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"
48 #include "llvm/Support/raw_ostream.h"
51 #include <forward_list>
57 #define DEBUG_TYPE "cfl-aa"
59 // Try to go from a Value* to a Function*. Never returns nullptr.
60 static Optional<Function *> parentFunctionOfValue(Value *);
62 // Returns possible functions called by the Inst* into the given
63 // SmallVectorImpl. Returns true if targets found, false otherwise.
64 // This is templated because InvokeInst/CallInst give us the same
65 // set of functions that we care about, and I don't like repeating
67 template <typename Inst>
68 static bool getPossibleTargets(Inst *, SmallVectorImpl<Function *> &);
70 // Some instructions need to have their users tracked. Instructions like
71 // `add` require you to get the users of the Instruction* itself, other
72 // instructions like `store` require you to get the users of the first
73 // operand. This function gets the "proper" value to track for each
74 // type of instruction we support.
75 static Optional<Value *> getTargetValue(Instruction *);
77 // There are certain instructions (i.e. FenceInst, etc.) that we ignore.
78 // This notes that we should ignore those.
79 static bool hasUsefulEdges(Instruction *);
81 const StratifiedIndex StratifiedLink::SetSentinel =
82 std::numeric_limits<StratifiedIndex>::max();
85 // StratifiedInfo Attribute things.
86 typedef unsigned StratifiedAttr;
87 LLVM_CONSTEXPR unsigned MaxStratifiedAttrIndex = NumStratifiedAttrs;
88 LLVM_CONSTEXPR unsigned AttrAllIndex = 0;
89 LLVM_CONSTEXPR unsigned AttrGlobalIndex = 1;
90 LLVM_CONSTEXPR unsigned AttrUnknownIndex = 2;
91 LLVM_CONSTEXPR unsigned AttrFirstArgIndex = 3;
92 LLVM_CONSTEXPR unsigned AttrLastArgIndex = MaxStratifiedAttrIndex;
93 LLVM_CONSTEXPR unsigned AttrMaxNumArgs = AttrLastArgIndex - AttrFirstArgIndex;
95 LLVM_CONSTEXPR StratifiedAttr AttrNone = 0;
96 LLVM_CONSTEXPR StratifiedAttr AttrUnknown = 1 << AttrUnknownIndex;
97 LLVM_CONSTEXPR StratifiedAttr AttrAll = ~AttrNone;
99 // \brief StratifiedSets call for knowledge of "direction", so this is how we
100 // represent that locally.
101 enum class Level { Same, Above, Below };
103 // \brief Edges can be one of four "weights" -- each weight must have an inverse
104 // weight (Assign has Assign; Reference has Dereference).
105 enum class EdgeType {
106 // The weight assigned when assigning from or to a value. For example, in:
107 // %b = getelementptr %a, 0
108 // ...The relationships are %b assign %a, and %a assign %b. This used to be
109 // two edges, but having a distinction bought us nothing.
112 // The edge used when we have an edge going from some handle to a Value.
113 // Examples of this include:
114 // %b = load %a (%b Dereference %a)
115 // %b = extractelement %a, 0 (%a Dereference %b)
118 // The edge used when our edge goes from a value to a handle that may have
119 // contained it at some point. Examples:
120 // %b = load %a (%a Reference %b)
121 // %b = extractelement %a, 0 (%b Reference %a)
125 // \brief Encodes the notion of a "use"
127 // \brief Which value the edge is coming from
130 // \brief Which value the edge is pointing to
133 // \brief Edge weight
136 // \brief Whether we aliased any external values along the way that may be
137 // invisible to the analysis (i.e. landingpad for exceptions, calls for
138 // interprocedural analysis, etc.)
139 StratifiedAttrs AdditionalAttrs;
141 Edge(Value *From, Value *To, EdgeType W, StratifiedAttrs A)
142 : From(From), To(To), Weight(W), AdditionalAttrs(A) {}
145 // \brief Information we have about a function and would like to keep around
146 struct FunctionInfo {
147 StratifiedSets<Value *> Sets;
148 // Lots of functions have < 4 returns. Adjust as necessary.
149 SmallVector<Value *, 4> ReturnedValues;
151 FunctionInfo(StratifiedSets<Value *> &&S, SmallVector<Value *, 4> &&RV)
152 : Sets(std::move(S)), ReturnedValues(std::move(RV)) {}
155 struct CFLAliasAnalysis;
157 struct FunctionHandle : public CallbackVH {
158 FunctionHandle(Function *Fn, CFLAliasAnalysis *CFLAA)
159 : CallbackVH(Fn), CFLAA(CFLAA) {
160 assert(Fn != nullptr);
161 assert(CFLAA != nullptr);
164 ~FunctionHandle() override {}
166 void deleted() override { removeSelfFromCache(); }
167 void allUsesReplacedWith(Value *) override { removeSelfFromCache(); }
170 CFLAliasAnalysis *CFLAA;
172 void removeSelfFromCache();
175 struct CFLAliasAnalysis : public ImmutablePass, public AliasAnalysis {
177 /// \brief Cached mapping of Functions to their StratifiedSets.
178 /// If a function's sets are currently being built, it is marked
179 /// in the cache as an Optional without a value. This way, if we
180 /// have any kind of recursion, it is discernable from a function
181 /// that simply has empty sets.
182 DenseMap<Function *, Optional<FunctionInfo>> Cache;
183 std::forward_list<FunctionHandle> Handles;
188 CFLAliasAnalysis() : ImmutablePass(ID) {
189 initializeCFLAliasAnalysisPass(*PassRegistry::getPassRegistry());
192 ~CFLAliasAnalysis() override {}
194 void getAnalysisUsage(AnalysisUsage &AU) const override {
195 AliasAnalysis::getAnalysisUsage(AU);
198 void *getAdjustedAnalysisPointer(const void *ID) override {
199 if (ID == &AliasAnalysis::ID)
200 return (AliasAnalysis *)this;
204 /// \brief Inserts the given Function into the cache.
205 void scan(Function *Fn);
207 void evict(Function *Fn) { Cache.erase(Fn); }
209 /// \brief Ensures that the given function is available in the cache.
210 /// Returns the appropriate entry from the cache.
211 const Optional<FunctionInfo> &ensureCached(Function *Fn) {
212 auto Iter = Cache.find(Fn);
213 if (Iter == Cache.end()) {
215 Iter = Cache.find(Fn);
216 assert(Iter != Cache.end());
217 assert(Iter->second.hasValue());
222 AliasResult query(const Location &LocA, const Location &LocB);
224 AliasResult alias(const Location &LocA, const Location &LocB) override {
225 if (LocA.Ptr == LocB.Ptr) {
226 if (LocA.Size == LocB.Size) {
233 // Comparisons between global variables and other constants should be
234 // handled by BasicAA.
235 // TODO: ConstantExpr handling -- CFLAA may report NoAlias when comparing
236 // a GlobalValue and ConstantExpr, but every query needs to have at least
237 // one Value tied to a Function, and neither GlobalValues nor ConstantExprs
239 if (isa<Constant>(LocA.Ptr) && isa<Constant>(LocB.Ptr)) {
240 return AliasAnalysis::alias(LocA, LocB);
243 AliasResult QueryResult = query(LocA, LocB);
244 if (QueryResult == MayAlias)
245 return AliasAnalysis::alias(LocA, LocB);
250 bool doInitialization(Module &M) override;
253 void FunctionHandle::removeSelfFromCache() {
254 assert(CFLAA != nullptr);
255 auto *Val = getValPtr();
256 CFLAA->evict(cast<Function>(Val));
260 // \brief Gets the edges our graph should have, based on an Instruction*
261 class GetEdgesVisitor : public InstVisitor<GetEdgesVisitor, void> {
262 CFLAliasAnalysis &AA;
263 SmallVectorImpl<Edge> &Output;
266 GetEdgesVisitor(CFLAliasAnalysis &AA, SmallVectorImpl<Edge> &Output)
267 : AA(AA), Output(Output) {}
269 void visitInstruction(Instruction &) {
270 llvm_unreachable("Unsupported instruction encountered");
273 void visitPtrToIntInst(PtrToIntInst &Inst) {
274 auto *Ptr = Inst.getOperand(0);
275 Output.push_back(Edge(Ptr, Ptr, EdgeType::Assign, AttrUnknown));
278 void visitIntToPtrInst(IntToPtrInst &Inst) {
280 Output.push_back(Edge(Ptr, Ptr, EdgeType::Assign, AttrUnknown));
283 void visitCastInst(CastInst &Inst) {
285 Edge(&Inst, Inst.getOperand(0), EdgeType::Assign, AttrNone));
288 void visitBinaryOperator(BinaryOperator &Inst) {
289 auto *Op1 = Inst.getOperand(0);
290 auto *Op2 = Inst.getOperand(1);
291 Output.push_back(Edge(&Inst, Op1, EdgeType::Assign, AttrNone));
292 Output.push_back(Edge(&Inst, Op2, EdgeType::Assign, AttrNone));
295 void visitAtomicCmpXchgInst(AtomicCmpXchgInst &Inst) {
296 auto *Ptr = Inst.getPointerOperand();
297 auto *Val = Inst.getNewValOperand();
298 Output.push_back(Edge(Ptr, Val, EdgeType::Dereference, AttrNone));
301 void visitAtomicRMWInst(AtomicRMWInst &Inst) {
302 auto *Ptr = Inst.getPointerOperand();
303 auto *Val = Inst.getValOperand();
304 Output.push_back(Edge(Ptr, Val, EdgeType::Dereference, AttrNone));
307 void visitPHINode(PHINode &Inst) {
308 for (Value *Val : Inst.incoming_values()) {
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));
543 void visitConstantExpr(ConstantExpr *CE) {
544 switch (CE->getOpcode()) {
546 llvm_unreachable("Unknown instruction type encountered!");
547 // Build the switch statement using the Instruction.def file.
548 #define HANDLE_INST(NUM, OPCODE, CLASS) \
549 case Instruction::OPCODE: \
550 visit##OPCODE(*(CLASS *)CE); \
552 #include "llvm/IR/Instruction.def"
557 // For a given instruction, we need to know which Value* to get the
558 // users of in order to build our graph. In some cases (i.e. add),
559 // we simply need the Instruction*. In other cases (i.e. store),
560 // finding the users of the Instruction* is useless; we need to find
561 // the users of the first operand. This handles determining which
562 // value to follow for us.
564 // Note: we *need* to keep this in sync with GetEdgesVisitor. Add
565 // something to GetEdgesVisitor, add it here -- remove something from
566 // GetEdgesVisitor, remove it here.
567 class GetTargetValueVisitor
568 : public InstVisitor<GetTargetValueVisitor, Value *> {
570 Value *visitInstruction(Instruction &Inst) { return &Inst; }
572 Value *visitStoreInst(StoreInst &Inst) { return Inst.getPointerOperand(); }
574 Value *visitAtomicCmpXchgInst(AtomicCmpXchgInst &Inst) {
575 return Inst.getPointerOperand();
578 Value *visitAtomicRMWInst(AtomicRMWInst &Inst) {
579 return Inst.getPointerOperand();
582 Value *visitInsertElementInst(InsertElementInst &Inst) {
583 return Inst.getOperand(0);
586 Value *visitInsertValueInst(InsertValueInst &Inst) {
587 return Inst.getAggregateOperand();
591 // Set building requires a weighted bidirectional graph.
592 template <typename EdgeTypeT> class WeightedBidirectionalGraph {
594 typedef std::size_t Node;
597 const static Node StartNode = Node(0);
603 Edge(const EdgeTypeT &W, const Node &N) : Weight(W), Other(N) {}
605 bool operator==(const Edge &E) const {
606 return Weight == E.Weight && Other == E.Other;
609 bool operator!=(const Edge &E) const { return !operator==(E); }
613 std::vector<Edge> Edges;
616 std::vector<NodeImpl> NodeImpls;
618 bool inbounds(Node NodeIndex) const { return NodeIndex < NodeImpls.size(); }
620 const NodeImpl &getNode(Node N) const { return NodeImpls[N]; }
621 NodeImpl &getNode(Node N) { return NodeImpls[N]; }
624 // ----- Various Edge iterators for the graph ----- //
626 // \brief Iterator for edges. Because this graph is bidirected, we don't
627 // allow modificaiton of the edges using this iterator. Additionally, the
628 // iterator becomes invalid if you add edges to or from the node you're
629 // getting the edges of.
630 struct EdgeIterator : public std::iterator<std::forward_iterator_tag,
631 std::tuple<EdgeTypeT, Node *>> {
632 EdgeIterator(const typename std::vector<Edge>::const_iterator &Iter)
635 EdgeIterator(NodeImpl &Impl) : Current(Impl.begin()) {}
637 EdgeIterator &operator++() {
642 EdgeIterator operator++(int) {
643 EdgeIterator Copy(Current);
648 std::tuple<EdgeTypeT, Node> &operator*() {
649 Store = std::make_tuple(Current->Weight, Current->Other);
653 bool operator==(const EdgeIterator &Other) const {
654 return Current == Other.Current;
657 bool operator!=(const EdgeIterator &Other) const {
658 return !operator==(Other);
662 typename std::vector<Edge>::const_iterator Current;
663 std::tuple<EdgeTypeT, Node> Store;
666 // Wrapper for EdgeIterator with begin()/end() calls.
667 struct EdgeIterable {
668 EdgeIterable(const std::vector<Edge> &Edges)
669 : BeginIter(Edges.begin()), EndIter(Edges.end()) {}
671 EdgeIterator begin() { return EdgeIterator(BeginIter); }
673 EdgeIterator end() { return EdgeIterator(EndIter); }
676 typename std::vector<Edge>::const_iterator BeginIter;
677 typename std::vector<Edge>::const_iterator EndIter;
680 // ----- Actual graph-related things ----- //
682 WeightedBidirectionalGraph() {}
684 WeightedBidirectionalGraph(WeightedBidirectionalGraph<EdgeTypeT> &&Other)
685 : NodeImpls(std::move(Other.NodeImpls)) {}
687 WeightedBidirectionalGraph<EdgeTypeT> &
688 operator=(WeightedBidirectionalGraph<EdgeTypeT> &&Other) {
689 NodeImpls = std::move(Other.NodeImpls);
694 auto Index = NodeImpls.size();
695 auto NewNode = Node(Index);
696 NodeImpls.push_back(NodeImpl());
700 void addEdge(Node From, Node To, const EdgeTypeT &Weight,
701 const EdgeTypeT &ReverseWeight) {
702 assert(inbounds(From));
703 assert(inbounds(To));
704 auto &FromNode = getNode(From);
705 auto &ToNode = getNode(To);
706 FromNode.Edges.push_back(Edge(Weight, To));
707 ToNode.Edges.push_back(Edge(ReverseWeight, From));
710 EdgeIterable edgesFor(const Node &N) const {
711 const auto &Node = getNode(N);
712 return EdgeIterable(Node.Edges);
715 bool empty() const { return NodeImpls.empty(); }
716 std::size_t size() const { return NodeImpls.size(); }
718 // \brief Gets an arbitrary node in the graph as a starting point for
720 Node getEntryNode() {
721 assert(inbounds(StartNode));
726 typedef WeightedBidirectionalGraph<std::pair<EdgeType, StratifiedAttrs>> GraphT;
727 typedef DenseMap<Value *, GraphT::Node> NodeMapT;
730 // -- Setting up/registering CFLAA pass -- //
731 char CFLAliasAnalysis::ID = 0;
733 INITIALIZE_AG_PASS(CFLAliasAnalysis, AliasAnalysis, "cfl-aa",
734 "CFL-Based AA implementation", false, true, false)
736 ImmutablePass *llvm::createCFLAliasAnalysisPass() {
737 return new CFLAliasAnalysis();
740 //===----------------------------------------------------------------------===//
741 // Function declarations that require types defined in the namespace above
742 //===----------------------------------------------------------------------===//
744 // Given an argument number, returns the appropriate Attr index to set.
745 static StratifiedAttr argNumberToAttrIndex(StratifiedAttr);
747 // Given a Value, potentially return which AttrIndex it maps to.
748 static Optional<StratifiedAttr> valueToAttrIndex(Value *Val);
750 // Gets the inverse of a given EdgeType.
751 static EdgeType flipWeight(EdgeType);
753 // Gets edges of the given Instruction*, writing them to the SmallVector*.
754 static void argsToEdges(CFLAliasAnalysis &, Instruction *,
755 SmallVectorImpl<Edge> &);
757 // Gets edges of the given ConstantExpr*, writing them to the SmallVector*.
758 static void argsToEdges(CFLAliasAnalysis &, ConstantExpr *,
759 SmallVectorImpl<Edge> &);
761 // Gets the "Level" that one should travel in StratifiedSets
762 // given an EdgeType.
763 static Level directionOfEdgeType(EdgeType);
765 // Builds the graph needed for constructing the StratifiedSets for the
767 static void buildGraphFrom(CFLAliasAnalysis &, Function *,
768 SmallVectorImpl<Value *> &, NodeMapT &, GraphT &);
770 // Gets the edges of a ConstantExpr as if it was an Instruction. This
771 // function also acts on any nested ConstantExprs, adding the edges
772 // of those to the given SmallVector as well.
773 static void constexprToEdges(CFLAliasAnalysis &, ConstantExpr &,
774 SmallVectorImpl<Edge> &);
776 // Given an Instruction, this will add it to the graph, along with any
777 // Instructions that are potentially only available from said Instruction
778 // For example, given the following line:
779 // %0 = load i16* getelementptr ([1 x i16]* @a, 0, 0), align 2
780 // addInstructionToGraph would add both the `load` and `getelementptr`
781 // instructions to the graph appropriately.
782 static void addInstructionToGraph(CFLAliasAnalysis &, Instruction &,
783 SmallVectorImpl<Value *> &, NodeMapT &,
786 // Notes whether it would be pointless to add the given Value to our sets.
787 static bool canSkipAddingToSets(Value *Val);
789 // Builds the graph + StratifiedSets for a function.
790 static FunctionInfo buildSetsFrom(CFLAliasAnalysis &, Function *);
792 static Optional<Function *> parentFunctionOfValue(Value *Val) {
793 if (auto *Inst = dyn_cast<Instruction>(Val)) {
794 auto *Bb = Inst->getParent();
795 return Bb->getParent();
798 if (auto *Arg = dyn_cast<Argument>(Val))
799 return Arg->getParent();
803 template <typename Inst>
804 static bool getPossibleTargets(Inst *Call,
805 SmallVectorImpl<Function *> &Output) {
806 if (auto *Fn = Call->getCalledFunction()) {
807 Output.push_back(Fn);
811 // TODO: If the call is indirect, we might be able to enumerate all potential
812 // targets of the call and return them, rather than just failing.
816 static Optional<Value *> getTargetValue(Instruction *Inst) {
817 GetTargetValueVisitor V;
818 return V.visit(Inst);
821 static bool hasUsefulEdges(Instruction *Inst) {
822 bool IsNonInvokeTerminator =
823 isa<TerminatorInst>(Inst) && !isa<InvokeInst>(Inst);
824 return !isa<CmpInst>(Inst) && !isa<FenceInst>(Inst) && !IsNonInvokeTerminator;
827 static bool hasUsefulEdges(ConstantExpr *CE) {
828 // ConstantExpr doens't have terminators, invokes, or fences, so only needs
829 // to check for compares.
830 return CE->getOpcode() != Instruction::ICmp &&
831 CE->getOpcode() != Instruction::FCmp;
834 static Optional<StratifiedAttr> valueToAttrIndex(Value *Val) {
835 if (isa<GlobalValue>(Val))
836 return AttrGlobalIndex;
838 if (auto *Arg = dyn_cast<Argument>(Val))
839 // Only pointer arguments should have the argument attribute,
840 // because things can't escape through scalars without us seeing a
841 // cast, and thus, interaction with them doesn't matter.
842 if (!Arg->hasNoAliasAttr() && Arg->getType()->isPointerTy())
843 return argNumberToAttrIndex(Arg->getArgNo());
847 static StratifiedAttr argNumberToAttrIndex(unsigned ArgNum) {
848 if (ArgNum >= AttrMaxNumArgs)
850 return ArgNum + AttrFirstArgIndex;
853 static EdgeType flipWeight(EdgeType Initial) {
855 case EdgeType::Assign:
856 return EdgeType::Assign;
857 case EdgeType::Dereference:
858 return EdgeType::Reference;
859 case EdgeType::Reference:
860 return EdgeType::Dereference;
862 llvm_unreachable("Incomplete coverage of EdgeType enum");
865 static void argsToEdges(CFLAliasAnalysis &Analysis, Instruction *Inst,
866 SmallVectorImpl<Edge> &Output) {
867 assert(hasUsefulEdges(Inst) &&
868 "Expected instructions to have 'useful' edges");
869 GetEdgesVisitor v(Analysis, Output);
873 static void argsToEdges(CFLAliasAnalysis &Analysis, ConstantExpr *CE,
874 SmallVectorImpl<Edge> &Output) {
875 assert(hasUsefulEdges(CE) && "Expected constant expr to have 'useful' edges");
876 GetEdgesVisitor v(Analysis, Output);
877 v.visitConstantExpr(CE);
880 static Level directionOfEdgeType(EdgeType Weight) {
882 case EdgeType::Reference:
884 case EdgeType::Dereference:
886 case EdgeType::Assign:
889 llvm_unreachable("Incomplete switch coverage");
892 static void constexprToEdges(CFLAliasAnalysis &Analysis,
893 ConstantExpr &CExprToCollapse,
894 SmallVectorImpl<Edge> &Results) {
895 SmallVector<ConstantExpr *, 4> Worklist;
896 Worklist.push_back(&CExprToCollapse);
898 SmallVector<Edge, 8> ConstexprEdges;
899 SmallPtrSet<ConstantExpr *, 4> Visited;
900 while (!Worklist.empty()) {
901 auto *CExpr = Worklist.pop_back_val();
903 if (!hasUsefulEdges(CExpr))
906 ConstexprEdges.clear();
907 argsToEdges(Analysis, CExpr, ConstexprEdges);
908 for (auto &Edge : ConstexprEdges) {
909 if (auto *Nested = dyn_cast<ConstantExpr>(Edge.From))
910 if (Visited.insert(Nested).second)
911 Worklist.push_back(Nested);
913 if (auto *Nested = dyn_cast<ConstantExpr>(Edge.To))
914 if (Visited.insert(Nested).second)
915 Worklist.push_back(Nested);
918 Results.append(ConstexprEdges.begin(), ConstexprEdges.end());
922 static void addInstructionToGraph(CFLAliasAnalysis &Analysis, Instruction &Inst,
923 SmallVectorImpl<Value *> &ReturnedValues,
924 NodeMapT &Map, GraphT &Graph) {
925 const auto findOrInsertNode = [&Map, &Graph](Value *Val) {
926 auto Pair = Map.insert(std::make_pair(Val, GraphT::Node()));
927 auto &Iter = Pair.first;
929 auto NewNode = Graph.addNode();
930 Iter->second = NewNode;
935 // We don't want the edges of most "return" instructions, but we *do* want
936 // to know what can be returned.
937 if (isa<ReturnInst>(&Inst))
938 ReturnedValues.push_back(&Inst);
940 if (!hasUsefulEdges(&Inst))
943 SmallVector<Edge, 8> Edges;
944 argsToEdges(Analysis, &Inst, Edges);
946 // In the case of an unused alloca (or similar), edges may be empty. Note
947 // that it exists so we can potentially answer NoAlias.
949 auto MaybeVal = getTargetValue(&Inst);
950 assert(MaybeVal.hasValue());
951 auto *Target = *MaybeVal;
952 findOrInsertNode(Target);
956 const auto addEdgeToGraph = [&Graph, &findOrInsertNode](const Edge &E) {
957 auto To = findOrInsertNode(E.To);
958 auto From = findOrInsertNode(E.From);
959 auto FlippedWeight = flipWeight(E.Weight);
960 auto Attrs = E.AdditionalAttrs;
961 Graph.addEdge(From, To, std::make_pair(E.Weight, Attrs),
962 std::make_pair(FlippedWeight, Attrs));
965 SmallVector<ConstantExpr *, 4> ConstantExprs;
966 for (const Edge &E : Edges) {
968 if (auto *Constexpr = dyn_cast<ConstantExpr>(E.To))
969 ConstantExprs.push_back(Constexpr);
970 if (auto *Constexpr = dyn_cast<ConstantExpr>(E.From))
971 ConstantExprs.push_back(Constexpr);
974 for (ConstantExpr *CE : ConstantExprs) {
976 constexprToEdges(Analysis, *CE, Edges);
977 std::for_each(Edges.begin(), Edges.end(), addEdgeToGraph);
981 // Aside: We may remove graph construction entirely, because it doesn't really
982 // buy us much that we don't already have. I'd like to add interprocedural
983 // analysis prior to this however, in case that somehow requires the graph
984 // produced by this for efficient execution
985 static void buildGraphFrom(CFLAliasAnalysis &Analysis, Function *Fn,
986 SmallVectorImpl<Value *> &ReturnedValues,
987 NodeMapT &Map, GraphT &Graph) {
988 for (auto &Bb : Fn->getBasicBlockList())
989 for (auto &Inst : Bb.getInstList())
990 addInstructionToGraph(Analysis, Inst, ReturnedValues, Map, Graph);
993 static bool canSkipAddingToSets(Value *Val) {
994 // Constants can share instances, which may falsely unify multiple
996 // store i32* null, i32** %ptr1
997 // store i32* null, i32** %ptr2
998 // clearly ptr1 and ptr2 should not be unified into the same set, so
999 // we should filter out the (potentially shared) instance to
1001 if (isa<Constant>(Val)) {
1002 bool Container = isa<ConstantVector>(Val) || isa<ConstantArray>(Val) ||
1003 isa<ConstantStruct>(Val);
1004 // TODO: Because all of these things are constant, we can determine whether
1005 // the data is *actually* mutable at graph building time. This will probably
1006 // come for free/cheap with offset awareness.
1007 bool CanStoreMutableData =
1008 isa<GlobalValue>(Val) || isa<ConstantExpr>(Val) || Container;
1009 return !CanStoreMutableData;
1015 static FunctionInfo buildSetsFrom(CFLAliasAnalysis &Analysis, Function *Fn) {
1018 SmallVector<Value *, 4> ReturnedValues;
1020 buildGraphFrom(Analysis, Fn, ReturnedValues, Map, Graph);
1022 DenseMap<GraphT::Node, Value *> NodeValueMap;
1023 NodeValueMap.resize(Map.size());
1024 for (const auto &Pair : Map)
1025 NodeValueMap.insert(std::make_pair(Pair.second, Pair.first));
1027 const auto findValueOrDie = [&NodeValueMap](GraphT::Node Node) {
1028 auto ValIter = NodeValueMap.find(Node);
1029 assert(ValIter != NodeValueMap.end());
1030 return ValIter->second;
1033 StratifiedSetsBuilder<Value *> Builder;
1035 SmallVector<GraphT::Node, 16> Worklist;
1036 for (auto &Pair : Map) {
1039 auto *Value = Pair.first;
1041 auto InitialNode = Pair.second;
1042 Worklist.push_back(InitialNode);
1043 while (!Worklist.empty()) {
1044 auto Node = Worklist.pop_back_val();
1045 auto *CurValue = findValueOrDie(Node);
1046 if (canSkipAddingToSets(CurValue))
1049 for (const auto &EdgeTuple : Graph.edgesFor(Node)) {
1050 auto Weight = std::get<0>(EdgeTuple);
1051 auto Label = Weight.first;
1052 auto &OtherNode = std::get<1>(EdgeTuple);
1053 auto *OtherValue = findValueOrDie(OtherNode);
1055 if (canSkipAddingToSets(OtherValue))
1059 switch (directionOfEdgeType(Label)) {
1061 Added = Builder.addAbove(CurValue, OtherValue);
1064 Added = Builder.addBelow(CurValue, OtherValue);
1067 Added = Builder.addWith(CurValue, OtherValue);
1071 auto Aliasing = Weight.second;
1072 if (auto MaybeCurIndex = valueToAttrIndex(CurValue))
1073 Aliasing.set(*MaybeCurIndex);
1074 if (auto MaybeOtherIndex = valueToAttrIndex(OtherValue))
1075 Aliasing.set(*MaybeOtherIndex);
1076 Builder.noteAttributes(CurValue, Aliasing);
1077 Builder.noteAttributes(OtherValue, Aliasing);
1080 Worklist.push_back(OtherNode);
1085 // There are times when we end up with parameters not in our graph (i.e. if
1086 // it's only used as the condition of a branch). Other bits of code depend on
1087 // things that were present during construction being present in the graph.
1088 // So, we add all present arguments here.
1089 for (auto &Arg : Fn->args()) {
1090 if (!Builder.add(&Arg))
1093 auto Attrs = valueToAttrIndex(&Arg);
1094 if (Attrs.hasValue())
1095 Builder.noteAttributes(&Arg, *Attrs);
1098 return FunctionInfo(Builder.build(), std::move(ReturnedValues));
1101 void CFLAliasAnalysis::scan(Function *Fn) {
1102 auto InsertPair = Cache.insert(std::make_pair(Fn, Optional<FunctionInfo>()));
1104 assert(InsertPair.second &&
1105 "Trying to scan a function that has already been cached");
1107 FunctionInfo Info(buildSetsFrom(*this, Fn));
1108 Cache[Fn] = std::move(Info);
1109 Handles.push_front(FunctionHandle(Fn, this));
1112 AliasAnalysis::AliasResult
1113 CFLAliasAnalysis::query(const AliasAnalysis::Location &LocA,
1114 const AliasAnalysis::Location &LocB) {
1115 auto *ValA = const_cast<Value *>(LocA.Ptr);
1116 auto *ValB = const_cast<Value *>(LocB.Ptr);
1118 Function *Fn = nullptr;
1119 auto MaybeFnA = parentFunctionOfValue(ValA);
1120 auto MaybeFnB = parentFunctionOfValue(ValB);
1121 if (!MaybeFnA.hasValue() && !MaybeFnB.hasValue()) {
1122 // The only times this is known to happen are when globals + InlineAsm
1124 DEBUG(dbgs() << "CFLAA: could not extract parent function information.\n");
1125 return AliasAnalysis::MayAlias;
1128 if (MaybeFnA.hasValue()) {
1130 assert((!MaybeFnB.hasValue() || *MaybeFnB == *MaybeFnA) &&
1131 "Interprocedural queries not supported");
1136 assert(Fn != nullptr);
1137 auto &MaybeInfo = ensureCached(Fn);
1138 assert(MaybeInfo.hasValue());
1140 auto &Sets = MaybeInfo->Sets;
1141 auto MaybeA = Sets.find(ValA);
1142 if (!MaybeA.hasValue())
1143 return AliasAnalysis::MayAlias;
1145 auto MaybeB = Sets.find(ValB);
1146 if (!MaybeB.hasValue())
1147 return AliasAnalysis::MayAlias;
1149 auto SetA = *MaybeA;
1150 auto SetB = *MaybeB;
1151 auto AttrsA = Sets.getLink(SetA.Index).Attrs;
1152 auto AttrsB = Sets.getLink(SetB.Index).Attrs;
1154 // Stratified set attributes are used as markets to signify whether a member
1155 // of a StratifiedSet (or a member of a set above the current set) has
1156 // interacted with either arguments or globals. "Interacted with" meaning
1157 // its value may be different depending on the value of an argument or
1158 // global. The thought behind this is that, because arguments and globals
1159 // may alias each other, if AttrsA and AttrsB have touched args/globals,
1160 // we must conservatively say that they alias. However, if at least one of
1161 // the sets has no values that could legally be altered by changing the value
1162 // of an argument or global, then we don't have to be as conservative.
1163 if (AttrsA.any() && AttrsB.any())
1164 return AliasAnalysis::MayAlias;
1166 // We currently unify things even if the accesses to them may not be in
1167 // bounds, so we can't return partial alias here because we don't
1168 // know whether the pointer is really within the object or not.
1169 // IE Given an out of bounds GEP and an alloca'd pointer, we may
1170 // unify the two. We can't return partial alias for this case.
1171 // Since we do not currently track enough information to
1174 if (SetA.Index == SetB.Index)
1175 return AliasAnalysis::MayAlias;
1177 return AliasAnalysis::NoAlias;
1180 bool CFLAliasAnalysis::doInitialization(Module &M) {
1181 InitializeAliasAnalysis(this, &M.getDataLayout());