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"
70 NumIters("anders-aa", "Number of iterations to reach convergence");
72 NumConstraints("anders-aa", "Number of constraints");
74 NumNodes("anders-aa", "Number of nodes");
76 NumEscapingFunctions("anders-aa", "Number of internal functions that escape");
78 NumIndirectCallees("anders-aa", "Number of indirect callees found");
80 class Andersens : public ModulePass, public AliasAnalysis,
81 private InstVisitor<Andersens> {
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;
89 Node *setValue(Value *V) {
90 assert(Val == 0 && "Value already set for this node!");
95 /// getValue - Return the LLVM value corresponding to this node.
97 Value *getValue() const { return Val; }
99 typedef std::vector<Node*>::const_iterator iterator;
100 iterator begin() const { return Pointees.begin(); }
101 iterator end() const { return Pointees.end(); }
103 /// addPointerTo - Add a pointer to the list of pointees of this node,
104 /// returning true if this caused a new pointer to be added, or false if
105 /// we already knew about the points-to relation.
106 bool addPointerTo(Node *N) {
107 std::vector<Node*>::iterator I = std::lower_bound(Pointees.begin(),
110 if (I != Pointees.end() && *I == N)
112 Pointees.insert(I, N);
116 /// intersects - Return true if the points-to set of this node intersects
117 /// with the points-to set of the specified node.
118 bool intersects(Node *N) const;
120 /// intersectsIgnoring - Return true if the points-to set of this node
121 /// intersects with the points-to set of the specified node on any nodes
122 /// except for the specified node to ignore.
123 bool intersectsIgnoring(Node *N, Node *Ignoring) const;
125 // Constraint application methods.
126 bool copyFrom(Node *N);
127 bool loadFrom(Node *N);
128 bool storeThrough(Node *N);
131 /// GraphNodes - This vector is populated as part of the object
132 /// identification stage of the analysis, which populates this vector with a
133 /// node for each memory object and fills in the ValueNodes map.
134 std::vector<Node> GraphNodes;
136 /// ValueNodes - This map indicates the Node that a particular Value* is
137 /// represented by. This contains entries for all pointers.
138 std::map<Value*, unsigned> ValueNodes;
140 /// ObjectNodes - This map contains entries for each memory object in the
141 /// program: globals, alloca's and mallocs.
142 std::map<Value*, unsigned> ObjectNodes;
144 /// ReturnNodes - This map contains an entry for each function in the
145 /// program that returns a value.
146 std::map<Function*, unsigned> ReturnNodes;
148 /// VarargNodes - This map contains the entry used to represent all pointers
149 /// passed through the varargs portion of a function call for a particular
150 /// function. An entry is not present in this map for functions that do not
151 /// take variable arguments.
152 std::map<Function*, unsigned> VarargNodes;
154 /// Constraint - Objects of this structure are used to represent the various
155 /// constraints identified by the algorithm. The constraints are 'copy',
156 /// for statements like "A = B", 'load' for statements like "A = *B", and
157 /// 'store' for statements like "*A = B".
159 enum ConstraintType { Copy, Load, Store } Type;
162 Constraint(ConstraintType Ty, Node *D, Node *S)
163 : Type(Ty), Dest(D), Src(S) {}
166 /// Constraints - This vector contains a list of all of the constraints
167 /// identified by the program.
168 std::vector<Constraint> Constraints;
170 /// EscapingInternalFunctions - This set contains all of the internal
171 /// functions that are found to escape from the program. If the address of
172 /// an internal function is passed to an external function or otherwise
173 /// escapes from the analyzed portion of the program, we must assume that
174 /// any pointer arguments can alias the universal node. This set keeps
175 /// track of those functions we are assuming to escape so far.
176 std::set<Function*> EscapingInternalFunctions;
178 /// IndirectCalls - This contains a list of all of the indirect call sites
179 /// in the program. Since the call graph is iteratively discovered, we may
180 /// need to add constraints to our graph as we find new targets of function
182 std::vector<CallSite> IndirectCalls;
184 /// IndirectCallees - For each call site in the indirect calls list, keep
185 /// track of the callees that we have discovered so far. As the analysis
186 /// proceeds, more callees are discovered, until the call graph finally
188 std::map<CallSite, std::vector<Function*> > IndirectCallees;
190 /// This enum defines the GraphNodes indices that correspond to important
199 bool runOnModule(Module &M) {
200 InitializeAliasAnalysis(this);
202 CollectConstraints(M);
203 DEBUG(PrintConstraints());
205 DEBUG(PrintPointsToGraph());
207 // Free the constraints list, as we don't need it to respond to alias
212 EscapingInternalFunctions.clear();
213 std::vector<Constraint>().swap(Constraints);
217 void releaseMemory() {
218 // FIXME: Until we have transitively required passes working correctly,
219 // this cannot be enabled! Otherwise, using -count-aa with the pass
220 // causes memory to be freed too early. :(
222 // The memory objects and ValueNodes data structures at the only ones that
223 // are still live after construction.
224 std::vector<Node>().swap(GraphNodes);
229 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
230 AliasAnalysis::getAnalysisUsage(AU);
231 AU.setPreservesAll(); // Does not transform code
234 //------------------------------------------------
235 // Implement the AliasAnalysis API
237 AliasResult alias(const Value *V1, unsigned V1Size,
238 const Value *V2, unsigned V2Size);
239 virtual ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size);
240 virtual ModRefResult getModRefInfo(CallSite CS1, CallSite CS2);
241 void getMustAliases(Value *P, std::vector<Value*> &RetVals);
242 bool pointsToConstantMemory(const Value *P);
244 virtual void deleteValue(Value *V) {
246 getAnalysis<AliasAnalysis>().deleteValue(V);
249 virtual void copyValue(Value *From, Value *To) {
250 ValueNodes[To] = ValueNodes[From];
251 getAnalysis<AliasAnalysis>().copyValue(From, To);
255 /// getNode - Return the node corresponding to the specified pointer scalar.
257 Node *getNode(Value *V) {
258 if (Constant *C = dyn_cast<Constant>(V))
259 if (!isa<GlobalValue>(C))
260 return getNodeForConstantPointer(C);
262 std::map<Value*, unsigned>::iterator I = ValueNodes.find(V);
263 if (I == ValueNodes.end()) {
267 assert(0 && "Value does not have a node in the points-to graph!");
269 return &GraphNodes[I->second];
272 /// getObject - Return the node corresponding to the memory object for the
273 /// specified global or allocation instruction.
274 Node *getObject(Value *V) {
275 std::map<Value*, unsigned>::iterator I = ObjectNodes.find(V);
276 assert(I != ObjectNodes.end() &&
277 "Value does not have an object in the points-to graph!");
278 return &GraphNodes[I->second];
281 /// getReturnNode - Return the node representing the return value for the
282 /// specified function.
283 Node *getReturnNode(Function *F) {
284 std::map<Function*, unsigned>::iterator I = ReturnNodes.find(F);
285 assert(I != ReturnNodes.end() && "Function does not return a value!");
286 return &GraphNodes[I->second];
289 /// getVarargNode - Return the node representing the variable arguments
290 /// formal for the specified function.
291 Node *getVarargNode(Function *F) {
292 std::map<Function*, unsigned>::iterator I = VarargNodes.find(F);
293 assert(I != VarargNodes.end() && "Function does not take var args!");
294 return &GraphNodes[I->second];
297 /// getNodeValue - Get the node for the specified LLVM value and set the
298 /// value for it to be the specified value.
299 Node *getNodeValue(Value &V) {
300 return getNode(&V)->setValue(&V);
303 void IdentifyObjects(Module &M);
304 void CollectConstraints(Module &M);
305 void SolveConstraints();
307 Node *getNodeForConstantPointer(Constant *C);
308 Node *getNodeForConstantPointerTarget(Constant *C);
309 void AddGlobalInitializerConstraints(Node *N, Constant *C);
311 void AddConstraintsForNonInternalLinkage(Function *F);
312 void AddConstraintsForCall(CallSite CS, Function *F);
313 bool AddConstraintsForExternalCall(CallSite CS, Function *F);
316 void PrintNode(Node *N);
317 void PrintConstraints();
318 void PrintPointsToGraph();
320 //===------------------------------------------------------------------===//
321 // Instruction visitation methods for adding constraints
323 friend class InstVisitor<Andersens>;
324 void visitReturnInst(ReturnInst &RI);
325 void visitInvokeInst(InvokeInst &II) { visitCallSite(CallSite(&II)); }
326 void visitCallInst(CallInst &CI) { visitCallSite(CallSite(&CI)); }
327 void visitCallSite(CallSite CS);
328 void visitAllocationInst(AllocationInst &AI);
329 void visitLoadInst(LoadInst &LI);
330 void visitStoreInst(StoreInst &SI);
331 void visitGetElementPtrInst(GetElementPtrInst &GEP);
332 void visitPHINode(PHINode &PN);
333 void visitCastInst(CastInst &CI);
334 void visitSetCondInst(SetCondInst &SCI) {} // NOOP!
335 void visitSelectInst(SelectInst &SI);
336 void visitVAArg(VAArgInst &I);
337 void visitInstruction(Instruction &I);
340 RegisterPass<Andersens> X("anders-aa",
341 "Andersen's Interprocedural Alias Analysis");
342 RegisterAnalysisGroup<AliasAnalysis> Y(X);
345 ModulePass *llvm::createAndersensPass() { return new Andersens(); }
347 //===----------------------------------------------------------------------===//
348 // AliasAnalysis Interface Implementation
349 //===----------------------------------------------------------------------===//
351 AliasAnalysis::AliasResult Andersens::alias(const Value *V1, unsigned V1Size,
352 const Value *V2, unsigned V2Size) {
353 Node *N1 = getNode(const_cast<Value*>(V1));
354 Node *N2 = getNode(const_cast<Value*>(V2));
356 // Check to see if the two pointers are known to not alias. They don't alias
357 // if their points-to sets do not intersect.
358 if (!N1->intersectsIgnoring(N2, &GraphNodes[NullObject]))
361 return AliasAnalysis::alias(V1, V1Size, V2, V2Size);
364 AliasAnalysis::ModRefResult
365 Andersens::getModRefInfo(CallSite CS, Value *P, unsigned Size) {
366 // The only thing useful that we can contribute for mod/ref information is
367 // when calling external function calls: if we know that memory never escapes
368 // from the program, it cannot be modified by an external call.
370 // NOTE: This is not really safe, at least not when the entire program is not
371 // available. The deal is that the external function could call back into the
372 // program and modify stuff. We ignore this technical niggle for now. This
373 // is, after all, a "research quality" implementation of Andersen's analysis.
374 if (Function *F = CS.getCalledFunction())
375 if (F->isExternal()) {
376 Node *N1 = getNode(P);
378 if (N1->begin() == N1->end())
379 return NoModRef; // P doesn't point to anything.
381 // Get the first pointee.
382 Node *FirstPointee = *N1->begin();
383 if (FirstPointee != &GraphNodes[UniversalSet])
384 return NoModRef; // P doesn't point to the universal set.
387 return AliasAnalysis::getModRefInfo(CS, P, Size);
390 AliasAnalysis::ModRefResult
391 Andersens::getModRefInfo(CallSite CS1, CallSite CS2) {
392 return AliasAnalysis::getModRefInfo(CS1,CS2);
395 /// getMustAlias - We can provide must alias information if we know that a
396 /// pointer can only point to a specific function or the null pointer.
397 /// Unfortunately we cannot determine must-alias information for global
398 /// variables or any other memory memory objects because we do not track whether
399 /// a pointer points to the beginning of an object or a field of it.
400 void Andersens::getMustAliases(Value *P, std::vector<Value*> &RetVals) {
401 Node *N = getNode(P);
402 Node::iterator I = N->begin();
404 // If there is exactly one element in the points-to set for the object...
407 Node *Pointee = *N->begin();
409 // If a function is the only object in the points-to set, then it must be
410 // the destination. Note that we can't handle global variables here,
411 // because we don't know if the pointer is actually pointing to a field of
412 // the global or to the beginning of it.
413 if (Value *V = Pointee->getValue()) {
414 if (Function *F = dyn_cast<Function>(V))
415 RetVals.push_back(F);
417 // If the object in the points-to set is the null object, then the null
418 // pointer is a must alias.
419 if (Pointee == &GraphNodes[NullObject])
420 RetVals.push_back(Constant::getNullValue(P->getType()));
425 AliasAnalysis::getMustAliases(P, RetVals);
428 /// pointsToConstantMemory - If we can determine that this pointer only points
429 /// to constant memory, return true. In practice, this means that if the
430 /// pointer can only point to constant globals, functions, or the null pointer,
433 bool Andersens::pointsToConstantMemory(const Value *P) {
434 Node *N = getNode((Value*)P);
435 for (Node::iterator I = N->begin(), E = N->end(); I != E; ++I) {
436 if (Value *V = (*I)->getValue()) {
437 if (!isa<GlobalValue>(V) || (isa<GlobalVariable>(V) &&
438 !cast<GlobalVariable>(V)->isConstant()))
439 return AliasAnalysis::pointsToConstantMemory(P);
441 if (*I != &GraphNodes[NullObject])
442 return AliasAnalysis::pointsToConstantMemory(P);
449 //===----------------------------------------------------------------------===//
450 // Object Identification Phase
451 //===----------------------------------------------------------------------===//
453 /// IdentifyObjects - This stage scans the program, adding an entry to the
454 /// GraphNodes list for each memory object in the program (global stack or
455 /// heap), and populates the ValueNodes and ObjectNodes maps for these objects.
457 void Andersens::IdentifyObjects(Module &M) {
458 unsigned NumObjects = 0;
460 // Object #0 is always the universal set: the object that we don't know
462 assert(NumObjects == UniversalSet && "Something changed!");
465 // Object #1 always represents the null pointer.
466 assert(NumObjects == NullPtr && "Something changed!");
469 // Object #2 always represents the null object (the object pointed to by null)
470 assert(NumObjects == NullObject && "Something changed!");
473 // Add all the globals first.
474 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
476 ObjectNodes[I] = NumObjects++;
477 ValueNodes[I] = NumObjects++;
480 // Add nodes for all of the functions and the instructions inside of them.
481 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
482 // The function itself is a memory object.
483 ValueNodes[F] = NumObjects++;
484 ObjectNodes[F] = NumObjects++;
485 if (isa<PointerType>(F->getFunctionType()->getReturnType()))
486 ReturnNodes[F] = NumObjects++;
487 if (F->getFunctionType()->isVarArg())
488 VarargNodes[F] = NumObjects++;
490 // Add nodes for all of the incoming pointer arguments.
491 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
493 if (isa<PointerType>(I->getType()))
494 ValueNodes[I] = NumObjects++;
496 // Scan the function body, creating a memory object for each heap/stack
497 // allocation in the body of the function and a node to represent all
498 // pointer values defined by instructions and used as operands.
499 for (inst_iterator II = inst_begin(F), E = inst_end(F); II != E; ++II) {
500 // If this is an heap or stack allocation, create a node for the memory
502 if (isa<PointerType>(II->getType())) {
503 ValueNodes[&*II] = NumObjects++;
504 if (AllocationInst *AI = dyn_cast<AllocationInst>(&*II))
505 ObjectNodes[AI] = NumObjects++;
510 // Now that we know how many objects to create, make them all now!
511 GraphNodes.resize(NumObjects);
512 NumNodes += NumObjects;
515 //===----------------------------------------------------------------------===//
516 // Constraint Identification Phase
517 //===----------------------------------------------------------------------===//
519 /// getNodeForConstantPointer - Return the node corresponding to the constant
521 Andersens::Node *Andersens::getNodeForConstantPointer(Constant *C) {
522 assert(isa<PointerType>(C->getType()) && "Not a constant pointer!");
524 if (isa<ConstantPointerNull>(C) || isa<UndefValue>(C))
525 return &GraphNodes[NullPtr];
526 else if (GlobalValue *GV = dyn_cast<GlobalValue>(C))
528 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
529 switch (CE->getOpcode()) {
530 case Instruction::GetElementPtr:
531 return getNodeForConstantPointer(CE->getOperand(0));
532 case Instruction::IntToPtr:
533 return &GraphNodes[UniversalSet];
534 case Instruction::BitCast:
535 return getNodeForConstantPointer(CE->getOperand(0));
537 std::cerr << "Constant Expr not yet handled: " << *CE << "\n";
541 assert(0 && "Unknown constant pointer!");
546 /// getNodeForConstantPointerTarget - Return the node POINTED TO by the
547 /// specified constant pointer.
548 Andersens::Node *Andersens::getNodeForConstantPointerTarget(Constant *C) {
549 assert(isa<PointerType>(C->getType()) && "Not a constant pointer!");
551 if (isa<ConstantPointerNull>(C))
552 return &GraphNodes[NullObject];
553 else if (GlobalValue *GV = dyn_cast<GlobalValue>(C))
554 return getObject(GV);
555 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
556 switch (CE->getOpcode()) {
557 case Instruction::GetElementPtr:
558 return getNodeForConstantPointerTarget(CE->getOperand(0));
559 case Instruction::IntToPtr:
560 return &GraphNodes[UniversalSet];
561 case Instruction::BitCast:
562 return getNodeForConstantPointerTarget(CE->getOperand(0));
564 std::cerr << "Constant Expr not yet handled: " << *CE << "\n";
568 assert(0 && "Unknown constant pointer!");
573 /// AddGlobalInitializerConstraints - Add inclusion constraints for the memory
574 /// object N, which contains values indicated by C.
575 void Andersens::AddGlobalInitializerConstraints(Node *N, Constant *C) {
576 if (C->getType()->isFirstClassType()) {
577 if (isa<PointerType>(C->getType()))
578 N->copyFrom(getNodeForConstantPointer(C));
580 } else if (C->isNullValue()) {
581 N->addPointerTo(&GraphNodes[NullObject]);
583 } else if (!isa<UndefValue>(C)) {
584 // If this is an array or struct, include constraints for each element.
585 assert(isa<ConstantArray>(C) || isa<ConstantStruct>(C));
586 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i)
587 AddGlobalInitializerConstraints(N, cast<Constant>(C->getOperand(i)));
591 /// AddConstraintsForNonInternalLinkage - If this function does not have
592 /// internal linkage, realize that we can't trust anything passed into or
593 /// returned by this function.
594 void Andersens::AddConstraintsForNonInternalLinkage(Function *F) {
595 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
596 if (isa<PointerType>(I->getType()))
597 // If this is an argument of an externally accessible function, the
598 // incoming pointer might point to anything.
599 Constraints.push_back(Constraint(Constraint::Copy, getNode(I),
600 &GraphNodes[UniversalSet]));
603 /// AddConstraintsForCall - If this is a call to a "known" function, add the
604 /// constraints and return true. If this is a call to an unknown function,
606 bool Andersens::AddConstraintsForExternalCall(CallSite CS, Function *F) {
607 assert(F->isExternal() && "Not an external function!");
609 // These functions don't induce any points-to constraints.
610 if (F->getName() == "atoi" || F->getName() == "atof" ||
611 F->getName() == "atol" || F->getName() == "atoll" ||
612 F->getName() == "remove" || F->getName() == "unlink" ||
613 F->getName() == "rename" || F->getName() == "memcmp" ||
614 F->getName() == "llvm.memset.i32" ||
615 F->getName() == "llvm.memset.i64" ||
616 F->getName() == "strcmp" || F->getName() == "strncmp" ||
617 F->getName() == "execl" || F->getName() == "execlp" ||
618 F->getName() == "execle" || F->getName() == "execv" ||
619 F->getName() == "execvp" || F->getName() == "chmod" ||
620 F->getName() == "puts" || F->getName() == "write" ||
621 F->getName() == "open" || F->getName() == "create" ||
622 F->getName() == "truncate" || F->getName() == "chdir" ||
623 F->getName() == "mkdir" || F->getName() == "rmdir" ||
624 F->getName() == "read" || F->getName() == "pipe" ||
625 F->getName() == "wait" || F->getName() == "time" ||
626 F->getName() == "stat" || F->getName() == "fstat" ||
627 F->getName() == "lstat" || F->getName() == "strtod" ||
628 F->getName() == "strtof" || F->getName() == "strtold" ||
629 F->getName() == "fopen" || F->getName() == "fdopen" ||
630 F->getName() == "freopen" ||
631 F->getName() == "fflush" || F->getName() == "feof" ||
632 F->getName() == "fileno" || F->getName() == "clearerr" ||
633 F->getName() == "rewind" || F->getName() == "ftell" ||
634 F->getName() == "ferror" || F->getName() == "fgetc" ||
635 F->getName() == "fgetc" || F->getName() == "_IO_getc" ||
636 F->getName() == "fwrite" || F->getName() == "fread" ||
637 F->getName() == "fgets" || F->getName() == "ungetc" ||
638 F->getName() == "fputc" ||
639 F->getName() == "fputs" || F->getName() == "putc" ||
640 F->getName() == "ftell" || F->getName() == "rewind" ||
641 F->getName() == "_IO_putc" || F->getName() == "fseek" ||
642 F->getName() == "fgetpos" || F->getName() == "fsetpos" ||
643 F->getName() == "printf" || F->getName() == "fprintf" ||
644 F->getName() == "sprintf" || F->getName() == "vprintf" ||
645 F->getName() == "vfprintf" || F->getName() == "vsprintf" ||
646 F->getName() == "scanf" || F->getName() == "fscanf" ||
647 F->getName() == "sscanf" || F->getName() == "__assert_fail" ||
648 F->getName() == "modf")
652 // These functions do induce points-to edges.
653 if (F->getName() == "llvm.memcpy.i32" || F->getName() == "llvm.memcpy.i64" ||
654 F->getName() == "llvm.memmove.i32" ||F->getName() == "llvm.memmove.i64" ||
655 F->getName() == "memmove") {
656 // Note: this is a poor approximation, this says Dest = Src, instead of
658 Constraints.push_back(Constraint(Constraint::Copy,
659 getNode(CS.getArgument(0)),
660 getNode(CS.getArgument(1))));
665 if (F->getName() == "realloc" || F->getName() == "strchr" ||
666 F->getName() == "strrchr" || F->getName() == "strstr" ||
667 F->getName() == "strtok") {
668 Constraints.push_back(Constraint(Constraint::Copy,
669 getNode(CS.getInstruction()),
670 getNode(CS.getArgument(0))));
679 /// CollectConstraints - This stage scans the program, adding a constraint to
680 /// the Constraints list for each instruction in the program that induces a
681 /// constraint, and setting up the initial points-to graph.
683 void Andersens::CollectConstraints(Module &M) {
684 // First, the universal set points to itself.
685 GraphNodes[UniversalSet].addPointerTo(&GraphNodes[UniversalSet]);
686 //Constraints.push_back(Constraint(Constraint::Load, &GraphNodes[UniversalSet],
687 // &GraphNodes[UniversalSet]));
688 Constraints.push_back(Constraint(Constraint::Store, &GraphNodes[UniversalSet],
689 &GraphNodes[UniversalSet]));
691 // Next, the null pointer points to the null object.
692 GraphNodes[NullPtr].addPointerTo(&GraphNodes[NullObject]);
694 // Next, add any constraints on global variables and their initializers.
695 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
697 // Associate the address of the global object as pointing to the memory for
698 // the global: &G = <G memory>
699 Node *Object = getObject(I);
701 getNodeValue(*I)->addPointerTo(Object);
703 if (I->hasInitializer()) {
704 AddGlobalInitializerConstraints(Object, I->getInitializer());
706 // If it doesn't have an initializer (i.e. it's defined in another
707 // translation unit), it points to the universal set.
708 Constraints.push_back(Constraint(Constraint::Copy, Object,
709 &GraphNodes[UniversalSet]));
713 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
714 // Make the function address point to the function object.
715 getNodeValue(*F)->addPointerTo(getObject(F)->setValue(F));
717 // Set up the return value node.
718 if (isa<PointerType>(F->getFunctionType()->getReturnType()))
719 getReturnNode(F)->setValue(F);
720 if (F->getFunctionType()->isVarArg())
721 getVarargNode(F)->setValue(F);
723 // Set up incoming argument nodes.
724 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
726 if (isa<PointerType>(I->getType()))
729 if (!F->hasInternalLinkage())
730 AddConstraintsForNonInternalLinkage(F);
732 if (!F->isExternal()) {
733 // Scan the function body, creating a memory object for each heap/stack
734 // allocation in the body of the function and a node to represent all
735 // pointer values defined by instructions and used as operands.
738 // External functions that return pointers return the universal set.
739 if (isa<PointerType>(F->getFunctionType()->getReturnType()))
740 Constraints.push_back(Constraint(Constraint::Copy,
742 &GraphNodes[UniversalSet]));
744 // Any pointers that are passed into the function have the universal set
746 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
748 if (isa<PointerType>(I->getType())) {
749 // Pointers passed into external functions could have anything stored
751 Constraints.push_back(Constraint(Constraint::Store, getNode(I),
752 &GraphNodes[UniversalSet]));
753 // Memory objects passed into external function calls can have the
754 // universal set point to them.
755 Constraints.push_back(Constraint(Constraint::Copy,
756 &GraphNodes[UniversalSet],
760 // If this is an external varargs function, it can also store pointers
761 // into any pointers passed through the varargs section.
762 if (F->getFunctionType()->isVarArg())
763 Constraints.push_back(Constraint(Constraint::Store, getVarargNode(F),
764 &GraphNodes[UniversalSet]));
767 NumConstraints += Constraints.size();
771 void Andersens::visitInstruction(Instruction &I) {
773 return; // This function is just a big assert.
775 if (isa<BinaryOperator>(I))
777 // Most instructions don't have any effect on pointer values.
778 switch (I.getOpcode()) {
779 case Instruction::Br:
780 case Instruction::Switch:
781 case Instruction::Unwind:
782 case Instruction::Unreachable:
783 case Instruction::Free:
784 case Instruction::Shl:
785 case Instruction::LShr:
786 case Instruction::AShr:
789 // Is this something we aren't handling yet?
790 std::cerr << "Unknown instruction: " << I;
795 void Andersens::visitAllocationInst(AllocationInst &AI) {
796 getNodeValue(AI)->addPointerTo(getObject(&AI)->setValue(&AI));
799 void Andersens::visitReturnInst(ReturnInst &RI) {
800 if (RI.getNumOperands() && isa<PointerType>(RI.getOperand(0)->getType()))
801 // return V --> <Copy/retval{F}/v>
802 Constraints.push_back(Constraint(Constraint::Copy,
803 getReturnNode(RI.getParent()->getParent()),
804 getNode(RI.getOperand(0))));
807 void Andersens::visitLoadInst(LoadInst &LI) {
808 if (isa<PointerType>(LI.getType()))
809 // P1 = load P2 --> <Load/P1/P2>
810 Constraints.push_back(Constraint(Constraint::Load, getNodeValue(LI),
811 getNode(LI.getOperand(0))));
814 void Andersens::visitStoreInst(StoreInst &SI) {
815 if (isa<PointerType>(SI.getOperand(0)->getType()))
816 // store P1, P2 --> <Store/P2/P1>
817 Constraints.push_back(Constraint(Constraint::Store,
818 getNode(SI.getOperand(1)),
819 getNode(SI.getOperand(0))));
822 void Andersens::visitGetElementPtrInst(GetElementPtrInst &GEP) {
823 // P1 = getelementptr P2, ... --> <Copy/P1/P2>
824 Constraints.push_back(Constraint(Constraint::Copy, getNodeValue(GEP),
825 getNode(GEP.getOperand(0))));
828 void Andersens::visitPHINode(PHINode &PN) {
829 if (isa<PointerType>(PN.getType())) {
830 Node *PNN = getNodeValue(PN);
831 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
832 // P1 = phi P2, P3 --> <Copy/P1/P2>, <Copy/P1/P3>, ...
833 Constraints.push_back(Constraint(Constraint::Copy, PNN,
834 getNode(PN.getIncomingValue(i))));
838 void Andersens::visitCastInst(CastInst &CI) {
839 Value *Op = CI.getOperand(0);
840 if (isa<PointerType>(CI.getType())) {
841 if (isa<PointerType>(Op->getType())) {
842 // P1 = cast P2 --> <Copy/P1/P2>
843 Constraints.push_back(Constraint(Constraint::Copy, getNodeValue(CI),
844 getNode(CI.getOperand(0))));
846 // P1 = cast int --> <Copy/P1/Univ>
848 Constraints.push_back(Constraint(Constraint::Copy, getNodeValue(CI),
849 &GraphNodes[UniversalSet]));
854 } else if (isa<PointerType>(Op->getType())) {
855 // int = cast P1 --> <Copy/Univ/P1>
857 Constraints.push_back(Constraint(Constraint::Copy,
858 &GraphNodes[UniversalSet],
859 getNode(CI.getOperand(0))));
861 getNode(CI.getOperand(0));
866 void Andersens::visitSelectInst(SelectInst &SI) {
867 if (isa<PointerType>(SI.getType())) {
868 Node *SIN = getNodeValue(SI);
869 // P1 = select C, P2, P3 ---> <Copy/P1/P2>, <Copy/P1/P3>
870 Constraints.push_back(Constraint(Constraint::Copy, SIN,
871 getNode(SI.getOperand(1))));
872 Constraints.push_back(Constraint(Constraint::Copy, SIN,
873 getNode(SI.getOperand(2))));
877 void Andersens::visitVAArg(VAArgInst &I) {
878 assert(0 && "vaarg not handled yet!");
881 /// AddConstraintsForCall - Add constraints for a call with actual arguments
882 /// specified by CS to the function specified by F. Note that the types of
883 /// arguments might not match up in the case where this is an indirect call and
884 /// the function pointer has been casted. If this is the case, do something
886 void Andersens::AddConstraintsForCall(CallSite CS, Function *F) {
887 // If this is a call to an external function, handle it directly to get some
888 // taste of context sensitivity.
889 if (F->isExternal() && AddConstraintsForExternalCall(CS, F))
892 if (isa<PointerType>(CS.getType())) {
893 Node *CSN = getNode(CS.getInstruction());
894 if (isa<PointerType>(F->getFunctionType()->getReturnType())) {
895 Constraints.push_back(Constraint(Constraint::Copy, CSN,
898 // If the function returns a non-pointer value, handle this just like we
899 // treat a nonpointer cast to pointer.
900 Constraints.push_back(Constraint(Constraint::Copy, CSN,
901 &GraphNodes[UniversalSet]));
903 } else if (isa<PointerType>(F->getFunctionType()->getReturnType())) {
904 Constraints.push_back(Constraint(Constraint::Copy,
905 &GraphNodes[UniversalSet],
909 Function::arg_iterator AI = F->arg_begin(), AE = F->arg_end();
910 CallSite::arg_iterator ArgI = CS.arg_begin(), ArgE = CS.arg_end();
911 for (; AI != AE && ArgI != ArgE; ++AI, ++ArgI)
912 if (isa<PointerType>(AI->getType())) {
913 if (isa<PointerType>((*ArgI)->getType())) {
914 // Copy the actual argument into the formal argument.
915 Constraints.push_back(Constraint(Constraint::Copy, getNode(AI),
918 Constraints.push_back(Constraint(Constraint::Copy, getNode(AI),
919 &GraphNodes[UniversalSet]));
921 } else if (isa<PointerType>((*ArgI)->getType())) {
922 Constraints.push_back(Constraint(Constraint::Copy,
923 &GraphNodes[UniversalSet],
927 // Copy all pointers passed through the varargs section to the varargs node.
928 if (F->getFunctionType()->isVarArg())
929 for (; ArgI != ArgE; ++ArgI)
930 if (isa<PointerType>((*ArgI)->getType()))
931 Constraints.push_back(Constraint(Constraint::Copy, getVarargNode(F),
933 // If more arguments are passed in than we track, just drop them on the floor.
936 void Andersens::visitCallSite(CallSite CS) {
937 if (isa<PointerType>(CS.getType()))
938 getNodeValue(*CS.getInstruction());
940 if (Function *F = CS.getCalledFunction()) {
941 AddConstraintsForCall(CS, F);
943 // We don't handle indirect call sites yet. Keep track of them for when we
944 // discover the call graph incrementally.
945 IndirectCalls.push_back(CS);
949 //===----------------------------------------------------------------------===//
950 // Constraint Solving Phase
951 //===----------------------------------------------------------------------===//
953 /// intersects - Return true if the points-to set of this node intersects
954 /// with the points-to set of the specified node.
955 bool Andersens::Node::intersects(Node *N) const {
956 iterator I1 = begin(), I2 = N->begin(), E1 = end(), E2 = N->end();
957 while (I1 != E1 && I2 != E2) {
958 if (*I1 == *I2) return true;
967 /// intersectsIgnoring - Return true if the points-to set of this node
968 /// intersects with the points-to set of the specified node on any nodes
969 /// except for the specified node to ignore.
970 bool Andersens::Node::intersectsIgnoring(Node *N, Node *Ignoring) const {
971 iterator I1 = begin(), I2 = N->begin(), E1 = end(), E2 = N->end();
972 while (I1 != E1 && I2 != E2) {
974 if (*I1 != Ignoring) return true;
976 } else if (*I1 < *I2)
984 // Copy constraint: all edges out of the source node get copied to the
985 // destination node. This returns true if a change is made.
986 bool Andersens::Node::copyFrom(Node *N) {
987 // Use a mostly linear-time merge since both of the lists are sorted.
988 bool Changed = false;
989 iterator I = N->begin(), E = N->end();
991 while (I != E && i != Pointees.size()) {
992 if (Pointees[i] < *I) {
994 } else if (Pointees[i] == *I) {
997 // We found a new element to copy over.
999 Pointees.insert(Pointees.begin()+i, *I);
1005 Pointees.insert(Pointees.end(), I, E);
1012 bool Andersens::Node::loadFrom(Node *N) {
1013 bool Changed = false;
1014 for (iterator I = N->begin(), E = N->end(); I != E; ++I)
1015 Changed |= copyFrom(*I);
1019 bool Andersens::Node::storeThrough(Node *N) {
1020 bool Changed = false;
1021 for (iterator I = begin(), E = end(); I != E; ++I)
1022 Changed |= (*I)->copyFrom(N);
1027 /// SolveConstraints - This stage iteratively processes the constraints list
1028 /// propagating constraints (adding edges to the Nodes in the points-to graph)
1029 /// until a fixed point is reached.
1031 void Andersens::SolveConstraints() {
1032 bool Changed = true;
1033 unsigned Iteration = 0;
1037 DOUT << "Starting iteration #" << Iteration++ << "!\n";
1039 // Loop over all of the constraints, applying them in turn.
1040 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
1041 Constraint &C = Constraints[i];
1043 case Constraint::Copy:
1044 Changed |= C.Dest->copyFrom(C.Src);
1046 case Constraint::Load:
1047 Changed |= C.Dest->loadFrom(C.Src);
1049 case Constraint::Store:
1050 Changed |= C.Dest->storeThrough(C.Src);
1053 assert(0 && "Unknown constraint!");
1058 // Check to see if any internal function's addresses have been passed to
1059 // external functions. If so, we have to assume that their incoming
1060 // arguments could be anything. If there are any internal functions in
1061 // the universal node that we don't know about, we must iterate.
1062 for (Node::iterator I = GraphNodes[UniversalSet].begin(),
1063 E = GraphNodes[UniversalSet].end(); I != E; ++I)
1064 if (Function *F = dyn_cast_or_null<Function>((*I)->getValue()))
1065 if (F->hasInternalLinkage() &&
1066 EscapingInternalFunctions.insert(F).second) {
1067 // We found a function that is just now escaping. Mark it as if it
1068 // didn't have internal linkage.
1069 AddConstraintsForNonInternalLinkage(F);
1070 DOUT << "Found escaping internal function: " << F->getName() <<"\n";
1071 ++NumEscapingFunctions;
1074 // Check to see if we have discovered any new callees of the indirect call
1075 // sites. If so, add constraints to the analysis.
1076 for (unsigned i = 0, e = IndirectCalls.size(); i != e; ++i) {
1077 CallSite CS = IndirectCalls[i];
1078 std::vector<Function*> &KnownCallees = IndirectCallees[CS];
1079 Node *CN = getNode(CS.getCalledValue());
1081 for (Node::iterator NI = CN->begin(), E = CN->end(); NI != E; ++NI)
1082 if (Function *F = dyn_cast_or_null<Function>((*NI)->getValue())) {
1083 std::vector<Function*>::iterator IP =
1084 std::lower_bound(KnownCallees.begin(), KnownCallees.end(), F);
1085 if (IP == KnownCallees.end() || *IP != F) {
1086 // Add the constraints for the call now.
1087 AddConstraintsForCall(CS, F);
1088 DOUT << "Found actual callee '"
1089 << F->getName() << "' for call: "
1090 << *CS.getInstruction() << "\n";
1091 ++NumIndirectCallees;
1092 KnownCallees.insert(IP, F);
1102 //===----------------------------------------------------------------------===//
1104 //===----------------------------------------------------------------------===//
1106 void Andersens::PrintNode(Node *N) {
1107 if (N == &GraphNodes[UniversalSet]) {
1108 std::cerr << "<universal>";
1110 } else if (N == &GraphNodes[NullPtr]) {
1111 std::cerr << "<nullptr>";
1113 } else if (N == &GraphNodes[NullObject]) {
1114 std::cerr << "<null>";
1118 assert(N->getValue() != 0 && "Never set node label!");
1119 Value *V = N->getValue();
1120 if (Function *F = dyn_cast<Function>(V)) {
1121 if (isa<PointerType>(F->getFunctionType()->getReturnType()) &&
1122 N == getReturnNode(F)) {
1123 std::cerr << F->getName() << ":retval";
1125 } else if (F->getFunctionType()->isVarArg() && N == getVarargNode(F)) {
1126 std::cerr << F->getName() << ":vararg";
1131 if (Instruction *I = dyn_cast<Instruction>(V))
1132 std::cerr << I->getParent()->getParent()->getName() << ":";
1133 else if (Argument *Arg = dyn_cast<Argument>(V))
1134 std::cerr << Arg->getParent()->getName() << ":";
1137 std::cerr << V->getName();
1139 std::cerr << "(unnamed)";
1141 if (isa<GlobalValue>(V) || isa<AllocationInst>(V))
1142 if (N == getObject(V))
1143 std::cerr << "<mem>";
1146 void Andersens::PrintConstraints() {
1147 std::cerr << "Constraints:\n";
1148 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
1149 std::cerr << " #" << i << ": ";
1150 Constraint &C = Constraints[i];
1151 if (C.Type == Constraint::Store)
1155 if (C.Type == Constraint::Load)
1162 void Andersens::PrintPointsToGraph() {
1163 std::cerr << "Points-to graph:\n";
1164 for (unsigned i = 0, e = GraphNodes.size(); i != e; ++i) {
1165 Node *N = &GraphNodes[i];
1166 std::cerr << "[" << (N->end() - N->begin()) << "] ";
1168 std::cerr << "\t--> ";
1169 for (Node::iterator I = N->begin(), E = N->end(); I != E; ++I) {
1170 if (I != N->begin()) std::cerr << ", ";