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"
67 STATISTIC(NumIters , "Number of iterations to reach convergence");
68 STATISTIC(NumConstraints , "Number of constraints");
69 STATISTIC(NumNodes , "Number of nodes");
70 STATISTIC(NumEscapingFunctions, "Number of internal functions that escape");
71 STATISTIC(NumIndirectCallees , "Number of indirect callees found");
74 class Andersens : public ModulePass, public AliasAnalysis,
75 private InstVisitor<Andersens> {
76 /// Node class - This class is used to represent a memory object in the
77 /// program, and is the primitive used to build the points-to graph.
79 std::vector<Node*> Pointees;
83 Node *setValue(Value *V) {
84 assert(Val == 0 && "Value already set for this node!");
89 /// getValue - Return the LLVM value corresponding to this node.
91 Value *getValue() const { return Val; }
93 typedef std::vector<Node*>::const_iterator iterator;
94 iterator begin() const { return Pointees.begin(); }
95 iterator end() const { return Pointees.end(); }
97 /// addPointerTo - Add a pointer to the list of pointees of this node,
98 /// returning true if this caused a new pointer to be added, or false if
99 /// we already knew about the points-to relation.
100 bool addPointerTo(Node *N) {
101 std::vector<Node*>::iterator I = std::lower_bound(Pointees.begin(),
104 if (I != Pointees.end() && *I == N)
106 Pointees.insert(I, N);
110 /// intersects - Return true if the points-to set of this node intersects
111 /// with the points-to set of the specified node.
112 bool intersects(Node *N) const;
114 /// intersectsIgnoring - Return true if the points-to set of this node
115 /// intersects with the points-to set of the specified node on any nodes
116 /// except for the specified node to ignore.
117 bool intersectsIgnoring(Node *N, Node *Ignoring) const;
119 // Constraint application methods.
120 bool copyFrom(Node *N);
121 bool loadFrom(Node *N);
122 bool storeThrough(Node *N);
125 /// GraphNodes - This vector is populated as part of the object
126 /// identification stage of the analysis, which populates this vector with a
127 /// node for each memory object and fills in the ValueNodes map.
128 std::vector<Node> GraphNodes;
130 /// ValueNodes - This map indicates the Node that a particular Value* is
131 /// represented by. This contains entries for all pointers.
132 std::map<Value*, unsigned> ValueNodes;
134 /// ObjectNodes - This map contains entries for each memory object in the
135 /// program: globals, alloca's and mallocs.
136 std::map<Value*, unsigned> ObjectNodes;
138 /// ReturnNodes - This map contains an entry for each function in the
139 /// program that returns a value.
140 std::map<Function*, unsigned> ReturnNodes;
142 /// VarargNodes - This map contains the entry used to represent all pointers
143 /// passed through the varargs portion of a function call for a particular
144 /// function. An entry is not present in this map for functions that do not
145 /// take variable arguments.
146 std::map<Function*, unsigned> VarargNodes;
148 /// Constraint - Objects of this structure are used to represent the various
149 /// constraints identified by the algorithm. The constraints are 'copy',
150 /// for statements like "A = B", 'load' for statements like "A = *B", and
151 /// 'store' for statements like "*A = B".
153 enum ConstraintType { Copy, Load, Store } Type;
156 Constraint(ConstraintType Ty, Node *D, Node *S)
157 : Type(Ty), Dest(D), Src(S) {}
160 /// Constraints - This vector contains a list of all of the constraints
161 /// identified by the program.
162 std::vector<Constraint> Constraints;
164 /// EscapingInternalFunctions - This set contains all of the internal
165 /// functions that are found to escape from the program. If the address of
166 /// an internal function is passed to an external function or otherwise
167 /// escapes from the analyzed portion of the program, we must assume that
168 /// any pointer arguments can alias the universal node. This set keeps
169 /// track of those functions we are assuming to escape so far.
170 std::set<Function*> EscapingInternalFunctions;
172 /// IndirectCalls - This contains a list of all of the indirect call sites
173 /// in the program. Since the call graph is iteratively discovered, we may
174 /// need to add constraints to our graph as we find new targets of function
176 std::vector<CallSite> IndirectCalls;
178 /// IndirectCallees - For each call site in the indirect calls list, keep
179 /// track of the callees that we have discovered so far. As the analysis
180 /// proceeds, more callees are discovered, until the call graph finally
182 std::map<CallSite, std::vector<Function*> > IndirectCallees;
184 /// This enum defines the GraphNodes indices that correspond to important
193 bool runOnModule(Module &M) {
194 InitializeAliasAnalysis(this);
196 CollectConstraints(M);
197 DEBUG(PrintConstraints());
199 DEBUG(PrintPointsToGraph());
201 // Free the constraints list, as we don't need it to respond to alias
206 EscapingInternalFunctions.clear();
207 std::vector<Constraint>().swap(Constraints);
211 void releaseMemory() {
212 // FIXME: Until we have transitively required passes working correctly,
213 // this cannot be enabled! Otherwise, using -count-aa with the pass
214 // causes memory to be freed too early. :(
216 // The memory objects and ValueNodes data structures at the only ones that
217 // are still live after construction.
218 std::vector<Node>().swap(GraphNodes);
223 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
224 AliasAnalysis::getAnalysisUsage(AU);
225 AU.setPreservesAll(); // Does not transform code
228 //------------------------------------------------
229 // Implement the AliasAnalysis API
231 AliasResult alias(const Value *V1, unsigned V1Size,
232 const Value *V2, unsigned V2Size);
233 virtual ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size);
234 virtual ModRefResult getModRefInfo(CallSite CS1, CallSite CS2);
235 void getMustAliases(Value *P, std::vector<Value*> &RetVals);
236 bool pointsToConstantMemory(const Value *P);
238 virtual void deleteValue(Value *V) {
240 getAnalysis<AliasAnalysis>().deleteValue(V);
243 virtual void copyValue(Value *From, Value *To) {
244 ValueNodes[To] = ValueNodes[From];
245 getAnalysis<AliasAnalysis>().copyValue(From, To);
249 /// getNode - Return the node corresponding to the specified pointer scalar.
251 Node *getNode(Value *V) {
252 if (Constant *C = dyn_cast<Constant>(V))
253 if (!isa<GlobalValue>(C))
254 return getNodeForConstantPointer(C);
256 std::map<Value*, unsigned>::iterator I = ValueNodes.find(V);
257 if (I == ValueNodes.end()) {
261 assert(0 && "Value does not have a node in the points-to graph!");
263 return &GraphNodes[I->second];
266 /// getObject - Return the node corresponding to the memory object for the
267 /// specified global or allocation instruction.
268 Node *getObject(Value *V) {
269 std::map<Value*, unsigned>::iterator I = ObjectNodes.find(V);
270 assert(I != ObjectNodes.end() &&
271 "Value does not have an object in the points-to graph!");
272 return &GraphNodes[I->second];
275 /// getReturnNode - Return the node representing the return value for the
276 /// specified function.
277 Node *getReturnNode(Function *F) {
278 std::map<Function*, unsigned>::iterator I = ReturnNodes.find(F);
279 assert(I != ReturnNodes.end() && "Function does not return a value!");
280 return &GraphNodes[I->second];
283 /// getVarargNode - Return the node representing the variable arguments
284 /// formal for the specified function.
285 Node *getVarargNode(Function *F) {
286 std::map<Function*, unsigned>::iterator I = VarargNodes.find(F);
287 assert(I != VarargNodes.end() && "Function does not take var args!");
288 return &GraphNodes[I->second];
291 /// getNodeValue - Get the node for the specified LLVM value and set the
292 /// value for it to be the specified value.
293 Node *getNodeValue(Value &V) {
294 return getNode(&V)->setValue(&V);
297 void IdentifyObjects(Module &M);
298 void CollectConstraints(Module &M);
299 void SolveConstraints();
301 Node *getNodeForConstantPointer(Constant *C);
302 Node *getNodeForConstantPointerTarget(Constant *C);
303 void AddGlobalInitializerConstraints(Node *N, Constant *C);
305 void AddConstraintsForNonInternalLinkage(Function *F);
306 void AddConstraintsForCall(CallSite CS, Function *F);
307 bool AddConstraintsForExternalCall(CallSite CS, Function *F);
310 void PrintNode(Node *N);
311 void PrintConstraints();
312 void PrintPointsToGraph();
314 //===------------------------------------------------------------------===//
315 // Instruction visitation methods for adding constraints
317 friend class InstVisitor<Andersens>;
318 void visitReturnInst(ReturnInst &RI);
319 void visitInvokeInst(InvokeInst &II) { visitCallSite(CallSite(&II)); }
320 void visitCallInst(CallInst &CI) { visitCallSite(CallSite(&CI)); }
321 void visitCallSite(CallSite CS);
322 void visitAllocationInst(AllocationInst &AI);
323 void visitLoadInst(LoadInst &LI);
324 void visitStoreInst(StoreInst &SI);
325 void visitGetElementPtrInst(GetElementPtrInst &GEP);
326 void visitPHINode(PHINode &PN);
327 void visitCastInst(CastInst &CI);
328 void visitICmpInst(ICmpInst &ICI) {} // NOOP!
329 void visitFCmpInst(FCmpInst &ICI) {} // NOOP!
330 void visitSelectInst(SelectInst &SI);
331 void visitVAArg(VAArgInst &I);
332 void visitInstruction(Instruction &I);
335 RegisterPass<Andersens> X("anders-aa",
336 "Andersen's Interprocedural Alias Analysis");
337 RegisterAnalysisGroup<AliasAnalysis> Y(X);
340 ModulePass *llvm::createAndersensPass() { return new Andersens(); }
342 //===----------------------------------------------------------------------===//
343 // AliasAnalysis Interface Implementation
344 //===----------------------------------------------------------------------===//
346 AliasAnalysis::AliasResult Andersens::alias(const Value *V1, unsigned V1Size,
347 const Value *V2, unsigned V2Size) {
348 Node *N1 = getNode(const_cast<Value*>(V1));
349 Node *N2 = getNode(const_cast<Value*>(V2));
351 // Check to see if the two pointers are known to not alias. They don't alias
352 // if their points-to sets do not intersect.
353 if (!N1->intersectsIgnoring(N2, &GraphNodes[NullObject]))
356 return AliasAnalysis::alias(V1, V1Size, V2, V2Size);
359 AliasAnalysis::ModRefResult
360 Andersens::getModRefInfo(CallSite CS, Value *P, unsigned Size) {
361 // The only thing useful that we can contribute for mod/ref information is
362 // when calling external function calls: if we know that memory never escapes
363 // from the program, it cannot be modified by an external call.
365 // NOTE: This is not really safe, at least not when the entire program is not
366 // available. The deal is that the external function could call back into the
367 // program and modify stuff. We ignore this technical niggle for now. This
368 // is, after all, a "research quality" implementation of Andersen's analysis.
369 if (Function *F = CS.getCalledFunction())
370 if (F->isExternal()) {
371 Node *N1 = getNode(P);
373 if (N1->begin() == N1->end())
374 return NoModRef; // P doesn't point to anything.
376 // Get the first pointee.
377 Node *FirstPointee = *N1->begin();
378 if (FirstPointee != &GraphNodes[UniversalSet])
379 return NoModRef; // P doesn't point to the universal set.
382 return AliasAnalysis::getModRefInfo(CS, P, Size);
385 AliasAnalysis::ModRefResult
386 Andersens::getModRefInfo(CallSite CS1, CallSite CS2) {
387 return AliasAnalysis::getModRefInfo(CS1,CS2);
390 /// getMustAlias - We can provide must alias information if we know that a
391 /// pointer can only point to a specific function or the null pointer.
392 /// Unfortunately we cannot determine must-alias information for global
393 /// variables or any other memory memory objects because we do not track whether
394 /// a pointer points to the beginning of an object or a field of it.
395 void Andersens::getMustAliases(Value *P, std::vector<Value*> &RetVals) {
396 Node *N = getNode(P);
397 Node::iterator I = N->begin();
399 // If there is exactly one element in the points-to set for the object...
402 Node *Pointee = *N->begin();
404 // If a function is the only object in the points-to set, then it must be
405 // the destination. Note that we can't handle global variables here,
406 // because we don't know if the pointer is actually pointing to a field of
407 // the global or to the beginning of it.
408 if (Value *V = Pointee->getValue()) {
409 if (Function *F = dyn_cast<Function>(V))
410 RetVals.push_back(F);
412 // If the object in the points-to set is the null object, then the null
413 // pointer is a must alias.
414 if (Pointee == &GraphNodes[NullObject])
415 RetVals.push_back(Constant::getNullValue(P->getType()));
420 AliasAnalysis::getMustAliases(P, RetVals);
423 /// pointsToConstantMemory - If we can determine that this pointer only points
424 /// to constant memory, return true. In practice, this means that if the
425 /// pointer can only point to constant globals, functions, or the null pointer,
428 bool Andersens::pointsToConstantMemory(const Value *P) {
429 Node *N = getNode((Value*)P);
430 for (Node::iterator I = N->begin(), E = N->end(); I != E; ++I) {
431 if (Value *V = (*I)->getValue()) {
432 if (!isa<GlobalValue>(V) || (isa<GlobalVariable>(V) &&
433 !cast<GlobalVariable>(V)->isConstant()))
434 return AliasAnalysis::pointsToConstantMemory(P);
436 if (*I != &GraphNodes[NullObject])
437 return AliasAnalysis::pointsToConstantMemory(P);
444 //===----------------------------------------------------------------------===//
445 // Object Identification Phase
446 //===----------------------------------------------------------------------===//
448 /// IdentifyObjects - This stage scans the program, adding an entry to the
449 /// GraphNodes list for each memory object in the program (global stack or
450 /// heap), and populates the ValueNodes and ObjectNodes maps for these objects.
452 void Andersens::IdentifyObjects(Module &M) {
453 unsigned NumObjects = 0;
455 // Object #0 is always the universal set: the object that we don't know
457 assert(NumObjects == UniversalSet && "Something changed!");
460 // Object #1 always represents the null pointer.
461 assert(NumObjects == NullPtr && "Something changed!");
464 // Object #2 always represents the null object (the object pointed to by null)
465 assert(NumObjects == NullObject && "Something changed!");
468 // Add all the globals first.
469 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
471 ObjectNodes[I] = NumObjects++;
472 ValueNodes[I] = NumObjects++;
475 // Add nodes for all of the functions and the instructions inside of them.
476 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
477 // The function itself is a memory object.
478 ValueNodes[F] = NumObjects++;
479 ObjectNodes[F] = NumObjects++;
480 if (isa<PointerType>(F->getFunctionType()->getReturnType()))
481 ReturnNodes[F] = NumObjects++;
482 if (F->getFunctionType()->isVarArg())
483 VarargNodes[F] = NumObjects++;
485 // Add nodes for all of the incoming pointer arguments.
486 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
488 if (isa<PointerType>(I->getType()))
489 ValueNodes[I] = NumObjects++;
491 // Scan the function body, creating a memory object for each heap/stack
492 // allocation in the body of the function and a node to represent all
493 // pointer values defined by instructions and used as operands.
494 for (inst_iterator II = inst_begin(F), E = inst_end(F); II != E; ++II) {
495 // If this is an heap or stack allocation, create a node for the memory
497 if (isa<PointerType>(II->getType())) {
498 ValueNodes[&*II] = NumObjects++;
499 if (AllocationInst *AI = dyn_cast<AllocationInst>(&*II))
500 ObjectNodes[AI] = NumObjects++;
505 // Now that we know how many objects to create, make them all now!
506 GraphNodes.resize(NumObjects);
507 NumNodes += NumObjects;
510 //===----------------------------------------------------------------------===//
511 // Constraint Identification Phase
512 //===----------------------------------------------------------------------===//
514 /// getNodeForConstantPointer - Return the node corresponding to the constant
516 Andersens::Node *Andersens::getNodeForConstantPointer(Constant *C) {
517 assert(isa<PointerType>(C->getType()) && "Not a constant pointer!");
519 if (isa<ConstantPointerNull>(C) || isa<UndefValue>(C))
520 return &GraphNodes[NullPtr];
521 else if (GlobalValue *GV = dyn_cast<GlobalValue>(C))
523 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
524 switch (CE->getOpcode()) {
525 case Instruction::GetElementPtr:
526 return getNodeForConstantPointer(CE->getOperand(0));
527 case Instruction::IntToPtr:
528 return &GraphNodes[UniversalSet];
529 case Instruction::BitCast:
530 return getNodeForConstantPointer(CE->getOperand(0));
532 cerr << "Constant Expr not yet handled: " << *CE << "\n";
536 assert(0 && "Unknown constant pointer!");
541 /// getNodeForConstantPointerTarget - Return the node POINTED TO by the
542 /// specified constant pointer.
543 Andersens::Node *Andersens::getNodeForConstantPointerTarget(Constant *C) {
544 assert(isa<PointerType>(C->getType()) && "Not a constant pointer!");
546 if (isa<ConstantPointerNull>(C))
547 return &GraphNodes[NullObject];
548 else if (GlobalValue *GV = dyn_cast<GlobalValue>(C))
549 return getObject(GV);
550 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
551 switch (CE->getOpcode()) {
552 case Instruction::GetElementPtr:
553 return getNodeForConstantPointerTarget(CE->getOperand(0));
554 case Instruction::IntToPtr:
555 return &GraphNodes[UniversalSet];
556 case Instruction::BitCast:
557 return getNodeForConstantPointerTarget(CE->getOperand(0));
559 cerr << "Constant Expr not yet handled: " << *CE << "\n";
563 assert(0 && "Unknown constant pointer!");
568 /// AddGlobalInitializerConstraints - Add inclusion constraints for the memory
569 /// object N, which contains values indicated by C.
570 void Andersens::AddGlobalInitializerConstraints(Node *N, Constant *C) {
571 if (C->getType()->isFirstClassType()) {
572 if (isa<PointerType>(C->getType()))
573 N->copyFrom(getNodeForConstantPointer(C));
575 } else if (C->isNullValue()) {
576 N->addPointerTo(&GraphNodes[NullObject]);
578 } else if (!isa<UndefValue>(C)) {
579 // If this is an array or struct, include constraints for each element.
580 assert(isa<ConstantArray>(C) || isa<ConstantStruct>(C));
581 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i)
582 AddGlobalInitializerConstraints(N, cast<Constant>(C->getOperand(i)));
586 /// AddConstraintsForNonInternalLinkage - If this function does not have
587 /// internal linkage, realize that we can't trust anything passed into or
588 /// returned by this function.
589 void Andersens::AddConstraintsForNonInternalLinkage(Function *F) {
590 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
591 if (isa<PointerType>(I->getType()))
592 // If this is an argument of an externally accessible function, the
593 // incoming pointer might point to anything.
594 Constraints.push_back(Constraint(Constraint::Copy, getNode(I),
595 &GraphNodes[UniversalSet]));
598 /// AddConstraintsForCall - If this is a call to a "known" function, add the
599 /// constraints and return true. If this is a call to an unknown function,
601 bool Andersens::AddConstraintsForExternalCall(CallSite CS, Function *F) {
602 assert(F->isExternal() && "Not an external function!");
604 // These functions don't induce any points-to constraints.
605 if (F->getName() == "atoi" || F->getName() == "atof" ||
606 F->getName() == "atol" || F->getName() == "atoll" ||
607 F->getName() == "remove" || F->getName() == "unlink" ||
608 F->getName() == "rename" || F->getName() == "memcmp" ||
609 F->getName() == "llvm.memset.i32" ||
610 F->getName() == "llvm.memset.i64" ||
611 F->getName() == "strcmp" || F->getName() == "strncmp" ||
612 F->getName() == "execl" || F->getName() == "execlp" ||
613 F->getName() == "execle" || F->getName() == "execv" ||
614 F->getName() == "execvp" || F->getName() == "chmod" ||
615 F->getName() == "puts" || F->getName() == "write" ||
616 F->getName() == "open" || F->getName() == "create" ||
617 F->getName() == "truncate" || F->getName() == "chdir" ||
618 F->getName() == "mkdir" || F->getName() == "rmdir" ||
619 F->getName() == "read" || F->getName() == "pipe" ||
620 F->getName() == "wait" || F->getName() == "time" ||
621 F->getName() == "stat" || F->getName() == "fstat" ||
622 F->getName() == "lstat" || F->getName() == "strtod" ||
623 F->getName() == "strtof" || F->getName() == "strtold" ||
624 F->getName() == "fopen" || F->getName() == "fdopen" ||
625 F->getName() == "freopen" ||
626 F->getName() == "fflush" || F->getName() == "feof" ||
627 F->getName() == "fileno" || F->getName() == "clearerr" ||
628 F->getName() == "rewind" || F->getName() == "ftell" ||
629 F->getName() == "ferror" || F->getName() == "fgetc" ||
630 F->getName() == "fgetc" || F->getName() == "_IO_getc" ||
631 F->getName() == "fwrite" || F->getName() == "fread" ||
632 F->getName() == "fgets" || F->getName() == "ungetc" ||
633 F->getName() == "fputc" ||
634 F->getName() == "fputs" || F->getName() == "putc" ||
635 F->getName() == "ftell" || F->getName() == "rewind" ||
636 F->getName() == "_IO_putc" || F->getName() == "fseek" ||
637 F->getName() == "fgetpos" || F->getName() == "fsetpos" ||
638 F->getName() == "printf" || F->getName() == "fprintf" ||
639 F->getName() == "sprintf" || F->getName() == "vprintf" ||
640 F->getName() == "vfprintf" || F->getName() == "vsprintf" ||
641 F->getName() == "scanf" || F->getName() == "fscanf" ||
642 F->getName() == "sscanf" || F->getName() == "__assert_fail" ||
643 F->getName() == "modf")
647 // These functions do induce points-to edges.
648 if (F->getName() == "llvm.memcpy.i32" || F->getName() == "llvm.memcpy.i64" ||
649 F->getName() == "llvm.memmove.i32" ||F->getName() == "llvm.memmove.i64" ||
650 F->getName() == "memmove") {
651 // Note: this is a poor approximation, this says Dest = Src, instead of
653 Constraints.push_back(Constraint(Constraint::Copy,
654 getNode(CS.getArgument(0)),
655 getNode(CS.getArgument(1))));
660 if (F->getName() == "realloc" || F->getName() == "strchr" ||
661 F->getName() == "strrchr" || F->getName() == "strstr" ||
662 F->getName() == "strtok") {
663 Constraints.push_back(Constraint(Constraint::Copy,
664 getNode(CS.getInstruction()),
665 getNode(CS.getArgument(0))));
674 /// CollectConstraints - This stage scans the program, adding a constraint to
675 /// the Constraints list for each instruction in the program that induces a
676 /// constraint, and setting up the initial points-to graph.
678 void Andersens::CollectConstraints(Module &M) {
679 // First, the universal set points to itself.
680 GraphNodes[UniversalSet].addPointerTo(&GraphNodes[UniversalSet]);
681 //Constraints.push_back(Constraint(Constraint::Load, &GraphNodes[UniversalSet],
682 // &GraphNodes[UniversalSet]));
683 Constraints.push_back(Constraint(Constraint::Store, &GraphNodes[UniversalSet],
684 &GraphNodes[UniversalSet]));
686 // Next, the null pointer points to the null object.
687 GraphNodes[NullPtr].addPointerTo(&GraphNodes[NullObject]);
689 // Next, add any constraints on global variables and their initializers.
690 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
692 // Associate the address of the global object as pointing to the memory for
693 // the global: &G = <G memory>
694 Node *Object = getObject(I);
696 getNodeValue(*I)->addPointerTo(Object);
698 if (I->hasInitializer()) {
699 AddGlobalInitializerConstraints(Object, I->getInitializer());
701 // If it doesn't have an initializer (i.e. it's defined in another
702 // translation unit), it points to the universal set.
703 Constraints.push_back(Constraint(Constraint::Copy, Object,
704 &GraphNodes[UniversalSet]));
708 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
709 // Make the function address point to the function object.
710 getNodeValue(*F)->addPointerTo(getObject(F)->setValue(F));
712 // Set up the return value node.
713 if (isa<PointerType>(F->getFunctionType()->getReturnType()))
714 getReturnNode(F)->setValue(F);
715 if (F->getFunctionType()->isVarArg())
716 getVarargNode(F)->setValue(F);
718 // Set up incoming argument nodes.
719 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
721 if (isa<PointerType>(I->getType()))
724 if (!F->hasInternalLinkage())
725 AddConstraintsForNonInternalLinkage(F);
727 if (!F->isExternal()) {
728 // Scan the function body, creating a memory object for each heap/stack
729 // allocation in the body of the function and a node to represent all
730 // pointer values defined by instructions and used as operands.
733 // External functions that return pointers return the universal set.
734 if (isa<PointerType>(F->getFunctionType()->getReturnType()))
735 Constraints.push_back(Constraint(Constraint::Copy,
737 &GraphNodes[UniversalSet]));
739 // Any pointers that are passed into the function have the universal set
741 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
743 if (isa<PointerType>(I->getType())) {
744 // Pointers passed into external functions could have anything stored
746 Constraints.push_back(Constraint(Constraint::Store, getNode(I),
747 &GraphNodes[UniversalSet]));
748 // Memory objects passed into external function calls can have the
749 // universal set point to them.
750 Constraints.push_back(Constraint(Constraint::Copy,
751 &GraphNodes[UniversalSet],
755 // If this is an external varargs function, it can also store pointers
756 // into any pointers passed through the varargs section.
757 if (F->getFunctionType()->isVarArg())
758 Constraints.push_back(Constraint(Constraint::Store, getVarargNode(F),
759 &GraphNodes[UniversalSet]));
762 NumConstraints += Constraints.size();
766 void Andersens::visitInstruction(Instruction &I) {
768 return; // This function is just a big assert.
770 if (isa<BinaryOperator>(I))
772 // Most instructions don't have any effect on pointer values.
773 switch (I.getOpcode()) {
774 case Instruction::Br:
775 case Instruction::Switch:
776 case Instruction::Unwind:
777 case Instruction::Unreachable:
778 case Instruction::Free:
779 case Instruction::Shl:
780 case Instruction::LShr:
781 case Instruction::AShr:
782 case Instruction::ICmp:
783 case Instruction::FCmp:
786 // Is this something we aren't handling yet?
787 cerr << "Unknown instruction: " << I;
792 void Andersens::visitAllocationInst(AllocationInst &AI) {
793 getNodeValue(AI)->addPointerTo(getObject(&AI)->setValue(&AI));
796 void Andersens::visitReturnInst(ReturnInst &RI) {
797 if (RI.getNumOperands() && isa<PointerType>(RI.getOperand(0)->getType()))
798 // return V --> <Copy/retval{F}/v>
799 Constraints.push_back(Constraint(Constraint::Copy,
800 getReturnNode(RI.getParent()->getParent()),
801 getNode(RI.getOperand(0))));
804 void Andersens::visitLoadInst(LoadInst &LI) {
805 if (isa<PointerType>(LI.getType()))
806 // P1 = load P2 --> <Load/P1/P2>
807 Constraints.push_back(Constraint(Constraint::Load, getNodeValue(LI),
808 getNode(LI.getOperand(0))));
811 void Andersens::visitStoreInst(StoreInst &SI) {
812 if (isa<PointerType>(SI.getOperand(0)->getType()))
813 // store P1, P2 --> <Store/P2/P1>
814 Constraints.push_back(Constraint(Constraint::Store,
815 getNode(SI.getOperand(1)),
816 getNode(SI.getOperand(0))));
819 void Andersens::visitGetElementPtrInst(GetElementPtrInst &GEP) {
820 // P1 = getelementptr P2, ... --> <Copy/P1/P2>
821 Constraints.push_back(Constraint(Constraint::Copy, getNodeValue(GEP),
822 getNode(GEP.getOperand(0))));
825 void Andersens::visitPHINode(PHINode &PN) {
826 if (isa<PointerType>(PN.getType())) {
827 Node *PNN = getNodeValue(PN);
828 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
829 // P1 = phi P2, P3 --> <Copy/P1/P2>, <Copy/P1/P3>, ...
830 Constraints.push_back(Constraint(Constraint::Copy, PNN,
831 getNode(PN.getIncomingValue(i))));
835 void Andersens::visitCastInst(CastInst &CI) {
836 Value *Op = CI.getOperand(0);
837 if (isa<PointerType>(CI.getType())) {
838 if (isa<PointerType>(Op->getType())) {
839 // P1 = cast P2 --> <Copy/P1/P2>
840 Constraints.push_back(Constraint(Constraint::Copy, getNodeValue(CI),
841 getNode(CI.getOperand(0))));
843 // P1 = cast int --> <Copy/P1/Univ>
845 Constraints.push_back(Constraint(Constraint::Copy, getNodeValue(CI),
846 &GraphNodes[UniversalSet]));
851 } else if (isa<PointerType>(Op->getType())) {
852 // int = cast P1 --> <Copy/Univ/P1>
854 Constraints.push_back(Constraint(Constraint::Copy,
855 &GraphNodes[UniversalSet],
856 getNode(CI.getOperand(0))));
858 getNode(CI.getOperand(0));
863 void Andersens::visitSelectInst(SelectInst &SI) {
864 if (isa<PointerType>(SI.getType())) {
865 Node *SIN = getNodeValue(SI);
866 // P1 = select C, P2, P3 ---> <Copy/P1/P2>, <Copy/P1/P3>
867 Constraints.push_back(Constraint(Constraint::Copy, SIN,
868 getNode(SI.getOperand(1))));
869 Constraints.push_back(Constraint(Constraint::Copy, SIN,
870 getNode(SI.getOperand(2))));
874 void Andersens::visitVAArg(VAArgInst &I) {
875 assert(0 && "vaarg not handled yet!");
878 /// AddConstraintsForCall - Add constraints for a call with actual arguments
879 /// specified by CS to the function specified by F. Note that the types of
880 /// arguments might not match up in the case where this is an indirect call and
881 /// the function pointer has been casted. If this is the case, do something
883 void Andersens::AddConstraintsForCall(CallSite CS, Function *F) {
884 // If this is a call to an external function, handle it directly to get some
885 // taste of context sensitivity.
886 if (F->isExternal() && AddConstraintsForExternalCall(CS, F))
889 if (isa<PointerType>(CS.getType())) {
890 Node *CSN = getNode(CS.getInstruction());
891 if (isa<PointerType>(F->getFunctionType()->getReturnType())) {
892 Constraints.push_back(Constraint(Constraint::Copy, CSN,
895 // If the function returns a non-pointer value, handle this just like we
896 // treat a nonpointer cast to pointer.
897 Constraints.push_back(Constraint(Constraint::Copy, CSN,
898 &GraphNodes[UniversalSet]));
900 } else if (isa<PointerType>(F->getFunctionType()->getReturnType())) {
901 Constraints.push_back(Constraint(Constraint::Copy,
902 &GraphNodes[UniversalSet],
906 Function::arg_iterator AI = F->arg_begin(), AE = F->arg_end();
907 CallSite::arg_iterator ArgI = CS.arg_begin(), ArgE = CS.arg_end();
908 for (; AI != AE && ArgI != ArgE; ++AI, ++ArgI)
909 if (isa<PointerType>(AI->getType())) {
910 if (isa<PointerType>((*ArgI)->getType())) {
911 // Copy the actual argument into the formal argument.
912 Constraints.push_back(Constraint(Constraint::Copy, getNode(AI),
915 Constraints.push_back(Constraint(Constraint::Copy, getNode(AI),
916 &GraphNodes[UniversalSet]));
918 } else if (isa<PointerType>((*ArgI)->getType())) {
919 Constraints.push_back(Constraint(Constraint::Copy,
920 &GraphNodes[UniversalSet],
924 // Copy all pointers passed through the varargs section to the varargs node.
925 if (F->getFunctionType()->isVarArg())
926 for (; ArgI != ArgE; ++ArgI)
927 if (isa<PointerType>((*ArgI)->getType()))
928 Constraints.push_back(Constraint(Constraint::Copy, getVarargNode(F),
930 // If more arguments are passed in than we track, just drop them on the floor.
933 void Andersens::visitCallSite(CallSite CS) {
934 if (isa<PointerType>(CS.getType()))
935 getNodeValue(*CS.getInstruction());
937 if (Function *F = CS.getCalledFunction()) {
938 AddConstraintsForCall(CS, F);
940 // We don't handle indirect call sites yet. Keep track of them for when we
941 // discover the call graph incrementally.
942 IndirectCalls.push_back(CS);
946 //===----------------------------------------------------------------------===//
947 // Constraint Solving Phase
948 //===----------------------------------------------------------------------===//
950 /// intersects - Return true if the points-to set of this node intersects
951 /// with the points-to set of the specified node.
952 bool Andersens::Node::intersects(Node *N) const {
953 iterator I1 = begin(), I2 = N->begin(), E1 = end(), E2 = N->end();
954 while (I1 != E1 && I2 != E2) {
955 if (*I1 == *I2) return true;
964 /// intersectsIgnoring - Return true if the points-to set of this node
965 /// intersects with the points-to set of the specified node on any nodes
966 /// except for the specified node to ignore.
967 bool Andersens::Node::intersectsIgnoring(Node *N, Node *Ignoring) const {
968 iterator I1 = begin(), I2 = N->begin(), E1 = end(), E2 = N->end();
969 while (I1 != E1 && I2 != E2) {
971 if (*I1 != Ignoring) return true;
973 } else if (*I1 < *I2)
981 // Copy constraint: all edges out of the source node get copied to the
982 // destination node. This returns true if a change is made.
983 bool Andersens::Node::copyFrom(Node *N) {
984 // Use a mostly linear-time merge since both of the lists are sorted.
985 bool Changed = false;
986 iterator I = N->begin(), E = N->end();
988 while (I != E && i != Pointees.size()) {
989 if (Pointees[i] < *I) {
991 } else if (Pointees[i] == *I) {
994 // We found a new element to copy over.
996 Pointees.insert(Pointees.begin()+i, *I);
1002 Pointees.insert(Pointees.end(), I, E);
1009 bool Andersens::Node::loadFrom(Node *N) {
1010 bool Changed = false;
1011 for (iterator I = N->begin(), E = N->end(); I != E; ++I)
1012 Changed |= copyFrom(*I);
1016 bool Andersens::Node::storeThrough(Node *N) {
1017 bool Changed = false;
1018 for (iterator I = begin(), E = end(); I != E; ++I)
1019 Changed |= (*I)->copyFrom(N);
1024 /// SolveConstraints - This stage iteratively processes the constraints list
1025 /// propagating constraints (adding edges to the Nodes in the points-to graph)
1026 /// until a fixed point is reached.
1028 void Andersens::SolveConstraints() {
1029 bool Changed = true;
1030 unsigned Iteration = 0;
1034 DOUT << "Starting iteration #" << Iteration++ << "!\n";
1036 // Loop over all of the constraints, applying them in turn.
1037 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
1038 Constraint &C = Constraints[i];
1040 case Constraint::Copy:
1041 Changed |= C.Dest->copyFrom(C.Src);
1043 case Constraint::Load:
1044 Changed |= C.Dest->loadFrom(C.Src);
1046 case Constraint::Store:
1047 Changed |= C.Dest->storeThrough(C.Src);
1050 assert(0 && "Unknown constraint!");
1055 // Check to see if any internal function's addresses have been passed to
1056 // external functions. If so, we have to assume that their incoming
1057 // arguments could be anything. If there are any internal functions in
1058 // the universal node that we don't know about, we must iterate.
1059 for (Node::iterator I = GraphNodes[UniversalSet].begin(),
1060 E = GraphNodes[UniversalSet].end(); I != E; ++I)
1061 if (Function *F = dyn_cast_or_null<Function>((*I)->getValue()))
1062 if (F->hasInternalLinkage() &&
1063 EscapingInternalFunctions.insert(F).second) {
1064 // We found a function that is just now escaping. Mark it as if it
1065 // didn't have internal linkage.
1066 AddConstraintsForNonInternalLinkage(F);
1067 DOUT << "Found escaping internal function: " << F->getName() <<"\n";
1068 ++NumEscapingFunctions;
1071 // Check to see if we have discovered any new callees of the indirect call
1072 // sites. If so, add constraints to the analysis.
1073 for (unsigned i = 0, e = IndirectCalls.size(); i != e; ++i) {
1074 CallSite CS = IndirectCalls[i];
1075 std::vector<Function*> &KnownCallees = IndirectCallees[CS];
1076 Node *CN = getNode(CS.getCalledValue());
1078 for (Node::iterator NI = CN->begin(), E = CN->end(); NI != E; ++NI)
1079 if (Function *F = dyn_cast_or_null<Function>((*NI)->getValue())) {
1080 std::vector<Function*>::iterator IP =
1081 std::lower_bound(KnownCallees.begin(), KnownCallees.end(), F);
1082 if (IP == KnownCallees.end() || *IP != F) {
1083 // Add the constraints for the call now.
1084 AddConstraintsForCall(CS, F);
1085 DOUT << "Found actual callee '"
1086 << F->getName() << "' for call: "
1087 << *CS.getInstruction() << "\n";
1088 ++NumIndirectCallees;
1089 KnownCallees.insert(IP, F);
1099 //===----------------------------------------------------------------------===//
1101 //===----------------------------------------------------------------------===//
1103 void Andersens::PrintNode(Node *N) {
1104 if (N == &GraphNodes[UniversalSet]) {
1105 cerr << "<universal>";
1107 } else if (N == &GraphNodes[NullPtr]) {
1108 cerr << "<nullptr>";
1110 } else if (N == &GraphNodes[NullObject]) {
1115 assert(N->getValue() != 0 && "Never set node label!");
1116 Value *V = N->getValue();
1117 if (Function *F = dyn_cast<Function>(V)) {
1118 if (isa<PointerType>(F->getFunctionType()->getReturnType()) &&
1119 N == getReturnNode(F)) {
1120 cerr << F->getName() << ":retval";
1122 } else if (F->getFunctionType()->isVarArg() && N == getVarargNode(F)) {
1123 cerr << F->getName() << ":vararg";
1128 if (Instruction *I = dyn_cast<Instruction>(V))
1129 cerr << I->getParent()->getParent()->getName() << ":";
1130 else if (Argument *Arg = dyn_cast<Argument>(V))
1131 cerr << Arg->getParent()->getName() << ":";
1134 cerr << V->getName();
1136 cerr << "(unnamed)";
1138 if (isa<GlobalValue>(V) || isa<AllocationInst>(V))
1139 if (N == getObject(V))
1143 void Andersens::PrintConstraints() {
1144 cerr << "Constraints:\n";
1145 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
1146 cerr << " #" << i << ": ";
1147 Constraint &C = Constraints[i];
1148 if (C.Type == Constraint::Store)
1152 if (C.Type == Constraint::Load)
1159 void Andersens::PrintPointsToGraph() {
1160 cerr << "Points-to graph:\n";
1161 for (unsigned i = 0, e = GraphNodes.size(); i != e; ++i) {
1162 Node *N = &GraphNodes[i];
1163 cerr << "[" << (N->end() - N->begin()) << "] ";
1166 for (Node::iterator I = N->begin(), E = N->end(); I != E; ++I) {
1167 if (I != N->begin()) cerr << ", ";