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/Compiler.h"
63 #include "llvm/Support/ErrorHandling.h"
64 #include "llvm/Support/InstIterator.h"
65 #include "llvm/Support/InstVisitor.h"
66 #include "llvm/Analysis/AliasAnalysis.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
88 STATISTIC(NumIters , "Number of iterations to reach convergence");
89 STATISTIC(NumConstraints, "Number of constraints");
90 STATISTIC(NumNodes , "Number of nodes");
91 STATISTIC(NumUnified , "Number of variables unified");
92 STATISTIC(NumErased , "Number of redundant constraints erased");
94 static const unsigned SelfRep = (unsigned)-1;
95 static const unsigned Unvisited = (unsigned)-1;
96 // Position of the function return node relative to the function node.
97 static const unsigned CallReturnPos = 1;
98 // Position of the function call node relative to the function node.
99 static const unsigned CallFirstArgPos = 2;
102 struct BitmapKeyInfo {
103 static inline SparseBitVector<> *getEmptyKey() {
104 return reinterpret_cast<SparseBitVector<> *>(-1);
106 static inline SparseBitVector<> *getTombstoneKey() {
107 return reinterpret_cast<SparseBitVector<> *>(-2);
109 static unsigned getHashValue(const SparseBitVector<> *bitmap) {
110 return bitmap->getHashValue();
112 static bool isEqual(const SparseBitVector<> *LHS,
113 const SparseBitVector<> *RHS) {
116 else if (LHS == getEmptyKey() || RHS == getEmptyKey()
117 || LHS == getTombstoneKey() || RHS == getTombstoneKey())
123 static bool isPod() { return true; }
126 class VISIBILITY_HIDDEN Andersens : public ModulePass, public AliasAnalysis,
127 private InstVisitor<Andersens> {
130 /// Constraint - Objects of this structure are used to represent the various
131 /// constraints identified by the algorithm. The constraints are 'copy',
132 /// for statements like "A = B", 'load' for statements like "A = *B",
133 /// 'store' for statements like "*A = B", and AddressOf for statements like
134 /// A = alloca; The Offset is applied as *(A + K) = B for stores,
135 /// A = *(B + K) for loads, and A = B + K for copies. It is
136 /// illegal on addressof constraints (because it is statically
137 /// resolvable to A = &C where C = B + K)
140 enum ConstraintType { Copy, Load, Store, AddressOf } Type;
145 Constraint(ConstraintType Ty, unsigned D, unsigned S, unsigned O = 0)
146 : Type(Ty), Dest(D), Src(S), Offset(O) {
147 assert((Offset == 0 || Ty != AddressOf) &&
148 "Offset is illegal on addressof constraints");
151 bool operator==(const Constraint &RHS) const {
152 return RHS.Type == Type
155 && RHS.Offset == Offset;
158 bool operator!=(const Constraint &RHS) const {
159 return !(*this == RHS);
162 bool operator<(const Constraint &RHS) const {
163 if (RHS.Type != Type)
164 return RHS.Type < Type;
165 else if (RHS.Dest != Dest)
166 return RHS.Dest < Dest;
167 else if (RHS.Src != Src)
168 return RHS.Src < Src;
169 return RHS.Offset < Offset;
173 // Information DenseSet requires implemented in order to be able to do
176 static inline std::pair<unsigned, unsigned> getEmptyKey() {
177 return std::make_pair(~0U, ~0U);
179 static inline std::pair<unsigned, unsigned> getTombstoneKey() {
180 return std::make_pair(~0U - 1, ~0U - 1);
182 static unsigned getHashValue(const std::pair<unsigned, unsigned> &P) {
183 return P.first ^ P.second;
185 static unsigned isEqual(const std::pair<unsigned, unsigned> &LHS,
186 const std::pair<unsigned, unsigned> &RHS) {
191 struct ConstraintKeyInfo {
192 static inline Constraint getEmptyKey() {
193 return Constraint(Constraint::Copy, ~0U, ~0U, ~0U);
195 static inline Constraint getTombstoneKey() {
196 return Constraint(Constraint::Copy, ~0U - 1, ~0U - 1, ~0U - 1);
198 static unsigned getHashValue(const Constraint &C) {
199 return C.Src ^ C.Dest ^ C.Type ^ C.Offset;
201 static bool isEqual(const Constraint &LHS,
202 const Constraint &RHS) {
203 return LHS.Type == RHS.Type && LHS.Dest == RHS.Dest
204 && LHS.Src == RHS.Src && LHS.Offset == RHS.Offset;
208 // Node class - This class is used to represent a node in the constraint
209 // graph. Due to various optimizations, it is not always the case that
210 // there is a mapping from a Node to a Value. In particular, we add
211 // artificial Node's that represent the set of pointed-to variables shared
212 // for each location equivalent Node.
215 static volatile sys::cas_flag Counter;
219 SparseBitVector<> *Edges;
220 SparseBitVector<> *PointsTo;
221 SparseBitVector<> *OldPointsTo;
222 std::list<Constraint> Constraints;
224 // Pointer and location equivalence labels
225 unsigned PointerEquivLabel;
226 unsigned LocationEquivLabel;
227 // Predecessor edges, both real and implicit
228 SparseBitVector<> *PredEdges;
229 SparseBitVector<> *ImplicitPredEdges;
230 // Set of nodes that point to us, only use for location equivalence.
231 SparseBitVector<> *PointedToBy;
232 // Number of incoming edges, used during variable substitution to early
233 // free the points-to sets
235 // True if our points-to set is in the Set2PEClass map
237 // True if our node has no indirect constraints (complex or otherwise)
239 // True if the node is address taken, *or* it is part of a group of nodes
240 // that must be kept together. This is set to true for functions and
241 // their arg nodes, which must be kept at the same position relative to
242 // their base function node.
245 // Nodes in cycles (or in equivalence classes) are united together using a
246 // standard union-find representation with path compression. NodeRep
247 // gives the index into GraphNodes for the representative Node.
250 // Modification timestamp. Assigned from Counter.
251 // Used for work list prioritization.
254 explicit Node(bool direct = true) :
255 Val(0), Edges(0), PointsTo(0), OldPointsTo(0),
256 PointerEquivLabel(0), LocationEquivLabel(0), PredEdges(0),
257 ImplicitPredEdges(0), PointedToBy(0), NumInEdges(0),
258 StoredInHash(false), Direct(direct), AddressTaken(false),
259 NodeRep(SelfRep), Timestamp(0) { }
261 Node *setValue(Value *V) {
262 assert(Val == 0 && "Value already set for this node!");
267 /// getValue - Return the LLVM value corresponding to this node.
269 Value *getValue() const { return Val; }
271 /// addPointerTo - Add a pointer to the list of pointees of this node,
272 /// returning true if this caused a new pointer to be added, or false if
273 /// we already knew about the points-to relation.
274 bool addPointerTo(unsigned Node) {
275 return PointsTo->test_and_set(Node);
278 /// intersects - Return true if the points-to set of this node intersects
279 /// with the points-to set of the specified node.
280 bool intersects(Node *N) const;
282 /// intersectsIgnoring - Return true if the points-to set of this node
283 /// intersects with the points-to set of the specified node on any nodes
284 /// except for the specified node to ignore.
285 bool intersectsIgnoring(Node *N, unsigned) const;
287 // Timestamp a node (used for work list prioritization)
289 Timestamp = sys::AtomicIncrement(&Counter);
294 return( (int) NodeRep < 0 );
298 struct WorkListElement {
301 WorkListElement(Node* n, unsigned t) : node(n), Timestamp(t) {}
303 // Note that we reverse the sense of the comparison because we
304 // actually want to give low timestamps the priority over high,
305 // whereas priority is typically interpreted as a greater value is
306 // given high priority.
307 bool operator<(const WorkListElement& that) const {
308 return( this->Timestamp > that.Timestamp );
312 // Priority-queue based work list specialized for Nodes.
314 std::priority_queue<WorkListElement> Q;
317 void insert(Node* n) {
318 Q.push( WorkListElement(n, n->Timestamp) );
321 // We automatically discard non-representative nodes and nodes
322 // that were in the work list twice (we keep a copy of the
323 // timestamp in the work list so we can detect this situation by
324 // comparing against the node's current timestamp).
326 while( !Q.empty() ) {
327 WorkListElement x = Q.top(); Q.pop();
328 Node* INode = x.node;
330 if( INode->isRep() &&
331 INode->Timestamp == x.Timestamp ) {
343 /// GraphNodes - This vector is populated as part of the object
344 /// identification stage of the analysis, which populates this vector with a
345 /// node for each memory object and fills in the ValueNodes map.
346 std::vector<Node> GraphNodes;
348 /// ValueNodes - This map indicates the Node that a particular Value* is
349 /// represented by. This contains entries for all pointers.
350 DenseMap<Value*, unsigned> ValueNodes;
352 /// ObjectNodes - This map contains entries for each memory object in the
353 /// program: globals, alloca's and mallocs.
354 DenseMap<Value*, unsigned> ObjectNodes;
356 /// ReturnNodes - This map contains an entry for each function in the
357 /// program that returns a value.
358 DenseMap<Function*, unsigned> ReturnNodes;
360 /// VarargNodes - This map contains the entry used to represent all pointers
361 /// passed through the varargs portion of a function call for a particular
362 /// function. An entry is not present in this map for functions that do not
363 /// take variable arguments.
364 DenseMap<Function*, unsigned> VarargNodes;
367 /// Constraints - This vector contains a list of all of the constraints
368 /// identified by the program.
369 std::vector<Constraint> Constraints;
371 // Map from graph node to maximum K value that is allowed (for functions,
372 // this is equivalent to the number of arguments + CallFirstArgPos)
373 std::map<unsigned, unsigned> MaxK;
375 /// This enum defines the GraphNodes indices that correspond to important
383 // Stack for Tarjan's
384 std::stack<unsigned> SCCStack;
385 // Map from Graph Node to DFS number
386 std::vector<unsigned> Node2DFS;
387 // Map from Graph Node to Deleted from graph.
388 std::vector<bool> Node2Deleted;
389 // Same as Node Maps, but implemented as std::map because it is faster to
391 std::map<unsigned, unsigned> Tarjan2DFS;
392 std::map<unsigned, bool> Tarjan2Deleted;
393 // Current DFS number
398 WorkList *CurrWL, *NextWL; // "current" and "next" work lists
400 // Offline variable substitution related things
402 // Temporary rep storage, used because we can't collapse SCC's in the
403 // predecessor graph by uniting the variables permanently, we can only do so
404 // for the successor graph.
405 std::vector<unsigned> VSSCCRep;
406 // Mapping from node to whether we have visited it during SCC finding yet.
407 std::vector<bool> Node2Visited;
408 // During variable substitution, we create unknowns to represent the unknown
409 // value that is a dereference of a variable. These nodes are known as
410 // "ref" nodes (since they represent the value of dereferences).
411 unsigned FirstRefNode;
412 // During HVN, we create represent address taken nodes as if they were
413 // unknown (since HVN, unlike HU, does not evaluate unions).
414 unsigned FirstAdrNode;
415 // Current pointer equivalence class number
417 // Mapping from points-to sets to equivalence classes
418 typedef DenseMap<SparseBitVector<> *, unsigned, BitmapKeyInfo> BitVectorMap;
419 BitVectorMap Set2PEClass;
420 // Mapping from pointer equivalences to the representative node. -1 if we
421 // have no representative node for this pointer equivalence class yet.
422 std::vector<int> PEClass2Node;
423 // Mapping from pointer equivalences to representative node. This includes
424 // pointer equivalent but not location equivalent variables. -1 if we have
425 // no representative node for this pointer equivalence class yet.
426 std::vector<int> PENLEClass2Node;
427 // Union/Find for HCD
428 std::vector<unsigned> HCDSCCRep;
429 // HCD's offline-detected cycles; "Statically DeTected"
430 // -1 if not part of such a cycle, otherwise a representative node.
431 std::vector<int> SDT;
432 // Whether to use SDT (UniteNodes can use it during solving, but not before)
437 Andersens() : ModulePass(&ID) {}
439 bool runOnModule(Module &M) {
440 InitializeAliasAnalysis(this);
442 CollectConstraints(M);
444 #define DEBUG_TYPE "anders-aa-constraints"
445 DEBUG(PrintConstraints());
447 #define DEBUG_TYPE "anders-aa"
449 DEBUG(PrintPointsToGraph());
451 // Free the constraints list, as we don't need it to respond to alias
453 std::vector<Constraint>().swap(Constraints);
454 //These are needed for Print() (-analyze in opt)
455 //ObjectNodes.clear();
456 //ReturnNodes.clear();
457 //VarargNodes.clear();
461 void releaseMemory() {
462 // FIXME: Until we have transitively required passes working correctly,
463 // this cannot be enabled! Otherwise, using -count-aa with the pass
464 // causes memory to be freed too early. :(
466 // The memory objects and ValueNodes data structures at the only ones that
467 // are still live after construction.
468 std::vector<Node>().swap(GraphNodes);
473 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
474 AliasAnalysis::getAnalysisUsage(AU);
475 AU.setPreservesAll(); // Does not transform code
478 //------------------------------------------------
479 // Implement the AliasAnalysis API
481 AliasResult alias(const Value *V1, unsigned V1Size,
482 const Value *V2, unsigned V2Size);
483 virtual ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size);
484 virtual ModRefResult getModRefInfo(CallSite CS1, CallSite CS2);
485 void getMustAliases(Value *P, std::vector<Value*> &RetVals);
486 bool pointsToConstantMemory(const Value *P);
488 virtual void deleteValue(Value *V) {
490 getAnalysis<AliasAnalysis>().deleteValue(V);
493 virtual void copyValue(Value *From, Value *To) {
494 ValueNodes[To] = ValueNodes[From];
495 getAnalysis<AliasAnalysis>().copyValue(From, To);
499 /// getNode - Return the node corresponding to the specified pointer scalar.
501 unsigned getNode(Value *V) {
502 if (Constant *C = dyn_cast<Constant>(V))
503 if (!isa<GlobalValue>(C))
504 return getNodeForConstantPointer(C);
506 DenseMap<Value*, unsigned>::iterator I = ValueNodes.find(V);
507 if (I == ValueNodes.end()) {
511 llvm_unreachable("Value does not have a node in the points-to graph!");
516 /// getObject - Return the node corresponding to the memory object for the
517 /// specified global or allocation instruction.
518 unsigned getObject(Value *V) const {
519 DenseMap<Value*, unsigned>::iterator I = ObjectNodes.find(V);
520 assert(I != ObjectNodes.end() &&
521 "Value does not have an object in the points-to graph!");
525 /// getReturnNode - Return the node representing the return value for the
526 /// specified function.
527 unsigned getReturnNode(Function *F) const {
528 DenseMap<Function*, unsigned>::iterator I = ReturnNodes.find(F);
529 assert(I != ReturnNodes.end() && "Function does not return a value!");
533 /// getVarargNode - Return the node representing the variable arguments
534 /// formal for the specified function.
535 unsigned getVarargNode(Function *F) const {
536 DenseMap<Function*, unsigned>::iterator I = VarargNodes.find(F);
537 assert(I != VarargNodes.end() && "Function does not take var args!");
541 /// getNodeValue - Get the node for the specified LLVM value and set the
542 /// value for it to be the specified value.
543 unsigned getNodeValue(Value &V) {
544 unsigned Index = getNode(&V);
545 GraphNodes[Index].setValue(&V);
549 unsigned UniteNodes(unsigned First, unsigned Second,
550 bool UnionByRank = true);
551 unsigned FindNode(unsigned Node);
552 unsigned FindNode(unsigned Node) const;
554 void IdentifyObjects(Module &M);
555 void CollectConstraints(Module &M);
556 bool AnalyzeUsesOfFunction(Value *);
557 void CreateConstraintGraph();
558 void OptimizeConstraints();
559 unsigned FindEquivalentNode(unsigned, unsigned);
560 void ClumpAddressTaken();
561 void RewriteConstraints();
565 void Search(unsigned Node);
566 void UnitePointerEquivalences();
567 void SolveConstraints();
568 bool QueryNode(unsigned Node);
569 void Condense(unsigned Node);
570 void HUValNum(unsigned Node);
571 void HVNValNum(unsigned Node);
572 unsigned getNodeForConstantPointer(Constant *C);
573 unsigned getNodeForConstantPointerTarget(Constant *C);
574 void AddGlobalInitializerConstraints(unsigned, Constant *C);
576 void AddConstraintsForNonInternalLinkage(Function *F);
577 void AddConstraintsForCall(CallSite CS, Function *F);
578 bool AddConstraintsForExternalCall(CallSite CS, Function *F);
581 void PrintNode(const Node *N) const;
582 void PrintConstraints() const ;
583 void PrintConstraint(const Constraint &) const;
584 void PrintLabels() const;
585 void PrintPointsToGraph() const;
587 //===------------------------------------------------------------------===//
588 // Instruction visitation methods for adding constraints
590 friend class InstVisitor<Andersens>;
591 void visitReturnInst(ReturnInst &RI);
592 void visitInvokeInst(InvokeInst &II) { visitCallSite(CallSite(&II)); }
593 void visitCallInst(CallInst &CI) { visitCallSite(CallSite(&CI)); }
594 void visitCallSite(CallSite CS);
595 void visitAllocationInst(AllocationInst &AI);
596 void visitLoadInst(LoadInst &LI);
597 void visitStoreInst(StoreInst &SI);
598 void visitGetElementPtrInst(GetElementPtrInst &GEP);
599 void visitPHINode(PHINode &PN);
600 void visitCastInst(CastInst &CI);
601 void visitICmpInst(ICmpInst &ICI) {} // NOOP!
602 void visitFCmpInst(FCmpInst &ICI) {} // NOOP!
603 void visitSelectInst(SelectInst &SI);
604 void visitVAArg(VAArgInst &I);
605 void visitInstruction(Instruction &I);
607 //===------------------------------------------------------------------===//
608 // Implement Analyize interface
610 void print(std::ostream &O, const Module* M) const {
611 PrintPointsToGraph();
616 char Andersens::ID = 0;
617 static RegisterPass<Andersens>
618 X("anders-aa", "Andersen's Interprocedural Alias Analysis", false, true);
619 static RegisterAnalysisGroup<AliasAnalysis> Y(X);
621 // Initialize Timestamp Counter (static).
622 volatile llvm::sys::cas_flag Andersens::Node::Counter = 0;
624 ModulePass *llvm::createAndersensPass() { return new Andersens(); }
626 //===----------------------------------------------------------------------===//
627 // AliasAnalysis Interface Implementation
628 //===----------------------------------------------------------------------===//
630 AliasAnalysis::AliasResult Andersens::alias(const Value *V1, unsigned V1Size,
631 const Value *V2, unsigned V2Size) {
632 Node *N1 = &GraphNodes[FindNode(getNode(const_cast<Value*>(V1)))];
633 Node *N2 = &GraphNodes[FindNode(getNode(const_cast<Value*>(V2)))];
635 // Check to see if the two pointers are known to not alias. They don't alias
636 // if their points-to sets do not intersect.
637 if (!N1->intersectsIgnoring(N2, NullObject))
640 return AliasAnalysis::alias(V1, V1Size, V2, V2Size);
643 AliasAnalysis::ModRefResult
644 Andersens::getModRefInfo(CallSite CS, Value *P, unsigned Size) {
645 // The only thing useful that we can contribute for mod/ref information is
646 // when calling external function calls: if we know that memory never escapes
647 // from the program, it cannot be modified by an external call.
649 // NOTE: This is not really safe, at least not when the entire program is not
650 // available. The deal is that the external function could call back into the
651 // program and modify stuff. We ignore this technical niggle for now. This
652 // is, after all, a "research quality" implementation of Andersen's analysis.
653 if (Function *F = CS.getCalledFunction())
654 if (F->isDeclaration()) {
655 Node *N1 = &GraphNodes[FindNode(getNode(P))];
657 if (N1->PointsTo->empty())
660 if (!UniversalSet->PointsTo->test(FindNode(getNode(P))))
661 return NoModRef; // Universal set does not contain P
663 if (!N1->PointsTo->test(UniversalSet))
664 return NoModRef; // P doesn't point to the universal set.
668 return AliasAnalysis::getModRefInfo(CS, P, Size);
671 AliasAnalysis::ModRefResult
672 Andersens::getModRefInfo(CallSite CS1, CallSite CS2) {
673 return AliasAnalysis::getModRefInfo(CS1,CS2);
676 /// getMustAlias - We can provide must alias information if we know that a
677 /// pointer can only point to a specific function or the null pointer.
678 /// Unfortunately we cannot determine must-alias information for global
679 /// variables or any other memory memory objects because we do not track whether
680 /// a pointer points to the beginning of an object or a field of it.
681 void Andersens::getMustAliases(Value *P, std::vector<Value*> &RetVals) {
682 Node *N = &GraphNodes[FindNode(getNode(P))];
683 if (N->PointsTo->count() == 1) {
684 Node *Pointee = &GraphNodes[N->PointsTo->find_first()];
685 // If a function is the only object in the points-to set, then it must be
686 // the destination. Note that we can't handle global variables here,
687 // because we don't know if the pointer is actually pointing to a field of
688 // the global or to the beginning of it.
689 if (Value *V = Pointee->getValue()) {
690 if (Function *F = dyn_cast<Function>(V))
691 RetVals.push_back(F);
693 // If the object in the points-to set is the null object, then the null
694 // pointer is a must alias.
695 if (Pointee == &GraphNodes[NullObject])
696 RetVals.push_back(Constant::getNullValue(P->getType()));
699 AliasAnalysis::getMustAliases(P, RetVals);
702 /// pointsToConstantMemory - If we can determine that this pointer only points
703 /// to constant memory, return true. In practice, this means that if the
704 /// pointer can only point to constant globals, functions, or the null pointer,
707 bool Andersens::pointsToConstantMemory(const Value *P) {
708 Node *N = &GraphNodes[FindNode(getNode(const_cast<Value*>(P)))];
711 for (SparseBitVector<>::iterator bi = N->PointsTo->begin();
712 bi != N->PointsTo->end();
715 Node *Pointee = &GraphNodes[i];
716 if (Value *V = Pointee->getValue()) {
717 if (!isa<GlobalValue>(V) || (isa<GlobalVariable>(V) &&
718 !cast<GlobalVariable>(V)->isConstant()))
719 return AliasAnalysis::pointsToConstantMemory(P);
722 return AliasAnalysis::pointsToConstantMemory(P);
729 //===----------------------------------------------------------------------===//
730 // Object Identification Phase
731 //===----------------------------------------------------------------------===//
733 /// IdentifyObjects - This stage scans the program, adding an entry to the
734 /// GraphNodes list for each memory object in the program (global stack or
735 /// heap), and populates the ValueNodes and ObjectNodes maps for these objects.
737 void Andersens::IdentifyObjects(Module &M) {
738 unsigned NumObjects = 0;
740 // Object #0 is always the universal set: the object that we don't know
742 assert(NumObjects == UniversalSet && "Something changed!");
745 // Object #1 always represents the null pointer.
746 assert(NumObjects == NullPtr && "Something changed!");
749 // Object #2 always represents the null object (the object pointed to by null)
750 assert(NumObjects == NullObject && "Something changed!");
753 // Add all the globals first.
754 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
756 ObjectNodes[I] = NumObjects++;
757 ValueNodes[I] = NumObjects++;
760 // Add nodes for all of the functions and the instructions inside of them.
761 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
762 // The function itself is a memory object.
763 unsigned First = NumObjects;
764 ValueNodes[F] = NumObjects++;
765 if (isa<PointerType>(F->getFunctionType()->getReturnType()))
766 ReturnNodes[F] = NumObjects++;
767 if (F->getFunctionType()->isVarArg())
768 VarargNodes[F] = NumObjects++;
771 // Add nodes for all of the incoming pointer arguments.
772 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
775 if (isa<PointerType>(I->getType()))
776 ValueNodes[I] = NumObjects++;
778 MaxK[First] = NumObjects - First;
780 // Scan the function body, creating a memory object for each heap/stack
781 // allocation in the body of the function and a node to represent all
782 // pointer values defined by instructions and used as operands.
783 for (inst_iterator II = inst_begin(F), E = inst_end(F); II != E; ++II) {
784 // If this is an heap or stack allocation, create a node for the memory
786 if (isa<PointerType>(II->getType())) {
787 ValueNodes[&*II] = NumObjects++;
788 if (AllocationInst *AI = dyn_cast<AllocationInst>(&*II))
789 ObjectNodes[AI] = NumObjects++;
792 // Calls to inline asm need to be added as well because the callee isn't
793 // referenced anywhere else.
794 if (CallInst *CI = dyn_cast<CallInst>(&*II)) {
795 Value *Callee = CI->getCalledValue();
796 if (isa<InlineAsm>(Callee))
797 ValueNodes[Callee] = NumObjects++;
802 // Now that we know how many objects to create, make them all now!
803 GraphNodes.resize(NumObjects);
804 NumNodes += NumObjects;
807 //===----------------------------------------------------------------------===//
808 // Constraint Identification Phase
809 //===----------------------------------------------------------------------===//
811 /// getNodeForConstantPointer - Return the node corresponding to the constant
813 unsigned Andersens::getNodeForConstantPointer(Constant *C) {
814 assert(isa<PointerType>(C->getType()) && "Not a constant pointer!");
816 if (isa<ConstantPointerNull>(C) || isa<UndefValue>(C))
818 else if (GlobalValue *GV = dyn_cast<GlobalValue>(C))
820 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
821 switch (CE->getOpcode()) {
822 case Instruction::GetElementPtr:
823 return getNodeForConstantPointer(CE->getOperand(0));
824 case Instruction::IntToPtr:
826 case Instruction::BitCast:
827 return getNodeForConstantPointer(CE->getOperand(0));
829 errs() << "Constant Expr not yet handled: " << *CE << "\n";
833 llvm_unreachable("Unknown constant pointer!");
838 /// getNodeForConstantPointerTarget - Return the node POINTED TO by the
839 /// specified constant pointer.
840 unsigned Andersens::getNodeForConstantPointerTarget(Constant *C) {
841 assert(isa<PointerType>(C->getType()) && "Not a constant pointer!");
843 if (isa<ConstantPointerNull>(C))
845 else if (GlobalValue *GV = dyn_cast<GlobalValue>(C))
846 return getObject(GV);
847 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
848 switch (CE->getOpcode()) {
849 case Instruction::GetElementPtr:
850 return getNodeForConstantPointerTarget(CE->getOperand(0));
851 case Instruction::IntToPtr:
853 case Instruction::BitCast:
854 return getNodeForConstantPointerTarget(CE->getOperand(0));
856 errs() << "Constant Expr not yet handled: " << *CE << "\n";
860 llvm_unreachable("Unknown constant pointer!");
865 /// AddGlobalInitializerConstraints - Add inclusion constraints for the memory
866 /// object N, which contains values indicated by C.
867 void Andersens::AddGlobalInitializerConstraints(unsigned NodeIndex,
869 if (C->getType()->isSingleValueType()) {
870 if (isa<PointerType>(C->getType()))
871 Constraints.push_back(Constraint(Constraint::Copy, NodeIndex,
872 getNodeForConstantPointer(C)));
873 } else if (C->isNullValue()) {
874 Constraints.push_back(Constraint(Constraint::Copy, NodeIndex,
877 } else if (!isa<UndefValue>(C)) {
878 // If this is an array or struct, include constraints for each element.
879 assert(isa<ConstantArray>(C) || isa<ConstantStruct>(C));
880 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i)
881 AddGlobalInitializerConstraints(NodeIndex,
882 cast<Constant>(C->getOperand(i)));
886 /// AddConstraintsForNonInternalLinkage - If this function does not have
887 /// internal linkage, realize that we can't trust anything passed into or
888 /// returned by this function.
889 void Andersens::AddConstraintsForNonInternalLinkage(Function *F) {
890 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
891 if (isa<PointerType>(I->getType()))
892 // If this is an argument of an externally accessible function, the
893 // incoming pointer might point to anything.
894 Constraints.push_back(Constraint(Constraint::Copy, getNode(I),
898 /// AddConstraintsForCall - If this is a call to a "known" function, add the
899 /// constraints and return true. If this is a call to an unknown function,
901 bool Andersens::AddConstraintsForExternalCall(CallSite CS, Function *F) {
902 assert(F->isDeclaration() && "Not an external function!");
904 // These functions don't induce any points-to constraints.
905 if (F->getName() == "atoi" || F->getName() == "atof" ||
906 F->getName() == "atol" || F->getName() == "atoll" ||
907 F->getName() == "remove" || F->getName() == "unlink" ||
908 F->getName() == "rename" || F->getName() == "memcmp" ||
909 F->getName() == "llvm.memset" ||
910 F->getName() == "strcmp" || F->getName() == "strncmp" ||
911 F->getName() == "execl" || F->getName() == "execlp" ||
912 F->getName() == "execle" || F->getName() == "execv" ||
913 F->getName() == "execvp" || F->getName() == "chmod" ||
914 F->getName() == "puts" || F->getName() == "write" ||
915 F->getName() == "open" || F->getName() == "create" ||
916 F->getName() == "truncate" || F->getName() == "chdir" ||
917 F->getName() == "mkdir" || F->getName() == "rmdir" ||
918 F->getName() == "read" || F->getName() == "pipe" ||
919 F->getName() == "wait" || F->getName() == "time" ||
920 F->getName() == "stat" || F->getName() == "fstat" ||
921 F->getName() == "lstat" || F->getName() == "strtod" ||
922 F->getName() == "strtof" || F->getName() == "strtold" ||
923 F->getName() == "fopen" || F->getName() == "fdopen" ||
924 F->getName() == "freopen" ||
925 F->getName() == "fflush" || F->getName() == "feof" ||
926 F->getName() == "fileno" || F->getName() == "clearerr" ||
927 F->getName() == "rewind" || F->getName() == "ftell" ||
928 F->getName() == "ferror" || F->getName() == "fgetc" ||
929 F->getName() == "fgetc" || F->getName() == "_IO_getc" ||
930 F->getName() == "fwrite" || F->getName() == "fread" ||
931 F->getName() == "fgets" || F->getName() == "ungetc" ||
932 F->getName() == "fputc" ||
933 F->getName() == "fputs" || F->getName() == "putc" ||
934 F->getName() == "ftell" || F->getName() == "rewind" ||
935 F->getName() == "_IO_putc" || F->getName() == "fseek" ||
936 F->getName() == "fgetpos" || F->getName() == "fsetpos" ||
937 F->getName() == "printf" || F->getName() == "fprintf" ||
938 F->getName() == "sprintf" || F->getName() == "vprintf" ||
939 F->getName() == "vfprintf" || F->getName() == "vsprintf" ||
940 F->getName() == "scanf" || F->getName() == "fscanf" ||
941 F->getName() == "sscanf" || F->getName() == "__assert_fail" ||
942 F->getName() == "modf")
946 // These functions do induce points-to edges.
947 if (F->getName() == "llvm.memcpy" ||
948 F->getName() == "llvm.memmove" ||
949 F->getName() == "memmove") {
951 const FunctionType *FTy = F->getFunctionType();
952 if (FTy->getNumParams() > 1 &&
953 isa<PointerType>(FTy->getParamType(0)) &&
954 isa<PointerType>(FTy->getParamType(1))) {
956 // *Dest = *Src, which requires an artificial graph node to represent the
957 // constraint. It is broken up into *Dest = temp, temp = *Src
958 unsigned FirstArg = getNode(CS.getArgument(0));
959 unsigned SecondArg = getNode(CS.getArgument(1));
960 unsigned TempArg = GraphNodes.size();
961 GraphNodes.push_back(Node());
962 Constraints.push_back(Constraint(Constraint::Store,
964 Constraints.push_back(Constraint(Constraint::Load,
965 TempArg, SecondArg));
966 // In addition, Dest = Src
967 Constraints.push_back(Constraint(Constraint::Copy,
968 FirstArg, SecondArg));
974 if (F->getName() == "realloc" || F->getName() == "strchr" ||
975 F->getName() == "strrchr" || F->getName() == "strstr" ||
976 F->getName() == "strtok") {
977 const FunctionType *FTy = F->getFunctionType();
978 if (FTy->getNumParams() > 0 &&
979 isa<PointerType>(FTy->getParamType(0))) {
980 Constraints.push_back(Constraint(Constraint::Copy,
981 getNode(CS.getInstruction()),
982 getNode(CS.getArgument(0))));
992 /// AnalyzeUsesOfFunction - Look at all of the users of the specified function.
993 /// If this is used by anything complex (i.e., the address escapes), return
995 bool Andersens::AnalyzeUsesOfFunction(Value *V) {
997 if (!isa<PointerType>(V->getType())) return true;
999 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
1000 if (dyn_cast<LoadInst>(*UI)) {
1002 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
1003 if (V == SI->getOperand(1)) {
1005 } else if (SI->getOperand(1)) {
1006 return true; // Storing the pointer
1008 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) {
1009 if (AnalyzeUsesOfFunction(GEP)) return true;
1010 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
1011 // Make sure that this is just the function being called, not that it is
1012 // passing into the function.
1013 for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i)
1014 if (CI->getOperand(i) == V) return true;
1015 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
1016 // Make sure that this is just the function being called, not that it is
1017 // passing into the function.
1018 for (unsigned i = 3, e = II->getNumOperands(); i != e; ++i)
1019 if (II->getOperand(i) == V) return true;
1020 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
1021 if (CE->getOpcode() == Instruction::GetElementPtr ||
1022 CE->getOpcode() == Instruction::BitCast) {
1023 if (AnalyzeUsesOfFunction(CE))
1028 } else if (ICmpInst *ICI = dyn_cast<ICmpInst>(*UI)) {
1029 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
1030 return true; // Allow comparison against null.
1031 } else if (dyn_cast<FreeInst>(*UI)) {
1039 /// CollectConstraints - This stage scans the program, adding a constraint to
1040 /// the Constraints list for each instruction in the program that induces a
1041 /// constraint, and setting up the initial points-to graph.
1043 void Andersens::CollectConstraints(Module &M) {
1044 // First, the universal set points to itself.
1045 Constraints.push_back(Constraint(Constraint::AddressOf, UniversalSet,
1047 Constraints.push_back(Constraint(Constraint::Store, UniversalSet,
1050 // Next, the null pointer points to the null object.
1051 Constraints.push_back(Constraint(Constraint::AddressOf, NullPtr, NullObject));
1053 // Next, add any constraints on global variables and their initializers.
1054 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
1056 // Associate the address of the global object as pointing to the memory for
1057 // the global: &G = <G memory>
1058 unsigned ObjectIndex = getObject(I);
1059 Node *Object = &GraphNodes[ObjectIndex];
1060 Object->setValue(I);
1061 Constraints.push_back(Constraint(Constraint::AddressOf, getNodeValue(*I),
1064 if (I->hasInitializer()) {
1065 AddGlobalInitializerConstraints(ObjectIndex, I->getInitializer());
1067 // If it doesn't have an initializer (i.e. it's defined in another
1068 // translation unit), it points to the universal set.
1069 Constraints.push_back(Constraint(Constraint::Copy, ObjectIndex,
1074 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
1075 // Set up the return value node.
1076 if (isa<PointerType>(F->getFunctionType()->getReturnType()))
1077 GraphNodes[getReturnNode(F)].setValue(F);
1078 if (F->getFunctionType()->isVarArg())
1079 GraphNodes[getVarargNode(F)].setValue(F);
1081 // Set up incoming argument nodes.
1082 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
1084 if (isa<PointerType>(I->getType()))
1087 // At some point we should just add constraints for the escaping functions
1088 // at solve time, but this slows down solving. For now, we simply mark
1089 // address taken functions as escaping and treat them as external.
1090 if (!F->hasLocalLinkage() || AnalyzeUsesOfFunction(F))
1091 AddConstraintsForNonInternalLinkage(F);
1093 if (!F->isDeclaration()) {
1094 // Scan the function body, creating a memory object for each heap/stack
1095 // allocation in the body of the function and a node to represent all
1096 // pointer values defined by instructions and used as operands.
1099 // External functions that return pointers return the universal set.
1100 if (isa<PointerType>(F->getFunctionType()->getReturnType()))
1101 Constraints.push_back(Constraint(Constraint::Copy,
1105 // Any pointers that are passed into the function have the universal set
1106 // stored into them.
1107 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
1109 if (isa<PointerType>(I->getType())) {
1110 // Pointers passed into external functions could have anything stored
1112 Constraints.push_back(Constraint(Constraint::Store, getNode(I),
1114 // Memory objects passed into external function calls can have the
1115 // universal set point to them.
1117 Constraints.push_back(Constraint(Constraint::Copy,
1121 Constraints.push_back(Constraint(Constraint::Copy,
1127 // If this is an external varargs function, it can also store pointers
1128 // into any pointers passed through the varargs section.
1129 if (F->getFunctionType()->isVarArg())
1130 Constraints.push_back(Constraint(Constraint::Store, getVarargNode(F),
1134 NumConstraints += Constraints.size();
1138 void Andersens::visitInstruction(Instruction &I) {
1140 return; // This function is just a big assert.
1142 if (isa<BinaryOperator>(I))
1144 // Most instructions don't have any effect on pointer values.
1145 switch (I.getOpcode()) {
1146 case Instruction::Br:
1147 case Instruction::Switch:
1148 case Instruction::Unwind:
1149 case Instruction::Unreachable:
1150 case Instruction::Free:
1151 case Instruction::ICmp:
1152 case Instruction::FCmp:
1155 // Is this something we aren't handling yet?
1156 errs() << "Unknown instruction: " << I;
1157 llvm_unreachable(0);
1161 void Andersens::visitAllocationInst(AllocationInst &AI) {
1162 unsigned ObjectIndex = getObject(&AI);
1163 GraphNodes[ObjectIndex].setValue(&AI);
1164 Constraints.push_back(Constraint(Constraint::AddressOf, getNodeValue(AI),
1168 void Andersens::visitReturnInst(ReturnInst &RI) {
1169 if (RI.getNumOperands() && isa<PointerType>(RI.getOperand(0)->getType()))
1170 // return V --> <Copy/retval{F}/v>
1171 Constraints.push_back(Constraint(Constraint::Copy,
1172 getReturnNode(RI.getParent()->getParent()),
1173 getNode(RI.getOperand(0))));
1176 void Andersens::visitLoadInst(LoadInst &LI) {
1177 if (isa<PointerType>(LI.getType()))
1178 // P1 = load P2 --> <Load/P1/P2>
1179 Constraints.push_back(Constraint(Constraint::Load, getNodeValue(LI),
1180 getNode(LI.getOperand(0))));
1183 void Andersens::visitStoreInst(StoreInst &SI) {
1184 if (isa<PointerType>(SI.getOperand(0)->getType()))
1185 // store P1, P2 --> <Store/P2/P1>
1186 Constraints.push_back(Constraint(Constraint::Store,
1187 getNode(SI.getOperand(1)),
1188 getNode(SI.getOperand(0))));
1191 void Andersens::visitGetElementPtrInst(GetElementPtrInst &GEP) {
1192 // P1 = getelementptr P2, ... --> <Copy/P1/P2>
1193 Constraints.push_back(Constraint(Constraint::Copy, getNodeValue(GEP),
1194 getNode(GEP.getOperand(0))));
1197 void Andersens::visitPHINode(PHINode &PN) {
1198 if (isa<PointerType>(PN.getType())) {
1199 unsigned PNN = getNodeValue(PN);
1200 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1201 // P1 = phi P2, P3 --> <Copy/P1/P2>, <Copy/P1/P3>, ...
1202 Constraints.push_back(Constraint(Constraint::Copy, PNN,
1203 getNode(PN.getIncomingValue(i))));
1207 void Andersens::visitCastInst(CastInst &CI) {
1208 Value *Op = CI.getOperand(0);
1209 if (isa<PointerType>(CI.getType())) {
1210 if (isa<PointerType>(Op->getType())) {
1211 // P1 = cast P2 --> <Copy/P1/P2>
1212 Constraints.push_back(Constraint(Constraint::Copy, getNodeValue(CI),
1213 getNode(CI.getOperand(0))));
1215 // P1 = cast int --> <Copy/P1/Univ>
1217 Constraints.push_back(Constraint(Constraint::Copy, getNodeValue(CI),
1223 } else if (isa<PointerType>(Op->getType())) {
1224 // int = cast P1 --> <Copy/Univ/P1>
1226 Constraints.push_back(Constraint(Constraint::Copy,
1228 getNode(CI.getOperand(0))));
1230 getNode(CI.getOperand(0));
1235 void Andersens::visitSelectInst(SelectInst &SI) {
1236 if (isa<PointerType>(SI.getType())) {
1237 unsigned SIN = getNodeValue(SI);
1238 // P1 = select C, P2, P3 ---> <Copy/P1/P2>, <Copy/P1/P3>
1239 Constraints.push_back(Constraint(Constraint::Copy, SIN,
1240 getNode(SI.getOperand(1))));
1241 Constraints.push_back(Constraint(Constraint::Copy, SIN,
1242 getNode(SI.getOperand(2))));
1246 void Andersens::visitVAArg(VAArgInst &I) {
1247 llvm_unreachable("vaarg not handled yet!");
1250 /// AddConstraintsForCall - Add constraints for a call with actual arguments
1251 /// specified by CS to the function specified by F. Note that the types of
1252 /// arguments might not match up in the case where this is an indirect call and
1253 /// the function pointer has been casted. If this is the case, do something
1255 void Andersens::AddConstraintsForCall(CallSite CS, Function *F) {
1256 Value *CallValue = CS.getCalledValue();
1257 bool IsDeref = F == NULL;
1259 // If this is a call to an external function, try to handle it directly to get
1260 // some taste of context sensitivity.
1261 if (F && F->isDeclaration() && AddConstraintsForExternalCall(CS, F))
1264 if (isa<PointerType>(CS.getType())) {
1265 unsigned CSN = getNode(CS.getInstruction());
1266 if (!F || isa<PointerType>(F->getFunctionType()->getReturnType())) {
1268 Constraints.push_back(Constraint(Constraint::Load, CSN,
1269 getNode(CallValue), CallReturnPos));
1271 Constraints.push_back(Constraint(Constraint::Copy, CSN,
1272 getNode(CallValue) + CallReturnPos));
1274 // If the function returns a non-pointer value, handle this just like we
1275 // treat a nonpointer cast to pointer.
1276 Constraints.push_back(Constraint(Constraint::Copy, CSN,
1279 } else if (F && isa<PointerType>(F->getFunctionType()->getReturnType())) {
1281 Constraints.push_back(Constraint(Constraint::Copy,
1283 getNode(CallValue) + CallReturnPos));
1285 Constraints.push_back(Constraint(Constraint::Copy,
1286 getNode(CallValue) + CallReturnPos,
1293 CallSite::arg_iterator ArgI = CS.arg_begin(), ArgE = CS.arg_end();
1294 bool external = !F || F->isDeclaration();
1297 Function::arg_iterator AI = F->arg_begin(), AE = F->arg_end();
1298 for (; AI != AE && ArgI != ArgE; ++AI, ++ArgI)
1301 if (external && isa<PointerType>((*ArgI)->getType()))
1303 // Add constraint that ArgI can now point to anything due to
1304 // escaping, as can everything it points to. The second portion of
1305 // this should be taken care of by universal = *universal
1306 Constraints.push_back(Constraint(Constraint::Copy,
1311 if (isa<PointerType>(AI->getType())) {
1312 if (isa<PointerType>((*ArgI)->getType())) {
1313 // Copy the actual argument into the formal argument.
1314 Constraints.push_back(Constraint(Constraint::Copy, getNode(AI),
1317 Constraints.push_back(Constraint(Constraint::Copy, getNode(AI),
1320 } else if (isa<PointerType>((*ArgI)->getType())) {
1322 Constraints.push_back(Constraint(Constraint::Copy,
1326 Constraints.push_back(Constraint(Constraint::Copy,
1334 unsigned ArgPos = CallFirstArgPos;
1335 for (; ArgI != ArgE; ++ArgI) {
1336 if (isa<PointerType>((*ArgI)->getType())) {
1337 // Copy the actual argument into the formal argument.
1338 Constraints.push_back(Constraint(Constraint::Store,
1340 getNode(*ArgI), ArgPos++));
1342 Constraints.push_back(Constraint(Constraint::Store,
1343 getNode (CallValue),
1344 UniversalSet, ArgPos++));
1348 // Copy all pointers passed through the varargs section to the varargs node.
1349 if (F && F->getFunctionType()->isVarArg())
1350 for (; ArgI != ArgE; ++ArgI)
1351 if (isa<PointerType>((*ArgI)->getType()))
1352 Constraints.push_back(Constraint(Constraint::Copy, getVarargNode(F),
1354 // If more arguments are passed in than we track, just drop them on the floor.
1357 void Andersens::visitCallSite(CallSite CS) {
1358 if (isa<PointerType>(CS.getType()))
1359 getNodeValue(*CS.getInstruction());
1361 if (Function *F = CS.getCalledFunction()) {
1362 AddConstraintsForCall(CS, F);
1364 AddConstraintsForCall(CS, NULL);
1368 //===----------------------------------------------------------------------===//
1369 // Constraint Solving Phase
1370 //===----------------------------------------------------------------------===//
1372 /// intersects - Return true if the points-to set of this node intersects
1373 /// with the points-to set of the specified node.
1374 bool Andersens::Node::intersects(Node *N) const {
1375 return PointsTo->intersects(N->PointsTo);
1378 /// intersectsIgnoring - Return true if the points-to set of this node
1379 /// intersects with the points-to set of the specified node on any nodes
1380 /// except for the specified node to ignore.
1381 bool Andersens::Node::intersectsIgnoring(Node *N, unsigned Ignoring) const {
1382 // TODO: If we are only going to call this with the same value for Ignoring,
1383 // we should move the special values out of the points-to bitmap.
1384 bool WeHadIt = PointsTo->test(Ignoring);
1385 bool NHadIt = N->PointsTo->test(Ignoring);
1386 bool Result = false;
1388 PointsTo->reset(Ignoring);
1390 N->PointsTo->reset(Ignoring);
1391 Result = PointsTo->intersects(N->PointsTo);
1393 PointsTo->set(Ignoring);
1395 N->PointsTo->set(Ignoring);
1400 /// Clump together address taken variables so that the points-to sets use up
1401 /// less space and can be operated on faster.
1403 void Andersens::ClumpAddressTaken() {
1405 #define DEBUG_TYPE "anders-aa-renumber"
1406 std::vector<unsigned> Translate;
1407 std::vector<Node> NewGraphNodes;
1409 Translate.resize(GraphNodes.size());
1410 unsigned NewPos = 0;
1412 for (unsigned i = 0; i < Constraints.size(); ++i) {
1413 Constraint &C = Constraints[i];
1414 if (C.Type == Constraint::AddressOf) {
1415 GraphNodes[C.Src].AddressTaken = true;
1418 for (unsigned i = 0; i < NumberSpecialNodes; ++i) {
1419 unsigned Pos = NewPos++;
1421 NewGraphNodes.push_back(GraphNodes[i]);
1422 DOUT << "Renumbering node " << i << " to node " << Pos << "\n";
1425 // I believe this ends up being faster than making two vectors and splicing
1427 for (unsigned i = NumberSpecialNodes; i < GraphNodes.size(); ++i) {
1428 if (GraphNodes[i].AddressTaken) {
1429 unsigned Pos = NewPos++;
1431 NewGraphNodes.push_back(GraphNodes[i]);
1432 DOUT << "Renumbering node " << i << " to node " << Pos << "\n";
1436 for (unsigned i = NumberSpecialNodes; i < GraphNodes.size(); ++i) {
1437 if (!GraphNodes[i].AddressTaken) {
1438 unsigned Pos = NewPos++;
1440 NewGraphNodes.push_back(GraphNodes[i]);
1441 DOUT << "Renumbering node " << i << " to node " << Pos << "\n";
1445 for (DenseMap<Value*, unsigned>::iterator Iter = ValueNodes.begin();
1446 Iter != ValueNodes.end();
1448 Iter->second = Translate[Iter->second];
1450 for (DenseMap<Value*, unsigned>::iterator Iter = ObjectNodes.begin();
1451 Iter != ObjectNodes.end();
1453 Iter->second = Translate[Iter->second];
1455 for (DenseMap<Function*, unsigned>::iterator Iter = ReturnNodes.begin();
1456 Iter != ReturnNodes.end();
1458 Iter->second = Translate[Iter->second];
1460 for (DenseMap<Function*, unsigned>::iterator Iter = VarargNodes.begin();
1461 Iter != VarargNodes.end();
1463 Iter->second = Translate[Iter->second];
1465 for (unsigned i = 0; i < Constraints.size(); ++i) {
1466 Constraint &C = Constraints[i];
1467 C.Src = Translate[C.Src];
1468 C.Dest = Translate[C.Dest];
1471 GraphNodes.swap(NewGraphNodes);
1473 #define DEBUG_TYPE "anders-aa"
1476 /// The technique used here is described in "Exploiting Pointer and Location
1477 /// Equivalence to Optimize Pointer Analysis. In the 14th International Static
1478 /// Analysis Symposium (SAS), August 2007." It is known as the "HVN" algorithm,
1479 /// and is equivalent to value numbering the collapsed constraint graph without
1480 /// evaluating unions. This is used as a pre-pass to HU in order to resolve
1481 /// first order pointer dereferences and speed up/reduce memory usage of HU.
1482 /// Running both is equivalent to HRU without the iteration
1483 /// HVN in more detail:
1484 /// Imagine the set of constraints was simply straight line code with no loops
1485 /// (we eliminate cycles, so there are no loops), such as:
1491 /// Applying value numbering to this code tells us:
1494 /// For HVN, this is as far as it goes. We assign new value numbers to every
1495 /// "address node", and every "reference node".
1496 /// To get the optimal result for this, we use a DFS + SCC (since all nodes in a
1497 /// cycle must have the same value number since the = operation is really
1498 /// inclusion, not overwrite), and value number nodes we receive points-to sets
1499 /// before we value our own node.
1500 /// The advantage of HU over HVN is that HU considers the inclusion property, so
1501 /// that if you have
1508 /// HU will determine that G == F == E. HVN will not, because it cannot prove
1509 /// that the points to information ends up being the same because they all
1510 /// receive &D from E anyway.
1512 void Andersens::HVN() {
1513 DOUT << "Beginning HVN\n";
1514 // Build a predecessor graph. This is like our constraint graph with the
1515 // edges going in the opposite direction, and there are edges for all the
1516 // constraints, instead of just copy constraints. We also build implicit
1517 // edges for constraints are implied but not explicit. I.E for the constraint
1518 // a = &b, we add implicit edges *a = b. This helps us capture more cycles
1519 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
1520 Constraint &C = Constraints[i];
1521 if (C.Type == Constraint::AddressOf) {
1522 GraphNodes[C.Src].AddressTaken = true;
1523 GraphNodes[C.Src].Direct = false;
1526 unsigned AdrNode = C.Src + FirstAdrNode;
1527 if (!GraphNodes[C.Dest].PredEdges)
1528 GraphNodes[C.Dest].PredEdges = new SparseBitVector<>;
1529 GraphNodes[C.Dest].PredEdges->set(AdrNode);
1532 unsigned RefNode = C.Dest + FirstRefNode;
1533 if (!GraphNodes[RefNode].ImplicitPredEdges)
1534 GraphNodes[RefNode].ImplicitPredEdges = new SparseBitVector<>;
1535 GraphNodes[RefNode].ImplicitPredEdges->set(C.Src);
1536 } else if (C.Type == Constraint::Load) {
1537 if (C.Offset == 0) {
1539 if (!GraphNodes[C.Dest].PredEdges)
1540 GraphNodes[C.Dest].PredEdges = new SparseBitVector<>;
1541 GraphNodes[C.Dest].PredEdges->set(C.Src + FirstRefNode);
1543 GraphNodes[C.Dest].Direct = false;
1545 } else if (C.Type == Constraint::Store) {
1546 if (C.Offset == 0) {
1548 unsigned RefNode = C.Dest + FirstRefNode;
1549 if (!GraphNodes[RefNode].PredEdges)
1550 GraphNodes[RefNode].PredEdges = new SparseBitVector<>;
1551 GraphNodes[RefNode].PredEdges->set(C.Src);
1554 // Dest = Src edge and *Dest = *Src edge
1555 if (!GraphNodes[C.Dest].PredEdges)
1556 GraphNodes[C.Dest].PredEdges = new SparseBitVector<>;
1557 GraphNodes[C.Dest].PredEdges->set(C.Src);
1558 unsigned RefNode = C.Dest + FirstRefNode;
1559 if (!GraphNodes[RefNode].ImplicitPredEdges)
1560 GraphNodes[RefNode].ImplicitPredEdges = new SparseBitVector<>;
1561 GraphNodes[RefNode].ImplicitPredEdges->set(C.Src + FirstRefNode);
1565 // Do SCC finding first to condense our predecessor graph
1567 Node2DFS.insert(Node2DFS.begin(), GraphNodes.size(), 0);
1568 Node2Deleted.insert(Node2Deleted.begin(), GraphNodes.size(), false);
1569 Node2Visited.insert(Node2Visited.begin(), GraphNodes.size(), false);
1571 for (unsigned i = 0; i < FirstRefNode; ++i) {
1572 unsigned Node = VSSCCRep[i];
1573 if (!Node2Visited[Node])
1576 for (BitVectorMap::iterator Iter = Set2PEClass.begin();
1577 Iter != Set2PEClass.end();
1580 Set2PEClass.clear();
1582 Node2Deleted.clear();
1583 Node2Visited.clear();
1584 DOUT << "Finished HVN\n";
1588 /// This is the workhorse of HVN value numbering. We combine SCC finding at the
1589 /// same time because it's easy.
1590 void Andersens::HVNValNum(unsigned NodeIndex) {
1591 unsigned MyDFS = DFSNumber++;
1592 Node *N = &GraphNodes[NodeIndex];
1593 Node2Visited[NodeIndex] = true;
1594 Node2DFS[NodeIndex] = MyDFS;
1596 // First process all our explicit edges
1598 for (SparseBitVector<>::iterator Iter = N->PredEdges->begin();
1599 Iter != N->PredEdges->end();
1601 unsigned j = VSSCCRep[*Iter];
1602 if (!Node2Deleted[j]) {
1603 if (!Node2Visited[j])
1605 if (Node2DFS[NodeIndex] > Node2DFS[j])
1606 Node2DFS[NodeIndex] = Node2DFS[j];
1610 // Now process all the implicit edges
1611 if (N->ImplicitPredEdges)
1612 for (SparseBitVector<>::iterator Iter = N->ImplicitPredEdges->begin();
1613 Iter != N->ImplicitPredEdges->end();
1615 unsigned j = VSSCCRep[*Iter];
1616 if (!Node2Deleted[j]) {
1617 if (!Node2Visited[j])
1619 if (Node2DFS[NodeIndex] > Node2DFS[j])
1620 Node2DFS[NodeIndex] = Node2DFS[j];
1624 // See if we found any cycles
1625 if (MyDFS == Node2DFS[NodeIndex]) {
1626 while (!SCCStack.empty() && Node2DFS[SCCStack.top()] >= MyDFS) {
1627 unsigned CycleNodeIndex = SCCStack.top();
1628 Node *CycleNode = &GraphNodes[CycleNodeIndex];
1629 VSSCCRep[CycleNodeIndex] = NodeIndex;
1631 N->Direct &= CycleNode->Direct;
1633 if (CycleNode->PredEdges) {
1635 N->PredEdges = new SparseBitVector<>;
1636 *(N->PredEdges) |= CycleNode->PredEdges;
1637 delete CycleNode->PredEdges;
1638 CycleNode->PredEdges = NULL;
1640 if (CycleNode->ImplicitPredEdges) {
1641 if (!N->ImplicitPredEdges)
1642 N->ImplicitPredEdges = new SparseBitVector<>;
1643 *(N->ImplicitPredEdges) |= CycleNode->ImplicitPredEdges;
1644 delete CycleNode->ImplicitPredEdges;
1645 CycleNode->ImplicitPredEdges = NULL;
1651 Node2Deleted[NodeIndex] = true;
1654 GraphNodes[NodeIndex].PointerEquivLabel = PEClass++;
1658 // Collect labels of successor nodes
1659 bool AllSame = true;
1660 unsigned First = ~0;
1661 SparseBitVector<> *Labels = new SparseBitVector<>;
1665 for (SparseBitVector<>::iterator Iter = N->PredEdges->begin();
1666 Iter != N->PredEdges->end();
1668 unsigned j = VSSCCRep[*Iter];
1669 unsigned Label = GraphNodes[j].PointerEquivLabel;
1670 // Ignore labels that are equal to us or non-pointers
1671 if (j == NodeIndex || Label == 0)
1673 if (First == (unsigned)~0)
1675 else if (First != Label)
1680 // We either have a non-pointer, a copy of an existing node, or a new node.
1681 // Assign the appropriate pointer equivalence label.
1682 if (Labels->empty()) {
1683 GraphNodes[NodeIndex].PointerEquivLabel = 0;
1684 } else if (AllSame) {
1685 GraphNodes[NodeIndex].PointerEquivLabel = First;
1687 GraphNodes[NodeIndex].PointerEquivLabel = Set2PEClass[Labels];
1688 if (GraphNodes[NodeIndex].PointerEquivLabel == 0) {
1689 unsigned EquivClass = PEClass++;
1690 Set2PEClass[Labels] = EquivClass;
1691 GraphNodes[NodeIndex].PointerEquivLabel = EquivClass;
1698 SCCStack.push(NodeIndex);
1702 /// The technique used here is described in "Exploiting Pointer and Location
1703 /// Equivalence to Optimize Pointer Analysis. In the 14th International Static
1704 /// Analysis Symposium (SAS), August 2007." It is known as the "HU" algorithm,
1705 /// and is equivalent to value numbering the collapsed constraint graph
1706 /// including evaluating unions.
1707 void Andersens::HU() {
1708 DOUT << "Beginning HU\n";
1709 // Build a predecessor graph. This is like our constraint graph with the
1710 // edges going in the opposite direction, and there are edges for all the
1711 // constraints, instead of just copy constraints. We also build implicit
1712 // edges for constraints are implied but not explicit. I.E for the constraint
1713 // a = &b, we add implicit edges *a = b. This helps us capture more cycles
1714 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
1715 Constraint &C = Constraints[i];
1716 if (C.Type == Constraint::AddressOf) {
1717 GraphNodes[C.Src].AddressTaken = true;
1718 GraphNodes[C.Src].Direct = false;
1720 GraphNodes[C.Dest].PointsTo->set(C.Src);
1722 unsigned RefNode = C.Dest + FirstRefNode;
1723 if (!GraphNodes[RefNode].ImplicitPredEdges)
1724 GraphNodes[RefNode].ImplicitPredEdges = new SparseBitVector<>;
1725 GraphNodes[RefNode].ImplicitPredEdges->set(C.Src);
1726 GraphNodes[C.Src].PointedToBy->set(C.Dest);
1727 } else if (C.Type == Constraint::Load) {
1728 if (C.Offset == 0) {
1730 if (!GraphNodes[C.Dest].PredEdges)
1731 GraphNodes[C.Dest].PredEdges = new SparseBitVector<>;
1732 GraphNodes[C.Dest].PredEdges->set(C.Src + FirstRefNode);
1734 GraphNodes[C.Dest].Direct = false;
1736 } else if (C.Type == Constraint::Store) {
1737 if (C.Offset == 0) {
1739 unsigned RefNode = C.Dest + FirstRefNode;
1740 if (!GraphNodes[RefNode].PredEdges)
1741 GraphNodes[RefNode].PredEdges = new SparseBitVector<>;
1742 GraphNodes[RefNode].PredEdges->set(C.Src);
1745 // Dest = Src edge and *Dest = *Src edg
1746 if (!GraphNodes[C.Dest].PredEdges)
1747 GraphNodes[C.Dest].PredEdges = new SparseBitVector<>;
1748 GraphNodes[C.Dest].PredEdges->set(C.Src);
1749 unsigned RefNode = C.Dest + FirstRefNode;
1750 if (!GraphNodes[RefNode].ImplicitPredEdges)
1751 GraphNodes[RefNode].ImplicitPredEdges = new SparseBitVector<>;
1752 GraphNodes[RefNode].ImplicitPredEdges->set(C.Src + FirstRefNode);
1756 // Do SCC finding first to condense our predecessor graph
1758 Node2DFS.insert(Node2DFS.begin(), GraphNodes.size(), 0);
1759 Node2Deleted.insert(Node2Deleted.begin(), GraphNodes.size(), false);
1760 Node2Visited.insert(Node2Visited.begin(), GraphNodes.size(), false);
1762 for (unsigned i = 0; i < FirstRefNode; ++i) {
1763 if (FindNode(i) == i) {
1764 unsigned Node = VSSCCRep[i];
1765 if (!Node2Visited[Node])
1770 // Reset tables for actual labeling
1772 Node2Visited.clear();
1773 Node2Deleted.clear();
1774 // Pre-grow our densemap so that we don't get really bad behavior
1775 Set2PEClass.resize(GraphNodes.size());
1777 // Visit the condensed graph and generate pointer equivalence labels.
1778 Node2Visited.insert(Node2Visited.begin(), GraphNodes.size(), false);
1779 for (unsigned i = 0; i < FirstRefNode; ++i) {
1780 if (FindNode(i) == i) {
1781 unsigned Node = VSSCCRep[i];
1782 if (!Node2Visited[Node])
1786 // PEClass nodes will be deleted by the deleting of N->PointsTo in our caller.
1787 Set2PEClass.clear();
1788 DOUT << "Finished HU\n";
1792 /// Implementation of standard Tarjan SCC algorithm as modified by Nuutilla.
1793 void Andersens::Condense(unsigned NodeIndex) {
1794 unsigned MyDFS = DFSNumber++;
1795 Node *N = &GraphNodes[NodeIndex];
1796 Node2Visited[NodeIndex] = true;
1797 Node2DFS[NodeIndex] = MyDFS;
1799 // First process all our explicit edges
1801 for (SparseBitVector<>::iterator Iter = N->PredEdges->begin();
1802 Iter != N->PredEdges->end();
1804 unsigned j = VSSCCRep[*Iter];
1805 if (!Node2Deleted[j]) {
1806 if (!Node2Visited[j])
1808 if (Node2DFS[NodeIndex] > Node2DFS[j])
1809 Node2DFS[NodeIndex] = Node2DFS[j];
1813 // Now process all the implicit edges
1814 if (N->ImplicitPredEdges)
1815 for (SparseBitVector<>::iterator Iter = N->ImplicitPredEdges->begin();
1816 Iter != N->ImplicitPredEdges->end();
1818 unsigned j = VSSCCRep[*Iter];
1819 if (!Node2Deleted[j]) {
1820 if (!Node2Visited[j])
1822 if (Node2DFS[NodeIndex] > Node2DFS[j])
1823 Node2DFS[NodeIndex] = Node2DFS[j];
1827 // See if we found any cycles
1828 if (MyDFS == Node2DFS[NodeIndex]) {
1829 while (!SCCStack.empty() && Node2DFS[SCCStack.top()] >= MyDFS) {
1830 unsigned CycleNodeIndex = SCCStack.top();
1831 Node *CycleNode = &GraphNodes[CycleNodeIndex];
1832 VSSCCRep[CycleNodeIndex] = NodeIndex;
1834 N->Direct &= CycleNode->Direct;
1836 *(N->PointsTo) |= CycleNode->PointsTo;
1837 delete CycleNode->PointsTo;
1838 CycleNode->PointsTo = NULL;
1839 if (CycleNode->PredEdges) {
1841 N->PredEdges = new SparseBitVector<>;
1842 *(N->PredEdges) |= CycleNode->PredEdges;
1843 delete CycleNode->PredEdges;
1844 CycleNode->PredEdges = NULL;
1846 if (CycleNode->ImplicitPredEdges) {
1847 if (!N->ImplicitPredEdges)
1848 N->ImplicitPredEdges = new SparseBitVector<>;
1849 *(N->ImplicitPredEdges) |= CycleNode->ImplicitPredEdges;
1850 delete CycleNode->ImplicitPredEdges;
1851 CycleNode->ImplicitPredEdges = NULL;
1856 Node2Deleted[NodeIndex] = true;
1858 // Set up number of incoming edges for other nodes
1860 for (SparseBitVector<>::iterator Iter = N->PredEdges->begin();
1861 Iter != N->PredEdges->end();
1863 ++GraphNodes[VSSCCRep[*Iter]].NumInEdges;
1865 SCCStack.push(NodeIndex);
1869 void Andersens::HUValNum(unsigned NodeIndex) {
1870 Node *N = &GraphNodes[NodeIndex];
1871 Node2Visited[NodeIndex] = true;
1873 // Eliminate dereferences of non-pointers for those non-pointers we have
1874 // already identified. These are ref nodes whose non-ref node:
1875 // 1. Has already been visited determined to point to nothing (and thus, a
1876 // dereference of it must point to nothing)
1877 // 2. Any direct node with no predecessor edges in our graph and with no
1878 // points-to set (since it can't point to anything either, being that it
1879 // receives no points-to sets and has none).
1880 if (NodeIndex >= FirstRefNode) {
1881 unsigned j = VSSCCRep[FindNode(NodeIndex - FirstRefNode)];
1882 if ((Node2Visited[j] && !GraphNodes[j].PointerEquivLabel)
1883 || (GraphNodes[j].Direct && !GraphNodes[j].PredEdges
1884 && GraphNodes[j].PointsTo->empty())){
1888 // Process all our explicit edges
1890 for (SparseBitVector<>::iterator Iter = N->PredEdges->begin();
1891 Iter != N->PredEdges->end();
1893 unsigned j = VSSCCRep[*Iter];
1894 if (!Node2Visited[j])
1897 // If this edge turned out to be the same as us, or got no pointer
1898 // equivalence label (and thus points to nothing) , just decrement our
1899 // incoming edges and continue.
1900 if (j == NodeIndex || GraphNodes[j].PointerEquivLabel == 0) {
1901 --GraphNodes[j].NumInEdges;
1905 *(N->PointsTo) |= GraphNodes[j].PointsTo;
1907 // If we didn't end up storing this in the hash, and we're done with all
1908 // the edges, we don't need the points-to set anymore.
1909 --GraphNodes[j].NumInEdges;
1910 if (!GraphNodes[j].NumInEdges && !GraphNodes[j].StoredInHash) {
1911 delete GraphNodes[j].PointsTo;
1912 GraphNodes[j].PointsTo = NULL;
1915 // If this isn't a direct node, generate a fresh variable.
1917 N->PointsTo->set(FirstRefNode + NodeIndex);
1920 // See If we have something equivalent to us, if not, generate a new
1921 // equivalence class.
1922 if (N->PointsTo->empty()) {
1927 N->PointerEquivLabel = Set2PEClass[N->PointsTo];
1928 if (N->PointerEquivLabel == 0) {
1929 unsigned EquivClass = PEClass++;
1930 N->StoredInHash = true;
1931 Set2PEClass[N->PointsTo] = EquivClass;
1932 N->PointerEquivLabel = EquivClass;
1935 N->PointerEquivLabel = PEClass++;
1940 /// Rewrite our list of constraints so that pointer equivalent nodes are
1941 /// replaced by their the pointer equivalence class representative.
1942 void Andersens::RewriteConstraints() {
1943 std::vector<Constraint> NewConstraints;
1944 DenseSet<Constraint, ConstraintKeyInfo> Seen;
1946 PEClass2Node.clear();
1947 PENLEClass2Node.clear();
1949 // We may have from 1 to Graphnodes + 1 equivalence classes.
1950 PEClass2Node.insert(PEClass2Node.begin(), GraphNodes.size() + 1, -1);
1951 PENLEClass2Node.insert(PENLEClass2Node.begin(), GraphNodes.size() + 1, -1);
1953 // Rewrite constraints, ignoring non-pointer constraints, uniting equivalent
1954 // nodes, and rewriting constraints to use the representative nodes.
1955 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
1956 Constraint &C = Constraints[i];
1957 unsigned RHSNode = FindNode(C.Src);
1958 unsigned LHSNode = FindNode(C.Dest);
1959 unsigned RHSLabel = GraphNodes[VSSCCRep[RHSNode]].PointerEquivLabel;
1960 unsigned LHSLabel = GraphNodes[VSSCCRep[LHSNode]].PointerEquivLabel;
1962 // First we try to eliminate constraints for things we can prove don't point
1964 if (LHSLabel == 0) {
1965 DEBUG(PrintNode(&GraphNodes[LHSNode]));
1966 DOUT << " is a non-pointer, ignoring constraint.\n";
1969 if (RHSLabel == 0) {
1970 DEBUG(PrintNode(&GraphNodes[RHSNode]));
1971 DOUT << " is a non-pointer, ignoring constraint.\n";
1974 // This constraint may be useless, and it may become useless as we translate
1976 if (C.Src == C.Dest && C.Type == Constraint::Copy)
1979 C.Src = FindEquivalentNode(RHSNode, RHSLabel);
1980 C.Dest = FindEquivalentNode(FindNode(LHSNode), LHSLabel);
1981 if ((C.Src == C.Dest && C.Type == Constraint::Copy)
1986 NewConstraints.push_back(C);
1988 Constraints.swap(NewConstraints);
1989 PEClass2Node.clear();
1992 /// See if we have a node that is pointer equivalent to the one being asked
1993 /// about, and if so, unite them and return the equivalent node. Otherwise,
1994 /// return the original node.
1995 unsigned Andersens::FindEquivalentNode(unsigned NodeIndex,
1996 unsigned NodeLabel) {
1997 if (!GraphNodes[NodeIndex].AddressTaken) {
1998 if (PEClass2Node[NodeLabel] != -1) {
1999 // We found an existing node with the same pointer label, so unify them.
2000 // We specifically request that Union-By-Rank not be used so that
2001 // PEClass2Node[NodeLabel] U= NodeIndex and not the other way around.
2002 return UniteNodes(PEClass2Node[NodeLabel], NodeIndex, false);
2004 PEClass2Node[NodeLabel] = NodeIndex;
2005 PENLEClass2Node[NodeLabel] = NodeIndex;
2007 } else if (PENLEClass2Node[NodeLabel] == -1) {
2008 PENLEClass2Node[NodeLabel] = NodeIndex;
2014 void Andersens::PrintLabels() const {
2015 for (unsigned i = 0; i < GraphNodes.size(); ++i) {
2016 if (i < FirstRefNode) {
2017 PrintNode(&GraphNodes[i]);
2018 } else if (i < FirstAdrNode) {
2020 PrintNode(&GraphNodes[i-FirstRefNode]);
2024 PrintNode(&GraphNodes[i-FirstAdrNode]);
2028 DOUT << " has pointer label " << GraphNodes[i].PointerEquivLabel
2029 << " and SCC rep " << VSSCCRep[i]
2030 << " and is " << (GraphNodes[i].Direct ? "Direct" : "Not direct")
2035 /// The technique used here is described in "The Ant and the
2036 /// Grasshopper: Fast and Accurate Pointer Analysis for Millions of
2037 /// Lines of Code. In Programming Language Design and Implementation
2038 /// (PLDI), June 2007." It is known as the "HCD" (Hybrid Cycle
2039 /// Detection) algorithm. It is called a hybrid because it performs an
2040 /// offline analysis and uses its results during the solving (online)
2041 /// phase. This is just the offline portion; the results of this
2042 /// operation are stored in SDT and are later used in SolveContraints()
2043 /// and UniteNodes().
2044 void Andersens::HCD() {
2045 DOUT << "Starting HCD.\n";
2046 HCDSCCRep.resize(GraphNodes.size());
2048 for (unsigned i = 0; i < GraphNodes.size(); ++i) {
2049 GraphNodes[i].Edges = new SparseBitVector<>;
2053 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
2054 Constraint &C = Constraints[i];
2055 assert (C.Src < GraphNodes.size() && C.Dest < GraphNodes.size());
2056 if (C.Type == Constraint::AddressOf) {
2058 } else if (C.Type == Constraint::Load) {
2060 GraphNodes[C.Dest].Edges->set(C.Src + FirstRefNode);
2061 } else if (C.Type == Constraint::Store) {
2063 GraphNodes[C.Dest + FirstRefNode].Edges->set(C.Src);
2065 GraphNodes[C.Dest].Edges->set(C.Src);
2069 Node2DFS.insert(Node2DFS.begin(), GraphNodes.size(), 0);
2070 Node2Deleted.insert(Node2Deleted.begin(), GraphNodes.size(), false);
2071 Node2Visited.insert(Node2Visited.begin(), GraphNodes.size(), false);
2072 SDT.insert(SDT.begin(), GraphNodes.size() / 2, -1);
2075 for (unsigned i = 0; i < GraphNodes.size(); ++i) {
2076 unsigned Node = HCDSCCRep[i];
2077 if (!Node2Deleted[Node])
2081 for (unsigned i = 0; i < GraphNodes.size(); ++i)
2082 if (GraphNodes[i].Edges != NULL) {
2083 delete GraphNodes[i].Edges;
2084 GraphNodes[i].Edges = NULL;
2087 while( !SCCStack.empty() )
2091 Node2Visited.clear();
2092 Node2Deleted.clear();
2094 DOUT << "HCD complete.\n";
2097 // Component of HCD:
2098 // Use Nuutila's variant of Tarjan's algorithm to detect
2099 // Strongly-Connected Components (SCCs). For non-trivial SCCs
2100 // containing ref nodes, insert the appropriate information in SDT.
2101 void Andersens::Search(unsigned Node) {
2102 unsigned MyDFS = DFSNumber++;
2104 Node2Visited[Node] = true;
2105 Node2DFS[Node] = MyDFS;
2107 for (SparseBitVector<>::iterator Iter = GraphNodes[Node].Edges->begin(),
2108 End = GraphNodes[Node].Edges->end();
2111 unsigned J = HCDSCCRep[*Iter];
2112 assert(GraphNodes[J].isRep() && "Debug check; must be representative");
2113 if (!Node2Deleted[J]) {
2114 if (!Node2Visited[J])
2116 if (Node2DFS[Node] > Node2DFS[J])
2117 Node2DFS[Node] = Node2DFS[J];
2121 if( MyDFS != Node2DFS[Node] ) {
2122 SCCStack.push(Node);
2126 // This node is the root of a SCC, so process it.
2128 // If the SCC is "non-trivial" (not a singleton) and contains a reference
2129 // node, we place this SCC into SDT. We unite the nodes in any case.
2130 if (!SCCStack.empty() && Node2DFS[SCCStack.top()] >= MyDFS) {
2131 SparseBitVector<> SCC;
2135 bool Ref = (Node >= FirstRefNode);
2137 Node2Deleted[Node] = true;
2140 unsigned P = SCCStack.top(); SCCStack.pop();
2141 Ref |= (P >= FirstRefNode);
2143 HCDSCCRep[P] = Node;
2144 } while (!SCCStack.empty() && Node2DFS[SCCStack.top()] >= MyDFS);
2147 unsigned Rep = SCC.find_first();
2148 assert(Rep < FirstRefNode && "The SCC didn't have a non-Ref node!");
2150 SparseBitVector<>::iterator i = SCC.begin();
2152 // Skip over the non-ref nodes
2153 while( *i < FirstRefNode )
2156 while( i != SCC.end() )
2157 SDT[ (*i++) - FirstRefNode ] = Rep;
2163 /// Optimize the constraints by performing offline variable substitution and
2164 /// other optimizations.
2165 void Andersens::OptimizeConstraints() {
2166 DOUT << "Beginning constraint optimization\n";
2170 // Function related nodes need to stay in the same relative position and can't
2171 // be location equivalent.
2172 for (std::map<unsigned, unsigned>::iterator Iter = MaxK.begin();
2175 for (unsigned i = Iter->first;
2176 i != Iter->first + Iter->second;
2178 GraphNodes[i].AddressTaken = true;
2179 GraphNodes[i].Direct = false;
2183 ClumpAddressTaken();
2184 FirstRefNode = GraphNodes.size();
2185 FirstAdrNode = FirstRefNode + GraphNodes.size();
2186 GraphNodes.insert(GraphNodes.end(), 2 * GraphNodes.size(),
2188 VSSCCRep.resize(GraphNodes.size());
2189 for (unsigned i = 0; i < GraphNodes.size(); ++i) {
2193 for (unsigned i = 0; i < GraphNodes.size(); ++i) {
2194 Node *N = &GraphNodes[i];
2195 delete N->PredEdges;
2196 N->PredEdges = NULL;
2197 delete N->ImplicitPredEdges;
2198 N->ImplicitPredEdges = NULL;
2201 #define DEBUG_TYPE "anders-aa-labels"
2202 DEBUG(PrintLabels());
2204 #define DEBUG_TYPE "anders-aa"
2205 RewriteConstraints();
2206 // Delete the adr nodes.
2207 GraphNodes.resize(FirstRefNode * 2);
2210 for (unsigned i = 0; i < GraphNodes.size(); ++i) {
2211 Node *N = &GraphNodes[i];
2212 if (FindNode(i) == i) {
2213 N->PointsTo = new SparseBitVector<>;
2214 N->PointedToBy = new SparseBitVector<>;
2218 N->PointerEquivLabel = 0;
2222 #define DEBUG_TYPE "anders-aa-labels"
2223 DEBUG(PrintLabels());
2225 #define DEBUG_TYPE "anders-aa"
2226 RewriteConstraints();
2227 for (unsigned i = 0; i < GraphNodes.size(); ++i) {
2228 if (FindNode(i) == i) {
2229 Node *N = &GraphNodes[i];
2232 delete N->PredEdges;
2233 N->PredEdges = NULL;
2234 delete N->ImplicitPredEdges;
2235 N->ImplicitPredEdges = NULL;
2236 delete N->PointedToBy;
2237 N->PointedToBy = NULL;
2241 // perform Hybrid Cycle Detection (HCD)
2245 // No longer any need for the upper half of GraphNodes (for ref nodes).
2246 GraphNodes.erase(GraphNodes.begin() + FirstRefNode, GraphNodes.end());
2250 DOUT << "Finished constraint optimization\n";
2255 /// Unite pointer but not location equivalent variables, now that the constraint
2257 void Andersens::UnitePointerEquivalences() {
2258 DOUT << "Uniting remaining pointer equivalences\n";
2259 for (unsigned i = 0; i < GraphNodes.size(); ++i) {
2260 if (GraphNodes[i].AddressTaken && GraphNodes[i].isRep()) {
2261 unsigned Label = GraphNodes[i].PointerEquivLabel;
2263 if (Label && PENLEClass2Node[Label] != -1)
2264 UniteNodes(i, PENLEClass2Node[Label]);
2267 DOUT << "Finished remaining pointer equivalences\n";
2268 PENLEClass2Node.clear();
2271 /// Create the constraint graph used for solving points-to analysis.
2273 void Andersens::CreateConstraintGraph() {
2274 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
2275 Constraint &C = Constraints[i];
2276 assert (C.Src < GraphNodes.size() && C.Dest < GraphNodes.size());
2277 if (C.Type == Constraint::AddressOf)
2278 GraphNodes[C.Dest].PointsTo->set(C.Src);
2279 else if (C.Type == Constraint::Load)
2280 GraphNodes[C.Src].Constraints.push_back(C);
2281 else if (C.Type == Constraint::Store)
2282 GraphNodes[C.Dest].Constraints.push_back(C);
2283 else if (C.Offset != 0)
2284 GraphNodes[C.Src].Constraints.push_back(C);
2286 GraphNodes[C.Src].Edges->set(C.Dest);
2290 // Perform DFS and cycle detection.
2291 bool Andersens::QueryNode(unsigned Node) {
2292 assert(GraphNodes[Node].isRep() && "Querying a non-rep node");
2293 unsigned OurDFS = ++DFSNumber;
2294 SparseBitVector<> ToErase;
2295 SparseBitVector<> NewEdges;
2296 Tarjan2DFS[Node] = OurDFS;
2298 // Changed denotes a change from a recursive call that we will bubble up.
2299 // Merged is set if we actually merge a node ourselves.
2300 bool Changed = false, Merged = false;
2302 for (SparseBitVector<>::iterator bi = GraphNodes[Node].Edges->begin();
2303 bi != GraphNodes[Node].Edges->end();
2305 unsigned RepNode = FindNode(*bi);
2306 // If this edge points to a non-representative node but we are
2307 // already planning to add an edge to its representative, we have no
2308 // need for this edge anymore.
2309 if (RepNode != *bi && NewEdges.test(RepNode)){
2314 // Continue about our DFS.
2315 if (!Tarjan2Deleted[RepNode]){
2316 if (Tarjan2DFS[RepNode] == 0) {
2317 Changed |= QueryNode(RepNode);
2318 // May have been changed by QueryNode
2319 RepNode = FindNode(RepNode);
2321 if (Tarjan2DFS[RepNode] < Tarjan2DFS[Node])
2322 Tarjan2DFS[Node] = Tarjan2DFS[RepNode];
2325 // We may have just discovered that this node is part of a cycle, in
2326 // which case we can also erase it.
2327 if (RepNode != *bi) {
2329 NewEdges.set(RepNode);
2333 GraphNodes[Node].Edges->intersectWithComplement(ToErase);
2334 GraphNodes[Node].Edges |= NewEdges;
2336 // If this node is a root of a non-trivial SCC, place it on our
2337 // worklist to be processed.
2338 if (OurDFS == Tarjan2DFS[Node]) {
2339 while (!SCCStack.empty() && Tarjan2DFS[SCCStack.top()] >= OurDFS) {
2340 Node = UniteNodes(Node, SCCStack.top());
2345 Tarjan2Deleted[Node] = true;
2348 NextWL->insert(&GraphNodes[Node]);
2350 SCCStack.push(Node);
2353 return(Changed | Merged);
2356 /// SolveConstraints - This stage iteratively processes the constraints list
2357 /// propagating constraints (adding edges to the Nodes in the points-to graph)
2358 /// until a fixed point is reached.
2360 /// We use a variant of the technique called "Lazy Cycle Detection", which is
2361 /// described in "The Ant and the Grasshopper: Fast and Accurate Pointer
2362 /// Analysis for Millions of Lines of Code. In Programming Language Design and
2363 /// Implementation (PLDI), June 2007."
2364 /// The paper describes performing cycle detection one node at a time, which can
2365 /// be expensive if there are no cycles, but there are long chains of nodes that
2366 /// it heuristically believes are cycles (because it will DFS from each node
2367 /// without state from previous nodes).
2368 /// Instead, we use the heuristic to build a worklist of nodes to check, then
2369 /// cycle detect them all at the same time to do this more cheaply. This
2370 /// catches cycles slightly later than the original technique did, but does it
2371 /// make significantly cheaper.
2373 void Andersens::SolveConstraints() {
2377 OptimizeConstraints();
2379 #define DEBUG_TYPE "anders-aa-constraints"
2380 DEBUG(PrintConstraints());
2382 #define DEBUG_TYPE "anders-aa"
2384 for (unsigned i = 0; i < GraphNodes.size(); ++i) {
2385 Node *N = &GraphNodes[i];
2386 N->PointsTo = new SparseBitVector<>;
2387 N->OldPointsTo = new SparseBitVector<>;
2388 N->Edges = new SparseBitVector<>;
2390 CreateConstraintGraph();
2391 UnitePointerEquivalences();
2392 assert(SCCStack.empty() && "SCC Stack should be empty by now!");
2394 Node2Deleted.clear();
2395 Node2DFS.insert(Node2DFS.begin(), GraphNodes.size(), 0);
2396 Node2Deleted.insert(Node2Deleted.begin(), GraphNodes.size(), false);
2398 DenseSet<Constraint, ConstraintKeyInfo> Seen;
2399 DenseSet<std::pair<unsigned,unsigned>, PairKeyInfo> EdgesChecked;
2401 // Order graph and add initial nodes to work list.
2402 for (unsigned i = 0; i < GraphNodes.size(); ++i) {
2403 Node *INode = &GraphNodes[i];
2405 // Add to work list if it's a representative and can contribute to the
2406 // calculation right now.
2407 if (INode->isRep() && !INode->PointsTo->empty()
2408 && (!INode->Edges->empty() || !INode->Constraints.empty())) {
2410 CurrWL->insert(INode);
2413 std::queue<unsigned int> TarjanWL;
2415 // "Rep and special variables" - in order for HCD to maintain conservative
2416 // results when !FULL_UNIVERSAL, we need to treat the special variables in
2417 // the same way that the !FULL_UNIVERSAL tweak does throughout the rest of
2418 // the analysis - it's ok to add edges from the special nodes, but never
2419 // *to* the special nodes.
2420 std::vector<unsigned int> RSV;
2422 while( !CurrWL->empty() ) {
2423 DOUT << "Starting iteration #" << ++NumIters << "\n";
2426 unsigned CurrNodeIndex;
2428 // Actual cycle checking code. We cycle check all of the lazy cycle
2429 // candidates from the last iteration in one go.
2430 if (!TarjanWL.empty()) {
2434 Tarjan2Deleted.clear();
2435 while (!TarjanWL.empty()) {
2436 unsigned int ToTarjan = TarjanWL.front();
2438 if (!Tarjan2Deleted[ToTarjan]
2439 && GraphNodes[ToTarjan].isRep()
2440 && Tarjan2DFS[ToTarjan] == 0)
2441 QueryNode(ToTarjan);
2445 // Add to work list if it's a representative and can contribute to the
2446 // calculation right now.
2447 while( (CurrNode = CurrWL->pop()) != NULL ) {
2448 CurrNodeIndex = CurrNode - &GraphNodes[0];
2452 // Figure out the changed points to bits
2453 SparseBitVector<> CurrPointsTo;
2454 CurrPointsTo.intersectWithComplement(CurrNode->PointsTo,
2455 CurrNode->OldPointsTo);
2456 if (CurrPointsTo.empty())
2459 *(CurrNode->OldPointsTo) |= CurrPointsTo;
2461 // Check the offline-computed equivalencies from HCD.
2465 if (SDT[CurrNodeIndex] >= 0) {
2467 Rep = FindNode(SDT[CurrNodeIndex]);
2472 for (SparseBitVector<>::iterator bi = CurrPointsTo.begin();
2473 bi != CurrPointsTo.end(); ++bi) {
2474 unsigned Node = FindNode(*bi);
2476 if (Node < NumberSpecialNodes) {
2477 RSV.push_back(Node);
2481 Rep = UniteNodes(Rep,Node);
2487 NextWL->insert(&GraphNodes[Rep]);
2489 if ( ! CurrNode->isRep() )
2495 /* Now process the constraints for this node. */
2496 for (std::list<Constraint>::iterator li = CurrNode->Constraints.begin();
2497 li != CurrNode->Constraints.end(); ) {
2498 li->Src = FindNode(li->Src);
2499 li->Dest = FindNode(li->Dest);
2501 // Delete redundant constraints
2502 if( Seen.count(*li) ) {
2503 std::list<Constraint>::iterator lk = li; li++;
2505 CurrNode->Constraints.erase(lk);
2511 // Src and Dest will be the vars we are going to process.
2512 // This may look a bit ugly, but what it does is allow us to process
2513 // both store and load constraints with the same code.
2514 // Load constraints say that every member of our RHS solution has K
2515 // added to it, and that variable gets an edge to LHS. We also union
2516 // RHS+K's solution into the LHS solution.
2517 // Store constraints say that every member of our LHS solution has K
2518 // added to it, and that variable gets an edge from RHS. We also union
2519 // RHS's solution into the LHS+K solution.
2522 unsigned K = li->Offset;
2523 unsigned CurrMember;
2524 if (li->Type == Constraint::Load) {
2527 } else if (li->Type == Constraint::Store) {
2531 // TODO Handle offseted copy constraint
2536 // See if we can use Hybrid Cycle Detection (that is, check
2537 // if it was a statically detected offline equivalence that
2538 // involves pointers; if so, remove the redundant constraints).
2539 if( SCC && K == 0 ) {
2543 if (GraphNodes[*Src].Edges->test_and_set(*Dest))
2544 if (GraphNodes[*Dest].PointsTo |= *(GraphNodes[*Src].PointsTo))
2545 NextWL->insert(&GraphNodes[*Dest]);
2547 for (unsigned i=0; i < RSV.size(); ++i) {
2548 CurrMember = RSV[i];
2550 if (*Dest < NumberSpecialNodes)
2552 if (GraphNodes[*Src].Edges->test_and_set(*Dest))
2553 if (GraphNodes[*Dest].PointsTo |= *(GraphNodes[*Src].PointsTo))
2554 NextWL->insert(&GraphNodes[*Dest]);
2557 // since all future elements of the points-to set will be
2558 // equivalent to the current ones, the complex constraints
2559 // become redundant.
2561 std::list<Constraint>::iterator lk = li; li++;
2563 // In this case, we can still erase the constraints when the
2564 // elements of the points-to sets are referenced by *Dest,
2565 // but not when they are referenced by *Src (i.e. for a Load
2566 // constraint). This is because if another special variable is
2567 // put into the points-to set later, we still need to add the
2568 // new edge from that special variable.
2569 if( lk->Type != Constraint::Load)
2571 GraphNodes[CurrNodeIndex].Constraints.erase(lk);
2573 const SparseBitVector<> &Solution = CurrPointsTo;
2575 for (SparseBitVector<>::iterator bi = Solution.begin();
2576 bi != Solution.end();
2580 // Need to increment the member by K since that is where we are
2581 // supposed to copy to/from. Note that in positive weight cycles,
2582 // which occur in address taking of fields, K can go past
2583 // MaxK[CurrMember] elements, even though that is all it could point
2585 if (K > 0 && K > MaxK[CurrMember])
2588 CurrMember = FindNode(CurrMember + K);
2590 // Add an edge to the graph, so we can just do regular
2591 // bitmap ior next time. It may also let us notice a cycle.
2593 if (*Dest < NumberSpecialNodes)
2596 if (GraphNodes[*Src].Edges->test_and_set(*Dest))
2597 if (GraphNodes[*Dest].PointsTo |= *(GraphNodes[*Src].PointsTo))
2598 NextWL->insert(&GraphNodes[*Dest]);
2604 SparseBitVector<> NewEdges;
2605 SparseBitVector<> ToErase;
2607 // Now all we have left to do is propagate points-to info along the
2608 // edges, erasing the redundant edges.
2609 for (SparseBitVector<>::iterator bi = CurrNode->Edges->begin();
2610 bi != CurrNode->Edges->end();
2613 unsigned DestVar = *bi;
2614 unsigned Rep = FindNode(DestVar);
2616 // If we ended up with this node as our destination, or we've already
2617 // got an edge for the representative, delete the current edge.
2618 if (Rep == CurrNodeIndex ||
2619 (Rep != DestVar && NewEdges.test(Rep))) {
2620 ToErase.set(DestVar);
2624 std::pair<unsigned,unsigned> edge(CurrNodeIndex,Rep);
2626 // This is where we do lazy cycle detection.
2627 // If this is a cycle candidate (equal points-to sets and this
2628 // particular edge has not been cycle-checked previously), add to the
2629 // list to check for cycles on the next iteration.
2630 if (!EdgesChecked.count(edge) &&
2631 *(GraphNodes[Rep].PointsTo) == *(CurrNode->PointsTo)) {
2632 EdgesChecked.insert(edge);
2635 // Union the points-to sets into the dest
2637 if (Rep >= NumberSpecialNodes)
2639 if (GraphNodes[Rep].PointsTo |= CurrPointsTo) {
2640 NextWL->insert(&GraphNodes[Rep]);
2642 // If this edge's destination was collapsed, rewrite the edge.
2643 if (Rep != DestVar) {
2644 ToErase.set(DestVar);
2648 CurrNode->Edges->intersectWithComplement(ToErase);
2649 CurrNode->Edges |= NewEdges;
2652 // Switch to other work list.
2653 WorkList* t = CurrWL; CurrWL = NextWL; NextWL = t;
2658 Node2Deleted.clear();
2659 for (unsigned i = 0; i < GraphNodes.size(); ++i) {
2660 Node *N = &GraphNodes[i];
2661 delete N->OldPointsTo;
2668 //===----------------------------------------------------------------------===//
2670 //===----------------------------------------------------------------------===//
2672 // Unite nodes First and Second, returning the one which is now the
2673 // representative node. First and Second are indexes into GraphNodes
2674 unsigned Andersens::UniteNodes(unsigned First, unsigned Second,
2676 assert (First < GraphNodes.size() && Second < GraphNodes.size() &&
2677 "Attempting to merge nodes that don't exist");
2679 Node *FirstNode = &GraphNodes[First];
2680 Node *SecondNode = &GraphNodes[Second];
2682 assert (SecondNode->isRep() && FirstNode->isRep() &&
2683 "Trying to unite two non-representative nodes!");
2684 if (First == Second)
2688 int RankFirst = (int) FirstNode ->NodeRep;
2689 int RankSecond = (int) SecondNode->NodeRep;
2691 // Rank starts at -1 and gets decremented as it increases.
2692 // Translation: higher rank, lower NodeRep value, which is always negative.
2693 if (RankFirst > RankSecond) {
2694 unsigned t = First; First = Second; Second = t;
2695 Node* tp = FirstNode; FirstNode = SecondNode; SecondNode = tp;
2696 } else if (RankFirst == RankSecond) {
2697 FirstNode->NodeRep = (unsigned) (RankFirst - 1);
2701 SecondNode->NodeRep = First;
2703 if (First >= NumberSpecialNodes)
2705 if (FirstNode->PointsTo && SecondNode->PointsTo)
2706 FirstNode->PointsTo |= *(SecondNode->PointsTo);
2707 if (FirstNode->Edges && SecondNode->Edges)
2708 FirstNode->Edges |= *(SecondNode->Edges);
2709 if (!SecondNode->Constraints.empty())
2710 FirstNode->Constraints.splice(FirstNode->Constraints.begin(),
2711 SecondNode->Constraints);
2712 if (FirstNode->OldPointsTo) {
2713 delete FirstNode->OldPointsTo;
2714 FirstNode->OldPointsTo = new SparseBitVector<>;
2717 // Destroy interesting parts of the merged-from node.
2718 delete SecondNode->OldPointsTo;
2719 delete SecondNode->Edges;
2720 delete SecondNode->PointsTo;
2721 SecondNode->Edges = NULL;
2722 SecondNode->PointsTo = NULL;
2723 SecondNode->OldPointsTo = NULL;
2726 DOUT << "Unified Node ";
2727 DEBUG(PrintNode(FirstNode));
2728 DOUT << " and Node ";
2729 DEBUG(PrintNode(SecondNode));
2733 if (SDT[Second] >= 0) {
2735 SDT[First] = SDT[Second];
2737 UniteNodes( FindNode(SDT[First]), FindNode(SDT[Second]) );
2738 First = FindNode(First);
2745 // Find the index into GraphNodes of the node representing Node, performing
2746 // path compression along the way
2747 unsigned Andersens::FindNode(unsigned NodeIndex) {
2748 assert (NodeIndex < GraphNodes.size()
2749 && "Attempting to find a node that can't exist");
2750 Node *N = &GraphNodes[NodeIndex];
2754 return (N->NodeRep = FindNode(N->NodeRep));
2757 // Find the index into GraphNodes of the node representing Node,
2758 // don't perform path compression along the way (for Print)
2759 unsigned Andersens::FindNode(unsigned NodeIndex) const {
2760 assert (NodeIndex < GraphNodes.size()
2761 && "Attempting to find a node that can't exist");
2762 const Node *N = &GraphNodes[NodeIndex];
2766 return FindNode(N->NodeRep);
2769 //===----------------------------------------------------------------------===//
2771 //===----------------------------------------------------------------------===//
2773 void Andersens::PrintNode(const Node *N) const {
2774 if (N == &GraphNodes[UniversalSet]) {
2775 errs() << "<universal>";
2777 } else if (N == &GraphNodes[NullPtr]) {
2778 errs() << "<nullptr>";
2780 } else if (N == &GraphNodes[NullObject]) {
2784 if (!N->getValue()) {
2785 errs() << "artificial" << (intptr_t) N;
2789 assert(N->getValue() != 0 && "Never set node label!");
2790 Value *V = N->getValue();
2791 if (Function *F = dyn_cast<Function>(V)) {
2792 if (isa<PointerType>(F->getFunctionType()->getReturnType()) &&
2793 N == &GraphNodes[getReturnNode(F)]) {
2794 errs() << F->getName() << ":retval";
2796 } else if (F->getFunctionType()->isVarArg() &&
2797 N == &GraphNodes[getVarargNode(F)]) {
2798 errs() << F->getName() << ":vararg";
2803 if (Instruction *I = dyn_cast<Instruction>(V))
2804 errs() << I->getParent()->getParent()->getName() << ":";
2805 else if (Argument *Arg = dyn_cast<Argument>(V))
2806 errs() << Arg->getParent()->getName() << ":";
2809 errs() << V->getName();
2811 errs() << "(unnamed)";
2813 if (isa<GlobalValue>(V) || isa<AllocationInst>(V))
2814 if (N == &GraphNodes[getObject(V)])
2817 void Andersens::PrintConstraint(const Constraint &C) const {
2818 if (C.Type == Constraint::Store) {
2823 PrintNode(&GraphNodes[C.Dest]);
2824 if (C.Type == Constraint::Store && C.Offset != 0)
2825 errs() << " + " << C.Offset << ")";
2827 if (C.Type == Constraint::Load) {
2832 else if (C.Type == Constraint::AddressOf)
2834 PrintNode(&GraphNodes[C.Src]);
2835 if (C.Offset != 0 && C.Type != Constraint::Store)
2836 errs() << " + " << C.Offset;
2837 if (C.Type == Constraint::Load && C.Offset != 0)
2842 void Andersens::PrintConstraints() const {
2843 errs() << "Constraints:\n";
2845 for (unsigned i = 0, e = Constraints.size(); i != e; ++i)
2846 PrintConstraint(Constraints[i]);
2849 void Andersens::PrintPointsToGraph() const {
2850 errs() << "Points-to graph:\n";
2851 for (unsigned i = 0, e = GraphNodes.size(); i != e; ++i) {
2852 const Node *N = &GraphNodes[i];
2853 if (FindNode(i) != i) {
2855 errs() << "\t--> same as ";
2856 PrintNode(&GraphNodes[FindNode(i)]);
2859 errs() << "[" << (N->PointsTo->count()) << "] ";
2864 for (SparseBitVector<>::iterator bi = N->PointsTo->begin();
2865 bi != N->PointsTo->end();
2869 PrintNode(&GraphNodes[*bi]);