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/Compiler.h"
59 #include "llvm/Support/InstIterator.h"
60 #include "llvm/Support/InstVisitor.h"
61 #include "llvm/Analysis/AliasAnalysis.h"
62 #include "llvm/Analysis/Passes.h"
63 #include "llvm/Support/Debug.h"
64 #include "llvm/ADT/Statistic.h"
69 STATISTIC(NumIters , "Number of iterations to reach convergence");
70 STATISTIC(NumConstraints , "Number of constraints");
71 STATISTIC(NumNodes , "Number of nodes");
72 STATISTIC(NumEscapingFunctions, "Number of internal functions that escape");
73 STATISTIC(NumIndirectCallees , "Number of indirect callees found");
76 class VISIBILITY_HIDDEN Andersens : public ModulePass, public AliasAnalysis,
77 private InstVisitor<Andersens> {
79 static char ID; // Class identification, replacement for typeinfo
80 Andersens() : ModulePass((intptr_t)&ID) {}
82 /// Node class - This class is used to represent a memory object in the
83 /// program, and is the primitive used to build the points-to graph.
85 std::vector<Node*> Pointees;
88 static const unsigned ID; // Pass identification, replacement for typeid
90 Node *setValue(Value *V) {
91 assert(Val == 0 && "Value already set for this node!");
96 /// getValue - Return the LLVM value corresponding to this node.
98 Value *getValue() const { return Val; }
100 typedef std::vector<Node*>::const_iterator iterator;
101 iterator begin() const { return Pointees.begin(); }
102 iterator end() const { return Pointees.end(); }
104 /// addPointerTo - Add a pointer to the list of pointees of this node,
105 /// returning true if this caused a new pointer to be added, or false if
106 /// we already knew about the points-to relation.
107 bool addPointerTo(Node *N) {
108 std::vector<Node*>::iterator I = std::lower_bound(Pointees.begin(),
111 if (I != Pointees.end() && *I == N)
113 Pointees.insert(I, N);
117 /// intersects - Return true if the points-to set of this node intersects
118 /// with the points-to set of the specified node.
119 bool intersects(Node *N) const;
121 /// intersectsIgnoring - Return true if the points-to set of this node
122 /// intersects with the points-to set of the specified node on any nodes
123 /// except for the specified node to ignore.
124 bool intersectsIgnoring(Node *N, Node *Ignoring) const;
126 // Constraint application methods.
127 bool copyFrom(Node *N);
128 bool loadFrom(Node *N);
129 bool storeThrough(Node *N);
132 /// GraphNodes - This vector is populated as part of the object
133 /// identification stage of the analysis, which populates this vector with a
134 /// node for each memory object and fills in the ValueNodes map.
135 std::vector<Node> GraphNodes;
137 /// ValueNodes - This map indicates the Node that a particular Value* is
138 /// represented by. This contains entries for all pointers.
139 std::map<Value*, unsigned> ValueNodes;
141 /// ObjectNodes - This map contains entries for each memory object in the
142 /// program: globals, alloca's and mallocs.
143 std::map<Value*, unsigned> ObjectNodes;
145 /// ReturnNodes - This map contains an entry for each function in the
146 /// program that returns a value.
147 std::map<Function*, unsigned> ReturnNodes;
149 /// VarargNodes - This map contains the entry used to represent all pointers
150 /// passed through the varargs portion of a function call for a particular
151 /// function. An entry is not present in this map for functions that do not
152 /// take variable arguments.
153 std::map<Function*, unsigned> VarargNodes;
155 /// Constraint - Objects of this structure are used to represent the various
156 /// constraints identified by the algorithm. The constraints are 'copy',
157 /// for statements like "A = B", 'load' for statements like "A = *B", and
158 /// 'store' for statements like "*A = B".
160 enum ConstraintType { Copy, Load, Store } Type;
163 Constraint(ConstraintType Ty, Node *D, Node *S)
164 : Type(Ty), Dest(D), Src(S) {}
167 /// Constraints - This vector contains a list of all of the constraints
168 /// identified by the program.
169 std::vector<Constraint> Constraints;
171 /// EscapingInternalFunctions - This set contains all of the internal
172 /// functions that are found to escape from the program. If the address of
173 /// an internal function is passed to an external function or otherwise
174 /// escapes from the analyzed portion of the program, we must assume that
175 /// any pointer arguments can alias the universal node. This set keeps
176 /// track of those functions we are assuming to escape so far.
177 std::set<Function*> EscapingInternalFunctions;
179 /// IndirectCalls - This contains a list of all of the indirect call sites
180 /// in the program. Since the call graph is iteratively discovered, we may
181 /// need to add constraints to our graph as we find new targets of function
183 std::vector<CallSite> IndirectCalls;
185 /// IndirectCallees - For each call site in the indirect calls list, keep
186 /// track of the callees that we have discovered so far. As the analysis
187 /// proceeds, more callees are discovered, until the call graph finally
189 std::map<CallSite, std::vector<Function*> > IndirectCallees;
191 /// This enum defines the GraphNodes indices that correspond to important
200 bool runOnModule(Module &M) {
201 InitializeAliasAnalysis(this);
203 CollectConstraints(M);
204 DEBUG(PrintConstraints());
206 DEBUG(PrintPointsToGraph());
208 // Free the constraints list, as we don't need it to respond to alias
213 EscapingInternalFunctions.clear();
214 std::vector<Constraint>().swap(Constraints);
218 void releaseMemory() {
219 // FIXME: Until we have transitively required passes working correctly,
220 // this cannot be enabled! Otherwise, using -count-aa with the pass
221 // causes memory to be freed too early. :(
223 // The memory objects and ValueNodes data structures at the only ones that
224 // are still live after construction.
225 std::vector<Node>().swap(GraphNodes);
230 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
231 AliasAnalysis::getAnalysisUsage(AU);
232 AU.setPreservesAll(); // Does not transform code
235 //------------------------------------------------
236 // Implement the AliasAnalysis API
238 AliasResult alias(const Value *V1, unsigned V1Size,
239 const Value *V2, unsigned V2Size);
240 virtual ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size);
241 virtual ModRefResult getModRefInfo(CallSite CS1, CallSite CS2);
242 void getMustAliases(Value *P, std::vector<Value*> &RetVals);
243 bool pointsToConstantMemory(const Value *P);
245 virtual void deleteValue(Value *V) {
247 getAnalysis<AliasAnalysis>().deleteValue(V);
250 virtual void copyValue(Value *From, Value *To) {
251 ValueNodes[To] = ValueNodes[From];
252 getAnalysis<AliasAnalysis>().copyValue(From, To);
256 /// getNode - Return the node corresponding to the specified pointer scalar.
258 Node *getNode(Value *V) {
259 if (Constant *C = dyn_cast<Constant>(V))
260 if (!isa<GlobalValue>(C))
261 return getNodeForConstantPointer(C);
263 std::map<Value*, unsigned>::iterator I = ValueNodes.find(V);
264 if (I == ValueNodes.end()) {
268 assert(0 && "Value does not have a node in the points-to graph!");
270 return &GraphNodes[I->second];
273 /// getObject - Return the node corresponding to the memory object for the
274 /// specified global or allocation instruction.
275 Node *getObject(Value *V) {
276 std::map<Value*, unsigned>::iterator I = ObjectNodes.find(V);
277 assert(I != ObjectNodes.end() &&
278 "Value does not have an object in the points-to graph!");
279 return &GraphNodes[I->second];
282 /// getReturnNode - Return the node representing the return value for the
283 /// specified function.
284 Node *getReturnNode(Function *F) {
285 std::map<Function*, unsigned>::iterator I = ReturnNodes.find(F);
286 assert(I != ReturnNodes.end() && "Function does not return a value!");
287 return &GraphNodes[I->second];
290 /// getVarargNode - Return the node representing the variable arguments
291 /// formal for the specified function.
292 Node *getVarargNode(Function *F) {
293 std::map<Function*, unsigned>::iterator I = VarargNodes.find(F);
294 assert(I != VarargNodes.end() && "Function does not take var args!");
295 return &GraphNodes[I->second];
298 /// getNodeValue - Get the node for the specified LLVM value and set the
299 /// value for it to be the specified value.
300 Node *getNodeValue(Value &V) {
301 return getNode(&V)->setValue(&V);
304 void IdentifyObjects(Module &M);
305 void CollectConstraints(Module &M);
306 void SolveConstraints();
308 Node *getNodeForConstantPointer(Constant *C);
309 Node *getNodeForConstantPointerTarget(Constant *C);
310 void AddGlobalInitializerConstraints(Node *N, Constant *C);
312 void AddConstraintsForNonInternalLinkage(Function *F);
313 void AddConstraintsForCall(CallSite CS, Function *F);
314 bool AddConstraintsForExternalCall(CallSite CS, Function *F);
317 void PrintNode(Node *N);
318 void PrintConstraints();
319 void PrintPointsToGraph();
321 //===------------------------------------------------------------------===//
322 // Instruction visitation methods for adding constraints
324 friend class InstVisitor<Andersens>;
325 void visitReturnInst(ReturnInst &RI);
326 void visitInvokeInst(InvokeInst &II) { visitCallSite(CallSite(&II)); }
327 void visitCallInst(CallInst &CI) { visitCallSite(CallSite(&CI)); }
328 void visitCallSite(CallSite CS);
329 void visitAllocationInst(AllocationInst &AI);
330 void visitLoadInst(LoadInst &LI);
331 void visitStoreInst(StoreInst &SI);
332 void visitGetElementPtrInst(GetElementPtrInst &GEP);
333 void visitPHINode(PHINode &PN);
334 void visitCastInst(CastInst &CI);
335 void visitICmpInst(ICmpInst &ICI) {} // NOOP!
336 void visitFCmpInst(FCmpInst &ICI) {} // NOOP!
337 void visitSelectInst(SelectInst &SI);
338 void visitVAArg(VAArgInst &I);
339 void visitInstruction(Instruction &I);
342 char Andersens::ID = 0;
343 RegisterPass<Andersens> X("anders-aa",
344 "Andersen's Interprocedural Alias Analysis");
345 RegisterAnalysisGroup<AliasAnalysis> Y(X);
348 ModulePass *llvm::createAndersensPass() { return new Andersens(); }
350 //===----------------------------------------------------------------------===//
351 // AliasAnalysis Interface Implementation
352 //===----------------------------------------------------------------------===//
354 AliasAnalysis::AliasResult Andersens::alias(const Value *V1, unsigned V1Size,
355 const Value *V2, unsigned V2Size) {
356 Node *N1 = getNode(const_cast<Value*>(V1));
357 Node *N2 = getNode(const_cast<Value*>(V2));
359 // Check to see if the two pointers are known to not alias. They don't alias
360 // if their points-to sets do not intersect.
361 if (!N1->intersectsIgnoring(N2, &GraphNodes[NullObject]))
364 return AliasAnalysis::alias(V1, V1Size, V2, V2Size);
367 AliasAnalysis::ModRefResult
368 Andersens::getModRefInfo(CallSite CS, Value *P, unsigned Size) {
369 // The only thing useful that we can contribute for mod/ref information is
370 // when calling external function calls: if we know that memory never escapes
371 // from the program, it cannot be modified by an external call.
373 // NOTE: This is not really safe, at least not when the entire program is not
374 // available. The deal is that the external function could call back into the
375 // program and modify stuff. We ignore this technical niggle for now. This
376 // is, after all, a "research quality" implementation of Andersen's analysis.
377 if (Function *F = CS.getCalledFunction())
378 if (F->isDeclaration()) {
379 Node *N1 = getNode(P);
381 if (N1->begin() == N1->end())
382 return NoModRef; // P doesn't point to anything.
384 // Get the first pointee.
385 Node *FirstPointee = *N1->begin();
386 if (FirstPointee != &GraphNodes[UniversalSet])
387 return NoModRef; // P doesn't point to the universal set.
390 return AliasAnalysis::getModRefInfo(CS, P, Size);
393 AliasAnalysis::ModRefResult
394 Andersens::getModRefInfo(CallSite CS1, CallSite CS2) {
395 return AliasAnalysis::getModRefInfo(CS1,CS2);
398 /// getMustAlias - We can provide must alias information if we know that a
399 /// pointer can only point to a specific function or the null pointer.
400 /// Unfortunately we cannot determine must-alias information for global
401 /// variables or any other memory memory objects because we do not track whether
402 /// a pointer points to the beginning of an object or a field of it.
403 void Andersens::getMustAliases(Value *P, std::vector<Value*> &RetVals) {
404 Node *N = getNode(P);
405 Node::iterator I = N->begin();
407 // If there is exactly one element in the points-to set for the object...
410 Node *Pointee = *N->begin();
412 // If a function is the only object in the points-to set, then it must be
413 // the destination. Note that we can't handle global variables here,
414 // because we don't know if the pointer is actually pointing to a field of
415 // the global or to the beginning of it.
416 if (Value *V = Pointee->getValue()) {
417 if (Function *F = dyn_cast<Function>(V))
418 RetVals.push_back(F);
420 // If the object in the points-to set is the null object, then the null
421 // pointer is a must alias.
422 if (Pointee == &GraphNodes[NullObject])
423 RetVals.push_back(Constant::getNullValue(P->getType()));
428 AliasAnalysis::getMustAliases(P, RetVals);
431 /// pointsToConstantMemory - If we can determine that this pointer only points
432 /// to constant memory, return true. In practice, this means that if the
433 /// pointer can only point to constant globals, functions, or the null pointer,
436 bool Andersens::pointsToConstantMemory(const Value *P) {
437 Node *N = getNode((Value*)P);
438 for (Node::iterator I = N->begin(), E = N->end(); I != E; ++I) {
439 if (Value *V = (*I)->getValue()) {
440 if (!isa<GlobalValue>(V) || (isa<GlobalVariable>(V) &&
441 !cast<GlobalVariable>(V)->isConstant()))
442 return AliasAnalysis::pointsToConstantMemory(P);
444 if (*I != &GraphNodes[NullObject])
445 return AliasAnalysis::pointsToConstantMemory(P);
452 //===----------------------------------------------------------------------===//
453 // Object Identification Phase
454 //===----------------------------------------------------------------------===//
456 /// IdentifyObjects - This stage scans the program, adding an entry to the
457 /// GraphNodes list for each memory object in the program (global stack or
458 /// heap), and populates the ValueNodes and ObjectNodes maps for these objects.
460 void Andersens::IdentifyObjects(Module &M) {
461 unsigned NumObjects = 0;
463 // Object #0 is always the universal set: the object that we don't know
465 assert(NumObjects == UniversalSet && "Something changed!");
468 // Object #1 always represents the null pointer.
469 assert(NumObjects == NullPtr && "Something changed!");
472 // Object #2 always represents the null object (the object pointed to by null)
473 assert(NumObjects == NullObject && "Something changed!");
476 // Add all the globals first.
477 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
479 ObjectNodes[I] = NumObjects++;
480 ValueNodes[I] = NumObjects++;
483 // Add nodes for all of the functions and the instructions inside of them.
484 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
485 // The function itself is a memory object.
486 ValueNodes[F] = NumObjects++;
487 ObjectNodes[F] = NumObjects++;
488 if (isa<PointerType>(F->getFunctionType()->getReturnType()))
489 ReturnNodes[F] = NumObjects++;
490 if (F->getFunctionType()->isVarArg())
491 VarargNodes[F] = NumObjects++;
493 // Add nodes for all of the incoming pointer arguments.
494 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
496 if (isa<PointerType>(I->getType()))
497 ValueNodes[I] = NumObjects++;
499 // Scan the function body, creating a memory object for each heap/stack
500 // allocation in the body of the function and a node to represent all
501 // pointer values defined by instructions and used as operands.
502 for (inst_iterator II = inst_begin(F), E = inst_end(F); II != E; ++II) {
503 // If this is an heap or stack allocation, create a node for the memory
505 if (isa<PointerType>(II->getType())) {
506 ValueNodes[&*II] = NumObjects++;
507 if (AllocationInst *AI = dyn_cast<AllocationInst>(&*II))
508 ObjectNodes[AI] = NumObjects++;
513 // Now that we know how many objects to create, make them all now!
514 GraphNodes.resize(NumObjects);
515 NumNodes += NumObjects;
518 //===----------------------------------------------------------------------===//
519 // Constraint Identification Phase
520 //===----------------------------------------------------------------------===//
522 /// getNodeForConstantPointer - Return the node corresponding to the constant
524 Andersens::Node *Andersens::getNodeForConstantPointer(Constant *C) {
525 assert(isa<PointerType>(C->getType()) && "Not a constant pointer!");
527 if (isa<ConstantPointerNull>(C) || isa<UndefValue>(C))
528 return &GraphNodes[NullPtr];
529 else if (GlobalValue *GV = dyn_cast<GlobalValue>(C))
531 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
532 switch (CE->getOpcode()) {
533 case Instruction::GetElementPtr:
534 return getNodeForConstantPointer(CE->getOperand(0));
535 case Instruction::IntToPtr:
536 return &GraphNodes[UniversalSet];
537 case Instruction::BitCast:
538 return getNodeForConstantPointer(CE->getOperand(0));
540 cerr << "Constant Expr not yet handled: " << *CE << "\n";
544 assert(0 && "Unknown constant pointer!");
549 /// getNodeForConstantPointerTarget - Return the node POINTED TO by the
550 /// specified constant pointer.
551 Andersens::Node *Andersens::getNodeForConstantPointerTarget(Constant *C) {
552 assert(isa<PointerType>(C->getType()) && "Not a constant pointer!");
554 if (isa<ConstantPointerNull>(C))
555 return &GraphNodes[NullObject];
556 else if (GlobalValue *GV = dyn_cast<GlobalValue>(C))
557 return getObject(GV);
558 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
559 switch (CE->getOpcode()) {
560 case Instruction::GetElementPtr:
561 return getNodeForConstantPointerTarget(CE->getOperand(0));
562 case Instruction::IntToPtr:
563 return &GraphNodes[UniversalSet];
564 case Instruction::BitCast:
565 return getNodeForConstantPointerTarget(CE->getOperand(0));
567 cerr << "Constant Expr not yet handled: " << *CE << "\n";
571 assert(0 && "Unknown constant pointer!");
576 /// AddGlobalInitializerConstraints - Add inclusion constraints for the memory
577 /// object N, which contains values indicated by C.
578 void Andersens::AddGlobalInitializerConstraints(Node *N, Constant *C) {
579 if (C->getType()->isFirstClassType()) {
580 if (isa<PointerType>(C->getType()))
581 N->copyFrom(getNodeForConstantPointer(C));
583 } else if (C->isNullValue()) {
584 N->addPointerTo(&GraphNodes[NullObject]);
586 } else if (!isa<UndefValue>(C)) {
587 // If this is an array or struct, include constraints for each element.
588 assert(isa<ConstantArray>(C) || isa<ConstantStruct>(C));
589 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i)
590 AddGlobalInitializerConstraints(N, cast<Constant>(C->getOperand(i)));
594 /// AddConstraintsForNonInternalLinkage - If this function does not have
595 /// internal linkage, realize that we can't trust anything passed into or
596 /// returned by this function.
597 void Andersens::AddConstraintsForNonInternalLinkage(Function *F) {
598 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
599 if (isa<PointerType>(I->getType()))
600 // If this is an argument of an externally accessible function, the
601 // incoming pointer might point to anything.
602 Constraints.push_back(Constraint(Constraint::Copy, getNode(I),
603 &GraphNodes[UniversalSet]));
606 /// AddConstraintsForCall - If this is a call to a "known" function, add the
607 /// constraints and return true. If this is a call to an unknown function,
609 bool Andersens::AddConstraintsForExternalCall(CallSite CS, Function *F) {
610 assert(F->isDeclaration() && "Not an external function!");
612 // These functions don't induce any points-to constraints.
613 if (F->getName() == "atoi" || F->getName() == "atof" ||
614 F->getName() == "atol" || F->getName() == "atoll" ||
615 F->getName() == "remove" || F->getName() == "unlink" ||
616 F->getName() == "rename" || F->getName() == "memcmp" ||
617 F->getName() == "llvm.memset.i32" ||
618 F->getName() == "llvm.memset.i64" ||
619 F->getName() == "strcmp" || F->getName() == "strncmp" ||
620 F->getName() == "execl" || F->getName() == "execlp" ||
621 F->getName() == "execle" || F->getName() == "execv" ||
622 F->getName() == "execvp" || F->getName() == "chmod" ||
623 F->getName() == "puts" || F->getName() == "write" ||
624 F->getName() == "open" || F->getName() == "create" ||
625 F->getName() == "truncate" || F->getName() == "chdir" ||
626 F->getName() == "mkdir" || F->getName() == "rmdir" ||
627 F->getName() == "read" || F->getName() == "pipe" ||
628 F->getName() == "wait" || F->getName() == "time" ||
629 F->getName() == "stat" || F->getName() == "fstat" ||
630 F->getName() == "lstat" || F->getName() == "strtod" ||
631 F->getName() == "strtof" || F->getName() == "strtold" ||
632 F->getName() == "fopen" || F->getName() == "fdopen" ||
633 F->getName() == "freopen" ||
634 F->getName() == "fflush" || F->getName() == "feof" ||
635 F->getName() == "fileno" || F->getName() == "clearerr" ||
636 F->getName() == "rewind" || F->getName() == "ftell" ||
637 F->getName() == "ferror" || F->getName() == "fgetc" ||
638 F->getName() == "fgetc" || F->getName() == "_IO_getc" ||
639 F->getName() == "fwrite" || F->getName() == "fread" ||
640 F->getName() == "fgets" || F->getName() == "ungetc" ||
641 F->getName() == "fputc" ||
642 F->getName() == "fputs" || F->getName() == "putc" ||
643 F->getName() == "ftell" || F->getName() == "rewind" ||
644 F->getName() == "_IO_putc" || F->getName() == "fseek" ||
645 F->getName() == "fgetpos" || F->getName() == "fsetpos" ||
646 F->getName() == "printf" || F->getName() == "fprintf" ||
647 F->getName() == "sprintf" || F->getName() == "vprintf" ||
648 F->getName() == "vfprintf" || F->getName() == "vsprintf" ||
649 F->getName() == "scanf" || F->getName() == "fscanf" ||
650 F->getName() == "sscanf" || F->getName() == "__assert_fail" ||
651 F->getName() == "modf")
655 // These functions do induce points-to edges.
656 if (F->getName() == "llvm.memcpy.i32" || F->getName() == "llvm.memcpy.i64" ||
657 F->getName() == "llvm.memmove.i32" ||F->getName() == "llvm.memmove.i64" ||
658 F->getName() == "memmove") {
659 // Note: this is a poor approximation, this says Dest = Src, instead of
661 Constraints.push_back(Constraint(Constraint::Copy,
662 getNode(CS.getArgument(0)),
663 getNode(CS.getArgument(1))));
668 if (F->getName() == "realloc" || F->getName() == "strchr" ||
669 F->getName() == "strrchr" || F->getName() == "strstr" ||
670 F->getName() == "strtok") {
671 Constraints.push_back(Constraint(Constraint::Copy,
672 getNode(CS.getInstruction()),
673 getNode(CS.getArgument(0))));
682 /// CollectConstraints - This stage scans the program, adding a constraint to
683 /// the Constraints list for each instruction in the program that induces a
684 /// constraint, and setting up the initial points-to graph.
686 void Andersens::CollectConstraints(Module &M) {
687 // First, the universal set points to itself.
688 GraphNodes[UniversalSet].addPointerTo(&GraphNodes[UniversalSet]);
689 //Constraints.push_back(Constraint(Constraint::Load, &GraphNodes[UniversalSet],
690 // &GraphNodes[UniversalSet]));
691 Constraints.push_back(Constraint(Constraint::Store, &GraphNodes[UniversalSet],
692 &GraphNodes[UniversalSet]));
694 // Next, the null pointer points to the null object.
695 GraphNodes[NullPtr].addPointerTo(&GraphNodes[NullObject]);
697 // Next, add any constraints on global variables and their initializers.
698 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
700 // Associate the address of the global object as pointing to the memory for
701 // the global: &G = <G memory>
702 Node *Object = getObject(I);
704 getNodeValue(*I)->addPointerTo(Object);
706 if (I->hasInitializer()) {
707 AddGlobalInitializerConstraints(Object, I->getInitializer());
709 // If it doesn't have an initializer (i.e. it's defined in another
710 // translation unit), it points to the universal set.
711 Constraints.push_back(Constraint(Constraint::Copy, Object,
712 &GraphNodes[UniversalSet]));
716 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
717 // Make the function address point to the function object.
718 getNodeValue(*F)->addPointerTo(getObject(F)->setValue(F));
720 // Set up the return value node.
721 if (isa<PointerType>(F->getFunctionType()->getReturnType()))
722 getReturnNode(F)->setValue(F);
723 if (F->getFunctionType()->isVarArg())
724 getVarargNode(F)->setValue(F);
726 // Set up incoming argument nodes.
727 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
729 if (isa<PointerType>(I->getType()))
732 if (!F->hasInternalLinkage())
733 AddConstraintsForNonInternalLinkage(F);
735 if (!F->isDeclaration()) {
736 // Scan the function body, creating a memory object for each heap/stack
737 // allocation in the body of the function and a node to represent all
738 // pointer values defined by instructions and used as operands.
741 // External functions that return pointers return the universal set.
742 if (isa<PointerType>(F->getFunctionType()->getReturnType()))
743 Constraints.push_back(Constraint(Constraint::Copy,
745 &GraphNodes[UniversalSet]));
747 // Any pointers that are passed into the function have the universal set
749 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
751 if (isa<PointerType>(I->getType())) {
752 // Pointers passed into external functions could have anything stored
754 Constraints.push_back(Constraint(Constraint::Store, getNode(I),
755 &GraphNodes[UniversalSet]));
756 // Memory objects passed into external function calls can have the
757 // universal set point to them.
758 Constraints.push_back(Constraint(Constraint::Copy,
759 &GraphNodes[UniversalSet],
763 // If this is an external varargs function, it can also store pointers
764 // into any pointers passed through the varargs section.
765 if (F->getFunctionType()->isVarArg())
766 Constraints.push_back(Constraint(Constraint::Store, getVarargNode(F),
767 &GraphNodes[UniversalSet]));
770 NumConstraints += Constraints.size();
774 void Andersens::visitInstruction(Instruction &I) {
776 return; // This function is just a big assert.
778 if (isa<BinaryOperator>(I))
780 // Most instructions don't have any effect on pointer values.
781 switch (I.getOpcode()) {
782 case Instruction::Br:
783 case Instruction::Switch:
784 case Instruction::Unwind:
785 case Instruction::Unreachable:
786 case Instruction::Free:
787 case Instruction::ICmp:
788 case Instruction::FCmp:
791 // Is this something we aren't handling yet?
792 cerr << "Unknown instruction: " << I;
797 void Andersens::visitAllocationInst(AllocationInst &AI) {
798 getNodeValue(AI)->addPointerTo(getObject(&AI)->setValue(&AI));
801 void Andersens::visitReturnInst(ReturnInst &RI) {
802 if (RI.getNumOperands() && isa<PointerType>(RI.getOperand(0)->getType()))
803 // return V --> <Copy/retval{F}/v>
804 Constraints.push_back(Constraint(Constraint::Copy,
805 getReturnNode(RI.getParent()->getParent()),
806 getNode(RI.getOperand(0))));
809 void Andersens::visitLoadInst(LoadInst &LI) {
810 if (isa<PointerType>(LI.getType()))
811 // P1 = load P2 --> <Load/P1/P2>
812 Constraints.push_back(Constraint(Constraint::Load, getNodeValue(LI),
813 getNode(LI.getOperand(0))));
816 void Andersens::visitStoreInst(StoreInst &SI) {
817 if (isa<PointerType>(SI.getOperand(0)->getType()))
818 // store P1, P2 --> <Store/P2/P1>
819 Constraints.push_back(Constraint(Constraint::Store,
820 getNode(SI.getOperand(1)),
821 getNode(SI.getOperand(0))));
824 void Andersens::visitGetElementPtrInst(GetElementPtrInst &GEP) {
825 // P1 = getelementptr P2, ... --> <Copy/P1/P2>
826 Constraints.push_back(Constraint(Constraint::Copy, getNodeValue(GEP),
827 getNode(GEP.getOperand(0))));
830 void Andersens::visitPHINode(PHINode &PN) {
831 if (isa<PointerType>(PN.getType())) {
832 Node *PNN = getNodeValue(PN);
833 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
834 // P1 = phi P2, P3 --> <Copy/P1/P2>, <Copy/P1/P3>, ...
835 Constraints.push_back(Constraint(Constraint::Copy, PNN,
836 getNode(PN.getIncomingValue(i))));
840 void Andersens::visitCastInst(CastInst &CI) {
841 Value *Op = CI.getOperand(0);
842 if (isa<PointerType>(CI.getType())) {
843 if (isa<PointerType>(Op->getType())) {
844 // P1 = cast P2 --> <Copy/P1/P2>
845 Constraints.push_back(Constraint(Constraint::Copy, getNodeValue(CI),
846 getNode(CI.getOperand(0))));
848 // P1 = cast int --> <Copy/P1/Univ>
850 Constraints.push_back(Constraint(Constraint::Copy, getNodeValue(CI),
851 &GraphNodes[UniversalSet]));
856 } else if (isa<PointerType>(Op->getType())) {
857 // int = cast P1 --> <Copy/Univ/P1>
859 Constraints.push_back(Constraint(Constraint::Copy,
860 &GraphNodes[UniversalSet],
861 getNode(CI.getOperand(0))));
863 getNode(CI.getOperand(0));
868 void Andersens::visitSelectInst(SelectInst &SI) {
869 if (isa<PointerType>(SI.getType())) {
870 Node *SIN = getNodeValue(SI);
871 // P1 = select C, P2, P3 ---> <Copy/P1/P2>, <Copy/P1/P3>
872 Constraints.push_back(Constraint(Constraint::Copy, SIN,
873 getNode(SI.getOperand(1))));
874 Constraints.push_back(Constraint(Constraint::Copy, SIN,
875 getNode(SI.getOperand(2))));
879 void Andersens::visitVAArg(VAArgInst &I) {
880 assert(0 && "vaarg not handled yet!");
883 /// AddConstraintsForCall - Add constraints for a call with actual arguments
884 /// specified by CS to the function specified by F. Note that the types of
885 /// arguments might not match up in the case where this is an indirect call and
886 /// the function pointer has been casted. If this is the case, do something
888 void Andersens::AddConstraintsForCall(CallSite CS, Function *F) {
889 // If this is a call to an external function, handle it directly to get some
890 // taste of context sensitivity.
891 if (F->isDeclaration() && AddConstraintsForExternalCall(CS, F))
894 if (isa<PointerType>(CS.getType())) {
895 Node *CSN = getNode(CS.getInstruction());
896 if (isa<PointerType>(F->getFunctionType()->getReturnType())) {
897 Constraints.push_back(Constraint(Constraint::Copy, CSN,
900 // If the function returns a non-pointer value, handle this just like we
901 // treat a nonpointer cast to pointer.
902 Constraints.push_back(Constraint(Constraint::Copy, CSN,
903 &GraphNodes[UniversalSet]));
905 } else if (isa<PointerType>(F->getFunctionType()->getReturnType())) {
906 Constraints.push_back(Constraint(Constraint::Copy,
907 &GraphNodes[UniversalSet],
911 Function::arg_iterator AI = F->arg_begin(), AE = F->arg_end();
912 CallSite::arg_iterator ArgI = CS.arg_begin(), ArgE = CS.arg_end();
913 for (; AI != AE && ArgI != ArgE; ++AI, ++ArgI)
914 if (isa<PointerType>(AI->getType())) {
915 if (isa<PointerType>((*ArgI)->getType())) {
916 // Copy the actual argument into the formal argument.
917 Constraints.push_back(Constraint(Constraint::Copy, getNode(AI),
920 Constraints.push_back(Constraint(Constraint::Copy, getNode(AI),
921 &GraphNodes[UniversalSet]));
923 } else if (isa<PointerType>((*ArgI)->getType())) {
924 Constraints.push_back(Constraint(Constraint::Copy,
925 &GraphNodes[UniversalSet],
929 // Copy all pointers passed through the varargs section to the varargs node.
930 if (F->getFunctionType()->isVarArg())
931 for (; ArgI != ArgE; ++ArgI)
932 if (isa<PointerType>((*ArgI)->getType()))
933 Constraints.push_back(Constraint(Constraint::Copy, getVarargNode(F),
935 // If more arguments are passed in than we track, just drop them on the floor.
938 void Andersens::visitCallSite(CallSite CS) {
939 if (isa<PointerType>(CS.getType()))
940 getNodeValue(*CS.getInstruction());
942 if (Function *F = CS.getCalledFunction()) {
943 AddConstraintsForCall(CS, F);
945 // We don't handle indirect call sites yet. Keep track of them for when we
946 // discover the call graph incrementally.
947 IndirectCalls.push_back(CS);
951 //===----------------------------------------------------------------------===//
952 // Constraint Solving Phase
953 //===----------------------------------------------------------------------===//
955 /// intersects - Return true if the points-to set of this node intersects
956 /// with the points-to set of the specified node.
957 bool Andersens::Node::intersects(Node *N) const {
958 iterator I1 = begin(), I2 = N->begin(), E1 = end(), E2 = N->end();
959 while (I1 != E1 && I2 != E2) {
960 if (*I1 == *I2) return true;
969 /// intersectsIgnoring - Return true if the points-to set of this node
970 /// intersects with the points-to set of the specified node on any nodes
971 /// except for the specified node to ignore.
972 bool Andersens::Node::intersectsIgnoring(Node *N, Node *Ignoring) const {
973 iterator I1 = begin(), I2 = N->begin(), E1 = end(), E2 = N->end();
974 while (I1 != E1 && I2 != E2) {
976 if (*I1 != Ignoring) return true;
978 } else if (*I1 < *I2)
986 // Copy constraint: all edges out of the source node get copied to the
987 // destination node. This returns true if a change is made.
988 bool Andersens::Node::copyFrom(Node *N) {
989 // Use a mostly linear-time merge since both of the lists are sorted.
990 bool Changed = false;
991 iterator I = N->begin(), E = N->end();
993 while (I != E && i != Pointees.size()) {
994 if (Pointees[i] < *I) {
996 } else if (Pointees[i] == *I) {
999 // We found a new element to copy over.
1001 Pointees.insert(Pointees.begin()+i, *I);
1007 Pointees.insert(Pointees.end(), I, E);
1014 bool Andersens::Node::loadFrom(Node *N) {
1015 bool Changed = false;
1016 for (iterator I = N->begin(), E = N->end(); I != E; ++I)
1017 Changed |= copyFrom(*I);
1021 bool Andersens::Node::storeThrough(Node *N) {
1022 bool Changed = false;
1023 for (iterator I = begin(), E = end(); I != E; ++I)
1024 Changed |= (*I)->copyFrom(N);
1029 /// SolveConstraints - This stage iteratively processes the constraints list
1030 /// propagating constraints (adding edges to the Nodes in the points-to graph)
1031 /// until a fixed point is reached.
1033 void Andersens::SolveConstraints() {
1034 bool Changed = true;
1035 unsigned Iteration = 0;
1039 DOUT << "Starting iteration #" << Iteration++ << "!\n";
1041 // Loop over all of the constraints, applying them in turn.
1042 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
1043 Constraint &C = Constraints[i];
1045 case Constraint::Copy:
1046 Changed |= C.Dest->copyFrom(C.Src);
1048 case Constraint::Load:
1049 Changed |= C.Dest->loadFrom(C.Src);
1051 case Constraint::Store:
1052 Changed |= C.Dest->storeThrough(C.Src);
1055 assert(0 && "Unknown constraint!");
1060 // Check to see if any internal function's addresses have been passed to
1061 // external functions. If so, we have to assume that their incoming
1062 // arguments could be anything. If there are any internal functions in
1063 // the universal node that we don't know about, we must iterate.
1064 for (Node::iterator I = GraphNodes[UniversalSet].begin(),
1065 E = GraphNodes[UniversalSet].end(); I != E; ++I)
1066 if (Function *F = dyn_cast_or_null<Function>((*I)->getValue()))
1067 if (F->hasInternalLinkage() &&
1068 EscapingInternalFunctions.insert(F).second) {
1069 // We found a function that is just now escaping. Mark it as if it
1070 // didn't have internal linkage.
1071 AddConstraintsForNonInternalLinkage(F);
1072 DOUT << "Found escaping internal function: " << F->getName() <<"\n";
1073 ++NumEscapingFunctions;
1076 // Check to see if we have discovered any new callees of the indirect call
1077 // sites. If so, add constraints to the analysis.
1078 for (unsigned i = 0, e = IndirectCalls.size(); i != e; ++i) {
1079 CallSite CS = IndirectCalls[i];
1080 std::vector<Function*> &KnownCallees = IndirectCallees[CS];
1081 Node *CN = getNode(CS.getCalledValue());
1083 for (Node::iterator NI = CN->begin(), E = CN->end(); NI != E; ++NI)
1084 if (Function *F = dyn_cast_or_null<Function>((*NI)->getValue())) {
1085 std::vector<Function*>::iterator IP =
1086 std::lower_bound(KnownCallees.begin(), KnownCallees.end(), F);
1087 if (IP == KnownCallees.end() || *IP != F) {
1088 // Add the constraints for the call now.
1089 AddConstraintsForCall(CS, F);
1090 DOUT << "Found actual callee '"
1091 << F->getName() << "' for call: "
1092 << *CS.getInstruction() << "\n";
1093 ++NumIndirectCallees;
1094 KnownCallees.insert(IP, F);
1104 //===----------------------------------------------------------------------===//
1106 //===----------------------------------------------------------------------===//
1108 void Andersens::PrintNode(Node *N) {
1109 if (N == &GraphNodes[UniversalSet]) {
1110 cerr << "<universal>";
1112 } else if (N == &GraphNodes[NullPtr]) {
1113 cerr << "<nullptr>";
1115 } else if (N == &GraphNodes[NullObject]) {
1120 assert(N->getValue() != 0 && "Never set node label!");
1121 Value *V = N->getValue();
1122 if (Function *F = dyn_cast<Function>(V)) {
1123 if (isa<PointerType>(F->getFunctionType()->getReturnType()) &&
1124 N == getReturnNode(F)) {
1125 cerr << F->getName() << ":retval";
1127 } else if (F->getFunctionType()->isVarArg() && N == getVarargNode(F)) {
1128 cerr << F->getName() << ":vararg";
1133 if (Instruction *I = dyn_cast<Instruction>(V))
1134 cerr << I->getParent()->getParent()->getName() << ":";
1135 else if (Argument *Arg = dyn_cast<Argument>(V))
1136 cerr << Arg->getParent()->getName() << ":";
1139 cerr << V->getName();
1141 cerr << "(unnamed)";
1143 if (isa<GlobalValue>(V) || isa<AllocationInst>(V))
1144 if (N == getObject(V))
1148 void Andersens::PrintConstraints() {
1149 cerr << "Constraints:\n";
1150 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
1151 cerr << " #" << i << ": ";
1152 Constraint &C = Constraints[i];
1153 if (C.Type == Constraint::Store)
1157 if (C.Type == Constraint::Load)
1164 void Andersens::PrintPointsToGraph() {
1165 cerr << "Points-to graph:\n";
1166 for (unsigned i = 0, e = GraphNodes.size(); i != e; ++i) {
1167 Node *N = &GraphNodes[i];
1168 cerr << "[" << (N->end() - N->begin()) << "] ";
1171 for (Node::iterator I = N->begin(), E = N->end(); I != E; ++I) {
1172 if (I != N->begin()) cerr << ", ";