1 //===- CFLAliasAnalysis.cpp - CFL-Based Alias Analysis Implementation ------==//
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 file implements a CFL-based context-insensitive alias analysis
11 // algorithm. It does not depend on types. The algorithm is a mixture of the one
12 // described in "Demand-driven alias analysis for C" by Xin Zheng and Radu
13 // Rugina, and "Fast algorithms for Dyck-CFL-reachability with applications to
14 // Alias Analysis" by Zhang Q, Lyu M R, Yuan H, and Su Z. -- to summarize the
15 // papers, we build a graph of the uses of a variable, where each node is a
16 // memory location, and each edge is an action that happened on that memory
17 // location. The "actions" can be one of Dereference, Reference, Assign, or
20 // Two variables are considered as aliasing iff you can reach one value's node
21 // from the other value's node and the language formed by concatenating all of
22 // the edge labels (actions) conforms to a context-free grammar.
24 // Because this algorithm requires a graph search on each query, we execute the
25 // algorithm outlined in "Fast algorithms..." (mentioned above)
26 // in order to transform the graph into sets of variables that may alias in
27 // ~nlogn time (n = number of variables.), which makes queries take constant
29 //===----------------------------------------------------------------------===//
31 #include "StratifiedSets.h"
32 #include "llvm/ADT/BitVector.h"
33 #include "llvm/ADT/DenseMap.h"
34 #include "llvm/ADT/None.h"
35 #include "llvm/ADT/Optional.h"
36 #include "llvm/Analysis/AliasAnalysis.h"
37 #include "llvm/Analysis/Passes.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/Function.h"
40 #include "llvm/IR/InstVisitor.h"
41 #include "llvm/IR/Instructions.h"
42 #include "llvm/IR/ValueHandle.h"
43 #include "llvm/Pass.h"
44 #include "llvm/Support/Allocator.h"
45 #include "llvm/Support/Compiler.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/ErrorHandling.h"
51 #include <forward_list>
56 #define DEBUG_TYPE "cfl-aa"
58 // Try to go from a Value* to a Function*. Never returns nullptr.
59 static Optional<Function *> parentFunctionOfValue(Value *);
61 // Returns possible functions called by the Inst* into the given
62 // SmallVectorImpl. Returns true if targets found, false otherwise.
63 // This is templated because InvokeInst/CallInst give us the same
64 // set of functions that we care about, and I don't like repeating
66 template <typename Inst>
67 static bool getPossibleTargets(Inst *, SmallVectorImpl<Function *> &);
69 // Some instructions need to have their users tracked. Instructions like
70 // `add` require you to get the users of the Instruction* itself, other
71 // instructions like `store` require you to get the users of the first
72 // operand. This function gets the "proper" value to track for each
73 // type of instruction we support.
74 static Optional<Value *> getTargetValue(Instruction *);
76 // There are certain instructions (i.e. FenceInst, etc.) that we ignore.
77 // This notes that we should ignore those.
78 static bool hasUsefulEdges(Instruction *);
80 const StratifiedIndex StratifiedLink::SetSentinel =
81 std::numeric_limits<StratifiedIndex>::max();
84 // StratifiedInfo Attribute things.
85 typedef unsigned StratifiedAttr;
86 LLVM_CONSTEXPR unsigned MaxStratifiedAttrIndex = NumStratifiedAttrs;
87 LLVM_CONSTEXPR unsigned AttrAllIndex = 0;
88 LLVM_CONSTEXPR unsigned AttrGlobalIndex = 1;
89 LLVM_CONSTEXPR unsigned AttrUnknownIndex = 2;
90 LLVM_CONSTEXPR unsigned AttrFirstArgIndex = 3;
91 LLVM_CONSTEXPR unsigned AttrLastArgIndex = MaxStratifiedAttrIndex;
92 LLVM_CONSTEXPR unsigned AttrMaxNumArgs = AttrLastArgIndex - AttrFirstArgIndex;
94 LLVM_CONSTEXPR StratifiedAttr AttrNone = 0;
95 LLVM_CONSTEXPR StratifiedAttr AttrUnknown = 1 << AttrUnknownIndex;
96 LLVM_CONSTEXPR StratifiedAttr AttrAll = ~AttrNone;
98 // \brief StratifiedSets call for knowledge of "direction", so this is how we
99 // represent that locally.
100 enum class Level { Same, Above, Below };
102 // \brief Edges can be one of four "weights" -- each weight must have an inverse
103 // weight (Assign has Assign; Reference has Dereference).
104 enum class EdgeType {
105 // The weight assigned when assigning from or to a value. For example, in:
106 // %b = getelementptr %a, 0
107 // ...The relationships are %b assign %a, and %a assign %b. This used to be
108 // two edges, but having a distinction bought us nothing.
111 // The edge used when we have an edge going from some handle to a Value.
112 // Examples of this include:
113 // %b = load %a (%b Dereference %a)
114 // %b = extractelement %a, 0 (%a Dereference %b)
117 // The edge used when our edge goes from a value to a handle that may have
118 // contained it at some point. Examples:
119 // %b = load %a (%a Reference %b)
120 // %b = extractelement %a, 0 (%b Reference %a)
124 // \brief Encodes the notion of a "use"
126 // \brief Which value the edge is coming from
129 // \brief Which value the edge is pointing to
132 // \brief Edge weight
135 // \brief Whether we aliased any external values along the way that may be
136 // invisible to the analysis (i.e. landingpad for exceptions, calls for
137 // interprocedural analysis, etc.)
138 StratifiedAttrs AdditionalAttrs;
140 Edge(Value *From, Value *To, EdgeType W, StratifiedAttrs A)
141 : From(From), To(To), Weight(W), AdditionalAttrs(A) {}
144 // \brief Information we have about a function and would like to keep around
145 struct FunctionInfo {
146 StratifiedSets<Value *> Sets;
147 // Lots of functions have < 4 returns. Adjust as necessary.
148 SmallVector<Value *, 4> ReturnedValues;
150 FunctionInfo(StratifiedSets<Value *> &&S, SmallVector<Value *, 4> &&RV)
151 : Sets(std::move(S)), ReturnedValues(std::move(RV)) {}
154 struct CFLAliasAnalysis;
156 struct FunctionHandle : public CallbackVH {
157 FunctionHandle(Function *Fn, CFLAliasAnalysis *CFLAA)
158 : CallbackVH(Fn), CFLAA(CFLAA) {
159 assert(Fn != nullptr);
160 assert(CFLAA != nullptr);
163 virtual ~FunctionHandle() {}
165 void deleted() override { removeSelfFromCache(); }
166 void allUsesReplacedWith(Value *) override { removeSelfFromCache(); }
169 CFLAliasAnalysis *CFLAA;
171 void removeSelfFromCache();
174 struct CFLAliasAnalysis : public ImmutablePass, public AliasAnalysis {
176 /// \brief Cached mapping of Functions to their StratifiedSets.
177 /// If a function's sets are currently being built, it is marked
178 /// in the cache as an Optional without a value. This way, if we
179 /// have any kind of recursion, it is discernable from a function
180 /// that simply has empty sets.
181 DenseMap<Function *, Optional<FunctionInfo>> Cache;
182 std::forward_list<FunctionHandle> Handles;
187 CFLAliasAnalysis() : ImmutablePass(ID) {
188 initializeCFLAliasAnalysisPass(*PassRegistry::getPassRegistry());
191 virtual ~CFLAliasAnalysis() {}
193 void getAnalysisUsage(AnalysisUsage &AU) const override {
194 AliasAnalysis::getAnalysisUsage(AU);
197 void *getAdjustedAnalysisPointer(const void *ID) override {
198 if (ID == &AliasAnalysis::ID)
199 return (AliasAnalysis *)this;
203 /// \brief Inserts the given Function into the cache.
204 void scan(Function *Fn);
206 void evict(Function *Fn) { Cache.erase(Fn); }
208 /// \brief Ensures that the given function is available in the cache.
209 /// Returns the appropriate entry from the cache.
210 const Optional<FunctionInfo> &ensureCached(Function *Fn) {
211 auto Iter = Cache.find(Fn);
212 if (Iter == Cache.end()) {
214 Iter = Cache.find(Fn);
215 assert(Iter != Cache.end());
216 assert(Iter->second.hasValue());
221 AliasResult query(const Location &LocA, const Location &LocB);
223 AliasResult alias(const Location &LocA, const Location &LocB) override {
224 if (LocA.Ptr == LocB.Ptr) {
225 if (LocA.Size == LocB.Size) {
232 // Comparisons between global variables and other constants should be
233 // handled by BasicAA.
234 // TODO: ConstantExpr handling -- CFLAA may report NoAlias when comparing
235 // a GlobalValue and ConstantExpr, but every query needs to have at least
236 // one Value tied to a Function, and neither GlobalValues nor ConstantExprs
238 if (isa<Constant>(LocA.Ptr) && isa<Constant>(LocB.Ptr)) {
239 return AliasAnalysis::alias(LocA, LocB);
242 AliasResult QueryResult = query(LocA, LocB);
243 if (QueryResult == MayAlias)
244 return AliasAnalysis::alias(LocA, LocB);
249 bool doInitialization(Module &M) override;
252 void FunctionHandle::removeSelfFromCache() {
253 assert(CFLAA != nullptr);
254 auto *Val = getValPtr();
255 CFLAA->evict(cast<Function>(Val));
259 // \brief Gets the edges our graph should have, based on an Instruction*
260 class GetEdgesVisitor : public InstVisitor<GetEdgesVisitor, void> {
261 CFLAliasAnalysis &AA;
262 SmallVectorImpl<Edge> &Output;
265 GetEdgesVisitor(CFLAliasAnalysis &AA, SmallVectorImpl<Edge> &Output)
266 : AA(AA), Output(Output) {}
268 void visitInstruction(Instruction &) {
269 llvm_unreachable("Unsupported instruction encountered");
272 void visitPtrToIntInst(PtrToIntInst &Inst) {
273 auto *Ptr = Inst.getOperand(0);
274 Output.push_back(Edge(Ptr, Ptr, EdgeType::Assign, AttrUnknown));
277 void visitIntToPtrInst(IntToPtrInst &Inst) {
279 Output.push_back(Edge(Ptr, Ptr, EdgeType::Assign, AttrUnknown));
282 void visitCastInst(CastInst &Inst) {
284 Edge(&Inst, Inst.getOperand(0), EdgeType::Assign, AttrNone));
287 void visitBinaryOperator(BinaryOperator &Inst) {
288 auto *Op1 = Inst.getOperand(0);
289 auto *Op2 = Inst.getOperand(1);
290 Output.push_back(Edge(&Inst, Op1, EdgeType::Assign, AttrNone));
291 Output.push_back(Edge(&Inst, Op2, EdgeType::Assign, AttrNone));
294 void visitAtomicCmpXchgInst(AtomicCmpXchgInst &Inst) {
295 auto *Ptr = Inst.getPointerOperand();
296 auto *Val = Inst.getNewValOperand();
297 Output.push_back(Edge(Ptr, Val, EdgeType::Dereference, AttrNone));
300 void visitAtomicRMWInst(AtomicRMWInst &Inst) {
301 auto *Ptr = Inst.getPointerOperand();
302 auto *Val = Inst.getValOperand();
303 Output.push_back(Edge(Ptr, Val, EdgeType::Dereference, AttrNone));
306 void visitPHINode(PHINode &Inst) {
307 for (unsigned I = 0, E = Inst.getNumIncomingValues(); I != E; ++I) {
308 Value *Val = Inst.getIncomingValue(I);
309 Output.push_back(Edge(&Inst, Val, EdgeType::Assign, AttrNone));
313 void visitGetElementPtrInst(GetElementPtrInst &Inst) {
314 auto *Op = Inst.getPointerOperand();
315 Output.push_back(Edge(&Inst, Op, EdgeType::Assign, AttrNone));
316 for (auto I = Inst.idx_begin(), E = Inst.idx_end(); I != E; ++I)
317 Output.push_back(Edge(&Inst, *I, EdgeType::Assign, AttrNone));
320 void visitSelectInst(SelectInst &Inst) {
321 // Condition is not processed here (The actual statement producing
322 // the condition result is processed elsewhere). For select, the
323 // condition is evaluated, but not loaded, stored, or assigned
324 // simply as a result of being the condition of a select.
326 auto *TrueVal = Inst.getTrueValue();
327 Output.push_back(Edge(&Inst, TrueVal, EdgeType::Assign, AttrNone));
328 auto *FalseVal = Inst.getFalseValue();
329 Output.push_back(Edge(&Inst, FalseVal, EdgeType::Assign, AttrNone));
332 void visitAllocaInst(AllocaInst &) {}
334 void visitLoadInst(LoadInst &Inst) {
335 auto *Ptr = Inst.getPointerOperand();
337 Output.push_back(Edge(Val, Ptr, EdgeType::Reference, AttrNone));
340 void visitStoreInst(StoreInst &Inst) {
341 auto *Ptr = Inst.getPointerOperand();
342 auto *Val = Inst.getValueOperand();
343 Output.push_back(Edge(Ptr, Val, EdgeType::Dereference, AttrNone));
346 void visitVAArgInst(VAArgInst &Inst) {
347 // We can't fully model va_arg here. For *Ptr = Inst.getOperand(0), it does
349 // 1. Loads a value from *((T*)*Ptr).
350 // 2. Increments (stores to) *Ptr by some target-specific amount.
351 // For now, we'll handle this like a landingpad instruction (by placing the
352 // result in its own group, and having that group alias externals).
354 Output.push_back(Edge(Val, Val, EdgeType::Assign, AttrAll));
357 static bool isFunctionExternal(Function *Fn) {
358 return Fn->isDeclaration() || !Fn->hasLocalLinkage();
361 // Gets whether the sets at Index1 above, below, or equal to the sets at
362 // Index2. Returns None if they are not in the same set chain.
363 static Optional<Level> getIndexRelation(const StratifiedSets<Value *> &Sets,
364 StratifiedIndex Index1,
365 StratifiedIndex Index2) {
366 if (Index1 == Index2)
369 const auto *Current = &Sets.getLink(Index1);
370 while (Current->hasBelow()) {
371 if (Current->Below == Index2)
373 Current = &Sets.getLink(Current->Below);
376 Current = &Sets.getLink(Index1);
377 while (Current->hasAbove()) {
378 if (Current->Above == Index2)
380 Current = &Sets.getLink(Current->Above);
387 tryInterproceduralAnalysis(const SmallVectorImpl<Function *> &Fns,
389 const iterator_range<User::op_iterator> &Args) {
390 const unsigned ExpectedMaxArgs = 8;
391 const unsigned MaxSupportedArgs = 50;
392 assert(Fns.size() > 0);
394 // I put this here to give us an upper bound on time taken by IPA. Is it
395 // really (realistically) needed? Keep in mind that we do have an n^2 algo.
396 if (std::distance(Args.begin(), Args.end()) > (int)MaxSupportedArgs)
399 // Exit early if we'll fail anyway
400 for (auto *Fn : Fns) {
401 if (isFunctionExternal(Fn) || Fn->isVarArg())
403 auto &MaybeInfo = AA.ensureCached(Fn);
404 if (!MaybeInfo.hasValue())
408 SmallVector<Value *, ExpectedMaxArgs> Arguments(Args.begin(), Args.end());
409 SmallVector<StratifiedInfo, ExpectedMaxArgs> Parameters;
410 for (auto *Fn : Fns) {
411 auto &Info = *AA.ensureCached(Fn);
412 auto &Sets = Info.Sets;
413 auto &RetVals = Info.ReturnedValues;
416 for (auto &Param : Fn->args()) {
417 auto MaybeInfo = Sets.find(&Param);
418 // Did a new parameter somehow get added to the function/slip by?
419 if (!MaybeInfo.hasValue())
421 Parameters.push_back(*MaybeInfo);
424 // Adding an edge from argument -> return value for each parameter that
425 // may alias the return value
426 for (unsigned I = 0, E = Parameters.size(); I != E; ++I) {
427 auto &ParamInfo = Parameters[I];
428 auto &ArgVal = Arguments[I];
429 bool AddEdge = false;
430 StratifiedAttrs Externals;
431 for (unsigned X = 0, XE = RetVals.size(); X != XE; ++X) {
432 auto MaybeInfo = Sets.find(RetVals[X]);
433 if (!MaybeInfo.hasValue())
436 auto &RetInfo = *MaybeInfo;
437 auto RetAttrs = Sets.getLink(RetInfo.Index).Attrs;
438 auto ParamAttrs = Sets.getLink(ParamInfo.Index).Attrs;
440 getIndexRelation(Sets, ParamInfo.Index, RetInfo.Index);
441 if (MaybeRelation.hasValue()) {
443 Externals |= RetAttrs | ParamAttrs;
447 Output.push_back(Edge(FuncValue, ArgVal, EdgeType::Assign,
448 StratifiedAttrs().flip()));
451 if (Parameters.size() != Arguments.size())
454 // Adding edges between arguments for arguments that may end up aliasing
455 // each other. This is necessary for functions such as
456 // void foo(int** a, int** b) { *a = *b; }
457 // (Technically, the proper sets for this would be those below
458 // Arguments[I] and Arguments[X], but our algorithm will produce
459 // extremely similar, and equally correct, results either way)
460 for (unsigned I = 0, E = Arguments.size(); I != E; ++I) {
461 auto &MainVal = Arguments[I];
462 auto &MainInfo = Parameters[I];
463 auto &MainAttrs = Sets.getLink(MainInfo.Index).Attrs;
464 for (unsigned X = I + 1; X != E; ++X) {
465 auto &SubInfo = Parameters[X];
466 auto &SubVal = Arguments[X];
467 auto &SubAttrs = Sets.getLink(SubInfo.Index).Attrs;
469 getIndexRelation(Sets, MainInfo.Index, SubInfo.Index);
471 if (!MaybeRelation.hasValue())
474 auto NewAttrs = SubAttrs | MainAttrs;
475 Output.push_back(Edge(MainVal, SubVal, EdgeType::Assign, NewAttrs));
482 template <typename InstT> void visitCallLikeInst(InstT &Inst) {
483 SmallVector<Function *, 4> Targets;
484 if (getPossibleTargets(&Inst, Targets)) {
485 if (tryInterproceduralAnalysis(Targets, &Inst, Inst.arg_operands()))
487 // Cleanup from interprocedural analysis
491 for (Value *V : Inst.arg_operands())
492 Output.push_back(Edge(&Inst, V, EdgeType::Assign, AttrAll));
495 void visitCallInst(CallInst &Inst) { visitCallLikeInst(Inst); }
497 void visitInvokeInst(InvokeInst &Inst) { visitCallLikeInst(Inst); }
499 // Because vectors/aggregates are immutable and unaddressable,
500 // there's nothing we can do to coax a value out of them, other
501 // than calling Extract{Element,Value}. We can effectively treat
502 // them as pointers to arbitrary memory locations we can store in
504 void visitExtractElementInst(ExtractElementInst &Inst) {
505 auto *Ptr = Inst.getVectorOperand();
507 Output.push_back(Edge(Val, Ptr, EdgeType::Reference, AttrNone));
510 void visitInsertElementInst(InsertElementInst &Inst) {
511 auto *Vec = Inst.getOperand(0);
512 auto *Val = Inst.getOperand(1);
513 Output.push_back(Edge(&Inst, Vec, EdgeType::Assign, AttrNone));
514 Output.push_back(Edge(&Inst, Val, EdgeType::Dereference, AttrNone));
517 void visitLandingPadInst(LandingPadInst &Inst) {
518 // Exceptions come from "nowhere", from our analysis' perspective.
519 // So we place the instruction its own group, noting that said group may
521 Output.push_back(Edge(&Inst, &Inst, EdgeType::Assign, AttrAll));
524 void visitInsertValueInst(InsertValueInst &Inst) {
525 auto *Agg = Inst.getOperand(0);
526 auto *Val = Inst.getOperand(1);
527 Output.push_back(Edge(&Inst, Agg, EdgeType::Assign, AttrNone));
528 Output.push_back(Edge(&Inst, Val, EdgeType::Dereference, AttrNone));
531 void visitExtractValueInst(ExtractValueInst &Inst) {
532 auto *Ptr = Inst.getAggregateOperand();
533 Output.push_back(Edge(&Inst, Ptr, EdgeType::Reference, AttrNone));
536 void visitShuffleVectorInst(ShuffleVectorInst &Inst) {
537 auto *From1 = Inst.getOperand(0);
538 auto *From2 = Inst.getOperand(1);
539 Output.push_back(Edge(&Inst, From1, EdgeType::Assign, AttrNone));
540 Output.push_back(Edge(&Inst, From2, EdgeType::Assign, AttrNone));
544 // For a given instruction, we need to know which Value* to get the
545 // users of in order to build our graph. In some cases (i.e. add),
546 // we simply need the Instruction*. In other cases (i.e. store),
547 // finding the users of the Instruction* is useless; we need to find
548 // the users of the first operand. This handles determining which
549 // value to follow for us.
551 // Note: we *need* to keep this in sync with GetEdgesVisitor. Add
552 // something to GetEdgesVisitor, add it here -- remove something from
553 // GetEdgesVisitor, remove it here.
554 class GetTargetValueVisitor
555 : public InstVisitor<GetTargetValueVisitor, Value *> {
557 Value *visitInstruction(Instruction &Inst) { return &Inst; }
559 Value *visitStoreInst(StoreInst &Inst) { return Inst.getPointerOperand(); }
561 Value *visitAtomicCmpXchgInst(AtomicCmpXchgInst &Inst) {
562 return Inst.getPointerOperand();
565 Value *visitAtomicRMWInst(AtomicRMWInst &Inst) {
566 return Inst.getPointerOperand();
569 Value *visitInsertElementInst(InsertElementInst &Inst) {
570 return Inst.getOperand(0);
573 Value *visitInsertValueInst(InsertValueInst &Inst) {
574 return Inst.getAggregateOperand();
578 // Set building requires a weighted bidirectional graph.
579 template <typename EdgeTypeT> class WeightedBidirectionalGraph {
581 typedef std::size_t Node;
584 const static Node StartNode = Node(0);
590 Edge(const EdgeTypeT &W, const Node &N) : Weight(W), Other(N) {}
592 bool operator==(const Edge &E) const {
593 return Weight == E.Weight && Other == E.Other;
596 bool operator!=(const Edge &E) const { return !operator==(E); }
600 std::vector<Edge> Edges;
603 std::vector<NodeImpl> NodeImpls;
605 bool inbounds(Node NodeIndex) const { return NodeIndex < NodeImpls.size(); }
607 const NodeImpl &getNode(Node N) const { return NodeImpls[N]; }
608 NodeImpl &getNode(Node N) { return NodeImpls[N]; }
611 // ----- Various Edge iterators for the graph ----- //
613 // \brief Iterator for edges. Because this graph is bidirected, we don't
614 // allow modificaiton of the edges using this iterator. Additionally, the
615 // iterator becomes invalid if you add edges to or from the node you're
616 // getting the edges of.
617 struct EdgeIterator : public std::iterator<std::forward_iterator_tag,
618 std::tuple<EdgeTypeT, Node *>> {
619 EdgeIterator(const typename std::vector<Edge>::const_iterator &Iter)
622 EdgeIterator(NodeImpl &Impl) : Current(Impl.begin()) {}
624 EdgeIterator &operator++() {
629 EdgeIterator operator++(int) {
630 EdgeIterator Copy(Current);
635 std::tuple<EdgeTypeT, Node> &operator*() {
636 Store = std::make_tuple(Current->Weight, Current->Other);
640 bool operator==(const EdgeIterator &Other) const {
641 return Current == Other.Current;
644 bool operator!=(const EdgeIterator &Other) const {
645 return !operator==(Other);
649 typename std::vector<Edge>::const_iterator Current;
650 std::tuple<EdgeTypeT, Node> Store;
653 // Wrapper for EdgeIterator with begin()/end() calls.
654 struct EdgeIterable {
655 EdgeIterable(const std::vector<Edge> &Edges)
656 : BeginIter(Edges.begin()), EndIter(Edges.end()) {}
658 EdgeIterator begin() { return EdgeIterator(BeginIter); }
660 EdgeIterator end() { return EdgeIterator(EndIter); }
663 typename std::vector<Edge>::const_iterator BeginIter;
664 typename std::vector<Edge>::const_iterator EndIter;
667 // ----- Actual graph-related things ----- //
669 WeightedBidirectionalGraph() {}
671 WeightedBidirectionalGraph(WeightedBidirectionalGraph<EdgeTypeT> &&Other)
672 : NodeImpls(std::move(Other.NodeImpls)) {}
674 WeightedBidirectionalGraph<EdgeTypeT> &
675 operator=(WeightedBidirectionalGraph<EdgeTypeT> &&Other) {
676 NodeImpls = std::move(Other.NodeImpls);
681 auto Index = NodeImpls.size();
682 auto NewNode = Node(Index);
683 NodeImpls.push_back(NodeImpl());
687 void addEdge(Node From, Node To, const EdgeTypeT &Weight,
688 const EdgeTypeT &ReverseWeight) {
689 assert(inbounds(From));
690 assert(inbounds(To));
691 auto &FromNode = getNode(From);
692 auto &ToNode = getNode(To);
693 FromNode.Edges.push_back(Edge(Weight, To));
694 ToNode.Edges.push_back(Edge(ReverseWeight, From));
697 EdgeIterable edgesFor(const Node &N) const {
698 const auto &Node = getNode(N);
699 return EdgeIterable(Node.Edges);
702 bool empty() const { return NodeImpls.empty(); }
703 std::size_t size() const { return NodeImpls.size(); }
705 // \brief Gets an arbitrary node in the graph as a starting point for
707 Node getEntryNode() {
708 assert(inbounds(StartNode));
713 typedef WeightedBidirectionalGraph<std::pair<EdgeType, StratifiedAttrs>> GraphT;
714 typedef DenseMap<Value *, GraphT::Node> NodeMapT;
717 // -- Setting up/registering CFLAA pass -- //
718 char CFLAliasAnalysis::ID = 0;
720 INITIALIZE_AG_PASS(CFLAliasAnalysis, AliasAnalysis, "cfl-aa",
721 "CFL-Based AA implementation", false, true, false)
723 ImmutablePass *llvm::createCFLAliasAnalysisPass() {
724 return new CFLAliasAnalysis();
727 //===----------------------------------------------------------------------===//
728 // Function declarations that require types defined in the namespace above
729 //===----------------------------------------------------------------------===//
731 // Given an argument number, returns the appropriate Attr index to set.
732 static StratifiedAttr argNumberToAttrIndex(StratifiedAttr);
734 // Given a Value, potentially return which AttrIndex it maps to.
735 static Optional<StratifiedAttr> valueToAttrIndex(Value *Val);
737 // Gets the inverse of a given EdgeType.
738 static EdgeType flipWeight(EdgeType);
740 // Gets edges of the given Instruction*, writing them to the SmallVector*.
741 static void argsToEdges(CFLAliasAnalysis &, Instruction *,
742 SmallVectorImpl<Edge> &);
744 // Gets the "Level" that one should travel in StratifiedSets
745 // given an EdgeType.
746 static Level directionOfEdgeType(EdgeType);
748 // Builds the graph needed for constructing the StratifiedSets for the
750 static void buildGraphFrom(CFLAliasAnalysis &, Function *,
751 SmallVectorImpl<Value *> &, NodeMapT &, GraphT &);
753 // Gets the edges of a ConstantExpr as if it was an Instruction. This
754 // function also acts on any nested ConstantExprs, adding the edges
755 // of those to the given SmallVector as well.
756 static void constexprToEdges(CFLAliasAnalysis &, ConstantExpr &,
757 SmallVectorImpl<Edge> &);
759 // Given an Instruction, this will add it to the graph, along with any
760 // Instructions that are potentially only available from said Instruction
761 // For example, given the following line:
762 // %0 = load i16* getelementptr ([1 x i16]* @a, 0, 0), align 2
763 // addInstructionToGraph would add both the `load` and `getelementptr`
764 // instructions to the graph appropriately.
765 static void addInstructionToGraph(CFLAliasAnalysis &, Instruction &,
766 SmallVectorImpl<Value *> &, NodeMapT &,
769 // Notes whether it would be pointless to add the given Value to our sets.
770 static bool canSkipAddingToSets(Value *Val);
772 // Builds the graph + StratifiedSets for a function.
773 static FunctionInfo buildSetsFrom(CFLAliasAnalysis &, Function *);
775 static Optional<Function *> parentFunctionOfValue(Value *Val) {
776 if (auto *Inst = dyn_cast<Instruction>(Val)) {
777 auto *Bb = Inst->getParent();
778 return Bb->getParent();
781 if (auto *Arg = dyn_cast<Argument>(Val))
782 return Arg->getParent();
786 template <typename Inst>
787 static bool getPossibleTargets(Inst *Call,
788 SmallVectorImpl<Function *> &Output) {
789 if (auto *Fn = Call->getCalledFunction()) {
790 Output.push_back(Fn);
794 // TODO: If the call is indirect, we might be able to enumerate all potential
795 // targets of the call and return them, rather than just failing.
799 static Optional<Value *> getTargetValue(Instruction *Inst) {
800 GetTargetValueVisitor V;
801 return V.visit(Inst);
804 static bool hasUsefulEdges(Instruction *Inst) {
805 bool IsNonInvokeTerminator =
806 isa<TerminatorInst>(Inst) && !isa<InvokeInst>(Inst);
807 return !isa<CmpInst>(Inst) && !isa<FenceInst>(Inst) && !IsNonInvokeTerminator;
810 static Optional<StratifiedAttr> valueToAttrIndex(Value *Val) {
811 if (isa<GlobalValue>(Val))
812 return AttrGlobalIndex;
814 if (auto *Arg = dyn_cast<Argument>(Val))
815 // Only pointer arguments should have the argument attribute,
816 // because things can't escape through scalars without us seeing a
817 // cast, and thus, interaction with them doesn't matter.
818 if (!Arg->hasNoAliasAttr() && Arg->getType()->isPointerTy())
819 return argNumberToAttrIndex(Arg->getArgNo());
823 static StratifiedAttr argNumberToAttrIndex(unsigned ArgNum) {
824 if (ArgNum >= AttrMaxNumArgs)
826 return ArgNum + AttrFirstArgIndex;
829 static EdgeType flipWeight(EdgeType Initial) {
831 case EdgeType::Assign:
832 return EdgeType::Assign;
833 case EdgeType::Dereference:
834 return EdgeType::Reference;
835 case EdgeType::Reference:
836 return EdgeType::Dereference;
838 llvm_unreachable("Incomplete coverage of EdgeType enum");
841 static void argsToEdges(CFLAliasAnalysis &Analysis, Instruction *Inst,
842 SmallVectorImpl<Edge> &Output) {
843 assert(hasUsefulEdges(Inst) &&
844 "Expected instructions to have 'useful' edges");
845 GetEdgesVisitor v(Analysis, Output);
849 static Level directionOfEdgeType(EdgeType Weight) {
851 case EdgeType::Reference:
853 case EdgeType::Dereference:
855 case EdgeType::Assign:
858 llvm_unreachable("Incomplete switch coverage");
861 static void constexprToEdges(CFLAliasAnalysis &Analysis,
862 ConstantExpr &CExprToCollapse,
863 SmallVectorImpl<Edge> &Results) {
864 SmallVector<ConstantExpr *, 4> Worklist;
865 Worklist.push_back(&CExprToCollapse);
867 SmallVector<Edge, 8> ConstexprEdges;
868 while (!Worklist.empty()) {
869 auto *CExpr = Worklist.pop_back_val();
870 std::unique_ptr<Instruction> Inst(CExpr->getAsInstruction());
872 if (!hasUsefulEdges(Inst.get()))
875 ConstexprEdges.clear();
876 argsToEdges(Analysis, Inst.get(), ConstexprEdges);
877 for (auto &Edge : ConstexprEdges) {
878 if (Edge.From == Inst.get())
880 else if (auto *Nested = dyn_cast<ConstantExpr>(Edge.From))
881 Worklist.push_back(Nested);
883 if (Edge.To == Inst.get())
885 else if (auto *Nested = dyn_cast<ConstantExpr>(Edge.To))
886 Worklist.push_back(Nested);
889 Results.append(ConstexprEdges.begin(), ConstexprEdges.end());
893 static void addInstructionToGraph(CFLAliasAnalysis &Analysis, Instruction &Inst,
894 SmallVectorImpl<Value *> &ReturnedValues,
895 NodeMapT &Map, GraphT &Graph) {
896 const auto findOrInsertNode = [&Map, &Graph](Value *Val) {
897 auto Pair = Map.insert(std::make_pair(Val, GraphT::Node()));
898 auto &Iter = Pair.first;
900 auto NewNode = Graph.addNode();
901 Iter->second = NewNode;
906 // We don't want the edges of most "return" instructions, but we *do* want
907 // to know what can be returned.
908 if (isa<ReturnInst>(&Inst))
909 ReturnedValues.push_back(&Inst);
911 if (!hasUsefulEdges(&Inst))
914 SmallVector<Edge, 8> Edges;
915 argsToEdges(Analysis, &Inst, Edges);
917 // In the case of an unused alloca (or similar), edges may be empty. Note
918 // that it exists so we can potentially answer NoAlias.
920 auto MaybeVal = getTargetValue(&Inst);
921 assert(MaybeVal.hasValue());
922 auto *Target = *MaybeVal;
923 findOrInsertNode(Target);
927 const auto addEdgeToGraph = [&Graph, &findOrInsertNode](const Edge &E) {
928 auto To = findOrInsertNode(E.To);
929 auto From = findOrInsertNode(E.From);
930 auto FlippedWeight = flipWeight(E.Weight);
931 auto Attrs = E.AdditionalAttrs;
932 Graph.addEdge(From, To, std::make_pair(E.Weight, Attrs),
933 std::make_pair(FlippedWeight, Attrs));
936 SmallVector<ConstantExpr *, 4> ConstantExprs;
937 for (const Edge &E : Edges) {
939 if (auto *Constexpr = dyn_cast<ConstantExpr>(E.To))
940 ConstantExprs.push_back(Constexpr);
941 if (auto *Constexpr = dyn_cast<ConstantExpr>(E.From))
942 ConstantExprs.push_back(Constexpr);
945 for (ConstantExpr *CE : ConstantExprs) {
947 constexprToEdges(Analysis, *CE, Edges);
948 std::for_each(Edges.begin(), Edges.end(), addEdgeToGraph);
952 // Aside: We may remove graph construction entirely, because it doesn't really
953 // buy us much that we don't already have. I'd like to add interprocedural
954 // analysis prior to this however, in case that somehow requires the graph
955 // produced by this for efficient execution
956 static void buildGraphFrom(CFLAliasAnalysis &Analysis, Function *Fn,
957 SmallVectorImpl<Value *> &ReturnedValues,
958 NodeMapT &Map, GraphT &Graph) {
959 for (auto &Bb : Fn->getBasicBlockList())
960 for (auto &Inst : Bb.getInstList())
961 addInstructionToGraph(Analysis, Inst, ReturnedValues, Map, Graph);
964 static bool canSkipAddingToSets(Value *Val) {
965 // Constants can share instances, which may falsely unify multiple
967 // store i32* null, i32** %ptr1
968 // store i32* null, i32** %ptr2
969 // clearly ptr1 and ptr2 should not be unified into the same set, so
970 // we should filter out the (potentially shared) instance to
972 if (isa<Constant>(Val)) {
973 bool Container = isa<ConstantVector>(Val) || isa<ConstantArray>(Val) ||
974 isa<ConstantStruct>(Val);
975 // TODO: Because all of these things are constant, we can determine whether
976 // the data is *actually* mutable at graph building time. This will probably
977 // come for free/cheap with offset awareness.
978 bool CanStoreMutableData =
979 isa<GlobalValue>(Val) || isa<ConstantExpr>(Val) || Container;
980 return !CanStoreMutableData;
986 static FunctionInfo buildSetsFrom(CFLAliasAnalysis &Analysis, Function *Fn) {
989 SmallVector<Value *, 4> ReturnedValues;
991 buildGraphFrom(Analysis, Fn, ReturnedValues, Map, Graph);
993 DenseMap<GraphT::Node, Value *> NodeValueMap;
994 NodeValueMap.resize(Map.size());
995 for (const auto &Pair : Map)
996 NodeValueMap.insert(std::make_pair(Pair.second, Pair.first));
998 const auto findValueOrDie = [&NodeValueMap](GraphT::Node Node) {
999 auto ValIter = NodeValueMap.find(Node);
1000 assert(ValIter != NodeValueMap.end());
1001 return ValIter->second;
1004 StratifiedSetsBuilder<Value *> Builder;
1006 SmallVector<GraphT::Node, 16> Worklist;
1007 for (auto &Pair : Map) {
1010 auto *Value = Pair.first;
1012 auto InitialNode = Pair.second;
1013 Worklist.push_back(InitialNode);
1014 while (!Worklist.empty()) {
1015 auto Node = Worklist.pop_back_val();
1016 auto *CurValue = findValueOrDie(Node);
1017 if (canSkipAddingToSets(CurValue))
1020 for (const auto &EdgeTuple : Graph.edgesFor(Node)) {
1021 auto Weight = std::get<0>(EdgeTuple);
1022 auto Label = Weight.first;
1023 auto &OtherNode = std::get<1>(EdgeTuple);
1024 auto *OtherValue = findValueOrDie(OtherNode);
1026 if (canSkipAddingToSets(OtherValue))
1030 switch (directionOfEdgeType(Label)) {
1032 Added = Builder.addAbove(CurValue, OtherValue);
1035 Added = Builder.addBelow(CurValue, OtherValue);
1038 Added = Builder.addWith(CurValue, OtherValue);
1042 auto Aliasing = Weight.second;
1043 if (auto MaybeCurIndex = valueToAttrIndex(CurValue))
1044 Aliasing.set(*MaybeCurIndex);
1045 if (auto MaybeOtherIndex = valueToAttrIndex(OtherValue))
1046 Aliasing.set(*MaybeOtherIndex);
1047 Builder.noteAttributes(CurValue, Aliasing);
1048 Builder.noteAttributes(OtherValue, Aliasing);
1051 Worklist.push_back(OtherNode);
1056 // There are times when we end up with parameters not in our graph (i.e. if
1057 // it's only used as the condition of a branch). Other bits of code depend on
1058 // things that were present during construction being present in the graph.
1059 // So, we add all present arguments here.
1060 for (auto &Arg : Fn->args()) {
1061 if (!Builder.add(&Arg))
1064 auto Attrs = valueToAttrIndex(&Arg);
1065 if (Attrs.hasValue())
1066 Builder.noteAttributes(&Arg, *Attrs);
1069 return FunctionInfo(Builder.build(), std::move(ReturnedValues));
1072 void CFLAliasAnalysis::scan(Function *Fn) {
1073 auto InsertPair = Cache.insert(std::make_pair(Fn, Optional<FunctionInfo>()));
1075 assert(InsertPair.second &&
1076 "Trying to scan a function that has already been cached");
1078 FunctionInfo Info(buildSetsFrom(*this, Fn));
1079 Cache[Fn] = std::move(Info);
1080 Handles.push_front(FunctionHandle(Fn, this));
1083 AliasAnalysis::AliasResult
1084 CFLAliasAnalysis::query(const AliasAnalysis::Location &LocA,
1085 const AliasAnalysis::Location &LocB) {
1086 auto *ValA = const_cast<Value *>(LocA.Ptr);
1087 auto *ValB = const_cast<Value *>(LocB.Ptr);
1089 Function *Fn = nullptr;
1090 auto MaybeFnA = parentFunctionOfValue(ValA);
1091 auto MaybeFnB = parentFunctionOfValue(ValB);
1092 if (!MaybeFnA.hasValue() && !MaybeFnB.hasValue()) {
1093 // The only times this is known to happen are when globals + InlineAsm
1095 DEBUG(dbgs() << "CFLAA: could not extract parent function information.\n");
1096 return AliasAnalysis::MayAlias;
1099 if (MaybeFnA.hasValue()) {
1101 assert((!MaybeFnB.hasValue() || *MaybeFnB == *MaybeFnA) &&
1102 "Interprocedural queries not supported");
1107 assert(Fn != nullptr);
1108 auto &MaybeInfo = ensureCached(Fn);
1109 assert(MaybeInfo.hasValue());
1111 auto &Sets = MaybeInfo->Sets;
1112 auto MaybeA = Sets.find(ValA);
1113 if (!MaybeA.hasValue())
1114 return AliasAnalysis::MayAlias;
1116 auto MaybeB = Sets.find(ValB);
1117 if (!MaybeB.hasValue())
1118 return AliasAnalysis::MayAlias;
1120 auto SetA = *MaybeA;
1121 auto SetB = *MaybeB;
1122 auto AttrsA = Sets.getLink(SetA.Index).Attrs;
1123 auto AttrsB = Sets.getLink(SetB.Index).Attrs;
1125 // Stratified set attributes are used as markets to signify whether a member
1126 // of a StratifiedSet (or a member of a set above the current set) has
1127 // interacted with either arguments or globals. "Interacted with" meaning
1128 // its value may be different depending on the value of an argument or
1129 // global. The thought behind this is that, because arguments and globals
1130 // may alias each other, if AttrsA and AttrsB have touched args/globals,
1131 // we must conservatively say that they alias. However, if at least one of
1132 // the sets has no values that could legally be altered by changing the value
1133 // of an argument or global, then we don't have to be as conservative.
1134 if (AttrsA.any() && AttrsB.any())
1135 return AliasAnalysis::MayAlias;
1137 // We currently unify things even if the accesses to them may not be in
1138 // bounds, so we can't return partial alias here because we don't
1139 // know whether the pointer is really within the object or not.
1140 // IE Given an out of bounds GEP and an alloca'd pointer, we may
1141 // unify the two. We can't return partial alias for this case.
1142 // Since we do not currently track enough information to
1145 if (SetA.Index == SetB.Index)
1146 return AliasAnalysis::MayAlias;
1148 return AliasAnalysis::NoAlias;
1151 bool CFLAliasAnalysis::doInitialization(Module &M) {
1152 InitializeAliasAnalysis(this, &M.getDataLayout());