1 //===- Andersens.cpp - Andersen's Interprocedural Alias Analysis ----------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines a very simple implementation of Andersen's interprocedural
11 // alias analysis. This implementation does not include any of the fancy
12 // features that make Andersen's reasonably efficient (like cycle elimination or
13 // variable substitution), but it should be useful for getting precision
14 // numbers and can be extended in the future.
16 // In pointer analysis terms, this is a subset-based, flow-insensitive,
17 // field-insensitive, and context-insensitive algorithm pointer algorithm.
19 // This algorithm is implemented as three stages:
20 // 1. Object identification.
21 // 2. Inclusion constraint identification.
22 // 3. Inclusion constraint solving.
24 // The object identification stage identifies all of the memory objects in the
25 // program, which includes globals, heap allocated objects, and stack allocated
28 // The inclusion constraint identification stage finds all inclusion constraints
29 // in the program by scanning the program, looking for pointer assignments and
30 // other statements that effect the points-to graph. For a statement like "A =
31 // B", this statement is processed to indicate that A can point to anything that
32 // B can point to. Constraints can handle copies, loads, and stores.
34 // The inclusion constraint solving phase iteratively propagates the inclusion
35 // constraints until a fixed point is reached. This is an O(N^3) algorithm.
37 // In the initial pass, all indirect function calls are completely ignored. As
38 // the analysis discovers new targets of function pointers, it iteratively
39 // resolves a precise (and conservative) call graph. Also related, this
40 // analysis initially assumes that all internal functions have known incoming
41 // pointers. If we find that an internal function's address escapes outside of
42 // the program, we update this assumption.
44 // Future Improvements:
45 // This implementation of Andersen's algorithm is extremely slow. To make it
46 // scale reasonably well, the inclusion constraints could be sorted (easy),
47 // offline variable substitution would be a huge win (straight-forward), and
48 // online cycle elimination (trickier) might help as well.
50 //===----------------------------------------------------------------------===//
52 #define DEBUG_TYPE "anders-aa"
53 #include "llvm/Constants.h"
54 #include "llvm/DerivedTypes.h"
55 #include "llvm/Instructions.h"
56 #include "llvm/Module.h"
57 #include "llvm/Pass.h"
58 #include "llvm/Support/InstIterator.h"
59 #include "llvm/Support/InstVisitor.h"
60 #include "llvm/Analysis/AliasAnalysis.h"
61 #include "llvm/Analysis/Passes.h"
62 #include "llvm/Support/Debug.h"
63 #include "llvm/ADT/Statistic.h"
69 NumIters("anders-aa", "Number of iterations to reach convergence");
71 NumConstraints("anders-aa", "Number of constraints");
73 NumNodes("anders-aa", "Number of nodes");
75 NumEscapingFunctions("anders-aa", "Number of internal functions that escape");
77 NumIndirectCallees("anders-aa", "Number of indirect callees found");
79 class Andersens : public ModulePass, public AliasAnalysis,
80 private InstVisitor<Andersens> {
81 /// Node class - This class is used to represent a memory object in the
82 /// program, and is the primitive used to build the points-to graph.
84 std::vector<Node*> Pointees;
88 Node *setValue(Value *V) {
89 assert(Val == 0 && "Value already set for this node!");
94 /// getValue - Return the LLVM value corresponding to this node.
96 Value *getValue() const { return Val; }
98 typedef std::vector<Node*>::const_iterator iterator;
99 iterator begin() const { return Pointees.begin(); }
100 iterator end() const { return Pointees.end(); }
102 /// addPointerTo - Add a pointer to the list of pointees of this node,
103 /// returning true if this caused a new pointer to be added, or false if
104 /// we already knew about the points-to relation.
105 bool addPointerTo(Node *N) {
106 std::vector<Node*>::iterator I = std::lower_bound(Pointees.begin(),
109 if (I != Pointees.end() && *I == N)
111 Pointees.insert(I, N);
115 /// intersects - Return true if the points-to set of this node intersects
116 /// with the points-to set of the specified node.
117 bool intersects(Node *N) const;
119 /// intersectsIgnoring - Return true if the points-to set of this node
120 /// intersects with the points-to set of the specified node on any nodes
121 /// except for the specified node to ignore.
122 bool intersectsIgnoring(Node *N, Node *Ignoring) const;
124 // Constraint application methods.
125 bool copyFrom(Node *N);
126 bool loadFrom(Node *N);
127 bool storeThrough(Node *N);
130 /// GraphNodes - This vector is populated as part of the object
131 /// identification stage of the analysis, which populates this vector with a
132 /// node for each memory object and fills in the ValueNodes map.
133 std::vector<Node> GraphNodes;
135 /// ValueNodes - This map indicates the Node that a particular Value* is
136 /// represented by. This contains entries for all pointers.
137 std::map<Value*, unsigned> ValueNodes;
139 /// ObjectNodes - This map contains entries for each memory object in the
140 /// program: globals, alloca's and mallocs.
141 std::map<Value*, unsigned> ObjectNodes;
143 /// ReturnNodes - This map contains an entry for each function in the
144 /// program that returns a value.
145 std::map<Function*, unsigned> ReturnNodes;
147 /// VarargNodes - This map contains the entry used to represent all pointers
148 /// passed through the varargs portion of a function call for a particular
149 /// function. An entry is not present in this map for functions that do not
150 /// take variable arguments.
151 std::map<Function*, unsigned> VarargNodes;
153 /// Constraint - Objects of this structure are used to represent the various
154 /// constraints identified by the algorithm. The constraints are 'copy',
155 /// for statements like "A = B", 'load' for statements like "A = *B", and
156 /// 'store' for statements like "*A = B".
158 enum ConstraintType { Copy, Load, Store } Type;
161 Constraint(ConstraintType Ty, Node *D, Node *S)
162 : Type(Ty), Dest(D), Src(S) {}
165 /// Constraints - This vector contains a list of all of the constraints
166 /// identified by the program.
167 std::vector<Constraint> Constraints;
169 /// EscapingInternalFunctions - This set contains all of the internal
170 /// functions that are found to escape from the program. If the address of
171 /// an internal function is passed to an external function or otherwise
172 /// escapes from the analyzed portion of the program, we must assume that
173 /// any pointer arguments can alias the universal node. This set keeps
174 /// track of those functions we are assuming to escape so far.
175 std::set<Function*> EscapingInternalFunctions;
177 /// IndirectCalls - This contains a list of all of the indirect call sites
178 /// in the program. Since the call graph is iteratively discovered, we may
179 /// need to add constraints to our graph as we find new targets of function
181 std::vector<CallSite> IndirectCalls;
183 /// IndirectCallees - For each call site in the indirect calls list, keep
184 /// track of the callees that we have discovered so far. As the analysis
185 /// proceeds, more callees are discovered, until the call graph finally
187 std::map<CallSite, std::vector<Function*> > IndirectCallees;
189 /// This enum defines the GraphNodes indices that correspond to important
198 bool runOnModule(Module &M) {
199 InitializeAliasAnalysis(this);
201 CollectConstraints(M);
202 DEBUG(PrintConstraints());
204 DEBUG(PrintPointsToGraph());
206 // Free the constraints list, as we don't need it to respond to alias
211 EscapingInternalFunctions.clear();
212 std::vector<Constraint>().swap(Constraints);
216 void releaseMemory() {
217 // FIXME: Until we have transitively required passes working correctly,
218 // this cannot be enabled! Otherwise, using -count-aa with the pass
219 // causes memory to be freed too early. :(
221 // The memory objects and ValueNodes data structures at the only ones that
222 // are still live after construction.
223 std::vector<Node>().swap(GraphNodes);
228 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
229 AliasAnalysis::getAnalysisUsage(AU);
230 AU.setPreservesAll(); // Does not transform code
233 //------------------------------------------------
234 // Implement the AliasAnalysis API
236 AliasResult alias(const Value *V1, unsigned V1Size,
237 const Value *V2, unsigned V2Size);
238 void getMustAliases(Value *P, std::vector<Value*> &RetVals);
239 bool pointsToConstantMemory(const Value *P);
241 virtual void deleteValue(Value *V) {
243 getAnalysis<AliasAnalysis>().deleteValue(V);
246 virtual void copyValue(Value *From, Value *To) {
247 ValueNodes[To] = ValueNodes[From];
248 getAnalysis<AliasAnalysis>().copyValue(From, To);
252 /// getNode - Return the node corresponding to the specified pointer scalar.
254 Node *getNode(Value *V) {
255 if (Constant *C = dyn_cast<Constant>(V))
256 if (!isa<GlobalValue>(C))
257 return getNodeForConstantPointer(C);
259 std::map<Value*, unsigned>::iterator I = ValueNodes.find(V);
260 if (I == ValueNodes.end()) {
262 assert(I != ValueNodes.end() &&
263 "Value does not have a node in the points-to graph!");
265 return &GraphNodes[I->second];
268 /// getObject - Return the node corresponding to the memory object for the
269 /// specified global or allocation instruction.
270 Node *getObject(Value *V) {
271 std::map<Value*, unsigned>::iterator I = ObjectNodes.find(V);
272 assert(I != ObjectNodes.end() &&
273 "Value does not have an object in the points-to graph!");
274 return &GraphNodes[I->second];
277 /// getReturnNode - Return the node representing the return value for the
278 /// specified function.
279 Node *getReturnNode(Function *F) {
280 std::map<Function*, unsigned>::iterator I = ReturnNodes.find(F);
281 assert(I != ReturnNodes.end() && "Function does not return a value!");
282 return &GraphNodes[I->second];
285 /// getVarargNode - Return the node representing the variable arguments
286 /// formal for the specified function.
287 Node *getVarargNode(Function *F) {
288 std::map<Function*, unsigned>::iterator I = VarargNodes.find(F);
289 assert(I != VarargNodes.end() && "Function does not take var args!");
290 return &GraphNodes[I->second];
293 /// getNodeValue - Get the node for the specified LLVM value and set the
294 /// value for it to be the specified value.
295 Node *getNodeValue(Value &V) {
296 return getNode(&V)->setValue(&V);
299 void IdentifyObjects(Module &M);
300 void CollectConstraints(Module &M);
301 void SolveConstraints();
303 Node *getNodeForConstantPointer(Constant *C);
304 Node *getNodeForConstantPointerTarget(Constant *C);
305 void AddGlobalInitializerConstraints(Node *N, Constant *C);
307 void AddConstraintsForNonInternalLinkage(Function *F);
308 bool AddConstraintsForExternalFunction(Function *F);
309 void AddConstraintsForCall(CallSite CS, Function *F);
312 void PrintNode(Node *N);
313 void PrintConstraints();
314 void PrintPointsToGraph();
316 //===------------------------------------------------------------------===//
317 // Instruction visitation methods for adding constraints
319 friend class InstVisitor<Andersens>;
320 void visitReturnInst(ReturnInst &RI);
321 void visitInvokeInst(InvokeInst &II) { visitCallSite(CallSite(&II)); }
322 void visitCallInst(CallInst &CI) { visitCallSite(CallSite(&CI)); }
323 void visitCallSite(CallSite CS);
324 void visitAllocationInst(AllocationInst &AI);
325 void visitLoadInst(LoadInst &LI);
326 void visitStoreInst(StoreInst &SI);
327 void visitGetElementPtrInst(GetElementPtrInst &GEP);
328 void visitPHINode(PHINode &PN);
329 void visitCastInst(CastInst &CI);
330 void visitSelectInst(SelectInst &SI);
331 void visitVANext(VANextInst &I);
332 void visitVAArg(VAArgInst &I);
333 void visitInstruction(Instruction &I);
336 RegisterOpt<Andersens> X("anders-aa",
337 "Andersen's Interprocedural Alias Analysis");
338 RegisterAnalysisGroup<AliasAnalysis, Andersens> Y;
341 ModulePass *llvm::createAndersensPass() { return new Andersens(); }
343 //===----------------------------------------------------------------------===//
344 // AliasAnalysis Interface Implementation
345 //===----------------------------------------------------------------------===//
347 AliasAnalysis::AliasResult Andersens::alias(const Value *V1, unsigned V1Size,
348 const Value *V2, unsigned V2Size) {
349 Node *N1 = getNode((Value*)V1);
350 Node *N2 = getNode((Value*)V2);
352 // Check to see if the two pointers are known to not alias. They don't alias
353 // if their points-to sets do not intersect.
354 if (!N1->intersectsIgnoring(N2, &GraphNodes[NullObject]))
357 return AliasAnalysis::alias(V1, V1Size, V2, V2Size);
361 /// getMustAlias - We can provide must alias information if we know that a
362 /// pointer can only point to a specific function or the null pointer.
363 /// Unfortunately we cannot determine must-alias information for global
364 /// variables or any other memory memory objects because we do not track whether
365 /// a pointer points to the beginning of an object or a field of it.
366 void Andersens::getMustAliases(Value *P, std::vector<Value*> &RetVals) {
367 Node *N = getNode(P);
368 Node::iterator I = N->begin();
370 // If there is exactly one element in the points-to set for the object...
373 Node *Pointee = *N->begin();
375 // If a function is the only object in the points-to set, then it must be
376 // the destination. Note that we can't handle global variables here,
377 // because we don't know if the pointer is actually pointing to a field of
378 // the global or to the beginning of it.
379 if (Value *V = Pointee->getValue()) {
380 if (Function *F = dyn_cast<Function>(V))
381 RetVals.push_back(F);
383 // If the object in the points-to set is the null object, then the null
384 // pointer is a must alias.
385 if (Pointee == &GraphNodes[NullObject])
386 RetVals.push_back(Constant::getNullValue(P->getType()));
391 AliasAnalysis::getMustAliases(P, RetVals);
394 /// pointsToConstantMemory - If we can determine that this pointer only points
395 /// to constant memory, return true. In practice, this means that if the
396 /// pointer can only point to constant globals, functions, or the null pointer,
399 bool Andersens::pointsToConstantMemory(const Value *P) {
400 Node *N = getNode((Value*)P);
401 for (Node::iterator I = N->begin(), E = N->end(); I != E; ++I) {
402 if (Value *V = (*I)->getValue()) {
403 if (!isa<GlobalValue>(V) || (isa<GlobalVariable>(V) &&
404 !cast<GlobalVariable>(V)->isConstant()))
405 return AliasAnalysis::pointsToConstantMemory(P);
407 if (*I != &GraphNodes[NullObject])
408 return AliasAnalysis::pointsToConstantMemory(P);
415 //===----------------------------------------------------------------------===//
416 // Object Identification Phase
417 //===----------------------------------------------------------------------===//
419 /// IdentifyObjects - This stage scans the program, adding an entry to the
420 /// GraphNodes list for each memory object in the program (global stack or
421 /// heap), and populates the ValueNodes and ObjectNodes maps for these objects.
423 void Andersens::IdentifyObjects(Module &M) {
424 unsigned NumObjects = 0;
426 // Object #0 is always the universal set: the object that we don't know
428 assert(NumObjects == UniversalSet && "Something changed!");
431 // Object #1 always represents the null pointer.
432 assert(NumObjects == NullPtr && "Something changed!");
435 // Object #2 always represents the null object (the object pointed to by null)
436 assert(NumObjects == NullObject && "Something changed!");
439 // Add all the globals first.
440 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
442 ObjectNodes[I] = NumObjects++;
443 ValueNodes[I] = NumObjects++;
446 // Add nodes for all of the functions and the instructions inside of them.
447 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
448 // The function itself is a memory object.
449 ValueNodes[F] = NumObjects++;
450 ObjectNodes[F] = NumObjects++;
451 if (isa<PointerType>(F->getFunctionType()->getReturnType()))
452 ReturnNodes[F] = NumObjects++;
453 if (F->getFunctionType()->isVarArg())
454 VarargNodes[F] = NumObjects++;
456 // Add nodes for all of the incoming pointer arguments.
457 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
459 if (isa<PointerType>(I->getType()))
460 ValueNodes[I] = NumObjects++;
462 // Scan the function body, creating a memory object for each heap/stack
463 // allocation in the body of the function and a node to represent all
464 // pointer values defined by instructions and used as operands.
465 for (inst_iterator II = inst_begin(F), E = inst_end(F); II != E; ++II) {
466 // If this is an heap or stack allocation, create a node for the memory
468 if (isa<PointerType>(II->getType())) {
469 ValueNodes[&*II] = NumObjects++;
470 if (AllocationInst *AI = dyn_cast<AllocationInst>(&*II))
471 ObjectNodes[AI] = NumObjects++;
476 // Now that we know how many objects to create, make them all now!
477 GraphNodes.resize(NumObjects);
478 NumNodes += NumObjects;
481 //===----------------------------------------------------------------------===//
482 // Constraint Identification Phase
483 //===----------------------------------------------------------------------===//
485 /// getNodeForConstantPointer - Return the node corresponding to the constant
487 Andersens::Node *Andersens::getNodeForConstantPointer(Constant *C) {
488 assert(isa<PointerType>(C->getType()) && "Not a constant pointer!");
490 if (isa<ConstantPointerNull>(C) || isa<UndefValue>(C))
491 return &GraphNodes[NullPtr];
492 else if (GlobalValue *GV = dyn_cast<GlobalValue>(C))
494 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
495 switch (CE->getOpcode()) {
496 case Instruction::GetElementPtr:
497 return getNodeForConstantPointer(CE->getOperand(0));
498 case Instruction::Cast:
499 if (isa<PointerType>(CE->getOperand(0)->getType()))
500 return getNodeForConstantPointer(CE->getOperand(0));
502 return &GraphNodes[UniversalSet];
504 std::cerr << "Constant Expr not yet handled: " << *CE << "\n";
508 assert(0 && "Unknown constant pointer!");
513 /// getNodeForConstantPointerTarget - Return the node POINTED TO by the
514 /// specified constant pointer.
515 Andersens::Node *Andersens::getNodeForConstantPointerTarget(Constant *C) {
516 assert(isa<PointerType>(C->getType()) && "Not a constant pointer!");
518 if (isa<ConstantPointerNull>(C))
519 return &GraphNodes[NullObject];
520 else if (GlobalValue *GV = dyn_cast<GlobalValue>(C))
521 return getObject(GV);
522 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
523 switch (CE->getOpcode()) {
524 case Instruction::GetElementPtr:
525 return getNodeForConstantPointerTarget(CE->getOperand(0));
526 case Instruction::Cast:
527 if (isa<PointerType>(CE->getOperand(0)->getType()))
528 return getNodeForConstantPointerTarget(CE->getOperand(0));
530 return &GraphNodes[UniversalSet];
532 std::cerr << "Constant Expr not yet handled: " << *CE << "\n";
536 assert(0 && "Unknown constant pointer!");
541 /// AddGlobalInitializerConstraints - Add inclusion constraints for the memory
542 /// object N, which contains values indicated by C.
543 void Andersens::AddGlobalInitializerConstraints(Node *N, Constant *C) {
544 if (C->getType()->isFirstClassType()) {
545 if (isa<PointerType>(C->getType()))
546 N->addPointerTo(getNodeForConstantPointer(C));
547 } else if (C->isNullValue()) {
548 N->addPointerTo(&GraphNodes[NullObject]);
551 // If this is an array or struct, include constraints for each element.
552 assert(isa<ConstantArray>(C) || isa<ConstantStruct>(C));
553 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i)
554 AddGlobalInitializerConstraints(N, cast<Constant>(C->getOperand(i)));
558 /// AddConstraintsForNonInternalLinkage - If this function does not have
559 /// internal linkage, realize that we can't trust anything passed into or
560 /// returned by this function.
561 void Andersens::AddConstraintsForNonInternalLinkage(Function *F) {
562 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
563 if (isa<PointerType>(I->getType()))
564 // If this is an argument of an externally accessible function, the
565 // incoming pointer might point to anything.
566 Constraints.push_back(Constraint(Constraint::Copy, getNode(I),
567 &GraphNodes[UniversalSet]));
570 /// AddConstraintsForExternalFunction - If this is a call to a "known" function,
571 /// add the constraints an return false. If this is a call to an unknown
572 /// function, return true.
573 bool Andersens::AddConstraintsForExternalFunction(Function *F) {
574 assert(F->isExternal() && "Not an external function!");
576 // These functions don't induce any points-to constraints.
577 if (F->getName() == "printf" || F->getName() == "fprintf" ||
578 F->getName() == "open" || F->getName() == "fopen" ||
579 F->getName() == "atoi" ||
580 F->getName() == "llvm.memset" || F->getName() == "memcmp" ||
581 F->getName() == "read" || F->getName() == "write")
584 // These functions do induce points-to edges.
585 if (F->getName() == "llvm.memcpy" || F->getName() == "llvm.memmove") {
586 Function::arg_iterator Dst = F->arg_begin(), Src = Dst;
587 // Note: this is a poor approximation, this says Dest = Src, instead of
590 Constraints.push_back(Constraint(Constraint::Copy, getNode(Dst),
600 /// CollectConstraints - This stage scans the program, adding a constraint to
601 /// the Constraints list for each instruction in the program that induces a
602 /// constraint, and setting up the initial points-to graph.
604 void Andersens::CollectConstraints(Module &M) {
605 // First, the universal set points to itself.
606 GraphNodes[UniversalSet].addPointerTo(&GraphNodes[UniversalSet]);
608 // Next, the null pointer points to the null object.
609 GraphNodes[NullPtr].addPointerTo(&GraphNodes[NullObject]);
611 // Next, add any constraints on global variables and their initializers.
612 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
614 // Associate the address of the global object as pointing to the memory for
615 // the global: &G = <G memory>
616 Node *Object = getObject(I);
618 getNodeValue(*I)->addPointerTo(Object);
620 if (I->hasInitializer()) {
621 AddGlobalInitializerConstraints(Object, I->getInitializer());
623 // If it doesn't have an initializer (i.e. it's defined in another
624 // translation unit), it points to the universal set.
625 Constraints.push_back(Constraint(Constraint::Copy, Object,
626 &GraphNodes[UniversalSet]));
630 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
631 // Make the function address point to the function object.
632 getNodeValue(*F)->addPointerTo(getObject(F)->setValue(F));
634 // Set up the return value node.
635 if (isa<PointerType>(F->getFunctionType()->getReturnType()))
636 getReturnNode(F)->setValue(F);
637 if (F->getFunctionType()->isVarArg())
638 getVarargNode(F)->setValue(F);
640 // Set up incoming argument nodes.
641 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
643 if (isa<PointerType>(I->getType()))
646 if (!F->hasInternalLinkage())
647 AddConstraintsForNonInternalLinkage(F);
649 if (!F->isExternal()) {
650 // Scan the function body, creating a memory object for each heap/stack
651 // allocation in the body of the function and a node to represent all
652 // pointer values defined by instructions and used as operands.
654 } else if (AddConstraintsForExternalFunction(F)) {
655 // If we don't "know" about this function, assume the worst.
657 // External functions that return pointers return the universal set.
658 if (isa<PointerType>(F->getFunctionType()->getReturnType()))
659 Constraints.push_back(Constraint(Constraint::Copy,
661 &GraphNodes[UniversalSet]));
663 // Any pointers that are passed into the function have the universal set
665 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
667 if (isa<PointerType>(I->getType())) {
668 // Pointers passed into external functions could have anything stored
670 Constraints.push_back(Constraint(Constraint::Store, getNode(I),
671 &GraphNodes[UniversalSet]));
672 // Memory objects passed into external function calls can have the
673 // universal set point to them.
674 Constraints.push_back(Constraint(Constraint::Copy,
675 &GraphNodes[UniversalSet],
679 // If this is an external varargs function, it can also store pointers
680 // into any pointers passed through the varargs section.
681 if (F->getFunctionType()->isVarArg())
682 Constraints.push_back(Constraint(Constraint::Store, getVarargNode(F),
683 &GraphNodes[UniversalSet]));
686 NumConstraints += Constraints.size();
690 void Andersens::visitInstruction(Instruction &I) {
692 return; // This function is just a big assert.
694 if (isa<BinaryOperator>(I))
696 // Most instructions don't have any effect on pointer values.
697 switch (I.getOpcode()) {
698 case Instruction::Br:
699 case Instruction::Switch:
700 case Instruction::Unwind:
701 case Instruction::Unreachable:
702 case Instruction::Free:
703 case Instruction::Shl:
704 case Instruction::Shr:
707 // Is this something we aren't handling yet?
708 std::cerr << "Unknown instruction: " << I;
713 void Andersens::visitAllocationInst(AllocationInst &AI) {
714 getNodeValue(AI)->addPointerTo(getObject(&AI)->setValue(&AI));
717 void Andersens::visitReturnInst(ReturnInst &RI) {
718 if (RI.getNumOperands() && isa<PointerType>(RI.getOperand(0)->getType()))
719 // return V --> <Copy/retval{F}/v>
720 Constraints.push_back(Constraint(Constraint::Copy,
721 getReturnNode(RI.getParent()->getParent()),
722 getNode(RI.getOperand(0))));
725 void Andersens::visitLoadInst(LoadInst &LI) {
726 if (isa<PointerType>(LI.getType()))
727 // P1 = load P2 --> <Load/P1/P2>
728 Constraints.push_back(Constraint(Constraint::Load, getNodeValue(LI),
729 getNode(LI.getOperand(0))));
732 void Andersens::visitStoreInst(StoreInst &SI) {
733 if (isa<PointerType>(SI.getOperand(0)->getType()))
734 // store P1, P2 --> <Store/P2/P1>
735 Constraints.push_back(Constraint(Constraint::Store,
736 getNode(SI.getOperand(1)),
737 getNode(SI.getOperand(0))));
740 void Andersens::visitGetElementPtrInst(GetElementPtrInst &GEP) {
741 // P1 = getelementptr P2, ... --> <Copy/P1/P2>
742 Constraints.push_back(Constraint(Constraint::Copy, getNodeValue(GEP),
743 getNode(GEP.getOperand(0))));
746 void Andersens::visitPHINode(PHINode &PN) {
747 if (isa<PointerType>(PN.getType())) {
748 Node *PNN = getNodeValue(PN);
749 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
750 // P1 = phi P2, P3 --> <Copy/P1/P2>, <Copy/P1/P3>, ...
751 Constraints.push_back(Constraint(Constraint::Copy, PNN,
752 getNode(PN.getIncomingValue(i))));
756 void Andersens::visitCastInst(CastInst &CI) {
757 Value *Op = CI.getOperand(0);
758 if (isa<PointerType>(CI.getType())) {
759 if (isa<PointerType>(Op->getType())) {
760 // P1 = cast P2 --> <Copy/P1/P2>
761 Constraints.push_back(Constraint(Constraint::Copy, getNodeValue(CI),
762 getNode(CI.getOperand(0))));
764 // P1 = cast int --> <Copy/P1/Univ>
765 Constraints.push_back(Constraint(Constraint::Copy, getNodeValue(CI),
766 &GraphNodes[UniversalSet]));
768 } else if (isa<PointerType>(Op->getType())) {
769 // int = cast P1 --> <Copy/Univ/P1>
770 Constraints.push_back(Constraint(Constraint::Copy,
771 &GraphNodes[UniversalSet],
772 getNode(CI.getOperand(0))));
776 void Andersens::visitSelectInst(SelectInst &SI) {
777 if (isa<PointerType>(SI.getType())) {
778 Node *SIN = getNodeValue(SI);
779 // P1 = select C, P2, P3 ---> <Copy/P1/P2>, <Copy/P1/P3>
780 Constraints.push_back(Constraint(Constraint::Copy, SIN,
781 getNode(SI.getOperand(1))));
782 Constraints.push_back(Constraint(Constraint::Copy, SIN,
783 getNode(SI.getOperand(2))));
787 void Andersens::visitVANext(VANextInst &I) {
789 assert(0 && "vanext not handled yet!");
791 void Andersens::visitVAArg(VAArgInst &I) {
792 assert(0 && "vaarg not handled yet!");
795 /// AddConstraintsForCall - Add constraints for a call with actual arguments
796 /// specified by CS to the function specified by F. Note that the types of
797 /// arguments might not match up in the case where this is an indirect call and
798 /// the function pointer has been casted. If this is the case, do something
800 void Andersens::AddConstraintsForCall(CallSite CS, Function *F) {
801 if (isa<PointerType>(CS.getType())) {
802 Node *CSN = getNode(CS.getInstruction());
803 if (isa<PointerType>(F->getFunctionType()->getReturnType())) {
804 Constraints.push_back(Constraint(Constraint::Copy, CSN,
807 // If the function returns a non-pointer value, handle this just like we
808 // treat a nonpointer cast to pointer.
809 Constraints.push_back(Constraint(Constraint::Copy, CSN,
810 &GraphNodes[UniversalSet]));
812 } else if (isa<PointerType>(F->getFunctionType()->getReturnType())) {
813 Constraints.push_back(Constraint(Constraint::Copy,
814 &GraphNodes[UniversalSet],
818 Function::arg_iterator AI = F->arg_begin(), AE = F->arg_end();
819 CallSite::arg_iterator ArgI = CS.arg_begin(), ArgE = CS.arg_end();
820 for (; AI != AE && ArgI != ArgE; ++AI, ++ArgI)
821 if (isa<PointerType>(AI->getType())) {
822 if (isa<PointerType>((*ArgI)->getType())) {
823 // Copy the actual argument into the formal argument.
824 Constraints.push_back(Constraint(Constraint::Copy, getNode(AI),
827 Constraints.push_back(Constraint(Constraint::Copy, getNode(AI),
828 &GraphNodes[UniversalSet]));
830 } else if (isa<PointerType>((*ArgI)->getType())) {
831 Constraints.push_back(Constraint(Constraint::Copy,
832 &GraphNodes[UniversalSet],
836 // Copy all pointers passed through the varargs section to the varargs node.
837 if (F->getFunctionType()->isVarArg())
838 for (; ArgI != ArgE; ++ArgI)
839 if (isa<PointerType>((*ArgI)->getType()))
840 Constraints.push_back(Constraint(Constraint::Copy, getVarargNode(F),
842 // If more arguments are passed in than we track, just drop them on the floor.
845 void Andersens::visitCallSite(CallSite CS) {
846 if (isa<PointerType>(CS.getType()))
847 getNodeValue(*CS.getInstruction());
849 if (Function *F = CS.getCalledFunction()) {
850 AddConstraintsForCall(CS, F);
852 // We don't handle indirect call sites yet. Keep track of them for when we
853 // discover the call graph incrementally.
854 IndirectCalls.push_back(CS);
858 //===----------------------------------------------------------------------===//
859 // Constraint Solving Phase
860 //===----------------------------------------------------------------------===//
862 /// intersects - Return true if the points-to set of this node intersects
863 /// with the points-to set of the specified node.
864 bool Andersens::Node::intersects(Node *N) const {
865 iterator I1 = begin(), I2 = N->begin(), E1 = end(), E2 = N->end();
866 while (I1 != E1 && I2 != E2) {
867 if (*I1 == *I2) return true;
876 /// intersectsIgnoring - Return true if the points-to set of this node
877 /// intersects with the points-to set of the specified node on any nodes
878 /// except for the specified node to ignore.
879 bool Andersens::Node::intersectsIgnoring(Node *N, Node *Ignoring) const {
880 iterator I1 = begin(), I2 = N->begin(), E1 = end(), E2 = N->end();
881 while (I1 != E1 && I2 != E2) {
883 if (*I1 != Ignoring) return true;
885 } else if (*I1 < *I2)
893 // Copy constraint: all edges out of the source node get copied to the
894 // destination node. This returns true if a change is made.
895 bool Andersens::Node::copyFrom(Node *N) {
896 // Use a mostly linear-time merge since both of the lists are sorted.
897 bool Changed = false;
898 iterator I = N->begin(), E = N->end();
900 while (I != E && i != Pointees.size()) {
901 if (Pointees[i] < *I) {
903 } else if (Pointees[i] == *I) {
906 // We found a new element to copy over.
908 Pointees.insert(Pointees.begin()+i, *I);
914 Pointees.insert(Pointees.end(), I, E);
921 bool Andersens::Node::loadFrom(Node *N) {
922 bool Changed = false;
923 for (iterator I = N->begin(), E = N->end(); I != E; ++I)
924 Changed |= copyFrom(*I);
928 bool Andersens::Node::storeThrough(Node *N) {
929 bool Changed = false;
930 for (iterator I = begin(), E = end(); I != E; ++I)
931 Changed |= (*I)->copyFrom(N);
936 /// SolveConstraints - This stage iteratively processes the constraints list
937 /// propagating constraints (adding edges to the Nodes in the points-to graph)
938 /// until a fixed point is reached.
940 void Andersens::SolveConstraints() {
942 unsigned Iteration = 0;
946 DEBUG(std::cerr << "Starting iteration #" << Iteration++ << "!\n");
948 // Loop over all of the constraints, applying them in turn.
949 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
950 Constraint &C = Constraints[i];
952 case Constraint::Copy:
953 Changed |= C.Dest->copyFrom(C.Src);
955 case Constraint::Load:
956 Changed |= C.Dest->loadFrom(C.Src);
958 case Constraint::Store:
959 Changed |= C.Dest->storeThrough(C.Src);
962 assert(0 && "Unknown constraint!");
967 // Check to see if any internal function's addresses have been passed to
968 // external functions. If so, we have to assume that their incoming
969 // arguments could be anything. If there are any internal functions in
970 // the universal node that we don't know about, we must iterate.
971 for (Node::iterator I = GraphNodes[UniversalSet].begin(),
972 E = GraphNodes[UniversalSet].end(); I != E; ++I)
973 if (Function *F = dyn_cast_or_null<Function>((*I)->getValue()))
974 if (F->hasInternalLinkage() &&
975 EscapingInternalFunctions.insert(F).second) {
976 // We found a function that is just now escaping. Mark it as if it
977 // didn't have internal linkage.
978 AddConstraintsForNonInternalLinkage(F);
979 DEBUG(std::cerr << "Found escaping internal function: "
980 << F->getName() << "\n");
981 ++NumEscapingFunctions;
984 // Check to see if we have discovered any new callees of the indirect call
985 // sites. If so, add constraints to the analysis.
986 for (unsigned i = 0, e = IndirectCalls.size(); i != e; ++i) {
987 CallSite CS = IndirectCalls[i];
988 std::vector<Function*> &KnownCallees = IndirectCallees[CS];
989 Node *CN = getNode(CS.getCalledValue());
991 for (Node::iterator NI = CN->begin(), E = CN->end(); NI != E; ++NI)
992 if (Function *F = dyn_cast_or_null<Function>((*NI)->getValue())) {
993 std::vector<Function*>::iterator IP =
994 std::lower_bound(KnownCallees.begin(), KnownCallees.end(), F);
995 if (IP == KnownCallees.end() || *IP != F) {
996 // Add the constraints for the call now.
997 AddConstraintsForCall(CS, F);
998 DEBUG(std::cerr << "Found actual callee '"
999 << F->getName() << "' for call: "
1000 << *CS.getInstruction() << "\n");
1001 ++NumIndirectCallees;
1002 KnownCallees.insert(IP, F);
1012 //===----------------------------------------------------------------------===//
1014 //===----------------------------------------------------------------------===//
1016 void Andersens::PrintNode(Node *N) {
1017 if (N == &GraphNodes[UniversalSet]) {
1018 std::cerr << "<universal>";
1020 } else if (N == &GraphNodes[NullPtr]) {
1021 std::cerr << "<nullptr>";
1023 } else if (N == &GraphNodes[NullObject]) {
1024 std::cerr << "<null>";
1028 assert(N->getValue() != 0 && "Never set node label!");
1029 Value *V = N->getValue();
1030 if (Function *F = dyn_cast<Function>(V)) {
1031 if (isa<PointerType>(F->getFunctionType()->getReturnType()) &&
1032 N == getReturnNode(F)) {
1033 std::cerr << F->getName() << ":retval";
1035 } else if (F->getFunctionType()->isVarArg() && N == getVarargNode(F)) {
1036 std::cerr << F->getName() << ":vararg";
1041 if (Instruction *I = dyn_cast<Instruction>(V))
1042 std::cerr << I->getParent()->getParent()->getName() << ":";
1043 else if (Argument *Arg = dyn_cast<Argument>(V))
1044 std::cerr << Arg->getParent()->getName() << ":";
1047 std::cerr << V->getName();
1049 std::cerr << "(unnamed)";
1051 if (isa<GlobalValue>(V) || isa<AllocationInst>(V))
1052 if (N == getObject(V))
1053 std::cerr << "<mem>";
1056 void Andersens::PrintConstraints() {
1057 std::cerr << "Constraints:\n";
1058 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
1059 std::cerr << " #" << i << ": ";
1060 Constraint &C = Constraints[i];
1061 if (C.Type == Constraint::Store)
1065 if (C.Type == Constraint::Load)
1072 void Andersens::PrintPointsToGraph() {
1073 std::cerr << "Points-to graph:\n";
1074 for (unsigned i = 0, e = GraphNodes.size(); i != e; ++i) {
1075 Node *N = &GraphNodes[i];
1076 std::cerr << "[" << (N->end() - N->begin()) << "] ";
1078 std::cerr << "\t--> ";
1079 for (Node::iterator I = N->begin(), E = N->end(); I != E; ++I) {
1080 if (I != N->begin()) std::cerr << ", ";