Reverting back to original 1.8 version so I can manually merge in patch.
[llvm-complete.git] / lib / Analysis / IPA / Andersens.cpp
blob1fc70e62a35758dd59dbd4aad192694c4e2103bb
1 //===- Andersens.cpp - Andersen's Interprocedural Alias Analysis ----------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
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.
7 //
8 //===----------------------------------------------------------------------===//
9 //
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
26 // objects.
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"
64 #include <set>
65 #include <iostream>
66 using namespace llvm;
68 namespace {
69 Statistic<>
70 NumIters("anders-aa", "Number of iterations to reach convergence");
71 Statistic<>
72 NumConstraints("anders-aa", "Number of constraints");
73 Statistic<>
74 NumNodes("anders-aa", "Number of nodes");
75 Statistic<>
76 NumEscapingFunctions("anders-aa", "Number of internal functions that escape");
77 Statistic<>
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.
84 class Node {
85 std::vector<Node*> Pointees;
86 Value *Val;
87 public:
88 Node() : Val(0) {}
89 Node *setValue(Value *V) {
90 assert(Val == 0 && "Value already set for this node!");
91 Val = V;
92 return this;
95 /// getValue - Return the LLVM value corresponding to this node.
96 ///
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(),
108 Pointees.end(),
110 if (I != Pointees.end() && *I == N)
111 return false;
112 Pointees.insert(I, N);
113 return true;
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".
158 struct Constraint {
159 enum ConstraintType { Copy, Load, Store } Type;
160 Node *Dest, *Src;
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
181 /// pointers.
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
187 /// stabilizes.
188 std::map<CallSite, std::vector<Function*> > IndirectCallees;
190 /// This enum defines the GraphNodes indices that correspond to important
191 /// fixed sets.
192 enum {
193 UniversalSet = 0,
194 NullPtr = 1,
195 NullObject = 2
198 public:
199 bool runOnModule(Module &M) {
200 InitializeAliasAnalysis(this);
201 IdentifyObjects(M);
202 CollectConstraints(M);
203 DEBUG(PrintConstraints());
204 SolveConstraints();
205 DEBUG(PrintPointsToGraph());
207 // Free the constraints list, as we don't need it to respond to alias
208 // requests.
209 ObjectNodes.clear();
210 ReturnNodes.clear();
211 VarargNodes.clear();
212 EscapingInternalFunctions.clear();
213 std::vector<Constraint>().swap(Constraints);
214 return false;
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. :(
221 #if 0
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);
225 ValueNodes.clear();
226 #endif
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 ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size);
240 void getMustAliases(Value *P, std::vector<Value*> &RetVals);
241 bool pointsToConstantMemory(const Value *P);
243 virtual void deleteValue(Value *V) {
244 ValueNodes.erase(V);
245 getAnalysis<AliasAnalysis>().deleteValue(V);
248 virtual void copyValue(Value *From, Value *To) {
249 ValueNodes[To] = ValueNodes[From];
250 getAnalysis<AliasAnalysis>().copyValue(From, To);
253 private:
254 /// getNode - Return the node corresponding to the specified pointer scalar.
256 Node *getNode(Value *V) {
257 if (Constant *C = dyn_cast<Constant>(V))
258 if (!isa<GlobalValue>(C))
259 return getNodeForConstantPointer(C);
261 std::map<Value*, unsigned>::iterator I = ValueNodes.find(V);
262 if (I == ValueNodes.end()) {
263 #ifndef NDEBUG
264 V->dump();
265 #endif
266 assert(0 && "Value does not have a node in the points-to graph!");
268 return &GraphNodes[I->second];
271 /// getObject - Return the node corresponding to the memory object for the
272 /// specified global or allocation instruction.
273 Node *getObject(Value *V) {
274 std::map<Value*, unsigned>::iterator I = ObjectNodes.find(V);
275 assert(I != ObjectNodes.end() &&
276 "Value does not have an object in the points-to graph!");
277 return &GraphNodes[I->second];
280 /// getReturnNode - Return the node representing the return value for the
281 /// specified function.
282 Node *getReturnNode(Function *F) {
283 std::map<Function*, unsigned>::iterator I = ReturnNodes.find(F);
284 assert(I != ReturnNodes.end() && "Function does not return a value!");
285 return &GraphNodes[I->second];
288 /// getVarargNode - Return the node representing the variable arguments
289 /// formal for the specified function.
290 Node *getVarargNode(Function *F) {
291 std::map<Function*, unsigned>::iterator I = VarargNodes.find(F);
292 assert(I != VarargNodes.end() && "Function does not take var args!");
293 return &GraphNodes[I->second];
296 /// getNodeValue - Get the node for the specified LLVM value and set the
297 /// value for it to be the specified value.
298 Node *getNodeValue(Value &V) {
299 return getNode(&V)->setValue(&V);
302 void IdentifyObjects(Module &M);
303 void CollectConstraints(Module &M);
304 void SolveConstraints();
306 Node *getNodeForConstantPointer(Constant *C);
307 Node *getNodeForConstantPointerTarget(Constant *C);
308 void AddGlobalInitializerConstraints(Node *N, Constant *C);
310 void AddConstraintsForNonInternalLinkage(Function *F);
311 void AddConstraintsForCall(CallSite CS, Function *F);
312 bool AddConstraintsForExternalCall(CallSite CS, Function *F);
315 void PrintNode(Node *N);
316 void PrintConstraints();
317 void PrintPointsToGraph();
319 //===------------------------------------------------------------------===//
320 // Instruction visitation methods for adding constraints
322 friend class InstVisitor<Andersens>;
323 void visitReturnInst(ReturnInst &RI);
324 void visitInvokeInst(InvokeInst &II) { visitCallSite(CallSite(&II)); }
325 void visitCallInst(CallInst &CI) { visitCallSite(CallSite(&CI)); }
326 void visitCallSite(CallSite CS);
327 void visitAllocationInst(AllocationInst &AI);
328 void visitLoadInst(LoadInst &LI);
329 void visitStoreInst(StoreInst &SI);
330 void visitGetElementPtrInst(GetElementPtrInst &GEP);
331 void visitPHINode(PHINode &PN);
332 void visitCastInst(CastInst &CI);
333 void visitSetCondInst(SetCondInst &SCI) {} // NOOP!
334 void visitSelectInst(SelectInst &SI);
335 void visitVAArg(VAArgInst &I);
336 void visitInstruction(Instruction &I);
339 RegisterOpt<Andersens> X("anders-aa",
340 "Andersen's Interprocedural Alias Analysis");
341 RegisterAnalysisGroup<AliasAnalysis, Andersens> Y;
344 ModulePass *llvm::createAndersensPass() { return new Andersens(); }
346 //===----------------------------------------------------------------------===//
347 // AliasAnalysis Interface Implementation
348 //===----------------------------------------------------------------------===//
350 AliasAnalysis::AliasResult Andersens::alias(const Value *V1, unsigned V1Size,
351 const Value *V2, unsigned V2Size) {
352 Node *N1 = getNode(const_cast<Value*>(V1));
353 Node *N2 = getNode(const_cast<Value*>(V2));
355 // Check to see if the two pointers are known to not alias. They don't alias
356 // if their points-to sets do not intersect.
357 if (!N1->intersectsIgnoring(N2, &GraphNodes[NullObject]))
358 return NoAlias;
360 return AliasAnalysis::alias(V1, V1Size, V2, V2Size);
363 AliasAnalysis::ModRefResult
364 Andersens::getModRefInfo(CallSite CS, Value *P, unsigned Size) {
365 // The only thing useful that we can contribute for mod/ref information is
366 // when calling external function calls: if we know that memory never escapes
367 // from the program, it cannot be modified by an external call.
369 // NOTE: This is not really safe, at least not when the entire program is not
370 // available. The deal is that the external function could call back into the
371 // program and modify stuff. We ignore this technical niggle for now. This
372 // is, after all, a "research quality" implementation of Andersen's analysis.
373 if (Function *F = CS.getCalledFunction())
374 if (F->isExternal()) {
375 Node *N1 = getNode(P);
376 bool PointsToUniversalSet = false;
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 /// 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();
398 if (I != N->end()) {
399 // If there is exactly one element in the points-to set for the object...
400 ++I;
401 if (I == N->end()) {
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);
411 } else {
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,
426 /// return true.
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);
435 } else {
436 if (*I != &GraphNodes[NullObject])
437 return AliasAnalysis::pointsToConstantMemory(P);
441 return true;
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
456 // anything about.
457 assert(NumObjects == UniversalSet && "Something changed!");
458 ++NumObjects;
460 // Object #1 always represents the null pointer.
461 assert(NumObjects == NullPtr && "Something changed!");
462 ++NumObjects;
464 // Object #2 always represents the null object (the object pointed to by null)
465 assert(NumObjects == NullObject && "Something changed!");
466 ++NumObjects;
468 // Add all the globals first.
469 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
470 I != E; ++I) {
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();
487 I != E; ++I)
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
496 // object.
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
515 /// pointer itself.
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))
522 return getNode(GV);
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::Cast:
528 if (isa<PointerType>(CE->getOperand(0)->getType()))
529 return getNodeForConstantPointer(CE->getOperand(0));
530 else
531 return &GraphNodes[UniversalSet];
532 default:
533 std::cerr << "Constant Expr not yet handled: " << *CE << "\n";
534 assert(0);
536 } else {
537 assert(0 && "Unknown constant pointer!");
539 return 0;
542 /// getNodeForConstantPointerTarget - Return the node POINTED TO by the
543 /// specified constant pointer.
544 Andersens::Node *Andersens::getNodeForConstantPointerTarget(Constant *C) {
545 assert(isa<PointerType>(C->getType()) && "Not a constant pointer!");
547 if (isa<ConstantPointerNull>(C))
548 return &GraphNodes[NullObject];
549 else if (GlobalValue *GV = dyn_cast<GlobalValue>(C))
550 return getObject(GV);
551 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
552 switch (CE->getOpcode()) {
553 case Instruction::GetElementPtr:
554 return getNodeForConstantPointerTarget(CE->getOperand(0));
555 case Instruction::Cast:
556 if (isa<PointerType>(CE->getOperand(0)->getType()))
557 return getNodeForConstantPointerTarget(CE->getOperand(0));
558 else
559 return &GraphNodes[UniversalSet];
560 default:
561 std::cerr << "Constant Expr not yet handled: " << *CE << "\n";
562 assert(0);
564 } else {
565 assert(0 && "Unknown constant pointer!");
567 return 0;
570 /// AddGlobalInitializerConstraints - Add inclusion constraints for the memory
571 /// object N, which contains values indicated by C.
572 void Andersens::AddGlobalInitializerConstraints(Node *N, Constant *C) {
573 if (C->getType()->isFirstClassType()) {
574 if (isa<PointerType>(C->getType()))
575 N->copyFrom(getNodeForConstantPointer(C));
577 } else if (C->isNullValue()) {
578 N->addPointerTo(&GraphNodes[NullObject]);
579 return;
580 } else if (!isa<UndefValue>(C)) {
581 // If this is an array or struct, include constraints for each element.
582 assert(isa<ConstantArray>(C) || isa<ConstantStruct>(C));
583 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i)
584 AddGlobalInitializerConstraints(N, cast<Constant>(C->getOperand(i)));
588 /// AddConstraintsForNonInternalLinkage - If this function does not have
589 /// internal linkage, realize that we can't trust anything passed into or
590 /// returned by this function.
591 void Andersens::AddConstraintsForNonInternalLinkage(Function *F) {
592 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
593 if (isa<PointerType>(I->getType()))
594 // If this is an argument of an externally accessible function, the
595 // incoming pointer might point to anything.
596 Constraints.push_back(Constraint(Constraint::Copy, getNode(I),
597 &GraphNodes[UniversalSet]));
600 /// AddConstraintsForCall - If this is a call to a "known" function, add the
601 /// constraints and return true. If this is a call to an unknown function,
602 /// return false.
603 bool Andersens::AddConstraintsForExternalCall(CallSite CS, Function *F) {
604 assert(F->isExternal() && "Not an external function!");
606 // These functions don't induce any points-to constraints.
607 if (F->getName() == "atoi" || F->getName() == "atof" ||
608 F->getName() == "atol" || F->getName() == "atoll" ||
609 F->getName() == "remove" || F->getName() == "unlink" ||
610 F->getName() == "rename" || F->getName() == "memcmp" ||
611 F->getName() == "llvm.memset.i32" ||
612 F->getName() == "llvm.memset.i64" ||
613 F->getName() == "strcmp" || F->getName() == "strncmp" ||
614 F->getName() == "execl" || F->getName() == "execlp" ||
615 F->getName() == "execle" || F->getName() == "execv" ||
616 F->getName() == "execvp" || F->getName() == "chmod" ||
617 F->getName() == "puts" || F->getName() == "write" ||
618 F->getName() == "open" || F->getName() == "create" ||
619 F->getName() == "truncate" || F->getName() == "chdir" ||
620 F->getName() == "mkdir" || F->getName() == "rmdir" ||
621 F->getName() == "read" || F->getName() == "pipe" ||
622 F->getName() == "wait" || F->getName() == "time" ||
623 F->getName() == "stat" || F->getName() == "fstat" ||
624 F->getName() == "lstat" || F->getName() == "strtod" ||
625 F->getName() == "strtof" || F->getName() == "strtold" ||
626 F->getName() == "fopen" || F->getName() == "fdopen" ||
627 F->getName() == "freopen" ||
628 F->getName() == "fflush" || F->getName() == "feof" ||
629 F->getName() == "fileno" || F->getName() == "clearerr" ||
630 F->getName() == "rewind" || F->getName() == "ftell" ||
631 F->getName() == "ferror" || F->getName() == "fgetc" ||
632 F->getName() == "fgetc" || F->getName() == "_IO_getc" ||
633 F->getName() == "fwrite" || F->getName() == "fread" ||
634 F->getName() == "fgets" || F->getName() == "ungetc" ||
635 F->getName() == "fputc" ||
636 F->getName() == "fputs" || F->getName() == "putc" ||
637 F->getName() == "ftell" || F->getName() == "rewind" ||
638 F->getName() == "_IO_putc" || F->getName() == "fseek" ||
639 F->getName() == "fgetpos" || F->getName() == "fsetpos" ||
640 F->getName() == "printf" || F->getName() == "fprintf" ||
641 F->getName() == "sprintf" || F->getName() == "vprintf" ||
642 F->getName() == "vfprintf" || F->getName() == "vsprintf" ||
643 F->getName() == "scanf" || F->getName() == "fscanf" ||
644 F->getName() == "sscanf" || F->getName() == "__assert_fail" ||
645 F->getName() == "modf")
646 return true;
649 // These functions do induce points-to edges.
650 if (F->getName() == "llvm.memcpy.i32" || F->getName() == "llvm.memcpy.i64" ||
651 F->getName() == "llvm.memmove.i32" ||F->getName() == "llvm.memmove.i64" ||
652 F->getName() == "memmove") {
653 // Note: this is a poor approximation, this says Dest = Src, instead of
654 // *Dest = *Src.
655 Constraints.push_back(Constraint(Constraint::Copy,
656 getNode(CS.getArgument(0)),
657 getNode(CS.getArgument(1))));
658 return true;
661 // Result = Arg0
662 if (F->getName() == "realloc" || F->getName() == "strchr" ||
663 F->getName() == "strrchr" || F->getName() == "strstr" ||
664 F->getName() == "strtok") {
665 Constraints.push_back(Constraint(Constraint::Copy,
666 getNode(CS.getInstruction()),
667 getNode(CS.getArgument(0))));
668 return true;
671 return false;
676 /// CollectConstraints - This stage scans the program, adding a constraint to
677 /// the Constraints list for each instruction in the program that induces a
678 /// constraint, and setting up the initial points-to graph.
680 void Andersens::CollectConstraints(Module &M) {
681 // First, the universal set points to itself.
682 GraphNodes[UniversalSet].addPointerTo(&GraphNodes[UniversalSet]);
683 //Constraints.push_back(Constraint(Constraint::Load, &GraphNodes[UniversalSet],
684 // &GraphNodes[UniversalSet]));
685 Constraints.push_back(Constraint(Constraint::Store, &GraphNodes[UniversalSet],
686 &GraphNodes[UniversalSet]));
688 // Next, the null pointer points to the null object.
689 GraphNodes[NullPtr].addPointerTo(&GraphNodes[NullObject]);
691 // Next, add any constraints on global variables and their initializers.
692 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
693 I != E; ++I) {
694 // Associate the address of the global object as pointing to the memory for
695 // the global: &G = <G memory>
696 Node *Object = getObject(I);
697 Object->setValue(I);
698 getNodeValue(*I)->addPointerTo(Object);
700 if (I->hasInitializer()) {
701 AddGlobalInitializerConstraints(Object, I->getInitializer());
702 } else {
703 // If it doesn't have an initializer (i.e. it's defined in another
704 // translation unit), it points to the universal set.
705 Constraints.push_back(Constraint(Constraint::Copy, Object,
706 &GraphNodes[UniversalSet]));
710 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
711 // Make the function address point to the function object.
712 getNodeValue(*F)->addPointerTo(getObject(F)->setValue(F));
714 // Set up the return value node.
715 if (isa<PointerType>(F->getFunctionType()->getReturnType()))
716 getReturnNode(F)->setValue(F);
717 if (F->getFunctionType()->isVarArg())
718 getVarargNode(F)->setValue(F);
720 // Set up incoming argument nodes.
721 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
722 I != E; ++I)
723 if (isa<PointerType>(I->getType()))
724 getNodeValue(*I);
726 if (!F->hasInternalLinkage())
727 AddConstraintsForNonInternalLinkage(F);
729 if (!F->isExternal()) {
730 // Scan the function body, creating a memory object for each heap/stack
731 // allocation in the body of the function and a node to represent all
732 // pointer values defined by instructions and used as operands.
733 visit(F);
734 } else {
735 // External functions that return pointers return the universal set.
736 if (isa<PointerType>(F->getFunctionType()->getReturnType()))
737 Constraints.push_back(Constraint(Constraint::Copy,
738 getReturnNode(F),
739 &GraphNodes[UniversalSet]));
741 // Any pointers that are passed into the function have the universal set
742 // stored into them.
743 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
744 I != E; ++I)
745 if (isa<PointerType>(I->getType())) {
746 // Pointers passed into external functions could have anything stored
747 // through them.
748 Constraints.push_back(Constraint(Constraint::Store, getNode(I),
749 &GraphNodes[UniversalSet]));
750 // Memory objects passed into external function calls can have the
751 // universal set point to them.
752 Constraints.push_back(Constraint(Constraint::Copy,
753 &GraphNodes[UniversalSet],
754 getNode(I)));
757 // If this is an external varargs function, it can also store pointers
758 // into any pointers passed through the varargs section.
759 if (F->getFunctionType()->isVarArg())
760 Constraints.push_back(Constraint(Constraint::Store, getVarargNode(F),
761 &GraphNodes[UniversalSet]));
764 NumConstraints += Constraints.size();
768 void Andersens::visitInstruction(Instruction &I) {
769 #ifdef NDEBUG
770 return; // This function is just a big assert.
771 #endif
772 if (isa<BinaryOperator>(I))
773 return;
774 // Most instructions don't have any effect on pointer values.
775 switch (I.getOpcode()) {
776 case Instruction::Br:
777 case Instruction::Switch:
778 case Instruction::Unwind:
779 case Instruction::Unreachable:
780 case Instruction::Free:
781 case Instruction::Shl:
782 case Instruction::Shr:
783 return;
784 default:
785 // Is this something we aren't handling yet?
786 std::cerr << "Unknown instruction: " << I;
787 abort();
791 void Andersens::visitAllocationInst(AllocationInst &AI) {
792 getNodeValue(AI)->addPointerTo(getObject(&AI)->setValue(&AI));
795 void Andersens::visitReturnInst(ReturnInst &RI) {
796 if (RI.getNumOperands() && isa<PointerType>(RI.getOperand(0)->getType()))
797 // return V --> <Copy/retval{F}/v>
798 Constraints.push_back(Constraint(Constraint::Copy,
799 getReturnNode(RI.getParent()->getParent()),
800 getNode(RI.getOperand(0))));
803 void Andersens::visitLoadInst(LoadInst &LI) {
804 if (isa<PointerType>(LI.getType()))
805 // P1 = load P2 --> <Load/P1/P2>
806 Constraints.push_back(Constraint(Constraint::Load, getNodeValue(LI),
807 getNode(LI.getOperand(0))));
810 void Andersens::visitStoreInst(StoreInst &SI) {
811 if (isa<PointerType>(SI.getOperand(0)->getType()))
812 // store P1, P2 --> <Store/P2/P1>
813 Constraints.push_back(Constraint(Constraint::Store,
814 getNode(SI.getOperand(1)),
815 getNode(SI.getOperand(0))));
818 void Andersens::visitGetElementPtrInst(GetElementPtrInst &GEP) {
819 // P1 = getelementptr P2, ... --> <Copy/P1/P2>
820 Constraints.push_back(Constraint(Constraint::Copy, getNodeValue(GEP),
821 getNode(GEP.getOperand(0))));
824 void Andersens::visitPHINode(PHINode &PN) {
825 if (isa<PointerType>(PN.getType())) {
826 Node *PNN = getNodeValue(PN);
827 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
828 // P1 = phi P2, P3 --> <Copy/P1/P2>, <Copy/P1/P3>, ...
829 Constraints.push_back(Constraint(Constraint::Copy, PNN,
830 getNode(PN.getIncomingValue(i))));
834 void Andersens::visitCastInst(CastInst &CI) {
835 Value *Op = CI.getOperand(0);
836 if (isa<PointerType>(CI.getType())) {
837 if (isa<PointerType>(Op->getType())) {
838 // P1 = cast P2 --> <Copy/P1/P2>
839 Constraints.push_back(Constraint(Constraint::Copy, getNodeValue(CI),
840 getNode(CI.getOperand(0))));
841 } else {
842 // P1 = cast int --> <Copy/P1/Univ>
843 #if 0
844 Constraints.push_back(Constraint(Constraint::Copy, getNodeValue(CI),
845 &GraphNodes[UniversalSet]));
846 #else
847 getNodeValue(CI);
848 #endif
850 } else if (isa<PointerType>(Op->getType())) {
851 // int = cast P1 --> <Copy/Univ/P1>
852 #if 0
853 Constraints.push_back(Constraint(Constraint::Copy,
854 &GraphNodes[UniversalSet],
855 getNode(CI.getOperand(0))));
856 #else
857 getNode(CI.getOperand(0));
858 #endif
862 void Andersens::visitSelectInst(SelectInst &SI) {
863 if (isa<PointerType>(SI.getType())) {
864 Node *SIN = getNodeValue(SI);
865 // P1 = select C, P2, P3 ---> <Copy/P1/P2>, <Copy/P1/P3>
866 Constraints.push_back(Constraint(Constraint::Copy, SIN,
867 getNode(SI.getOperand(1))));
868 Constraints.push_back(Constraint(Constraint::Copy, SIN,
869 getNode(SI.getOperand(2))));
873 void Andersens::visitVAArg(VAArgInst &I) {
874 assert(0 && "vaarg not handled yet!");
877 /// AddConstraintsForCall - Add constraints for a call with actual arguments
878 /// specified by CS to the function specified by F. Note that the types of
879 /// arguments might not match up in the case where this is an indirect call and
880 /// the function pointer has been casted. If this is the case, do something
881 /// reasonable.
882 void Andersens::AddConstraintsForCall(CallSite CS, Function *F) {
883 // If this is a call to an external function, handle it directly to get some
884 // taste of context sensitivity.
885 if (F->isExternal() && AddConstraintsForExternalCall(CS, F))
886 return;
888 if (isa<PointerType>(CS.getType())) {
889 Node *CSN = getNode(CS.getInstruction());
890 if (isa<PointerType>(F->getFunctionType()->getReturnType())) {
891 Constraints.push_back(Constraint(Constraint::Copy, CSN,
892 getReturnNode(F)));
893 } else {
894 // If the function returns a non-pointer value, handle this just like we
895 // treat a nonpointer cast to pointer.
896 Constraints.push_back(Constraint(Constraint::Copy, CSN,
897 &GraphNodes[UniversalSet]));
899 } else if (isa<PointerType>(F->getFunctionType()->getReturnType())) {
900 Constraints.push_back(Constraint(Constraint::Copy,
901 &GraphNodes[UniversalSet],
902 getReturnNode(F)));
905 Function::arg_iterator AI = F->arg_begin(), AE = F->arg_end();
906 CallSite::arg_iterator ArgI = CS.arg_begin(), ArgE = CS.arg_end();
907 for (; AI != AE && ArgI != ArgE; ++AI, ++ArgI)
908 if (isa<PointerType>(AI->getType())) {
909 if (isa<PointerType>((*ArgI)->getType())) {
910 // Copy the actual argument into the formal argument.
911 Constraints.push_back(Constraint(Constraint::Copy, getNode(AI),
912 getNode(*ArgI)));
913 } else {
914 Constraints.push_back(Constraint(Constraint::Copy, getNode(AI),
915 &GraphNodes[UniversalSet]));
917 } else if (isa<PointerType>((*ArgI)->getType())) {
918 Constraints.push_back(Constraint(Constraint::Copy,
919 &GraphNodes[UniversalSet],
920 getNode(*ArgI)));
923 // Copy all pointers passed through the varargs section to the varargs node.
924 if (F->getFunctionType()->isVarArg())
925 for (; ArgI != ArgE; ++ArgI)
926 if (isa<PointerType>((*ArgI)->getType()))
927 Constraints.push_back(Constraint(Constraint::Copy, getVarargNode(F),
928 getNode(*ArgI)));
929 // If more arguments are passed in than we track, just drop them on the floor.
932 void Andersens::visitCallSite(CallSite CS) {
933 if (isa<PointerType>(CS.getType()))
934 getNodeValue(*CS.getInstruction());
936 if (Function *F = CS.getCalledFunction()) {
937 AddConstraintsForCall(CS, F);
938 } else {
939 // We don't handle indirect call sites yet. Keep track of them for when we
940 // discover the call graph incrementally.
941 IndirectCalls.push_back(CS);
945 //===----------------------------------------------------------------------===//
946 // Constraint Solving Phase
947 //===----------------------------------------------------------------------===//
949 /// intersects - Return true if the points-to set of this node intersects
950 /// with the points-to set of the specified node.
951 bool Andersens::Node::intersects(Node *N) const {
952 iterator I1 = begin(), I2 = N->begin(), E1 = end(), E2 = N->end();
953 while (I1 != E1 && I2 != E2) {
954 if (*I1 == *I2) return true;
955 if (*I1 < *I2)
956 ++I1;
957 else
958 ++I2;
960 return false;
963 /// intersectsIgnoring - Return true if the points-to set of this node
964 /// intersects with the points-to set of the specified node on any nodes
965 /// except for the specified node to ignore.
966 bool Andersens::Node::intersectsIgnoring(Node *N, Node *Ignoring) const {
967 iterator I1 = begin(), I2 = N->begin(), E1 = end(), E2 = N->end();
968 while (I1 != E1 && I2 != E2) {
969 if (*I1 == *I2) {
970 if (*I1 != Ignoring) return true;
971 ++I1; ++I2;
972 } else if (*I1 < *I2)
973 ++I1;
974 else
975 ++I2;
977 return false;
980 // Copy constraint: all edges out of the source node get copied to the
981 // destination node. This returns true if a change is made.
982 bool Andersens::Node::copyFrom(Node *N) {
983 // Use a mostly linear-time merge since both of the lists are sorted.
984 bool Changed = false;
985 iterator I = N->begin(), E = N->end();
986 unsigned i = 0;
987 while (I != E && i != Pointees.size()) {
988 if (Pointees[i] < *I) {
989 ++i;
990 } else if (Pointees[i] == *I) {
991 ++i; ++I;
992 } else {
993 // We found a new element to copy over.
994 Changed = true;
995 Pointees.insert(Pointees.begin()+i, *I);
996 ++i; ++I;
1000 if (I != E) {
1001 Pointees.insert(Pointees.end(), I, E);
1002 Changed = true;
1005 return Changed;
1008 bool Andersens::Node::loadFrom(Node *N) {
1009 bool Changed = false;
1010 for (iterator I = N->begin(), E = N->end(); I != E; ++I)
1011 Changed |= copyFrom(*I);
1012 return Changed;
1015 bool Andersens::Node::storeThrough(Node *N) {
1016 bool Changed = false;
1017 for (iterator I = begin(), E = end(); I != E; ++I)
1018 Changed |= (*I)->copyFrom(N);
1019 return Changed;
1023 /// SolveConstraints - This stage iteratively processes the constraints list
1024 /// propagating constraints (adding edges to the Nodes in the points-to graph)
1025 /// until a fixed point is reached.
1027 void Andersens::SolveConstraints() {
1028 bool Changed = true;
1029 unsigned Iteration = 0;
1030 while (Changed) {
1031 Changed = false;
1032 ++NumIters;
1033 DEBUG(std::cerr << "Starting iteration #" << Iteration++ << "!\n");
1035 // Loop over all of the constraints, applying them in turn.
1036 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
1037 Constraint &C = Constraints[i];
1038 switch (C.Type) {
1039 case Constraint::Copy:
1040 Changed |= C.Dest->copyFrom(C.Src);
1041 break;
1042 case Constraint::Load:
1043 Changed |= C.Dest->loadFrom(C.Src);
1044 break;
1045 case Constraint::Store:
1046 Changed |= C.Dest->storeThrough(C.Src);
1047 break;
1048 default:
1049 assert(0 && "Unknown constraint!");
1053 if (Changed) {
1054 // Check to see if any internal function's addresses have been passed to
1055 // external functions. If so, we have to assume that their incoming
1056 // arguments could be anything. If there are any internal functions in
1057 // the universal node that we don't know about, we must iterate.
1058 for (Node::iterator I = GraphNodes[UniversalSet].begin(),
1059 E = GraphNodes[UniversalSet].end(); I != E; ++I)
1060 if (Function *F = dyn_cast_or_null<Function>((*I)->getValue()))
1061 if (F->hasInternalLinkage() &&
1062 EscapingInternalFunctions.insert(F).second) {
1063 // We found a function that is just now escaping. Mark it as if it
1064 // didn't have internal linkage.
1065 AddConstraintsForNonInternalLinkage(F);
1066 DEBUG(std::cerr << "Found escaping internal function: "
1067 << 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 DEBUG(std::cerr << "Found actual callee '"
1086 << F->getName() << "' for call: "
1087 << *CS.getInstruction() << "\n");
1088 ++NumIndirectCallees;
1089 KnownCallees.insert(IP, F);
1099 //===----------------------------------------------------------------------===//
1100 // Debugging Output
1101 //===----------------------------------------------------------------------===//
1103 void Andersens::PrintNode(Node *N) {
1104 if (N == &GraphNodes[UniversalSet]) {
1105 std::cerr << "<universal>";
1106 return;
1107 } else if (N == &GraphNodes[NullPtr]) {
1108 std::cerr << "<nullptr>";
1109 return;
1110 } else if (N == &GraphNodes[NullObject]) {
1111 std::cerr << "<null>";
1112 return;
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 std::cerr << F->getName() << ":retval";
1121 return;
1122 } else if (F->getFunctionType()->isVarArg() && N == getVarargNode(F)) {
1123 std::cerr << F->getName() << ":vararg";
1124 return;
1128 if (Instruction *I = dyn_cast<Instruction>(V))
1129 std::cerr << I->getParent()->getParent()->getName() << ":";
1130 else if (Argument *Arg = dyn_cast<Argument>(V))
1131 std::cerr << Arg->getParent()->getName() << ":";
1133 if (V->hasName())
1134 std::cerr << V->getName();
1135 else
1136 std::cerr << "(unnamed)";
1138 if (isa<GlobalValue>(V) || isa<AllocationInst>(V))
1139 if (N == getObject(V))
1140 std::cerr << "<mem>";
1143 void Andersens::PrintConstraints() {
1144 std::cerr << "Constraints:\n";
1145 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
1146 std::cerr << " #" << i << ": ";
1147 Constraint &C = Constraints[i];
1148 if (C.Type == Constraint::Store)
1149 std::cerr << "*";
1150 PrintNode(C.Dest);
1151 std::cerr << " = ";
1152 if (C.Type == Constraint::Load)
1153 std::cerr << "*";
1154 PrintNode(C.Src);
1155 std::cerr << "\n";
1159 void Andersens::PrintPointsToGraph() {
1160 std::cerr << "Points-to graph:\n";
1161 for (unsigned i = 0, e = GraphNodes.size(); i != e; ++i) {
1162 Node *N = &GraphNodes[i];
1163 std::cerr << "[" << (N->end() - N->begin()) << "] ";
1164 PrintNode(N);
1165 std::cerr << "\t--> ";
1166 for (Node::iterator I = N->begin(), E = N->end(); I != E; ++I) {
1167 if (I != N->begin()) std::cerr << ", ";
1168 PrintNode(*I);
1170 std::cerr << "\n";