOnModule(Module &M) {
993 Result.reset(new GlobalsAAResult(GlobalsAAResult::analyzeModule(
994 M, getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
995 getAnalysis<CallGraphWrapperPass>().getCallGraph())));
999 bool GlobalsAAWrapperPass::doFinalization(Module &M) {
1004 void GlobalsAAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
1005 AU.setPreservesAll();
1006 AU.addRequired<CallGraphWrapperPass>();
1007 AU.addRequired<TargetLibraryInfoWrapperPass>();