1 //===- Andersens.cpp - Andersen's Interprocedural Alias Analysis ----------===//
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 defines an implementation of Andersen's interprocedural alias
13 // In pointer analysis terms, this is a subset-based, flow-insensitive,
14 // field-sensitive, and context-insensitive algorithm pointer algorithm.
16 // This algorithm is implemented as three stages:
17 // 1. Object identification.
18 // 2. Inclusion constraint identification.
19 // 3. Offline constraint graph optimization
20 // 4. Inclusion constraint solving.
22 // The object identification stage identifies all of the memory objects in the
23 // program, which includes globals, heap allocated objects, and stack allocated
26 // The inclusion constraint identification stage finds all inclusion constraints
27 // in the program by scanning the program, looking for pointer assignments and
28 // other statements that effect the points-to graph. For a statement like "A =
29 // B", this statement is processed to indicate that A can point to anything that
30 // B can point to. Constraints can handle copies, loads, and stores, and
33 // The offline constraint graph optimization portion includes offline variable
34 // substitution algorithms intended to compute pointer and location
35 // equivalences. Pointer equivalences are those pointers that will have the
36 // same points-to sets, and location equivalences are those variables that
37 // always appear together in points-to sets. It also includes an offline
38 // cycle detection algorithm that allows cycles to be collapsed sooner
41 // The inclusion constraint solving phase iteratively propagates the inclusion
42 // constraints until a fixed point is reached. This is an O(N^3) algorithm.
44 // Function constraints are handled as if they were structs with X fields.
45 // Thus, an access to argument X of function Y is an access to node index
46 // getNode(Y) + X. This representation allows handling of indirect calls
47 // without any issues. To wit, an indirect call Y(a,b) is equivalent to
48 // *(Y + 1) = a, *(Y + 2) = b.
49 // The return node for a function is always located at getNode(F) +
50 // CallReturnPos. The arguments start at getNode(F) + CallArgPos.
52 // Future Improvements:
54 //===----------------------------------------------------------------------===//
56 #define DEBUG_TYPE "anders-aa"
57 #include "llvm/Constants.h"
58 #include "llvm/DerivedTypes.h"
59 #include "llvm/Instructions.h"
60 #include "llvm/Module.h"
61 #include "llvm/Pass.h"
62 #include "llvm/Support/ErrorHandling.h"
63 #include "llvm/Support/InstIterator.h"
64 #include "llvm/Support/InstVisitor.h"
65 #include "llvm/Analysis/AliasAnalysis.h"
66 #include "llvm/Analysis/MallocFreeHelper.h"
67 #include "llvm/Analysis/Passes.h"
68 #include "llvm/Support/Debug.h"
69 #include "llvm/System/Atomic.h"
70 #include "llvm/ADT/Statistic.h"
71 #include "llvm/ADT/SparseBitVector.h"
72 #include "llvm/ADT/DenseSet.h"
81 // Determining the actual set of nodes the universal set can consist of is very
82 // expensive because it means propagating around very large sets. We rely on
83 // other analysis being able to determine which nodes can never be pointed to in
84 // order to disambiguate further than "points-to anything".
85 #define FULL_UNIVERSAL 0
89 STATISTIC(NumIters , "Number of iterations to reach convergence");
91 STATISTIC(NumConstraints, "Number of constraints");
92 STATISTIC(NumNodes , "Number of nodes");
93 STATISTIC(NumUnified , "Number of variables unified");
94 STATISTIC(NumErased , "Number of redundant constraints erased");
96 static const unsigned SelfRep = (unsigned)-1;
97 static const unsigned Unvisited = (unsigned)-1;
98 // Position of the function return node relative to the function node.
99 static const unsigned CallReturnPos = 1;
100 // Position of the function call node relative to the function node.
101 static const unsigned CallFirstArgPos = 2;
104 struct BitmapKeyInfo {
105 static inline SparseBitVector<> *getEmptyKey() {
106 return reinterpret_cast<SparseBitVector<> *>(-1);
108 static inline SparseBitVector<> *getTombstoneKey() {
109 return reinterpret_cast<SparseBitVector<> *>(-2);
111 static unsigned getHashValue(const SparseBitVector<> *bitmap) {
112 return bitmap->getHashValue();
114 static bool isEqual(const SparseBitVector<> *LHS,
115 const SparseBitVector<> *RHS) {
118 else if (LHS == getEmptyKey() || RHS == getEmptyKey()
119 || LHS == getTombstoneKey() || RHS == getTombstoneKey())
125 static bool isPod() { return true; }
128 class Andersens : public ModulePass, public AliasAnalysis,
129 private InstVisitor<Andersens> {
132 /// Constraint - Objects of this structure are used to represent the various
133 /// constraints identified by the algorithm. The constraints are 'copy',
134 /// for statements like "A = B", 'load' for statements like "A = *B",
135 /// 'store' for statements like "*A = B", and AddressOf for statements like
136 /// A = alloca; The Offset is applied as *(A + K) = B for stores,
137 /// A = *(B + K) for loads, and A = B + K for copies. It is
138 /// illegal on addressof constraints (because it is statically
139 /// resolvable to A = &C where C = B + K)
142 enum ConstraintType { Copy, Load, Store, AddressOf } Type;
147 Constraint(ConstraintType Ty, unsigned D, unsigned S, unsigned O = 0)
148 : Type(Ty), Dest(D), Src(S), Offset(O) {
149 assert((Offset == 0 || Ty != AddressOf) &&
150 "Offset is illegal on addressof constraints");
153 bool operator==(const Constraint &RHS) const {
154 return RHS.Type == Type
157 && RHS.Offset == Offset;
160 bool operator!=(const Constraint &RHS) const {
161 return !(*this == RHS);
164 bool operator<(const Constraint &RHS) const {
165 if (RHS.Type != Type)
166 return RHS.Type < Type;
167 else if (RHS.Dest != Dest)
168 return RHS.Dest < Dest;
169 else if (RHS.Src != Src)
170 return RHS.Src < Src;
171 return RHS.Offset < Offset;
175 // Information DenseSet requires implemented in order to be able to do
178 static inline std::pair<unsigned, unsigned> getEmptyKey() {
179 return std::make_pair(~0U, ~0U);
181 static inline std::pair<unsigned, unsigned> getTombstoneKey() {
182 return std::make_pair(~0U - 1, ~0U - 1);
184 static unsigned getHashValue(const std::pair<unsigned, unsigned> &P) {
185 return P.first ^ P.second;
187 static unsigned isEqual(const std::pair<unsigned, unsigned> &LHS,
188 const std::pair<unsigned, unsigned> &RHS) {
193 struct ConstraintKeyInfo {
194 static inline Constraint getEmptyKey() {
195 return Constraint(Constraint::Copy, ~0U, ~0U, ~0U);
197 static inline Constraint getTombstoneKey() {
198 return Constraint(Constraint::Copy, ~0U - 1, ~0U - 1, ~0U - 1);
200 static unsigned getHashValue(const Constraint &C) {
201 return C.Src ^ C.Dest ^ C.Type ^ C.Offset;
203 static bool isEqual(const Constraint &LHS,
204 const Constraint &RHS) {
205 return LHS.Type == RHS.Type && LHS.Dest == RHS.Dest
206 && LHS.Src == RHS.Src && LHS.Offset == RHS.Offset;
210 // Node class - This class is used to represent a node in the constraint
211 // graph. Due to various optimizations, it is not always the case that
212 // there is a mapping from a Node to a Value. In particular, we add
213 // artificial Node's that represent the set of pointed-to variables shared
214 // for each location equivalent Node.
217 static volatile sys::cas_flag Counter;
221 SparseBitVector<> *Edges;
222 SparseBitVector<> *PointsTo;
223 SparseBitVector<> *OldPointsTo;
224 std::list<Constraint> Constraints;
226 // Pointer and location equivalence labels
227 unsigned PointerEquivLabel;
228 unsigned LocationEquivLabel;
229 // Predecessor edges, both real and implicit
230 SparseBitVector<> *PredEdges;
231 SparseBitVector<> *ImplicitPredEdges;
232 // Set of nodes that point to us, only use for location equivalence.
233 SparseBitVector<> *PointedToBy;
234 // Number of incoming edges, used during variable substitution to early
235 // free the points-to sets
237 // True if our points-to set is in the Set2PEClass map
239 // True if our node has no indirect constraints (complex or otherwise)
241 // True if the node is address taken, *or* it is part of a group of nodes
242 // that must be kept together. This is set to true for functions and
243 // their arg nodes, which must be kept at the same position relative to
244 // their base function node.
247 // Nodes in cycles (or in equivalence classes) are united together using a
248 // standard union-find representation with path compression. NodeRep
249 // gives the index into GraphNodes for the representative Node.
252 // Modification timestamp. Assigned from Counter.
253 // Used for work list prioritization.
256 explicit Node(bool direct = true) :
257 Val(0), Edges(0), PointsTo(0), OldPointsTo(0),
258 PointerEquivLabel(0), LocationEquivLabel(0), PredEdges(0),
259 ImplicitPredEdges(0), PointedToBy(0), NumInEdges(0),
260 StoredInHash(false), Direct(direct), AddressTaken(false),
261 NodeRep(SelfRep), Timestamp(0) { }
263 Node *setValue(Value *V) {
264 assert(Val == 0 && "Value already set for this node!");
269 /// getValue - Return the LLVM value corresponding to this node.
271 Value *getValue() const { return Val; }
273 /// addPointerTo - Add a pointer to the list of pointees of this node,
274 /// returning true if this caused a new pointer to be added, or false if
275 /// we already knew about the points-to relation.
276 bool addPointerTo(unsigned Node) {
277 return PointsTo->test_and_set(Node);
280 /// intersects - Return true if the points-to set of this node intersects
281 /// with the points-to set of the specified node.
282 bool intersects(Node *N) const;
284 /// intersectsIgnoring - Return true if the points-to set of this node
285 /// intersects with the points-to set of the specified node on any nodes
286 /// except for the specified node to ignore.
287 bool intersectsIgnoring(Node *N, unsigned) const;
289 // Timestamp a node (used for work list prioritization)
291 Timestamp = sys::AtomicIncrement(&Counter);
296 return( (int) NodeRep < 0 );
300 struct WorkListElement {
303 WorkListElement(Node* n, unsigned t) : node(n), Timestamp(t) {}
305 // Note that we reverse the sense of the comparison because we
306 // actually want to give low timestamps the priority over high,
307 // whereas priority is typically interpreted as a greater value is
308 // given high priority.
309 bool operator<(const WorkListElement& that) const {
310 return( this->Timestamp > that.Timestamp );
314 // Priority-queue based work list specialized for Nodes.
316 std::priority_queue<WorkListElement> Q;
319 void insert(Node* n) {
320 Q.push( WorkListElement(n, n->Timestamp) );
323 // We automatically discard non-representative nodes and nodes
324 // that were in the work list twice (we keep a copy of the
325 // timestamp in the work list so we can detect this situation by
326 // comparing against the node's current timestamp).
328 while( !Q.empty() ) {
329 WorkListElement x = Q.top(); Q.pop();
330 Node* INode = x.node;
332 if( INode->isRep() &&
333 INode->Timestamp == x.Timestamp ) {
345 /// GraphNodes - This vector is populated as part of the object
346 /// identification stage of the analysis, which populates this vector with a
347 /// node for each memory object and fills in the ValueNodes map.
348 std::vector<Node> GraphNodes;
350 /// ValueNodes - This map indicates the Node that a particular Value* is
351 /// represented by. This contains entries for all pointers.
352 DenseMap<Value*, unsigned> ValueNodes;
354 /// ObjectNodes - This map contains entries for each memory object in the
355 /// program: globals, alloca's and mallocs.
356 DenseMap<Value*, unsigned> ObjectNodes;
358 /// ReturnNodes - This map contains an entry for each function in the
359 /// program that returns a value.
360 DenseMap<Function*, unsigned> ReturnNodes;
362 /// VarargNodes - This map contains the entry used to represent all pointers
363 /// passed through the varargs portion of a function call for a particular
364 /// function. An entry is not present in this map for functions that do not
365 /// take variable arguments.
366 DenseMap<Function*, unsigned> VarargNodes;
369 /// Constraints - This vector contains a list of all of the constraints
370 /// identified by the program.
371 std::vector<Constraint> Constraints;
373 // Map from graph node to maximum K value that is allowed (for functions,
374 // this is equivalent to the number of arguments + CallFirstArgPos)
375 std::map<unsigned, unsigned> MaxK;
377 /// This enum defines the GraphNodes indices that correspond to important
385 // Stack for Tarjan's
386 std::stack<unsigned> SCCStack;
387 // Map from Graph Node to DFS number
388 std::vector<unsigned> Node2DFS;
389 // Map from Graph Node to Deleted from graph.
390 std::vector<bool> Node2Deleted;
391 // Same as Node Maps, but implemented as std::map because it is faster to
393 std::map<unsigned, unsigned> Tarjan2DFS;
394 std::map<unsigned, bool> Tarjan2Deleted;
395 // Current DFS number
400 WorkList *CurrWL, *NextWL; // "current" and "next" work lists
402 // Offline variable substitution related things
404 // Temporary rep storage, used because we can't collapse SCC's in the
405 // predecessor graph by uniting the variables permanently, we can only do so
406 // for the successor graph.
407 std::vector<unsigned> VSSCCRep;
408 // Mapping from node to whether we have visited it during SCC finding yet.
409 std::vector<bool> Node2Visited;
410 // During variable substitution, we create unknowns to represent the unknown
411 // value that is a dereference of a variable. These nodes are known as
412 // "ref" nodes (since they represent the value of dereferences).
413 unsigned FirstRefNode;
414 // During HVN, we create represent address taken nodes as if they were
415 // unknown (since HVN, unlike HU, does not evaluate unions).
416 unsigned FirstAdrNode;
417 // Current pointer equivalence class number
419 // Mapping from points-to sets to equivalence classes
420 typedef DenseMap<SparseBitVector<> *, unsigned, BitmapKeyInfo> BitVectorMap;
421 BitVectorMap Set2PEClass;
422 // Mapping from pointer equivalences to the representative node. -1 if we
423 // have no representative node for this pointer equivalence class yet.
424 std::vector<int> PEClass2Node;
425 // Mapping from pointer equivalences to representative node. This includes
426 // pointer equivalent but not location equivalent variables. -1 if we have
427 // no representative node for this pointer equivalence class yet.
428 std::vector<int> PENLEClass2Node;
429 // Union/Find for HCD
430 std::vector<unsigned> HCDSCCRep;
431 // HCD's offline-detected cycles; "Statically DeTected"
432 // -1 if not part of such a cycle, otherwise a representative node.
433 std::vector<int> SDT;
434 // Whether to use SDT (UniteNodes can use it during solving, but not before)
439 Andersens() : ModulePass(&ID) {}
441 bool runOnModule(Module &M) {
442 InitializeAliasAnalysis(this);
444 CollectConstraints(M);
446 #define DEBUG_TYPE "anders-aa-constraints"
447 DEBUG(PrintConstraints());
449 #define DEBUG_TYPE "anders-aa"
451 DEBUG(PrintPointsToGraph());
453 // Free the constraints list, as we don't need it to respond to alias
455 std::vector<Constraint>().swap(Constraints);
456 //These are needed for Print() (-analyze in opt)
457 //ObjectNodes.clear();
458 //ReturnNodes.clear();
459 //VarargNodes.clear();
463 void releaseMemory() {
464 // FIXME: Until we have transitively required passes working correctly,
465 // this cannot be enabled! Otherwise, using -count-aa with the pass
466 // causes memory to be freed too early. :(
468 // The memory objects and ValueNodes data structures at the only ones that
469 // are still live after construction.
470 std::vector<Node>().swap(GraphNodes);
475 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
476 AliasAnalysis::getAnalysisUsage(AU);
477 AU.setPreservesAll(); // Does not transform code
480 //------------------------------------------------
481 // Implement the AliasAnalysis API
483 AliasResult alias(const Value *V1, unsigned V1Size,
484 const Value *V2, unsigned V2Size);
485 virtual ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size);
486 virtual ModRefResult getModRefInfo(CallSite CS1, CallSite CS2);
487 void getMustAliases(Value *P, std::vector<Value*> &RetVals);
488 bool pointsToConstantMemory(const Value *P);
490 virtual void deleteValue(Value *V) {
492 getAnalysis<AliasAnalysis>().deleteValue(V);
495 virtual void copyValue(Value *From, Value *To) {
496 ValueNodes[To] = ValueNodes[From];
497 getAnalysis<AliasAnalysis>().copyValue(From, To);
501 /// getNode - Return the node corresponding to the specified pointer scalar.
503 unsigned getNode(Value *V) {
504 if (Constant *C = dyn_cast<Constant>(V))
505 if (!isa<GlobalValue>(C))
506 return getNodeForConstantPointer(C);
508 DenseMap<Value*, unsigned>::iterator I = ValueNodes.find(V);
509 if (I == ValueNodes.end()) {
513 llvm_unreachable("Value does not have a node in the points-to graph!");
518 /// getObject - Return the node corresponding to the memory object for the
519 /// specified global or allocation instruction.
520 unsigned getObject(Value *V) const {
521 DenseMap<Value*, unsigned>::iterator I = ObjectNodes.find(V);
522 assert(I != ObjectNodes.end() &&
523 "Value does not have an object in the points-to graph!");
527 /// getReturnNode - Return the node representing the return value for the
528 /// specified function.
529 unsigned getReturnNode(Function *F) const {
530 DenseMap<Function*, unsigned>::iterator I = ReturnNodes.find(F);
531 assert(I != ReturnNodes.end() && "Function does not return a value!");
535 /// getVarargNode - Return the node representing the variable arguments
536 /// formal for the specified function.
537 unsigned getVarargNode(Function *F) const {
538 DenseMap<Function*, unsigned>::iterator I = VarargNodes.find(F);
539 assert(I != VarargNodes.end() && "Function does not take var args!");
543 /// getNodeValue - Get the node for the specified LLVM value and set the
544 /// value for it to be the specified value.
545 unsigned getNodeValue(Value &V) {
546 unsigned Index = getNode(&V);
547 GraphNodes[Index].setValue(&V);
551 unsigned UniteNodes(unsigned First, unsigned Second,
552 bool UnionByRank = true);
553 unsigned FindNode(unsigned Node);
554 unsigned FindNode(unsigned Node) const;
556 void IdentifyObjects(Module &M);
557 void CollectConstraints(Module &M);
558 bool AnalyzeUsesOfFunction(Value *);
559 void CreateConstraintGraph();
560 void OptimizeConstraints();
561 unsigned FindEquivalentNode(unsigned, unsigned);
562 void ClumpAddressTaken();
563 void RewriteConstraints();
567 void Search(unsigned Node);
568 void UnitePointerEquivalences();
569 void SolveConstraints();
570 bool QueryNode(unsigned Node);
571 void Condense(unsigned Node);
572 void HUValNum(unsigned Node);
573 void HVNValNum(unsigned Node);
574 unsigned getNodeForConstantPointer(Constant *C);
575 unsigned getNodeForConstantPointerTarget(Constant *C);
576 void AddGlobalInitializerConstraints(unsigned, Constant *C);
578 void AddConstraintsForNonInternalLinkage(Function *F);
579 void AddConstraintsForCall(CallSite CS, Function *F);
580 bool AddConstraintsForExternalCall(CallSite CS, Function *F);
583 void PrintNode(const Node *N) const;
584 void PrintConstraints() const ;
585 void PrintConstraint(const Constraint &) const;
586 void PrintLabels() const;
587 void PrintPointsToGraph() const;
589 //===------------------------------------------------------------------===//
590 // Instruction visitation methods for adding constraints
592 friend class InstVisitor<Andersens>;
593 void visitReturnInst(ReturnInst &RI);
594 void visitInvokeInst(InvokeInst &II) { visitCallSite(CallSite(&II)); }
595 void visitCallInst(CallInst &CI) {
596 if (isMalloc(&CI)) visitAlloc(CI);
597 else visitCallSite(CallSite(&CI));
599 void visitCallSite(CallSite CS);
600 void visitAllocaInst(AllocaInst &I);
601 void visitAlloc(Instruction &I);
602 void visitLoadInst(LoadInst &LI);
603 void visitStoreInst(StoreInst &SI);
604 void visitGetElementPtrInst(GetElementPtrInst &GEP);
605 void visitPHINode(PHINode &PN);
606 void visitCastInst(CastInst &CI);
607 void visitICmpInst(ICmpInst &ICI) {} // NOOP!
608 void visitFCmpInst(FCmpInst &ICI) {} // NOOP!
609 void visitSelectInst(SelectInst &SI);
610 void visitVAArg(VAArgInst &I);
611 void visitInstruction(Instruction &I);
613 //===------------------------------------------------------------------===//
614 // Implement Analyize interface
616 void print(raw_ostream &O, const Module*) const {
617 PrintPointsToGraph();
622 char Andersens::ID = 0;
623 static RegisterPass<Andersens>
624 X("anders-aa", "Andersen's Interprocedural Alias Analysis (experimental)",
626 static RegisterAnalysisGroup<AliasAnalysis> Y(X);
628 // Initialize Timestamp Counter (static).
629 volatile llvm::sys::cas_flag Andersens::Node::Counter = 0;
631 ModulePass *llvm::createAndersensPass() { return new Andersens(); }
633 //===----------------------------------------------------------------------===//
634 // AliasAnalysis Interface Implementation
635 //===----------------------------------------------------------------------===//
637 AliasAnalysis::AliasResult Andersens::alias(const Value *V1, unsigned V1Size,
638 const Value *V2, unsigned V2Size) {
639 Node *N1 = &GraphNodes[FindNode(getNode(const_cast<Value*>(V1)))];
640 Node *N2 = &GraphNodes[FindNode(getNode(const_cast<Value*>(V2)))];
642 // Check to see if the two pointers are known to not alias. They don't alias
643 // if their points-to sets do not intersect.
644 if (!N1->intersectsIgnoring(N2, NullObject))
647 return AliasAnalysis::alias(V1, V1Size, V2, V2Size);
650 AliasAnalysis::ModRefResult
651 Andersens::getModRefInfo(CallSite CS, Value *P, unsigned Size) {
652 // The only thing useful that we can contribute for mod/ref information is
653 // when calling external function calls: if we know that memory never escapes
654 // from the program, it cannot be modified by an external call.
656 // NOTE: This is not really safe, at least not when the entire program is not
657 // available. The deal is that the external function could call back into the
658 // program and modify stuff. We ignore this technical niggle for now. This
659 // is, after all, a "research quality" implementation of Andersen's analysis.
660 if (Function *F = CS.getCalledFunction())
661 if (F->isDeclaration()) {
662 Node *N1 = &GraphNodes[FindNode(getNode(P))];
664 if (N1->PointsTo->empty())
667 if (!UniversalSet->PointsTo->test(FindNode(getNode(P))))
668 return NoModRef; // Universal set does not contain P
670 if (!N1->PointsTo->test(UniversalSet))
671 return NoModRef; // P doesn't point to the universal set.
675 return AliasAnalysis::getModRefInfo(CS, P, Size);
678 AliasAnalysis::ModRefResult
679 Andersens::getModRefInfo(CallSite CS1, CallSite CS2) {
680 return AliasAnalysis::getModRefInfo(CS1,CS2);
683 /// getMustAlias - We can provide must alias information if we know that a
684 /// pointer can only point to a specific function or the null pointer.
685 /// Unfortunately we cannot determine must-alias information for global
686 /// variables or any other memory memory objects because we do not track whether
687 /// a pointer points to the beginning of an object or a field of it.
688 void Andersens::getMustAliases(Value *P, std::vector<Value*> &RetVals) {
689 Node *N = &GraphNodes[FindNode(getNode(P))];
690 if (N->PointsTo->count() == 1) {
691 Node *Pointee = &GraphNodes[N->PointsTo->find_first()];
692 // If a function is the only object in the points-to set, then it must be
693 // the destination. Note that we can't handle global variables here,
694 // because we don't know if the pointer is actually pointing to a field of
695 // the global or to the beginning of it.
696 if (Value *V = Pointee->getValue()) {
697 if (Function *F = dyn_cast<Function>(V))
698 RetVals.push_back(F);
700 // If the object in the points-to set is the null object, then the null
701 // pointer is a must alias.
702 if (Pointee == &GraphNodes[NullObject])
703 RetVals.push_back(Constant::getNullValue(P->getType()));
706 AliasAnalysis::getMustAliases(P, RetVals);
709 /// pointsToConstantMemory - If we can determine that this pointer only points
710 /// to constant memory, return true. In practice, this means that if the
711 /// pointer can only point to constant globals, functions, or the null pointer,
714 bool Andersens::pointsToConstantMemory(const Value *P) {
715 Node *N = &GraphNodes[FindNode(getNode(const_cast<Value*>(P)))];
718 for (SparseBitVector<>::iterator bi = N->PointsTo->begin();
719 bi != N->PointsTo->end();
722 Node *Pointee = &GraphNodes[i];
723 if (Value *V = Pointee->getValue()) {
724 if (!isa<GlobalValue>(V) || (isa<GlobalVariable>(V) &&
725 !cast<GlobalVariable>(V)->isConstant()))
726 return AliasAnalysis::pointsToConstantMemory(P);
729 return AliasAnalysis::pointsToConstantMemory(P);
736 //===----------------------------------------------------------------------===//
737 // Object Identification Phase
738 //===----------------------------------------------------------------------===//
740 /// IdentifyObjects - This stage scans the program, adding an entry to the
741 /// GraphNodes list for each memory object in the program (global stack or
742 /// heap), and populates the ValueNodes and ObjectNodes maps for these objects.
744 void Andersens::IdentifyObjects(Module &M) {
745 unsigned NumObjects = 0;
747 // Object #0 is always the universal set: the object that we don't know
749 assert(NumObjects == UniversalSet && "Something changed!");
752 // Object #1 always represents the null pointer.
753 assert(NumObjects == NullPtr && "Something changed!");
756 // Object #2 always represents the null object (the object pointed to by null)
757 assert(NumObjects == NullObject && "Something changed!");
760 // Add all the globals first.
761 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
763 ObjectNodes[I] = NumObjects++;
764 ValueNodes[I] = NumObjects++;
767 // Add nodes for all of the functions and the instructions inside of them.
768 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
769 // The function itself is a memory object.
770 unsigned First = NumObjects;
771 ValueNodes[F] = NumObjects++;
772 if (isa<PointerType>(F->getFunctionType()->getReturnType()))
773 ReturnNodes[F] = NumObjects++;
774 if (F->getFunctionType()->isVarArg())
775 VarargNodes[F] = NumObjects++;
778 // Add nodes for all of the incoming pointer arguments.
779 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
782 if (isa<PointerType>(I->getType()))
783 ValueNodes[I] = NumObjects++;
785 MaxK[First] = NumObjects - First;
787 // Scan the function body, creating a memory object for each heap/stack
788 // allocation in the body of the function and a node to represent all
789 // pointer values defined by instructions and used as operands.
790 for (inst_iterator II = inst_begin(F), E = inst_end(F); II != E; ++II) {
791 // If this is an heap or stack allocation, create a node for the memory
793 if (isa<PointerType>(II->getType())) {
794 ValueNodes[&*II] = NumObjects++;
795 if (AllocaInst *AI = dyn_cast<AllocaInst>(&*II))
796 ObjectNodes[AI] = NumObjects++;
797 else if (isMalloc(&*II))
798 ObjectNodes[&*II] = NumObjects++;
801 // Calls to inline asm need to be added as well because the callee isn't
802 // referenced anywhere else.
803 if (CallInst *CI = dyn_cast<CallInst>(&*II)) {
804 Value *Callee = CI->getCalledValue();
805 if (isa<InlineAsm>(Callee))
806 ValueNodes[Callee] = NumObjects++;
811 // Now that we know how many objects to create, make them all now!
812 GraphNodes.resize(NumObjects);
813 NumNodes += NumObjects;
816 //===----------------------------------------------------------------------===//
817 // Constraint Identification Phase
818 //===----------------------------------------------------------------------===//
820 /// getNodeForConstantPointer - Return the node corresponding to the constant
822 unsigned Andersens::getNodeForConstantPointer(Constant *C) {
823 assert(isa<PointerType>(C->getType()) && "Not a constant pointer!");
825 if (isa<ConstantPointerNull>(C) || isa<UndefValue>(C))
827 else if (GlobalValue *GV = dyn_cast<GlobalValue>(C))
829 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
830 switch (CE->getOpcode()) {
831 case Instruction::GetElementPtr:
832 return getNodeForConstantPointer(CE->getOperand(0));
833 case Instruction::IntToPtr:
835 case Instruction::BitCast:
836 return getNodeForConstantPointer(CE->getOperand(0));
838 errs() << "Constant Expr not yet handled: " << *CE << "\n";
842 llvm_unreachable("Unknown constant pointer!");
847 /// getNodeForConstantPointerTarget - Return the node POINTED TO by the
848 /// specified constant pointer.
849 unsigned Andersens::getNodeForConstantPointerTarget(Constant *C) {
850 assert(isa<PointerType>(C->getType()) && "Not a constant pointer!");
852 if (isa<ConstantPointerNull>(C))
854 else if (GlobalValue *GV = dyn_cast<GlobalValue>(C))
855 return getObject(GV);
856 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
857 switch (CE->getOpcode()) {
858 case Instruction::GetElementPtr:
859 return getNodeForConstantPointerTarget(CE->getOperand(0));
860 case Instruction::IntToPtr:
862 case Instruction::BitCast:
863 return getNodeForConstantPointerTarget(CE->getOperand(0));
865 errs() << "Constant Expr not yet handled: " << *CE << "\n";
869 llvm_unreachable("Unknown constant pointer!");
874 /// AddGlobalInitializerConstraints - Add inclusion constraints for the memory
875 /// object N, which contains values indicated by C.
876 void Andersens::AddGlobalInitializerConstraints(unsigned NodeIndex,
878 if (C->getType()->isSingleValueType()) {
879 if (isa<PointerType>(C->getType()))
880 Constraints.push_back(Constraint(Constraint::Copy, NodeIndex,
881 getNodeForConstantPointer(C)));
882 } else if (C->isNullValue()) {
883 Constraints.push_back(Constraint(Constraint::Copy, NodeIndex,
886 } else if (!isa<UndefValue>(C)) {
887 // If this is an array or struct, include constraints for each element.
888 assert(isa<ConstantArray>(C) || isa<ConstantStruct>(C));
889 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i)
890 AddGlobalInitializerConstraints(NodeIndex,
891 cast<Constant>(C->getOperand(i)));
895 /// AddConstraintsForNonInternalLinkage - If this function does not have
896 /// internal linkage, realize that we can't trust anything passed into or
897 /// returned by this function.
898 void Andersens::AddConstraintsForNonInternalLinkage(Function *F) {
899 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
900 if (isa<PointerType>(I->getType()))
901 // If this is an argument of an externally accessible function, the
902 // incoming pointer might point to anything.
903 Constraints.push_back(Constraint(Constraint::Copy, getNode(I),
907 /// AddConstraintsForCall - If this is a call to a "known" function, add the
908 /// constraints and return true. If this is a call to an unknown function,
910 bool Andersens::AddConstraintsForExternalCall(CallSite CS, Function *F) {
911 assert(F->isDeclaration() && "Not an external function!");
913 // These functions don't induce any points-to constraints.
914 if (F->getName() == "atoi" || F->getName() == "atof" ||
915 F->getName() == "atol" || F->getName() == "atoll" ||
916 F->getName() == "remove" || F->getName() == "unlink" ||
917 F->getName() == "rename" || F->getName() == "memcmp" ||
918 F->getName() == "llvm.memset" ||
919 F->getName() == "strcmp" || F->getName() == "strncmp" ||
920 F->getName() == "execl" || F->getName() == "execlp" ||
921 F->getName() == "execle" || F->getName() == "execv" ||
922 F->getName() == "execvp" || F->getName() == "chmod" ||
923 F->getName() == "puts" || F->getName() == "write" ||
924 F->getName() == "open" || F->getName() == "create" ||
925 F->getName() == "truncate" || F->getName() == "chdir" ||
926 F->getName() == "mkdir" || F->getName() == "rmdir" ||
927 F->getName() == "read" || F->getName() == "pipe" ||
928 F->getName() == "wait" || F->getName() == "time" ||
929 F->getName() == "stat" || F->getName() == "fstat" ||
930 F->getName() == "lstat" || F->getName() == "strtod" ||
931 F->getName() == "strtof" || F->getName() == "strtold" ||
932 F->getName() == "fopen" || F->getName() == "fdopen" ||
933 F->getName() == "freopen" ||
934 F->getName() == "fflush" || F->getName() == "feof" ||
935 F->getName() == "fileno" || F->getName() == "clearerr" ||
936 F->getName() == "rewind" || F->getName() == "ftell" ||
937 F->getName() == "ferror" || F->getName() == "fgetc" ||
938 F->getName() == "fgetc" || F->getName() == "_IO_getc" ||
939 F->getName() == "fwrite" || F->getName() == "fread" ||
940 F->getName() == "fgets" || F->getName() == "ungetc" ||
941 F->getName() == "fputc" ||
942 F->getName() == "fputs" || F->getName() == "putc" ||
943 F->getName() == "ftell" || F->getName() == "rewind" ||
944 F->getName() == "_IO_putc" || F->getName() == "fseek" ||
945 F->getName() == "fgetpos" || F->getName() == "fsetpos" ||
946 F->getName() == "printf" || F->getName() == "fprintf" ||
947 F->getName() == "sprintf" || F->getName() == "vprintf" ||
948 F->getName() == "vfprintf" || F->getName() == "vsprintf" ||
949 F->getName() == "scanf" || F->getName() == "fscanf" ||
950 F->getName() == "sscanf" || F->getName() == "__assert_fail" ||
951 F->getName() == "modf")
955 // These functions do induce points-to edges.
956 if (F->getName() == "llvm.memcpy" ||
957 F->getName() == "llvm.memmove" ||
958 F->getName() == "memmove") {
960 const FunctionType *FTy = F->getFunctionType();
961 if (FTy->getNumParams() > 1 &&
962 isa<PointerType>(FTy->getParamType(0)) &&
963 isa<PointerType>(FTy->getParamType(1))) {
965 // *Dest = *Src, which requires an artificial graph node to represent the
966 // constraint. It is broken up into *Dest = temp, temp = *Src
967 unsigned FirstArg = getNode(CS.getArgument(0));
968 unsigned SecondArg = getNode(CS.getArgument(1));
969 unsigned TempArg = GraphNodes.size();
970 GraphNodes.push_back(Node());
971 Constraints.push_back(Constraint(Constraint::Store,
973 Constraints.push_back(Constraint(Constraint::Load,
974 TempArg, SecondArg));
975 // In addition, Dest = Src
976 Constraints.push_back(Constraint(Constraint::Copy,
977 FirstArg, SecondArg));
983 if (F->getName() == "realloc" || F->getName() == "strchr" ||
984 F->getName() == "strrchr" || F->getName() == "strstr" ||
985 F->getName() == "strtok") {
986 const FunctionType *FTy = F->getFunctionType();
987 if (FTy->getNumParams() > 0 &&
988 isa<PointerType>(FTy->getParamType(0))) {
989 Constraints.push_back(Constraint(Constraint::Copy,
990 getNode(CS.getInstruction()),
991 getNode(CS.getArgument(0))));
1001 /// AnalyzeUsesOfFunction - Look at all of the users of the specified function.
1002 /// If this is used by anything complex (i.e., the address escapes), return
1004 bool Andersens::AnalyzeUsesOfFunction(Value *V) {
1006 if (!isa<PointerType>(V->getType())) return true;
1008 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
1009 if (isa<LoadInst>(*UI)) {
1011 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
1012 if (V == SI->getOperand(1)) {
1014 } else if (SI->getOperand(1)) {
1015 return true; // Storing the pointer
1017 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) {
1018 if (AnalyzeUsesOfFunction(GEP)) return true;
1019 } else if (isFreeCall(*UI)) {
1021 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
1022 // Make sure that this is just the function being called, not that it is
1023 // passing into the function.
1024 for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i)
1025 if (CI->getOperand(i) == V) return true;
1026 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
1027 // Make sure that this is just the function being called, not that it is
1028 // passing into the function.
1029 for (unsigned i = 3, e = II->getNumOperands(); i != e; ++i)
1030 if (II->getOperand(i) == V) return true;
1031 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
1032 if (CE->getOpcode() == Instruction::GetElementPtr ||
1033 CE->getOpcode() == Instruction::BitCast) {
1034 if (AnalyzeUsesOfFunction(CE))
1039 } else if (ICmpInst *ICI = dyn_cast<ICmpInst>(*UI)) {
1040 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
1041 return true; // Allow comparison against null.
1048 /// CollectConstraints - This stage scans the program, adding a constraint to
1049 /// the Constraints list for each instruction in the program that induces a
1050 /// constraint, and setting up the initial points-to graph.
1052 void Andersens::CollectConstraints(Module &M) {
1053 // First, the universal set points to itself.
1054 Constraints.push_back(Constraint(Constraint::AddressOf, UniversalSet,
1056 Constraints.push_back(Constraint(Constraint::Store, UniversalSet,
1059 // Next, the null pointer points to the null object.
1060 Constraints.push_back(Constraint(Constraint::AddressOf, NullPtr, NullObject));
1062 // Next, add any constraints on global variables and their initializers.
1063 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
1065 // Associate the address of the global object as pointing to the memory for
1066 // the global: &G = <G memory>
1067 unsigned ObjectIndex = getObject(I);
1068 Node *Object = &GraphNodes[ObjectIndex];
1069 Object->setValue(I);
1070 Constraints.push_back(Constraint(Constraint::AddressOf, getNodeValue(*I),
1073 if (I->hasDefinitiveInitializer()) {
1074 AddGlobalInitializerConstraints(ObjectIndex, I->getInitializer());
1076 // If it doesn't have an initializer (i.e. it's defined in another
1077 // translation unit), it points to the universal set.
1078 Constraints.push_back(Constraint(Constraint::Copy, ObjectIndex,
1083 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
1084 // Set up the return value node.
1085 if (isa<PointerType>(F->getFunctionType()->getReturnType()))
1086 GraphNodes[getReturnNode(F)].setValue(F);
1087 if (F->getFunctionType()->isVarArg())
1088 GraphNodes[getVarargNode(F)].setValue(F);
1090 // Set up incoming argument nodes.
1091 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
1093 if (isa<PointerType>(I->getType()))
1096 // At some point we should just add constraints for the escaping functions
1097 // at solve time, but this slows down solving. For now, we simply mark
1098 // address taken functions as escaping and treat them as external.
1099 if (!F->hasLocalLinkage() || AnalyzeUsesOfFunction(F))
1100 AddConstraintsForNonInternalLinkage(F);
1102 if (!F->isDeclaration()) {
1103 // Scan the function body, creating a memory object for each heap/stack
1104 // allocation in the body of the function and a node to represent all
1105 // pointer values defined by instructions and used as operands.
1108 // External functions that return pointers return the universal set.
1109 if (isa<PointerType>(F->getFunctionType()->getReturnType()))
1110 Constraints.push_back(Constraint(Constraint::Copy,
1114 // Any pointers that are passed into the function have the universal set
1115 // stored into them.
1116 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
1118 if (isa<PointerType>(I->getType())) {
1119 // Pointers passed into external functions could have anything stored
1121 Constraints.push_back(Constraint(Constraint::Store, getNode(I),
1123 // Memory objects passed into external function calls can have the
1124 // universal set point to them.
1126 Constraints.push_back(Constraint(Constraint::Copy,
1130 Constraints.push_back(Constraint(Constraint::Copy,
1136 // If this is an external varargs function, it can also store pointers
1137 // into any pointers passed through the varargs section.
1138 if (F->getFunctionType()->isVarArg())
1139 Constraints.push_back(Constraint(Constraint::Store, getVarargNode(F),
1143 NumConstraints += Constraints.size();
1147 void Andersens::visitInstruction(Instruction &I) {
1149 return; // This function is just a big assert.
1151 if (isa<BinaryOperator>(I))
1153 // Most instructions don't have any effect on pointer values.
1154 switch (I.getOpcode()) {
1155 case Instruction::Br:
1156 case Instruction::Switch:
1157 case Instruction::Unwind:
1158 case Instruction::Unreachable:
1159 case Instruction::ICmp:
1160 case Instruction::FCmp:
1163 // Is this something we aren't handling yet?
1164 errs() << "Unknown instruction: " << I;
1165 llvm_unreachable(0);
1169 void Andersens::visitAllocaInst(AllocaInst &I) {
1173 void Andersens::visitAlloc(Instruction &I) {
1174 unsigned ObjectIndex = getObject(&I);
1175 GraphNodes[ObjectIndex].setValue(&I);
1176 Constraints.push_back(Constraint(Constraint::AddressOf, getNodeValue(I),
1180 void Andersens::visitReturnInst(ReturnInst &RI) {
1181 if (RI.getNumOperands() && isa<PointerType>(RI.getOperand(0)->getType()))
1182 // return V --> <Copy/retval{F}/v>
1183 Constraints.push_back(Constraint(Constraint::Copy,
1184 getReturnNode(RI.getParent()->getParent()),
1185 getNode(RI.getOperand(0))));
1188 void Andersens::visitLoadInst(LoadInst &LI) {
1189 if (isa<PointerType>(LI.getType()))
1190 // P1 = load P2 --> <Load/P1/P2>
1191 Constraints.push_back(Constraint(Constraint::Load, getNodeValue(LI),
1192 getNode(LI.getOperand(0))));
1195 void Andersens::visitStoreInst(StoreInst &SI) {
1196 if (isa<PointerType>(SI.getOperand(0)->getType()))
1197 // store P1, P2 --> <Store/P2/P1>
1198 Constraints.push_back(Constraint(Constraint::Store,
1199 getNode(SI.getOperand(1)),
1200 getNode(SI.getOperand(0))));
1203 void Andersens::visitGetElementPtrInst(GetElementPtrInst &GEP) {
1204 // P1 = getelementptr P2, ... --> <Copy/P1/P2>
1205 Constraints.push_back(Constraint(Constraint::Copy, getNodeValue(GEP),
1206 getNode(GEP.getOperand(0))));
1209 void Andersens::visitPHINode(PHINode &PN) {
1210 if (isa<PointerType>(PN.getType())) {
1211 unsigned PNN = getNodeValue(PN);
1212 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1213 // P1 = phi P2, P3 --> <Copy/P1/P2>, <Copy/P1/P3>, ...
1214 Constraints.push_back(Constraint(Constraint::Copy, PNN,
1215 getNode(PN.getIncomingValue(i))));
1219 void Andersens::visitCastInst(CastInst &CI) {
1220 Value *Op = CI.getOperand(0);
1221 if (isa<PointerType>(CI.getType())) {
1222 if (isa<PointerType>(Op->getType())) {
1223 // P1 = cast P2 --> <Copy/P1/P2>
1224 Constraints.push_back(Constraint(Constraint::Copy, getNodeValue(CI),
1225 getNode(CI.getOperand(0))));
1227 // P1 = cast int --> <Copy/P1/Univ>
1229 Constraints.push_back(Constraint(Constraint::Copy, getNodeValue(CI),
1235 } else if (isa<PointerType>(Op->getType())) {
1236 // int = cast P1 --> <Copy/Univ/P1>
1238 Constraints.push_back(Constraint(Constraint::Copy,
1240 getNode(CI.getOperand(0))));
1242 getNode(CI.getOperand(0));
1247 void Andersens::visitSelectInst(SelectInst &SI) {
1248 if (isa<PointerType>(SI.getType())) {
1249 unsigned SIN = getNodeValue(SI);
1250 // P1 = select C, P2, P3 ---> <Copy/P1/P2>, <Copy/P1/P3>
1251 Constraints.push_back(Constraint(Constraint::Copy, SIN,
1252 getNode(SI.getOperand(1))));
1253 Constraints.push_back(Constraint(Constraint::Copy, SIN,
1254 getNode(SI.getOperand(2))));
1258 void Andersens::visitVAArg(VAArgInst &I) {
1259 llvm_unreachable("vaarg not handled yet!");
1262 /// AddConstraintsForCall - Add constraints for a call with actual arguments
1263 /// specified by CS to the function specified by F. Note that the types of
1264 /// arguments might not match up in the case where this is an indirect call and
1265 /// the function pointer has been casted. If this is the case, do something
1267 void Andersens::AddConstraintsForCall(CallSite CS, Function *F) {
1268 Value *CallValue = CS.getCalledValue();
1269 bool IsDeref = F == NULL;
1271 // If this is a call to an external function, try to handle it directly to get
1272 // some taste of context sensitivity.
1273 if (F && F->isDeclaration() && AddConstraintsForExternalCall(CS, F))
1276 if (isa<PointerType>(CS.getType())) {
1277 unsigned CSN = getNode(CS.getInstruction());
1278 if (!F || isa<PointerType>(F->getFunctionType()->getReturnType())) {
1280 Constraints.push_back(Constraint(Constraint::Load, CSN,
1281 getNode(CallValue), CallReturnPos));
1283 Constraints.push_back(Constraint(Constraint::Copy, CSN,
1284 getNode(CallValue) + CallReturnPos));
1286 // If the function returns a non-pointer value, handle this just like we
1287 // treat a nonpointer cast to pointer.
1288 Constraints.push_back(Constraint(Constraint::Copy, CSN,
1291 } else if (F && isa<PointerType>(F->getFunctionType()->getReturnType())) {
1293 Constraints.push_back(Constraint(Constraint::Copy,
1295 getNode(CallValue) + CallReturnPos));
1297 Constraints.push_back(Constraint(Constraint::Copy,
1298 getNode(CallValue) + CallReturnPos,
1305 CallSite::arg_iterator ArgI = CS.arg_begin(), ArgE = CS.arg_end();
1306 bool external = !F || F->isDeclaration();
1309 Function::arg_iterator AI = F->arg_begin(), AE = F->arg_end();
1310 for (; AI != AE && ArgI != ArgE; ++AI, ++ArgI)
1313 if (external && isa<PointerType>((*ArgI)->getType()))
1315 // Add constraint that ArgI can now point to anything due to
1316 // escaping, as can everything it points to. The second portion of
1317 // this should be taken care of by universal = *universal
1318 Constraints.push_back(Constraint(Constraint::Copy,
1323 if (isa<PointerType>(AI->getType())) {
1324 if (isa<PointerType>((*ArgI)->getType())) {
1325 // Copy the actual argument into the formal argument.
1326 Constraints.push_back(Constraint(Constraint::Copy, getNode(AI),
1329 Constraints.push_back(Constraint(Constraint::Copy, getNode(AI),
1332 } else if (isa<PointerType>((*ArgI)->getType())) {
1334 Constraints.push_back(Constraint(Constraint::Copy,
1338 Constraints.push_back(Constraint(Constraint::Copy,
1346 unsigned ArgPos = CallFirstArgPos;
1347 for (; ArgI != ArgE; ++ArgI) {
1348 if (isa<PointerType>((*ArgI)->getType())) {
1349 // Copy the actual argument into the formal argument.
1350 Constraints.push_back(Constraint(Constraint::Store,
1352 getNode(*ArgI), ArgPos++));
1354 Constraints.push_back(Constraint(Constraint::Store,
1355 getNode (CallValue),
1356 UniversalSet, ArgPos++));
1360 // Copy all pointers passed through the varargs section to the varargs node.
1361 if (F && F->getFunctionType()->isVarArg())
1362 for (; ArgI != ArgE; ++ArgI)
1363 if (isa<PointerType>((*ArgI)->getType()))
1364 Constraints.push_back(Constraint(Constraint::Copy, getVarargNode(F),
1366 // If more arguments are passed in than we track, just drop them on the floor.
1369 void Andersens::visitCallSite(CallSite CS) {
1370 if (isa<PointerType>(CS.getType()))
1371 getNodeValue(*CS.getInstruction());
1373 if (Function *F = CS.getCalledFunction()) {
1374 AddConstraintsForCall(CS, F);
1376 AddConstraintsForCall(CS, NULL);
1380 //===----------------------------------------------------------------------===//
1381 // Constraint Solving Phase
1382 //===----------------------------------------------------------------------===//
1384 /// intersects - Return true if the points-to set of this node intersects
1385 /// with the points-to set of the specified node.
1386 bool Andersens::Node::intersects(Node *N) const {
1387 return PointsTo->intersects(N->PointsTo);
1390 /// intersectsIgnoring - Return true if the points-to set of this node
1391 /// intersects with the points-to set of the specified node on any nodes
1392 /// except for the specified node to ignore.
1393 bool Andersens::Node::intersectsIgnoring(Node *N, unsigned Ignoring) const {
1394 // TODO: If we are only going to call this with the same value for Ignoring,
1395 // we should move the special values out of the points-to bitmap.
1396 bool WeHadIt = PointsTo->test(Ignoring);
1397 bool NHadIt = N->PointsTo->test(Ignoring);
1398 bool Result = false;
1400 PointsTo->reset(Ignoring);
1402 N->PointsTo->reset(Ignoring);
1403 Result = PointsTo->intersects(N->PointsTo);
1405 PointsTo->set(Ignoring);
1407 N->PointsTo->set(Ignoring);
1412 /// Clump together address taken variables so that the points-to sets use up
1413 /// less space and can be operated on faster.
1415 void Andersens::ClumpAddressTaken() {
1417 #define DEBUG_TYPE "anders-aa-renumber"
1418 std::vector<unsigned> Translate;
1419 std::vector<Node> NewGraphNodes;
1421 Translate.resize(GraphNodes.size());
1422 unsigned NewPos = 0;
1424 for (unsigned i = 0; i < Constraints.size(); ++i) {
1425 Constraint &C = Constraints[i];
1426 if (C.Type == Constraint::AddressOf) {
1427 GraphNodes[C.Src].AddressTaken = true;
1430 for (unsigned i = 0; i < NumberSpecialNodes; ++i) {
1431 unsigned Pos = NewPos++;
1433 NewGraphNodes.push_back(GraphNodes[i]);
1434 DEBUG(errs() << "Renumbering node " << i << " to node " << Pos << "\n");
1437 // I believe this ends up being faster than making two vectors and splicing
1439 for (unsigned i = NumberSpecialNodes; i < GraphNodes.size(); ++i) {
1440 if (GraphNodes[i].AddressTaken) {
1441 unsigned Pos = NewPos++;
1443 NewGraphNodes.push_back(GraphNodes[i]);
1444 DEBUG(errs() << "Renumbering node " << i << " to node " << Pos << "\n");
1448 for (unsigned i = NumberSpecialNodes; i < GraphNodes.size(); ++i) {
1449 if (!GraphNodes[i].AddressTaken) {
1450 unsigned Pos = NewPos++;
1452 NewGraphNodes.push_back(GraphNodes[i]);
1453 DEBUG(errs() << "Renumbering node " << i << " to node " << Pos << "\n");
1457 for (DenseMap<Value*, unsigned>::iterator Iter = ValueNodes.begin();
1458 Iter != ValueNodes.end();
1460 Iter->second = Translate[Iter->second];
1462 for (DenseMap<Value*, unsigned>::iterator Iter = ObjectNodes.begin();
1463 Iter != ObjectNodes.end();
1465 Iter->second = Translate[Iter->second];
1467 for (DenseMap<Function*, unsigned>::iterator Iter = ReturnNodes.begin();
1468 Iter != ReturnNodes.end();
1470 Iter->second = Translate[Iter->second];
1472 for (DenseMap<Function*, unsigned>::iterator Iter = VarargNodes.begin();
1473 Iter != VarargNodes.end();
1475 Iter->second = Translate[Iter->second];
1477 for (unsigned i = 0; i < Constraints.size(); ++i) {
1478 Constraint &C = Constraints[i];
1479 C.Src = Translate[C.Src];
1480 C.Dest = Translate[C.Dest];
1483 GraphNodes.swap(NewGraphNodes);
1485 #define DEBUG_TYPE "anders-aa"
1488 /// The technique used here is described in "Exploiting Pointer and Location
1489 /// Equivalence to Optimize Pointer Analysis. In the 14th International Static
1490 /// Analysis Symposium (SAS), August 2007." It is known as the "HVN" algorithm,
1491 /// and is equivalent to value numbering the collapsed constraint graph without
1492 /// evaluating unions. This is used as a pre-pass to HU in order to resolve
1493 /// first order pointer dereferences and speed up/reduce memory usage of HU.
1494 /// Running both is equivalent to HRU without the iteration
1495 /// HVN in more detail:
1496 /// Imagine the set of constraints was simply straight line code with no loops
1497 /// (we eliminate cycles, so there are no loops), such as:
1503 /// Applying value numbering to this code tells us:
1506 /// For HVN, this is as far as it goes. We assign new value numbers to every
1507 /// "address node", and every "reference node".
1508 /// To get the optimal result for this, we use a DFS + SCC (since all nodes in a
1509 /// cycle must have the same value number since the = operation is really
1510 /// inclusion, not overwrite), and value number nodes we receive points-to sets
1511 /// before we value our own node.
1512 /// The advantage of HU over HVN is that HU considers the inclusion property, so
1513 /// that if you have
1520 /// HU will determine that G == F == E. HVN will not, because it cannot prove
1521 /// that the points to information ends up being the same because they all
1522 /// receive &D from E anyway.
1524 void Andersens::HVN() {
1525 DEBUG(errs() << "Beginning HVN\n");
1526 // Build a predecessor graph. This is like our constraint graph with the
1527 // edges going in the opposite direction, and there are edges for all the
1528 // constraints, instead of just copy constraints. We also build implicit
1529 // edges for constraints are implied but not explicit. I.E for the constraint
1530 // a = &b, we add implicit edges *a = b. This helps us capture more cycles
1531 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
1532 Constraint &C = Constraints[i];
1533 if (C.Type == Constraint::AddressOf) {
1534 GraphNodes[C.Src].AddressTaken = true;
1535 GraphNodes[C.Src].Direct = false;
1538 unsigned AdrNode = C.Src + FirstAdrNode;
1539 if (!GraphNodes[C.Dest].PredEdges)
1540 GraphNodes[C.Dest].PredEdges = new SparseBitVector<>;
1541 GraphNodes[C.Dest].PredEdges->set(AdrNode);
1544 unsigned RefNode = C.Dest + FirstRefNode;
1545 if (!GraphNodes[RefNode].ImplicitPredEdges)
1546 GraphNodes[RefNode].ImplicitPredEdges = new SparseBitVector<>;
1547 GraphNodes[RefNode].ImplicitPredEdges->set(C.Src);
1548 } else if (C.Type == Constraint::Load) {
1549 if (C.Offset == 0) {
1551 if (!GraphNodes[C.Dest].PredEdges)
1552 GraphNodes[C.Dest].PredEdges = new SparseBitVector<>;
1553 GraphNodes[C.Dest].PredEdges->set(C.Src + FirstRefNode);
1555 GraphNodes[C.Dest].Direct = false;
1557 } else if (C.Type == Constraint::Store) {
1558 if (C.Offset == 0) {
1560 unsigned RefNode = C.Dest + FirstRefNode;
1561 if (!GraphNodes[RefNode].PredEdges)
1562 GraphNodes[RefNode].PredEdges = new SparseBitVector<>;
1563 GraphNodes[RefNode].PredEdges->set(C.Src);
1566 // Dest = Src edge and *Dest = *Src edge
1567 if (!GraphNodes[C.Dest].PredEdges)
1568 GraphNodes[C.Dest].PredEdges = new SparseBitVector<>;
1569 GraphNodes[C.Dest].PredEdges->set(C.Src);
1570 unsigned RefNode = C.Dest + FirstRefNode;
1571 if (!GraphNodes[RefNode].ImplicitPredEdges)
1572 GraphNodes[RefNode].ImplicitPredEdges = new SparseBitVector<>;
1573 GraphNodes[RefNode].ImplicitPredEdges->set(C.Src + FirstRefNode);
1577 // Do SCC finding first to condense our predecessor graph
1579 Node2DFS.insert(Node2DFS.begin(), GraphNodes.size(), 0);
1580 Node2Deleted.insert(Node2Deleted.begin(), GraphNodes.size(), false);
1581 Node2Visited.insert(Node2Visited.begin(), GraphNodes.size(), false);
1583 for (unsigned i = 0; i < FirstRefNode; ++i) {
1584 unsigned Node = VSSCCRep[i];
1585 if (!Node2Visited[Node])
1588 for (BitVectorMap::iterator Iter = Set2PEClass.begin();
1589 Iter != Set2PEClass.end();
1592 Set2PEClass.clear();
1594 Node2Deleted.clear();
1595 Node2Visited.clear();
1596 DEBUG(errs() << "Finished HVN\n");
1600 /// This is the workhorse of HVN value numbering. We combine SCC finding at the
1601 /// same time because it's easy.
1602 void Andersens::HVNValNum(unsigned NodeIndex) {
1603 unsigned MyDFS = DFSNumber++;
1604 Node *N = &GraphNodes[NodeIndex];
1605 Node2Visited[NodeIndex] = true;
1606 Node2DFS[NodeIndex] = MyDFS;
1608 // First process all our explicit edges
1610 for (SparseBitVector<>::iterator Iter = N->PredEdges->begin();
1611 Iter != N->PredEdges->end();
1613 unsigned j = VSSCCRep[*Iter];
1614 if (!Node2Deleted[j]) {
1615 if (!Node2Visited[j])
1617 if (Node2DFS[NodeIndex] > Node2DFS[j])
1618 Node2DFS[NodeIndex] = Node2DFS[j];
1622 // Now process all the implicit edges
1623 if (N->ImplicitPredEdges)
1624 for (SparseBitVector<>::iterator Iter = N->ImplicitPredEdges->begin();
1625 Iter != N->ImplicitPredEdges->end();
1627 unsigned j = VSSCCRep[*Iter];
1628 if (!Node2Deleted[j]) {
1629 if (!Node2Visited[j])
1631 if (Node2DFS[NodeIndex] > Node2DFS[j])
1632 Node2DFS[NodeIndex] = Node2DFS[j];
1636 // See if we found any cycles
1637 if (MyDFS == Node2DFS[NodeIndex]) {
1638 while (!SCCStack.empty() && Node2DFS[SCCStack.top()] >= MyDFS) {
1639 unsigned CycleNodeIndex = SCCStack.top();
1640 Node *CycleNode = &GraphNodes[CycleNodeIndex];
1641 VSSCCRep[CycleNodeIndex] = NodeIndex;
1643 N->Direct &= CycleNode->Direct;
1645 if (CycleNode->PredEdges) {
1647 N->PredEdges = new SparseBitVector<>;
1648 *(N->PredEdges) |= CycleNode->PredEdges;
1649 delete CycleNode->PredEdges;
1650 CycleNode->PredEdges = NULL;
1652 if (CycleNode->ImplicitPredEdges) {
1653 if (!N->ImplicitPredEdges)
1654 N->ImplicitPredEdges = new SparseBitVector<>;
1655 *(N->ImplicitPredEdges) |= CycleNode->ImplicitPredEdges;
1656 delete CycleNode->ImplicitPredEdges;
1657 CycleNode->ImplicitPredEdges = NULL;
1663 Node2Deleted[NodeIndex] = true;
1666 GraphNodes[NodeIndex].PointerEquivLabel = PEClass++;
1670 // Collect labels of successor nodes
1671 bool AllSame = true;
1672 unsigned First = ~0;
1673 SparseBitVector<> *Labels = new SparseBitVector<>;
1677 for (SparseBitVector<>::iterator Iter = N->PredEdges->begin();
1678 Iter != N->PredEdges->end();
1680 unsigned j = VSSCCRep[*Iter];
1681 unsigned Label = GraphNodes[j].PointerEquivLabel;
1682 // Ignore labels that are equal to us or non-pointers
1683 if (j == NodeIndex || Label == 0)
1685 if (First == (unsigned)~0)
1687 else if (First != Label)
1692 // We either have a non-pointer, a copy of an existing node, or a new node.
1693 // Assign the appropriate pointer equivalence label.
1694 if (Labels->empty()) {
1695 GraphNodes[NodeIndex].PointerEquivLabel = 0;
1696 } else if (AllSame) {
1697 GraphNodes[NodeIndex].PointerEquivLabel = First;
1699 GraphNodes[NodeIndex].PointerEquivLabel = Set2PEClass[Labels];
1700 if (GraphNodes[NodeIndex].PointerEquivLabel == 0) {
1701 unsigned EquivClass = PEClass++;
1702 Set2PEClass[Labels] = EquivClass;
1703 GraphNodes[NodeIndex].PointerEquivLabel = EquivClass;
1710 SCCStack.push(NodeIndex);
1714 /// The technique used here is described in "Exploiting Pointer and Location
1715 /// Equivalence to Optimize Pointer Analysis. In the 14th International Static
1716 /// Analysis Symposium (SAS), August 2007." It is known as the "HU" algorithm,
1717 /// and is equivalent to value numbering the collapsed constraint graph
1718 /// including evaluating unions.
1719 void Andersens::HU() {
1720 DEBUG(errs() << "Beginning HU\n");
1721 // Build a predecessor graph. This is like our constraint graph with the
1722 // edges going in the opposite direction, and there are edges for all the
1723 // constraints, instead of just copy constraints. We also build implicit
1724 // edges for constraints are implied but not explicit. I.E for the constraint
1725 // a = &b, we add implicit edges *a = b. This helps us capture more cycles
1726 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
1727 Constraint &C = Constraints[i];
1728 if (C.Type == Constraint::AddressOf) {
1729 GraphNodes[C.Src].AddressTaken = true;
1730 GraphNodes[C.Src].Direct = false;
1732 GraphNodes[C.Dest].PointsTo->set(C.Src);
1734 unsigned RefNode = C.Dest + FirstRefNode;
1735 if (!GraphNodes[RefNode].ImplicitPredEdges)
1736 GraphNodes[RefNode].ImplicitPredEdges = new SparseBitVector<>;
1737 GraphNodes[RefNode].ImplicitPredEdges->set(C.Src);
1738 GraphNodes[C.Src].PointedToBy->set(C.Dest);
1739 } else if (C.Type == Constraint::Load) {
1740 if (C.Offset == 0) {
1742 if (!GraphNodes[C.Dest].PredEdges)
1743 GraphNodes[C.Dest].PredEdges = new SparseBitVector<>;
1744 GraphNodes[C.Dest].PredEdges->set(C.Src + FirstRefNode);
1746 GraphNodes[C.Dest].Direct = false;
1748 } else if (C.Type == Constraint::Store) {
1749 if (C.Offset == 0) {
1751 unsigned RefNode = C.Dest + FirstRefNode;
1752 if (!GraphNodes[RefNode].PredEdges)
1753 GraphNodes[RefNode].PredEdges = new SparseBitVector<>;
1754 GraphNodes[RefNode].PredEdges->set(C.Src);
1757 // Dest = Src edge and *Dest = *Src edg
1758 if (!GraphNodes[C.Dest].PredEdges)
1759 GraphNodes[C.Dest].PredEdges = new SparseBitVector<>;
1760 GraphNodes[C.Dest].PredEdges->set(C.Src);
1761 unsigned RefNode = C.Dest + FirstRefNode;
1762 if (!GraphNodes[RefNode].ImplicitPredEdges)
1763 GraphNodes[RefNode].ImplicitPredEdges = new SparseBitVector<>;
1764 GraphNodes[RefNode].ImplicitPredEdges->set(C.Src + FirstRefNode);
1768 // Do SCC finding first to condense our predecessor graph
1770 Node2DFS.insert(Node2DFS.begin(), GraphNodes.size(), 0);
1771 Node2Deleted.insert(Node2Deleted.begin(), GraphNodes.size(), false);
1772 Node2Visited.insert(Node2Visited.begin(), GraphNodes.size(), false);
1774 for (unsigned i = 0; i < FirstRefNode; ++i) {
1775 if (FindNode(i) == i) {
1776 unsigned Node = VSSCCRep[i];
1777 if (!Node2Visited[Node])
1782 // Reset tables for actual labeling
1784 Node2Visited.clear();
1785 Node2Deleted.clear();
1786 // Pre-grow our densemap so that we don't get really bad behavior
1787 Set2PEClass.resize(GraphNodes.size());
1789 // Visit the condensed graph and generate pointer equivalence labels.
1790 Node2Visited.insert(Node2Visited.begin(), GraphNodes.size(), false);
1791 for (unsigned i = 0; i < FirstRefNode; ++i) {
1792 if (FindNode(i) == i) {
1793 unsigned Node = VSSCCRep[i];
1794 if (!Node2Visited[Node])
1798 // PEClass nodes will be deleted by the deleting of N->PointsTo in our caller.
1799 Set2PEClass.clear();
1800 DEBUG(errs() << "Finished HU\n");
1804 /// Implementation of standard Tarjan SCC algorithm as modified by Nuutilla.
1805 void Andersens::Condense(unsigned NodeIndex) {
1806 unsigned MyDFS = DFSNumber++;
1807 Node *N = &GraphNodes[NodeIndex];
1808 Node2Visited[NodeIndex] = true;
1809 Node2DFS[NodeIndex] = MyDFS;
1811 // First process all our explicit edges
1813 for (SparseBitVector<>::iterator Iter = N->PredEdges->begin();
1814 Iter != N->PredEdges->end();
1816 unsigned j = VSSCCRep[*Iter];
1817 if (!Node2Deleted[j]) {
1818 if (!Node2Visited[j])
1820 if (Node2DFS[NodeIndex] > Node2DFS[j])
1821 Node2DFS[NodeIndex] = Node2DFS[j];
1825 // Now process all the implicit edges
1826 if (N->ImplicitPredEdges)
1827 for (SparseBitVector<>::iterator Iter = N->ImplicitPredEdges->begin();
1828 Iter != N->ImplicitPredEdges->end();
1830 unsigned j = VSSCCRep[*Iter];
1831 if (!Node2Deleted[j]) {
1832 if (!Node2Visited[j])
1834 if (Node2DFS[NodeIndex] > Node2DFS[j])
1835 Node2DFS[NodeIndex] = Node2DFS[j];
1839 // See if we found any cycles
1840 if (MyDFS == Node2DFS[NodeIndex]) {
1841 while (!SCCStack.empty() && Node2DFS[SCCStack.top()] >= MyDFS) {
1842 unsigned CycleNodeIndex = SCCStack.top();
1843 Node *CycleNode = &GraphNodes[CycleNodeIndex];
1844 VSSCCRep[CycleNodeIndex] = NodeIndex;
1846 N->Direct &= CycleNode->Direct;
1848 *(N->PointsTo) |= CycleNode->PointsTo;
1849 delete CycleNode->PointsTo;
1850 CycleNode->PointsTo = NULL;
1851 if (CycleNode->PredEdges) {
1853 N->PredEdges = new SparseBitVector<>;
1854 *(N->PredEdges) |= CycleNode->PredEdges;
1855 delete CycleNode->PredEdges;
1856 CycleNode->PredEdges = NULL;
1858 if (CycleNode->ImplicitPredEdges) {
1859 if (!N->ImplicitPredEdges)
1860 N->ImplicitPredEdges = new SparseBitVector<>;
1861 *(N->ImplicitPredEdges) |= CycleNode->ImplicitPredEdges;
1862 delete CycleNode->ImplicitPredEdges;
1863 CycleNode->ImplicitPredEdges = NULL;
1868 Node2Deleted[NodeIndex] = true;
1870 // Set up number of incoming edges for other nodes
1872 for (SparseBitVector<>::iterator Iter = N->PredEdges->begin();
1873 Iter != N->PredEdges->end();
1875 ++GraphNodes[VSSCCRep[*Iter]].NumInEdges;
1877 SCCStack.push(NodeIndex);
1881 void Andersens::HUValNum(unsigned NodeIndex) {
1882 Node *N = &GraphNodes[NodeIndex];
1883 Node2Visited[NodeIndex] = true;
1885 // Eliminate dereferences of non-pointers for those non-pointers we have
1886 // already identified. These are ref nodes whose non-ref node:
1887 // 1. Has already been visited determined to point to nothing (and thus, a
1888 // dereference of it must point to nothing)
1889 // 2. Any direct node with no predecessor edges in our graph and with no
1890 // points-to set (since it can't point to anything either, being that it
1891 // receives no points-to sets and has none).
1892 if (NodeIndex >= FirstRefNode) {
1893 unsigned j = VSSCCRep[FindNode(NodeIndex - FirstRefNode)];
1894 if ((Node2Visited[j] && !GraphNodes[j].PointerEquivLabel)
1895 || (GraphNodes[j].Direct && !GraphNodes[j].PredEdges
1896 && GraphNodes[j].PointsTo->empty())){
1900 // Process all our explicit edges
1902 for (SparseBitVector<>::iterator Iter = N->PredEdges->begin();
1903 Iter != N->PredEdges->end();
1905 unsigned j = VSSCCRep[*Iter];
1906 if (!Node2Visited[j])
1909 // If this edge turned out to be the same as us, or got no pointer
1910 // equivalence label (and thus points to nothing) , just decrement our
1911 // incoming edges and continue.
1912 if (j == NodeIndex || GraphNodes[j].PointerEquivLabel == 0) {
1913 --GraphNodes[j].NumInEdges;
1917 *(N->PointsTo) |= GraphNodes[j].PointsTo;
1919 // If we didn't end up storing this in the hash, and we're done with all
1920 // the edges, we don't need the points-to set anymore.
1921 --GraphNodes[j].NumInEdges;
1922 if (!GraphNodes[j].NumInEdges && !GraphNodes[j].StoredInHash) {
1923 delete GraphNodes[j].PointsTo;
1924 GraphNodes[j].PointsTo = NULL;
1927 // If this isn't a direct node, generate a fresh variable.
1929 N->PointsTo->set(FirstRefNode + NodeIndex);
1932 // See If we have something equivalent to us, if not, generate a new
1933 // equivalence class.
1934 if (N->PointsTo->empty()) {
1939 N->PointerEquivLabel = Set2PEClass[N->PointsTo];
1940 if (N->PointerEquivLabel == 0) {
1941 unsigned EquivClass = PEClass++;
1942 N->StoredInHash = true;
1943 Set2PEClass[N->PointsTo] = EquivClass;
1944 N->PointerEquivLabel = EquivClass;
1947 N->PointerEquivLabel = PEClass++;
1952 /// Rewrite our list of constraints so that pointer equivalent nodes are
1953 /// replaced by their the pointer equivalence class representative.
1954 void Andersens::RewriteConstraints() {
1955 std::vector<Constraint> NewConstraints;
1956 DenseSet<Constraint, ConstraintKeyInfo> Seen;
1958 PEClass2Node.clear();
1959 PENLEClass2Node.clear();
1961 // We may have from 1 to Graphnodes + 1 equivalence classes.
1962 PEClass2Node.insert(PEClass2Node.begin(), GraphNodes.size() + 1, -1);
1963 PENLEClass2Node.insert(PENLEClass2Node.begin(), GraphNodes.size() + 1, -1);
1965 // Rewrite constraints, ignoring non-pointer constraints, uniting equivalent
1966 // nodes, and rewriting constraints to use the representative nodes.
1967 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
1968 Constraint &C = Constraints[i];
1969 unsigned RHSNode = FindNode(C.Src);
1970 unsigned LHSNode = FindNode(C.Dest);
1971 unsigned RHSLabel = GraphNodes[VSSCCRep[RHSNode]].PointerEquivLabel;
1972 unsigned LHSLabel = GraphNodes[VSSCCRep[LHSNode]].PointerEquivLabel;
1974 // First we try to eliminate constraints for things we can prove don't point
1976 if (LHSLabel == 0) {
1977 DEBUG(PrintNode(&GraphNodes[LHSNode]));
1978 DEBUG(errs() << " is a non-pointer, ignoring constraint.\n");
1981 if (RHSLabel == 0) {
1982 DEBUG(PrintNode(&GraphNodes[RHSNode]));
1983 DEBUG(errs() << " is a non-pointer, ignoring constraint.\n");
1986 // This constraint may be useless, and it may become useless as we translate
1988 if (C.Src == C.Dest && C.Type == Constraint::Copy)
1991 C.Src = FindEquivalentNode(RHSNode, RHSLabel);
1992 C.Dest = FindEquivalentNode(FindNode(LHSNode), LHSLabel);
1993 if ((C.Src == C.Dest && C.Type == Constraint::Copy)
1998 NewConstraints.push_back(C);
2000 Constraints.swap(NewConstraints);
2001 PEClass2Node.clear();
2004 /// See if we have a node that is pointer equivalent to the one being asked
2005 /// about, and if so, unite them and return the equivalent node. Otherwise,
2006 /// return the original node.
2007 unsigned Andersens::FindEquivalentNode(unsigned NodeIndex,
2008 unsigned NodeLabel) {
2009 if (!GraphNodes[NodeIndex].AddressTaken) {
2010 if (PEClass2Node[NodeLabel] != -1) {
2011 // We found an existing node with the same pointer label, so unify them.
2012 // We specifically request that Union-By-Rank not be used so that
2013 // PEClass2Node[NodeLabel] U= NodeIndex and not the other way around.
2014 return UniteNodes(PEClass2Node[NodeLabel], NodeIndex, false);
2016 PEClass2Node[NodeLabel] = NodeIndex;
2017 PENLEClass2Node[NodeLabel] = NodeIndex;
2019 } else if (PENLEClass2Node[NodeLabel] == -1) {
2020 PENLEClass2Node[NodeLabel] = NodeIndex;
2026 void Andersens::PrintLabels() const {
2027 for (unsigned i = 0; i < GraphNodes.size(); ++i) {
2028 if (i < FirstRefNode) {
2029 PrintNode(&GraphNodes[i]);
2030 } else if (i < FirstAdrNode) {
2031 DEBUG(errs() << "REF(");
2032 PrintNode(&GraphNodes[i-FirstRefNode]);
2033 DEBUG(errs() <<")");
2035 DEBUG(errs() << "ADR(");
2036 PrintNode(&GraphNodes[i-FirstAdrNode]);
2037 DEBUG(errs() <<")");
2040 DEBUG(errs() << " has pointer label " << GraphNodes[i].PointerEquivLabel
2041 << " and SCC rep " << VSSCCRep[i]
2042 << " and is " << (GraphNodes[i].Direct ? "Direct" : "Not direct")
2047 /// The technique used here is described in "The Ant and the
2048 /// Grasshopper: Fast and Accurate Pointer Analysis for Millions of
2049 /// Lines of Code. In Programming Language Design and Implementation
2050 /// (PLDI), June 2007." It is known as the "HCD" (Hybrid Cycle
2051 /// Detection) algorithm. It is called a hybrid because it performs an
2052 /// offline analysis and uses its results during the solving (online)
2053 /// phase. This is just the offline portion; the results of this
2054 /// operation are stored in SDT and are later used in SolveContraints()
2055 /// and UniteNodes().
2056 void Andersens::HCD() {
2057 DEBUG(errs() << "Starting HCD.\n");
2058 HCDSCCRep.resize(GraphNodes.size());
2060 for (unsigned i = 0; i < GraphNodes.size(); ++i) {
2061 GraphNodes[i].Edges = new SparseBitVector<>;
2065 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
2066 Constraint &C = Constraints[i];
2067 assert (C.Src < GraphNodes.size() && C.Dest < GraphNodes.size());
2068 if (C.Type == Constraint::AddressOf) {
2070 } else if (C.Type == Constraint::Load) {
2072 GraphNodes[C.Dest].Edges->set(C.Src + FirstRefNode);
2073 } else if (C.Type == Constraint::Store) {
2075 GraphNodes[C.Dest + FirstRefNode].Edges->set(C.Src);
2077 GraphNodes[C.Dest].Edges->set(C.Src);
2081 Node2DFS.insert(Node2DFS.begin(), GraphNodes.size(), 0);
2082 Node2Deleted.insert(Node2Deleted.begin(), GraphNodes.size(), false);
2083 Node2Visited.insert(Node2Visited.begin(), GraphNodes.size(), false);
2084 SDT.insert(SDT.begin(), GraphNodes.size() / 2, -1);
2087 for (unsigned i = 0; i < GraphNodes.size(); ++i) {
2088 unsigned Node = HCDSCCRep[i];
2089 if (!Node2Deleted[Node])
2093 for (unsigned i = 0; i < GraphNodes.size(); ++i)
2094 if (GraphNodes[i].Edges != NULL) {
2095 delete GraphNodes[i].Edges;
2096 GraphNodes[i].Edges = NULL;
2099 while( !SCCStack.empty() )
2103 Node2Visited.clear();
2104 Node2Deleted.clear();
2106 DEBUG(errs() << "HCD complete.\n");
2109 // Component of HCD:
2110 // Use Nuutila's variant of Tarjan's algorithm to detect
2111 // Strongly-Connected Components (SCCs). For non-trivial SCCs
2112 // containing ref nodes, insert the appropriate information in SDT.
2113 void Andersens::Search(unsigned Node) {
2114 unsigned MyDFS = DFSNumber++;
2116 Node2Visited[Node] = true;
2117 Node2DFS[Node] = MyDFS;
2119 for (SparseBitVector<>::iterator Iter = GraphNodes[Node].Edges->begin(),
2120 End = GraphNodes[Node].Edges->end();
2123 unsigned J = HCDSCCRep[*Iter];
2124 assert(GraphNodes[J].isRep() && "Debug check; must be representative");
2125 if (!Node2Deleted[J]) {
2126 if (!Node2Visited[J])
2128 if (Node2DFS[Node] > Node2DFS[J])
2129 Node2DFS[Node] = Node2DFS[J];
2133 if( MyDFS != Node2DFS[Node] ) {
2134 SCCStack.push(Node);
2138 // This node is the root of a SCC, so process it.
2140 // If the SCC is "non-trivial" (not a singleton) and contains a reference
2141 // node, we place this SCC into SDT. We unite the nodes in any case.
2142 if (!SCCStack.empty() && Node2DFS[SCCStack.top()] >= MyDFS) {
2143 SparseBitVector<> SCC;
2147 bool Ref = (Node >= FirstRefNode);
2149 Node2Deleted[Node] = true;
2152 unsigned P = SCCStack.top(); SCCStack.pop();
2153 Ref |= (P >= FirstRefNode);
2155 HCDSCCRep[P] = Node;
2156 } while (!SCCStack.empty() && Node2DFS[SCCStack.top()] >= MyDFS);
2159 unsigned Rep = SCC.find_first();
2160 assert(Rep < FirstRefNode && "The SCC didn't have a non-Ref node!");
2162 SparseBitVector<>::iterator i = SCC.begin();
2164 // Skip over the non-ref nodes
2165 while( *i < FirstRefNode )
2168 while( i != SCC.end() )
2169 SDT[ (*i++) - FirstRefNode ] = Rep;
2175 /// Optimize the constraints by performing offline variable substitution and
2176 /// other optimizations.
2177 void Andersens::OptimizeConstraints() {
2178 DEBUG(errs() << "Beginning constraint optimization\n");
2182 // Function related nodes need to stay in the same relative position and can't
2183 // be location equivalent.
2184 for (std::map<unsigned, unsigned>::iterator Iter = MaxK.begin();
2187 for (unsigned i = Iter->first;
2188 i != Iter->first + Iter->second;
2190 GraphNodes[i].AddressTaken = true;
2191 GraphNodes[i].Direct = false;
2195 ClumpAddressTaken();
2196 FirstRefNode = GraphNodes.size();
2197 FirstAdrNode = FirstRefNode + GraphNodes.size();
2198 GraphNodes.insert(GraphNodes.end(), 2 * GraphNodes.size(),
2200 VSSCCRep.resize(GraphNodes.size());
2201 for (unsigned i = 0; i < GraphNodes.size(); ++i) {
2205 for (unsigned i = 0; i < GraphNodes.size(); ++i) {
2206 Node *N = &GraphNodes[i];
2207 delete N->PredEdges;
2208 N->PredEdges = NULL;
2209 delete N->ImplicitPredEdges;
2210 N->ImplicitPredEdges = NULL;
2213 #define DEBUG_TYPE "anders-aa-labels"
2214 DEBUG(PrintLabels());
2216 #define DEBUG_TYPE "anders-aa"
2217 RewriteConstraints();
2218 // Delete the adr nodes.
2219 GraphNodes.resize(FirstRefNode * 2);
2222 for (unsigned i = 0; i < GraphNodes.size(); ++i) {
2223 Node *N = &GraphNodes[i];
2224 if (FindNode(i) == i) {
2225 N->PointsTo = new SparseBitVector<>;
2226 N->PointedToBy = new SparseBitVector<>;
2230 N->PointerEquivLabel = 0;
2234 #define DEBUG_TYPE "anders-aa-labels"
2235 DEBUG(PrintLabels());
2237 #define DEBUG_TYPE "anders-aa"
2238 RewriteConstraints();
2239 for (unsigned i = 0; i < GraphNodes.size(); ++i) {
2240 if (FindNode(i) == i) {
2241 Node *N = &GraphNodes[i];
2244 delete N->PredEdges;
2245 N->PredEdges = NULL;
2246 delete N->ImplicitPredEdges;
2247 N->ImplicitPredEdges = NULL;
2248 delete N->PointedToBy;
2249 N->PointedToBy = NULL;
2253 // perform Hybrid Cycle Detection (HCD)
2257 // No longer any need for the upper half of GraphNodes (for ref nodes).
2258 GraphNodes.erase(GraphNodes.begin() + FirstRefNode, GraphNodes.end());
2262 DEBUG(errs() << "Finished constraint optimization\n");
2267 /// Unite pointer but not location equivalent variables, now that the constraint
2269 void Andersens::UnitePointerEquivalences() {
2270 DEBUG(errs() << "Uniting remaining pointer equivalences\n");
2271 for (unsigned i = 0; i < GraphNodes.size(); ++i) {
2272 if (GraphNodes[i].AddressTaken && GraphNodes[i].isRep()) {
2273 unsigned Label = GraphNodes[i].PointerEquivLabel;
2275 if (Label && PENLEClass2Node[Label] != -1)
2276 UniteNodes(i, PENLEClass2Node[Label]);
2279 DEBUG(errs() << "Finished remaining pointer equivalences\n");
2280 PENLEClass2Node.clear();
2283 /// Create the constraint graph used for solving points-to analysis.
2285 void Andersens::CreateConstraintGraph() {
2286 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
2287 Constraint &C = Constraints[i];
2288 assert (C.Src < GraphNodes.size() && C.Dest < GraphNodes.size());
2289 if (C.Type == Constraint::AddressOf)
2290 GraphNodes[C.Dest].PointsTo->set(C.Src);
2291 else if (C.Type == Constraint::Load)
2292 GraphNodes[C.Src].Constraints.push_back(C);
2293 else if (C.Type == Constraint::Store)
2294 GraphNodes[C.Dest].Constraints.push_back(C);
2295 else if (C.Offset != 0)
2296 GraphNodes[C.Src].Constraints.push_back(C);
2298 GraphNodes[C.Src].Edges->set(C.Dest);
2302 // Perform DFS and cycle detection.
2303 bool Andersens::QueryNode(unsigned Node) {
2304 assert(GraphNodes[Node].isRep() && "Querying a non-rep node");
2305 unsigned OurDFS = ++DFSNumber;
2306 SparseBitVector<> ToErase;
2307 SparseBitVector<> NewEdges;
2308 Tarjan2DFS[Node] = OurDFS;
2310 // Changed denotes a change from a recursive call that we will bubble up.
2311 // Merged is set if we actually merge a node ourselves.
2312 bool Changed = false, Merged = false;
2314 for (SparseBitVector<>::iterator bi = GraphNodes[Node].Edges->begin();
2315 bi != GraphNodes[Node].Edges->end();
2317 unsigned RepNode = FindNode(*bi);
2318 // If this edge points to a non-representative node but we are
2319 // already planning to add an edge to its representative, we have no
2320 // need for this edge anymore.
2321 if (RepNode != *bi && NewEdges.test(RepNode)){
2326 // Continue about our DFS.
2327 if (!Tarjan2Deleted[RepNode]){
2328 if (Tarjan2DFS[RepNode] == 0) {
2329 Changed |= QueryNode(RepNode);
2330 // May have been changed by QueryNode
2331 RepNode = FindNode(RepNode);
2333 if (Tarjan2DFS[RepNode] < Tarjan2DFS[Node])
2334 Tarjan2DFS[Node] = Tarjan2DFS[RepNode];
2337 // We may have just discovered that this node is part of a cycle, in
2338 // which case we can also erase it.
2339 if (RepNode != *bi) {
2341 NewEdges.set(RepNode);
2345 GraphNodes[Node].Edges->intersectWithComplement(ToErase);
2346 GraphNodes[Node].Edges |= NewEdges;
2348 // If this node is a root of a non-trivial SCC, place it on our
2349 // worklist to be processed.
2350 if (OurDFS == Tarjan2DFS[Node]) {
2351 while (!SCCStack.empty() && Tarjan2DFS[SCCStack.top()] >= OurDFS) {
2352 Node = UniteNodes(Node, SCCStack.top());
2357 Tarjan2Deleted[Node] = true;
2360 NextWL->insert(&GraphNodes[Node]);
2362 SCCStack.push(Node);
2365 return(Changed | Merged);
2368 /// SolveConstraints - This stage iteratively processes the constraints list
2369 /// propagating constraints (adding edges to the Nodes in the points-to graph)
2370 /// until a fixed point is reached.
2372 /// We use a variant of the technique called "Lazy Cycle Detection", which is
2373 /// described in "The Ant and the Grasshopper: Fast and Accurate Pointer
2374 /// Analysis for Millions of Lines of Code. In Programming Language Design and
2375 /// Implementation (PLDI), June 2007."
2376 /// The paper describes performing cycle detection one node at a time, which can
2377 /// be expensive if there are no cycles, but there are long chains of nodes that
2378 /// it heuristically believes are cycles (because it will DFS from each node
2379 /// without state from previous nodes).
2380 /// Instead, we use the heuristic to build a worklist of nodes to check, then
2381 /// cycle detect them all at the same time to do this more cheaply. This
2382 /// catches cycles slightly later than the original technique did, but does it
2383 /// make significantly cheaper.
2385 void Andersens::SolveConstraints() {
2389 OptimizeConstraints();
2391 #define DEBUG_TYPE "anders-aa-constraints"
2392 DEBUG(PrintConstraints());
2394 #define DEBUG_TYPE "anders-aa"
2396 for (unsigned i = 0; i < GraphNodes.size(); ++i) {
2397 Node *N = &GraphNodes[i];
2398 N->PointsTo = new SparseBitVector<>;
2399 N->OldPointsTo = new SparseBitVector<>;
2400 N->Edges = new SparseBitVector<>;
2402 CreateConstraintGraph();
2403 UnitePointerEquivalences();
2404 assert(SCCStack.empty() && "SCC Stack should be empty by now!");
2406 Node2Deleted.clear();
2407 Node2DFS.insert(Node2DFS.begin(), GraphNodes.size(), 0);
2408 Node2Deleted.insert(Node2Deleted.begin(), GraphNodes.size(), false);
2410 DenseSet<Constraint, ConstraintKeyInfo> Seen;
2411 DenseSet<std::pair<unsigned,unsigned>, PairKeyInfo> EdgesChecked;
2413 // Order graph and add initial nodes to work list.
2414 for (unsigned i = 0; i < GraphNodes.size(); ++i) {
2415 Node *INode = &GraphNodes[i];
2417 // Add to work list if it's a representative and can contribute to the
2418 // calculation right now.
2419 if (INode->isRep() && !INode->PointsTo->empty()
2420 && (!INode->Edges->empty() || !INode->Constraints.empty())) {
2422 CurrWL->insert(INode);
2425 std::queue<unsigned int> TarjanWL;
2427 // "Rep and special variables" - in order for HCD to maintain conservative
2428 // results when !FULL_UNIVERSAL, we need to treat the special variables in
2429 // the same way that the !FULL_UNIVERSAL tweak does throughout the rest of
2430 // the analysis - it's ok to add edges from the special nodes, but never
2431 // *to* the special nodes.
2432 std::vector<unsigned int> RSV;
2434 while( !CurrWL->empty() ) {
2435 DEBUG(errs() << "Starting iteration #" << ++NumIters << "\n");
2438 unsigned CurrNodeIndex;
2440 // Actual cycle checking code. We cycle check all of the lazy cycle
2441 // candidates from the last iteration in one go.
2442 if (!TarjanWL.empty()) {
2446 Tarjan2Deleted.clear();
2447 while (!TarjanWL.empty()) {
2448 unsigned int ToTarjan = TarjanWL.front();
2450 if (!Tarjan2Deleted[ToTarjan]
2451 && GraphNodes[ToTarjan].isRep()
2452 && Tarjan2DFS[ToTarjan] == 0)
2453 QueryNode(ToTarjan);
2457 // Add to work list if it's a representative and can contribute to the
2458 // calculation right now.
2459 while( (CurrNode = CurrWL->pop()) != NULL ) {
2460 CurrNodeIndex = CurrNode - &GraphNodes[0];
2464 // Figure out the changed points to bits
2465 SparseBitVector<> CurrPointsTo;
2466 CurrPointsTo.intersectWithComplement(CurrNode->PointsTo,
2467 CurrNode->OldPointsTo);
2468 if (CurrPointsTo.empty())
2471 *(CurrNode->OldPointsTo) |= CurrPointsTo;
2473 // Check the offline-computed equivalencies from HCD.
2477 if (SDT[CurrNodeIndex] >= 0) {
2479 Rep = FindNode(SDT[CurrNodeIndex]);
2484 for (SparseBitVector<>::iterator bi = CurrPointsTo.begin();
2485 bi != CurrPointsTo.end(); ++bi) {
2486 unsigned Node = FindNode(*bi);
2488 if (Node < NumberSpecialNodes) {
2489 RSV.push_back(Node);
2493 Rep = UniteNodes(Rep,Node);
2499 NextWL->insert(&GraphNodes[Rep]);
2501 if ( ! CurrNode->isRep() )
2507 /* Now process the constraints for this node. */
2508 for (std::list<Constraint>::iterator li = CurrNode->Constraints.begin();
2509 li != CurrNode->Constraints.end(); ) {
2510 li->Src = FindNode(li->Src);
2511 li->Dest = FindNode(li->Dest);
2513 // Delete redundant constraints
2514 if( Seen.count(*li) ) {
2515 std::list<Constraint>::iterator lk = li; li++;
2517 CurrNode->Constraints.erase(lk);
2523 // Src and Dest will be the vars we are going to process.
2524 // This may look a bit ugly, but what it does is allow us to process
2525 // both store and load constraints with the same code.
2526 // Load constraints say that every member of our RHS solution has K
2527 // added to it, and that variable gets an edge to LHS. We also union
2528 // RHS+K's solution into the LHS solution.
2529 // Store constraints say that every member of our LHS solution has K
2530 // added to it, and that variable gets an edge from RHS. We also union
2531 // RHS's solution into the LHS+K solution.
2534 unsigned K = li->Offset;
2535 unsigned CurrMember;
2536 if (li->Type == Constraint::Load) {
2539 } else if (li->Type == Constraint::Store) {
2543 // TODO Handle offseted copy constraint
2548 // See if we can use Hybrid Cycle Detection (that is, check
2549 // if it was a statically detected offline equivalence that
2550 // involves pointers; if so, remove the redundant constraints).
2551 if( SCC && K == 0 ) {
2555 if (GraphNodes[*Src].Edges->test_and_set(*Dest))
2556 if (GraphNodes[*Dest].PointsTo |= *(GraphNodes[*Src].PointsTo))
2557 NextWL->insert(&GraphNodes[*Dest]);
2559 for (unsigned i=0; i < RSV.size(); ++i) {
2560 CurrMember = RSV[i];
2562 if (*Dest < NumberSpecialNodes)
2564 if (GraphNodes[*Src].Edges->test_and_set(*Dest))
2565 if (GraphNodes[*Dest].PointsTo |= *(GraphNodes[*Src].PointsTo))
2566 NextWL->insert(&GraphNodes[*Dest]);
2569 // since all future elements of the points-to set will be
2570 // equivalent to the current ones, the complex constraints
2571 // become redundant.
2573 std::list<Constraint>::iterator lk = li; li++;
2575 // In this case, we can still erase the constraints when the
2576 // elements of the points-to sets are referenced by *Dest,
2577 // but not when they are referenced by *Src (i.e. for a Load
2578 // constraint). This is because if another special variable is
2579 // put into the points-to set later, we still need to add the
2580 // new edge from that special variable.
2581 if( lk->Type != Constraint::Load)
2583 GraphNodes[CurrNodeIndex].Constraints.erase(lk);
2585 const SparseBitVector<> &Solution = CurrPointsTo;
2587 for (SparseBitVector<>::iterator bi = Solution.begin();
2588 bi != Solution.end();
2592 // Need to increment the member by K since that is where we are
2593 // supposed to copy to/from. Note that in positive weight cycles,
2594 // which occur in address taking of fields, K can go past
2595 // MaxK[CurrMember] elements, even though that is all it could point
2597 if (K > 0 && K > MaxK[CurrMember])
2600 CurrMember = FindNode(CurrMember + K);
2602 // Add an edge to the graph, so we can just do regular
2603 // bitmap ior next time. It may also let us notice a cycle.
2605 if (*Dest < NumberSpecialNodes)
2608 if (GraphNodes[*Src].Edges->test_and_set(*Dest))
2609 if (GraphNodes[*Dest].PointsTo |= *(GraphNodes[*Src].PointsTo))
2610 NextWL->insert(&GraphNodes[*Dest]);
2616 SparseBitVector<> NewEdges;
2617 SparseBitVector<> ToErase;
2619 // Now all we have left to do is propagate points-to info along the
2620 // edges, erasing the redundant edges.
2621 for (SparseBitVector<>::iterator bi = CurrNode->Edges->begin();
2622 bi != CurrNode->Edges->end();
2625 unsigned DestVar = *bi;
2626 unsigned Rep = FindNode(DestVar);
2628 // If we ended up with this node as our destination, or we've already
2629 // got an edge for the representative, delete the current edge.
2630 if (Rep == CurrNodeIndex ||
2631 (Rep != DestVar && NewEdges.test(Rep))) {
2632 ToErase.set(DestVar);
2636 std::pair<unsigned,unsigned> edge(CurrNodeIndex,Rep);
2638 // This is where we do lazy cycle detection.
2639 // If this is a cycle candidate (equal points-to sets and this
2640 // particular edge has not been cycle-checked previously), add to the
2641 // list to check for cycles on the next iteration.
2642 if (!EdgesChecked.count(edge) &&
2643 *(GraphNodes[Rep].PointsTo) == *(CurrNode->PointsTo)) {
2644 EdgesChecked.insert(edge);
2647 // Union the points-to sets into the dest
2649 if (Rep >= NumberSpecialNodes)
2651 if (GraphNodes[Rep].PointsTo |= CurrPointsTo) {
2652 NextWL->insert(&GraphNodes[Rep]);
2654 // If this edge's destination was collapsed, rewrite the edge.
2655 if (Rep != DestVar) {
2656 ToErase.set(DestVar);
2660 CurrNode->Edges->intersectWithComplement(ToErase);
2661 CurrNode->Edges |= NewEdges;
2664 // Switch to other work list.
2665 WorkList* t = CurrWL; CurrWL = NextWL; NextWL = t;
2670 Node2Deleted.clear();
2671 for (unsigned i = 0; i < GraphNodes.size(); ++i) {
2672 Node *N = &GraphNodes[i];
2673 delete N->OldPointsTo;
2680 //===----------------------------------------------------------------------===//
2682 //===----------------------------------------------------------------------===//
2684 // Unite nodes First and Second, returning the one which is now the
2685 // representative node. First and Second are indexes into GraphNodes
2686 unsigned Andersens::UniteNodes(unsigned First, unsigned Second,
2688 assert (First < GraphNodes.size() && Second < GraphNodes.size() &&
2689 "Attempting to merge nodes that don't exist");
2691 Node *FirstNode = &GraphNodes[First];
2692 Node *SecondNode = &GraphNodes[Second];
2694 assert (SecondNode->isRep() && FirstNode->isRep() &&
2695 "Trying to unite two non-representative nodes!");
2696 if (First == Second)
2700 int RankFirst = (int) FirstNode ->NodeRep;
2701 int RankSecond = (int) SecondNode->NodeRep;
2703 // Rank starts at -1 and gets decremented as it increases.
2704 // Translation: higher rank, lower NodeRep value, which is always negative.
2705 if (RankFirst > RankSecond) {
2706 unsigned t = First; First = Second; Second = t;
2707 Node* tp = FirstNode; FirstNode = SecondNode; SecondNode = tp;
2708 } else if (RankFirst == RankSecond) {
2709 FirstNode->NodeRep = (unsigned) (RankFirst - 1);
2713 SecondNode->NodeRep = First;
2715 if (First >= NumberSpecialNodes)
2717 if (FirstNode->PointsTo && SecondNode->PointsTo)
2718 FirstNode->PointsTo |= *(SecondNode->PointsTo);
2719 if (FirstNode->Edges && SecondNode->Edges)
2720 FirstNode->Edges |= *(SecondNode->Edges);
2721 if (!SecondNode->Constraints.empty())
2722 FirstNode->Constraints.splice(FirstNode->Constraints.begin(),
2723 SecondNode->Constraints);
2724 if (FirstNode->OldPointsTo) {
2725 delete FirstNode->OldPointsTo;
2726 FirstNode->OldPointsTo = new SparseBitVector<>;
2729 // Destroy interesting parts of the merged-from node.
2730 delete SecondNode->OldPointsTo;
2731 delete SecondNode->Edges;
2732 delete SecondNode->PointsTo;
2733 SecondNode->Edges = NULL;
2734 SecondNode->PointsTo = NULL;
2735 SecondNode->OldPointsTo = NULL;
2738 DEBUG(errs() << "Unified Node ");
2739 DEBUG(PrintNode(FirstNode));
2740 DEBUG(errs() << " and Node ");
2741 DEBUG(PrintNode(SecondNode));
2742 DEBUG(errs() << "\n");
2745 if (SDT[Second] >= 0) {
2747 SDT[First] = SDT[Second];
2749 UniteNodes( FindNode(SDT[First]), FindNode(SDT[Second]) );
2750 First = FindNode(First);
2757 // Find the index into GraphNodes of the node representing Node, performing
2758 // path compression along the way
2759 unsigned Andersens::FindNode(unsigned NodeIndex) {
2760 assert (NodeIndex < GraphNodes.size()
2761 && "Attempting to find a node that can't exist");
2762 Node *N = &GraphNodes[NodeIndex];
2766 return (N->NodeRep = FindNode(N->NodeRep));
2769 // Find the index into GraphNodes of the node representing Node,
2770 // don't perform path compression along the way (for Print)
2771 unsigned Andersens::FindNode(unsigned NodeIndex) const {
2772 assert (NodeIndex < GraphNodes.size()
2773 && "Attempting to find a node that can't exist");
2774 const Node *N = &GraphNodes[NodeIndex];
2778 return FindNode(N->NodeRep);
2781 //===----------------------------------------------------------------------===//
2783 //===----------------------------------------------------------------------===//
2785 void Andersens::PrintNode(const Node *N) const {
2786 if (N == &GraphNodes[UniversalSet]) {
2787 errs() << "<universal>";
2789 } else if (N == &GraphNodes[NullPtr]) {
2790 errs() << "<nullptr>";
2792 } else if (N == &GraphNodes[NullObject]) {
2796 if (!N->getValue()) {
2797 errs() << "artificial" << (intptr_t) N;
2801 assert(N->getValue() != 0 && "Never set node label!");
2802 Value *V = N->getValue();
2803 if (Function *F = dyn_cast<Function>(V)) {
2804 if (isa<PointerType>(F->getFunctionType()->getReturnType()) &&
2805 N == &GraphNodes[getReturnNode(F)]) {
2806 errs() << F->getName() << ":retval";
2808 } else if (F->getFunctionType()->isVarArg() &&
2809 N == &GraphNodes[getVarargNode(F)]) {
2810 errs() << F->getName() << ":vararg";
2815 if (Instruction *I = dyn_cast<Instruction>(V))
2816 errs() << I->getParent()->getParent()->getName() << ":";
2817 else if (Argument *Arg = dyn_cast<Argument>(V))
2818 errs() << Arg->getParent()->getName() << ":";
2821 errs() << V->getName();
2823 errs() << "(unnamed)";
2825 if (isa<GlobalValue>(V) || isa<AllocaInst>(V) || isMalloc(V))
2826 if (N == &GraphNodes[getObject(V)])
2829 void Andersens::PrintConstraint(const Constraint &C) const {
2830 if (C.Type == Constraint::Store) {
2835 PrintNode(&GraphNodes[C.Dest]);
2836 if (C.Type == Constraint::Store && C.Offset != 0)
2837 errs() << " + " << C.Offset << ")";
2839 if (C.Type == Constraint::Load) {
2844 else if (C.Type == Constraint::AddressOf)
2846 PrintNode(&GraphNodes[C.Src]);
2847 if (C.Offset != 0 && C.Type != Constraint::Store)
2848 errs() << " + " << C.Offset;
2849 if (C.Type == Constraint::Load && C.Offset != 0)
2854 void Andersens::PrintConstraints() const {
2855 errs() << "Constraints:\n";
2857 for (unsigned i = 0, e = Constraints.size(); i != e; ++i)
2858 PrintConstraint(Constraints[i]);
2861 void Andersens::PrintPointsToGraph() const {
2862 errs() << "Points-to graph:\n";
2863 for (unsigned i = 0, e = GraphNodes.size(); i != e; ++i) {
2864 const Node *N = &GraphNodes[i];
2865 if (FindNode(i) != i) {
2867 errs() << "\t--> same as ";
2868 PrintNode(&GraphNodes[FindNode(i)]);
2871 errs() << "[" << (N->PointsTo->count()) << "] ";
2876 for (SparseBitVector<>::iterator bi = N->PointsTo->begin();
2877 bi != N->PointsTo->end();
2881 PrintNode(&GraphNodes[*bi]);