1 //===- Local.cpp - Compute a local data structure graph for a function ----===//
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 // Compute the local version of the data structure graph for a function. The
11 // external interface to this file is the DSGraph constructor.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Analysis/DataStructure/DataStructure.h"
16 #include "llvm/Analysis/DataStructure/DSGraph.h"
17 #include "llvm/Constants.h"
18 #include "llvm/DerivedTypes.h"
19 #include "llvm/Instructions.h"
20 #include "llvm/Intrinsics.h"
21 #include "llvm/Support/GetElementPtrTypeIterator.h"
22 #include "llvm/Support/InstVisitor.h"
23 #include "llvm/Target/TargetData.h"
24 #include "Support/CommandLine.h"
25 #include "Support/Debug.h"
26 #include "Support/Timer.h"
28 // FIXME: This should eventually be a FunctionPass that is automatically
29 // aggregated into a Pass.
31 #include "llvm/Module.h"
35 static RegisterAnalysis<LocalDataStructures>
36 X("datastructure", "Local Data Structure Analysis");
39 TrackIntegersAsPointers("dsa-track-integers", cl::Hidden,
40 cl::desc("If this is set, track integers as potential pointers"));
44 // isPointerType - Return true if this type is big enough to hold a pointer.
45 bool isPointerType(const Type *Ty) {
46 if (isa<PointerType>(Ty))
48 else if (TrackIntegersAsPointers && Ty->isPrimitiveType() &&Ty->isInteger())
49 return Ty->getPrimitiveSize() >= PointerSize;
58 DisableDirectCallOpt("disable-direct-call-dsopt", cl::Hidden,
59 cl::desc("Disable direct call optimization in "
60 "DSGraph construction"));
62 DisableFieldSensitivity("disable-ds-field-sensitivity", cl::Hidden,
63 cl::desc("Disable field sensitivity in DSGraphs"));
65 //===--------------------------------------------------------------------===//
67 //===--------------------------------------------------------------------===//
69 /// This class is the builder class that constructs the local data structure
70 /// graph by performing a single pass over the function in question.
72 class GraphBuilder : InstVisitor<GraphBuilder> {
74 DSNodeHandle *RetNode; // Node that gets returned...
75 DSScalarMap &ScalarMap;
76 std::vector<DSCallSite> *FunctionCalls;
79 GraphBuilder(Function &f, DSGraph &g, DSNodeHandle &retNode,
80 std::vector<DSCallSite> &fc)
81 : G(g), RetNode(&retNode), ScalarMap(G.getScalarMap()),
84 // Create scalar nodes for all pointer arguments...
85 for (Function::aiterator I = f.abegin(), E = f.aend(); I != E; ++I)
86 if (isPointerType(I->getType()))
89 visit(f); // Single pass over the function
92 // GraphBuilder ctor for working on the globals graph
93 GraphBuilder(DSGraph &g)
94 : G(g), RetNode(0), ScalarMap(G.getScalarMap()), FunctionCalls(0) {
97 void mergeInGlobalInitializer(GlobalVariable *GV);
100 // Visitor functions, used to handle each instruction type we encounter...
101 friend class InstVisitor<GraphBuilder>;
102 void visitMallocInst(MallocInst &MI) { handleAlloc(MI, true); }
103 void visitAllocaInst(AllocaInst &AI) { handleAlloc(AI, false); }
104 void handleAlloc(AllocationInst &AI, bool isHeap);
106 void visitPHINode(PHINode &PN);
108 void visitGetElementPtrInst(User &GEP);
109 void visitReturnInst(ReturnInst &RI);
110 void visitLoadInst(LoadInst &LI);
111 void visitStoreInst(StoreInst &SI);
112 void visitCallInst(CallInst &CI);
113 void visitInvokeInst(InvokeInst &II);
114 void visitSetCondInst(SetCondInst &SCI) {} // SetEQ & friends are ignored
115 void visitFreeInst(FreeInst &FI);
116 void visitCastInst(CastInst &CI);
117 void visitInstruction(Instruction &I);
119 void visitCallSite(CallSite CS);
120 void visitVANextInst(VANextInst &I);
121 void visitVAArgInst(VAArgInst &I);
123 void MergeConstantInitIntoNode(DSNodeHandle &NH, Constant *C);
125 // Helper functions used to implement the visitation functions...
127 /// createNode - Create a new DSNode, ensuring that it is properly added to
130 DSNode *createNode(const Type *Ty = 0) {
131 DSNode *N = new DSNode(Ty, &G); // Create the node
132 if (DisableFieldSensitivity) {
133 // Create node handle referring to the old node so that it is
134 // immediately removed from the graph when the node handle is destroyed.
135 DSNodeHandle OldNNH = N;
136 N->foldNodeCompletely();
137 if (DSNode *FN = N->getForwardNode())
143 /// setDestTo - Set the ScalarMap entry for the specified value to point to
144 /// the specified destination. If the Value already points to a node, make
145 /// sure to merge the two destinations together.
147 void setDestTo(Value &V, const DSNodeHandle &NH);
149 /// getValueDest - Return the DSNode that the actual value points to.
151 DSNodeHandle getValueDest(Value &V);
153 /// getLink - This method is used to return the specified link in the
154 /// specified node if one exists. If a link does not already exist (it's
155 /// null), then we create a new node, link it, then return it.
157 DSNodeHandle &getLink(const DSNodeHandle &Node, unsigned Link = 0);
163 //===----------------------------------------------------------------------===//
164 // DSGraph constructor - Simply use the GraphBuilder to construct the local
166 DSGraph::DSGraph(const TargetData &td, Function &F, DSGraph *GG)
167 : GlobalsGraph(GG), TD(td) {
168 PrintAuxCalls = false;
170 DEBUG(std::cerr << " [Loc] Calculating graph for: " << F.getName() << "\n");
172 // Use the graph builder to construct the local version of the graph
173 GraphBuilder B(F, *this, ReturnNodes[&F], FunctionCalls);
175 Timer::addPeakMemoryMeasurement();
178 // Remove all integral constants from the scalarmap!
179 for (DSScalarMap::iterator I = ScalarMap.begin(); I != ScalarMap.end();)
180 if (isa<ConstantIntegral>(I->first))
181 ScalarMap.erase(I++);
185 // If there are any constant globals referenced in this function, merge their
186 // initializers into the local graph from the globals graph.
187 if (ScalarMap.global_begin() != ScalarMap.global_end()) {
188 ReachabilityCloner RC(*this, *GG, 0);
190 for (DSScalarMap::global_iterator I = ScalarMap.global_begin();
191 I != ScalarMap.global_end(); ++I)
192 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(*I))
193 if (!GV->isExternal() && GV->isConstant())
194 RC.merge(ScalarMap[GV], GG->ScalarMap[GV]);
197 markIncompleteNodes(DSGraph::MarkFormalArgs);
199 // Remove any nodes made dead due to merging...
200 removeDeadNodes(DSGraph::KeepUnreachableGlobals);
204 //===----------------------------------------------------------------------===//
205 // Helper method implementations...
208 /// getValueDest - Return the DSNode that the actual value points to.
210 DSNodeHandle GraphBuilder::getValueDest(Value &Val) {
212 if (V == Constant::getNullValue(V->getType()))
213 return 0; // Null doesn't point to anything, don't add to ScalarMap!
215 DSNodeHandle &NH = ScalarMap[V];
217 return NH; // Already have a node? Just return it...
219 // Otherwise we need to create a new node to point to.
220 // Check first for constant expressions that must be traversed to
221 // extract the actual value.
223 if (GlobalValue* GV = dyn_cast<GlobalValue>(V)) {
224 // Create a new global node for this global variable...
225 N = createNode(GV->getType()->getElementType());
227 } else if (Constant *C = dyn_cast<Constant>(V)) {
228 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
229 if (CE->getOpcode() == Instruction::Cast)
230 NH = getValueDest(*CE->getOperand(0));
231 else if (CE->getOpcode() == Instruction::GetElementPtr) {
232 visitGetElementPtrInst(*CE);
233 DSScalarMap::iterator I = ScalarMap.find(CE);
234 assert(I != ScalarMap.end() && "GEP didn't get processed right?");
237 // This returns a conservative unknown node for any unhandled ConstExpr
238 return NH = createNode()->setUnknownNodeMarker();
240 if (NH.getNode() == 0) { // (getelementptr null, X) returns null
246 } else if (ConstantIntegral *CI = dyn_cast<ConstantIntegral>(C)) {
247 // Random constants are unknown mem
248 return NH = createNode()->setUnknownNodeMarker();
250 assert(0 && "Unknown constant type!");
252 N = createNode(); // just create a shadow node
254 // Otherwise just create a shadow node
258 NH.setTo(N, 0); // Remember that we are pointing to it...
263 /// getLink - This method is used to return the specified link in the
264 /// specified node if one exists. If a link does not already exist (it's
265 /// null), then we create a new node, link it, then return it. We must
266 /// specify the type of the Node field we are accessing so that we know what
267 /// type should be linked to if we need to create a new node.
269 DSNodeHandle &GraphBuilder::getLink(const DSNodeHandle &node, unsigned LinkNo) {
270 DSNodeHandle &Node = const_cast<DSNodeHandle&>(node);
271 DSNodeHandle &Link = Node.getLink(LinkNo);
272 if (!Link.getNode()) {
273 // If the link hasn't been created yet, make and return a new shadow node
280 /// setDestTo - Set the ScalarMap entry for the specified value to point to the
281 /// specified destination. If the Value already points to a node, make sure to
282 /// merge the two destinations together.
284 void GraphBuilder::setDestTo(Value &V, const DSNodeHandle &NH) {
285 ScalarMap[&V].mergeWith(NH);
289 //===----------------------------------------------------------------------===//
290 // Specific instruction type handler implementations...
293 /// Alloca & Malloc instruction implementation - Simply create a new memory
294 /// object, pointing the scalar to it.
296 void GraphBuilder::handleAlloc(AllocationInst &AI, bool isHeap) {
297 DSNode *N = createNode();
299 N->setHeapNodeMarker();
301 N->setAllocaNodeMarker();
305 // PHINode - Make the scalar for the PHI node point to all of the things the
306 // incoming values point to... which effectively causes them to be merged.
308 void GraphBuilder::visitPHINode(PHINode &PN) {
309 if (!isPointerType(PN.getType())) return; // Only pointer PHIs
311 DSNodeHandle &PNDest = ScalarMap[&PN];
312 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
313 PNDest.mergeWith(getValueDest(*PN.getIncomingValue(i)));
316 void GraphBuilder::visitGetElementPtrInst(User &GEP) {
317 DSNodeHandle Value = getValueDest(*GEP.getOperand(0));
318 if (Value.getNode() == 0) return;
320 // As a special case, if all of the index operands of GEP are constant zeros,
321 // handle this just like we handle casts (ie, don't do much).
322 bool AllZeros = true;
323 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i)
324 if (GEP.getOperand(i) !=
325 Constant::getNullValue(GEP.getOperand(i)->getType())) {
330 // If all of the indices are zero, the result points to the operand without
331 // applying the type.
333 setDestTo(GEP, Value);
338 const PointerType *PTy = cast<PointerType>(GEP.getOperand(0)->getType());
339 const Type *CurTy = PTy->getElementType();
341 if (Value.getNode()->mergeTypeInfo(CurTy, Value.getOffset())) {
342 // If the node had to be folded... exit quickly
343 setDestTo(GEP, Value); // GEP result points to folded node
347 const TargetData &TD = Value.getNode()->getTargetData();
350 // Handle the pointer index specially...
351 if (GEP.getNumOperands() > 1 &&
352 (!isa<Constant>(GEP.getOperand(1)) ||
353 !cast<Constant>(GEP.getOperand(1))->isNullValue())) {
355 // If we already know this is an array being accessed, don't do anything...
356 if (!TopTypeRec.isArray) {
357 TopTypeRec.isArray = true;
359 // If we are treating some inner field pointer as an array, fold the node
360 // up because we cannot handle it right. This can come because of
361 // something like this: &((&Pt->X)[1]) == &Pt->Y
363 if (Value.getOffset()) {
364 // Value is now the pointer we want to GEP to be...
365 Value.getNode()->foldNodeCompletely();
366 setDestTo(GEP, Value); // GEP result points to folded node
369 // This is a pointer to the first byte of the node. Make sure that we
370 // are pointing to the outter most type in the node.
371 // FIXME: We need to check one more case here...
377 // All of these subscripts are indexing INTO the elements we have...
379 for (gep_type_iterator I = gep_type_begin(GEP), E = gep_type_end(GEP);
381 if (const StructType *STy = dyn_cast<StructType>(*I)) {
382 unsigned FieldNo = cast<ConstantUInt>(I.getOperand())->getValue();
383 Offset += TD.getStructLayout(STy)->MemberOffsets[FieldNo];
384 } else if (const PointerType *PTy = dyn_cast<PointerType>(*I)) {
385 if (!isa<Constant>(I.getOperand()) ||
386 !cast<Constant>(I.getOperand())->isNullValue())
387 Value.getNode()->setArrayMarker();
392 if (const SequentialType *STy = cast<SequentialType>(*I)) {
393 CurTy = STy->getElementType();
394 if (ConstantSInt *CS = dyn_cast<ConstantSInt>(GEP.getOperand(i))) {
395 Offset += CS->getValue()*TD.getTypeSize(CurTy);
397 // Variable index into a node. We must merge all of the elements of the
398 // sequential type here.
399 if (isa<PointerType>(STy))
400 std::cerr << "Pointer indexing not handled yet!\n";
402 const ArrayType *ATy = cast<ArrayType>(STy);
403 unsigned ElSize = TD.getTypeSize(CurTy);
404 DSNode *N = Value.getNode();
405 assert(N && "Value must have a node!");
406 unsigned RawOffset = Offset+Value.getOffset();
408 // Loop over all of the elements of the array, merging them into the
410 for (unsigned i = 1, e = ATy->getNumElements(); i != e; ++i)
411 // Merge all of the byte components of this array element
412 for (unsigned j = 0; j != ElSize; ++j)
413 N->mergeIndexes(RawOffset+j, RawOffset+i*ElSize+j);
419 // Add in the offset calculated...
420 Value.setOffset(Value.getOffset()+Offset);
422 // Value is now the pointer we want to GEP to be...
423 setDestTo(GEP, Value);
426 void GraphBuilder::visitLoadInst(LoadInst &LI) {
427 DSNodeHandle Ptr = getValueDest(*LI.getOperand(0));
428 if (Ptr.getNode() == 0) return;
430 // Make that the node is read from...
431 Ptr.getNode()->setReadMarker();
433 // Ensure a typerecord exists...
434 Ptr.getNode()->mergeTypeInfo(LI.getType(), Ptr.getOffset(), false);
436 if (isPointerType(LI.getType()))
437 setDestTo(LI, getLink(Ptr));
440 void GraphBuilder::visitStoreInst(StoreInst &SI) {
441 const Type *StoredTy = SI.getOperand(0)->getType();
442 DSNodeHandle Dest = getValueDest(*SI.getOperand(1));
443 if (Dest.isNull()) return;
445 // Mark that the node is written to...
446 Dest.getNode()->setModifiedMarker();
448 // Ensure a type-record exists...
449 Dest.getNode()->mergeTypeInfo(StoredTy, Dest.getOffset());
451 // Avoid adding edges from null, or processing non-"pointer" stores
452 if (isPointerType(StoredTy))
453 Dest.addEdgeTo(getValueDest(*SI.getOperand(0)));
456 void GraphBuilder::visitReturnInst(ReturnInst &RI) {
457 if (RI.getNumOperands() && isPointerType(RI.getOperand(0)->getType()))
458 RetNode->mergeWith(getValueDest(*RI.getOperand(0)));
461 void GraphBuilder::visitVANextInst(VANextInst &I) {
462 getValueDest(*I.getOperand(0)).mergeWith(getValueDest(I));
465 void GraphBuilder::visitVAArgInst(VAArgInst &I) {
466 DSNodeHandle Ptr = getValueDest(*I.getOperand(0));
467 if (Ptr.isNull()) return;
469 // Make that the node is read from.
470 Ptr.getNode()->setReadMarker();
472 // Ensure a typerecord exists...
473 Ptr.getNode()->mergeTypeInfo(I.getType(), Ptr.getOffset(), false);
475 if (isPointerType(I.getType()))
476 setDestTo(I, getLink(Ptr));
480 void GraphBuilder::visitCallInst(CallInst &CI) {
484 void GraphBuilder::visitInvokeInst(InvokeInst &II) {
488 void GraphBuilder::visitCallSite(CallSite CS) {
489 Value *Callee = CS.getCalledValue();
491 // Special case handling of certain libc allocation functions here.
492 if (Function *F = dyn_cast<Function>(Callee))
494 switch (F->getIntrinsicID()) {
495 case Intrinsic::vastart:
496 getValueDest(*CS.getInstruction()).getNode()->setAllocaNodeMarker();
498 case Intrinsic::vacopy:
499 getValueDest(*CS.getInstruction()).
500 mergeWith(getValueDest(**(CS.arg_begin())));
502 case Intrinsic::vaend:
504 case Intrinsic::memmove:
505 case Intrinsic::memcpy: {
506 // Merge the first & second arguments, and mark the memory read and
508 DSNodeHandle RetNH = getValueDest(**CS.arg_begin());
509 RetNH.mergeWith(getValueDest(**(CS.arg_begin()+1)));
510 if (DSNode *N = RetNH.getNode())
511 N->setModifiedMarker()->setReadMarker();
514 case Intrinsic::memset:
515 // Mark the memory modified.
516 if (DSNode *N = getValueDest(**CS.arg_begin()).getNode())
517 N->setModifiedMarker();
520 if (F->getName() == "calloc" || F->getName() == "posix_memalign" ||
521 F->getName() == "memalign" || F->getName() == "valloc") {
522 setDestTo(*CS.getInstruction(),
523 createNode()->setHeapNodeMarker()->setModifiedMarker());
525 } else if (F->getName() == "realloc") {
526 DSNodeHandle RetNH = getValueDest(*CS.getInstruction());
527 RetNH.mergeWith(getValueDest(**CS.arg_begin()));
528 if (DSNode *N = RetNH.getNode())
529 N->setHeapNodeMarker()->setModifiedMarker()->setReadMarker();
531 } else if (F->getName() == "memmove") {
532 // Merge the first & second arguments, and mark the memory read and
534 DSNodeHandle RetNH = getValueDest(**CS.arg_begin());
535 RetNH.mergeWith(getValueDest(**(CS.arg_begin()+1)));
536 if (DSNode *N = RetNH.getNode())
537 N->setModifiedMarker()->setReadMarker();
540 } else if (F->getName() == "atoi" || F->getName() == "atof" ||
541 F->getName() == "atol" || F->getName() == "atoll" ||
542 F->getName() == "remove" || F->getName() == "unlink" ||
543 F->getName() == "rename" || F->getName() == "memcmp" ||
544 F->getName() == "strcmp" || F->getName() == "strncmp" ||
545 F->getName() == "execl" || F->getName() == "execlp" ||
546 F->getName() == "execle" || F->getName() == "execv" ||
547 F->getName() == "execvp" || F->getName() == "chmod" ||
548 F->getName() == "puts" || F->getName() == "write" ||
549 F->getName() == "open" || F->getName() == "create" ||
550 F->getName() == "truncate" || F->getName() == "chdir" ||
551 F->getName() == "mkdir" || F->getName() == "rmdir") {
552 // These functions read all of their pointer operands.
553 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
555 if (isPointerType((*AI)->getType()))
556 if (DSNode *N = getValueDest(**AI).getNode())
560 } else if (F->getName() == "read" || F->getName() == "pipe" ||
561 F->getName() == "wait" || F->getName() == "time") {
562 // These functions write all of their pointer operands.
563 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
565 if (isPointerType((*AI)->getType()))
566 if (DSNode *N = getValueDest(**AI).getNode())
567 N->setModifiedMarker();
570 } else if (F->getName() == "stat" || F->getName() == "fstat" ||
571 F->getName() == "lstat") {
572 // These functions read their first operand if its a pointer.
573 CallSite::arg_iterator AI = CS.arg_begin();
574 if (isPointerType((*AI)->getType())) {
575 DSNodeHandle Path = getValueDest(**AI);
576 if (DSNode *N = Path.getNode()) N->setReadMarker();
579 // Then they write into the stat buffer.
580 DSNodeHandle StatBuf = getValueDest(**++AI);
581 if (DSNode *N = StatBuf.getNode()) {
582 N->setModifiedMarker();
583 const Type *StatTy = F->getFunctionType()->getParamType(1);
584 if (const PointerType *PTy = dyn_cast<PointerType>(StatTy))
585 N->mergeTypeInfo(PTy->getElementType(), StatBuf.getOffset());
588 } else if (F->getName() == "strtod" || F->getName() == "strtof" ||
589 F->getName() == "strtold") {
590 // These functions read the first pointer
591 if (DSNode *Str = getValueDest(**CS.arg_begin()).getNode()) {
592 Str->setReadMarker();
593 // If the second parameter is passed, it will point to the first
595 const DSNodeHandle &EndPtrNH = getValueDest(**(CS.arg_begin()+1));
596 if (DSNode *End = EndPtrNH.getNode()) {
597 End->mergeTypeInfo(PointerType::get(Type::SByteTy),
598 EndPtrNH.getOffset(), false);
599 End->setModifiedMarker();
600 DSNodeHandle &Link = getLink(EndPtrNH);
601 Link.mergeWith(getValueDest(**CS.arg_begin()));
606 } else if (F->getName() == "fopen" || F->getName() == "fdopen" ||
607 F->getName() == "freopen") {
608 // These functions read all of their pointer operands.
609 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
611 if (isPointerType((*AI)->getType()))
612 if (DSNode *N = getValueDest(**AI).getNode())
615 // fopen allocates in an unknown way and writes to the file
616 // descriptor. Also, merge the allocated type into the node.
617 DSNodeHandle Result = getValueDest(*CS.getInstruction());
618 if (DSNode *N = Result.getNode()) {
619 N->setModifiedMarker()->setUnknownNodeMarker();
620 const Type *RetTy = F->getFunctionType()->getReturnType();
621 if (const PointerType *PTy = dyn_cast<PointerType>(RetTy))
622 N->mergeTypeInfo(PTy->getElementType(), Result.getOffset());
625 // If this is freopen, merge the file descriptor passed in with the
627 if (F->getName() == "freopen")
628 Result.mergeWith(getValueDest(**--CS.arg_end()));
631 } else if (F->getName() == "fclose" && CS.arg_end()-CS.arg_begin() ==1){
632 // fclose reads and deallocates the memory in an unknown way for the
633 // file descriptor. It merges the FILE type into the descriptor.
634 DSNodeHandle H = getValueDest(**CS.arg_begin());
635 if (DSNode *N = H.getNode()) {
636 N->setReadMarker()->setUnknownNodeMarker();
637 const Type *ArgTy = F->getFunctionType()->getParamType(0);
638 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
639 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
642 } else if (CS.arg_end()-CS.arg_begin() == 1 &&
643 (F->getName() == "fflush" || F->getName() == "feof" ||
644 F->getName() == "fileno" || F->getName() == "clearerr" ||
645 F->getName() == "rewind" || F->getName() == "ftell" ||
646 F->getName() == "ferror" || F->getName() == "fgetc" ||
647 F->getName() == "fgetc" || F->getName() == "_IO_getc")) {
648 // fflush reads and writes the memory for the file descriptor. It
649 // merges the FILE type into the descriptor.
650 DSNodeHandle H = getValueDest(**CS.arg_begin());
651 if (DSNode *N = H.getNode()) {
652 N->setReadMarker()->setModifiedMarker();
654 const Type *ArgTy = F->getFunctionType()->getParamType(0);
655 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
656 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
659 } else if (CS.arg_end()-CS.arg_begin() == 4 &&
660 (F->getName() == "fwrite" || F->getName() == "fread")) {
661 // fread writes the first operand, fwrite reads it. They both
662 // read/write the FILE descriptor, and merges the FILE type.
663 DSNodeHandle H = getValueDest(**--CS.arg_end());
664 if (DSNode *N = H.getNode()) {
665 N->setReadMarker()->setModifiedMarker();
666 const Type *ArgTy = F->getFunctionType()->getParamType(3);
667 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
668 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
671 H = getValueDest(**CS.arg_begin());
672 if (DSNode *N = H.getNode())
673 if (F->getName() == "fwrite")
676 N->setModifiedMarker();
678 } else if (F->getName() == "fgets" && CS.arg_end()-CS.arg_begin() == 3){
679 // fgets reads and writes the memory for the file descriptor. It
680 // merges the FILE type into the descriptor, and writes to the
681 // argument. It returns the argument as well.
682 CallSite::arg_iterator AI = CS.arg_begin();
683 DSNodeHandle H = getValueDest(**AI);
684 if (DSNode *N = H.getNode())
685 N->setModifiedMarker(); // Writes buffer
686 H.mergeWith(getValueDest(*CS.getInstruction())); // Returns buffer
689 // Reads and writes file descriptor, merge in FILE type.
690 H = getValueDest(**AI);
691 if (DSNode *N = H.getNode()) {
692 N->setReadMarker()->setModifiedMarker();
693 const Type *ArgTy = F->getFunctionType()->getParamType(2);
694 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
695 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
698 } else if (F->getName() == "ungetc" || F->getName() == "fputc" ||
699 F->getName() == "fputs" || F->getName() == "putc" ||
700 F->getName() == "ftell" || F->getName() == "rewind" ||
701 F->getName() == "_IO_putc") {
702 // These functions read and write the memory for the file descriptor,
703 // which is passes as the last argument.
704 DSNodeHandle H = getValueDest(**--CS.arg_end());
705 if (DSNode *N = H.getNode()) {
706 N->setReadMarker()->setModifiedMarker();
707 const Type *ArgTy = *--F->getFunctionType()->param_end();
708 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
709 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
712 // Any pointer arguments are read.
713 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
715 if (isPointerType((*AI)->getType()))
716 if (DSNode *N = getValueDest(**AI).getNode())
719 } else if (F->getName() == "fseek" || F->getName() == "fgetpos" ||
720 F->getName() == "fsetpos") {
721 // These functions read and write the memory for the file descriptor,
722 // and read/write all other arguments.
723 DSNodeHandle H = getValueDest(**CS.arg_begin());
724 if (DSNode *N = H.getNode()) {
725 const Type *ArgTy = *--F->getFunctionType()->param_end();
726 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
727 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
730 // Any pointer arguments are read.
731 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
733 if (isPointerType((*AI)->getType()))
734 if (DSNode *N = getValueDest(**AI).getNode())
735 N->setReadMarker()->setModifiedMarker();
737 } else if (F->getName() == "printf" || F->getName() == "fprintf" ||
738 F->getName() == "sprintf") {
739 CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
741 if (F->getName() == "fprintf") {
742 // fprintf reads and writes the FILE argument, and applies the type
744 DSNodeHandle H = getValueDest(**AI);
745 if (DSNode *N = H.getNode()) {
746 N->setModifiedMarker();
747 const Type *ArgTy = (*AI)->getType();
748 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
749 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
751 } else if (F->getName() == "sprintf") {
752 // sprintf writes the first string argument.
753 DSNodeHandle H = getValueDest(**AI++);
754 if (DSNode *N = H.getNode()) {
755 N->setModifiedMarker();
756 const Type *ArgTy = (*AI)->getType();
757 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
758 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
762 for (; AI != E; ++AI) {
763 // printf reads all pointer arguments.
764 if (isPointerType((*AI)->getType()))
765 if (DSNode *N = getValueDest(**AI).getNode())
769 } else if (F->getName() == "vprintf" || F->getName() == "vfprintf" ||
770 F->getName() == "vsprintf") {
771 CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
773 if (F->getName() == "vfprintf") {
774 // ffprintf reads and writes the FILE argument, and applies the type
776 DSNodeHandle H = getValueDest(**AI);
777 if (DSNode *N = H.getNode()) {
778 N->setModifiedMarker()->setReadMarker();
779 const Type *ArgTy = (*AI)->getType();
780 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
781 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
784 } else if (F->getName() == "vsprintf") {
785 // vsprintf writes the first string argument.
786 DSNodeHandle H = getValueDest(**AI++);
787 if (DSNode *N = H.getNode()) {
788 N->setModifiedMarker();
789 const Type *ArgTy = (*AI)->getType();
790 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
791 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
797 if (isPointerType((*AI)->getType()))
798 if (DSNode *N = getValueDest(**AI).getNode())
803 // Read the valist, and the pointed-to objects.
804 if (AI != E && isPointerType((*AI)->getType())) {
805 const DSNodeHandle &VAList = getValueDest(**AI);
806 if (DSNode *N = VAList.getNode()) {
808 N->mergeTypeInfo(PointerType::get(Type::SByteTy),
809 VAList.getOffset(), false);
811 DSNodeHandle &VAListObjs = getLink(VAList);
812 VAListObjs.getNode()->setReadMarker();
817 } else if (F->getName() == "scanf" || F->getName() == "fscanf" ||
818 F->getName() == "sscanf") {
819 CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
821 if (F->getName() == "fscanf") {
822 // fscanf reads and writes the FILE argument, and applies the type
824 DSNodeHandle H = getValueDest(**AI);
825 if (DSNode *N = H.getNode()) {
827 const Type *ArgTy = (*AI)->getType();
828 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
829 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
831 } else if (F->getName() == "sscanf") {
832 // sscanf reads the first string argument.
833 DSNodeHandle H = getValueDest(**AI++);
834 if (DSNode *N = H.getNode()) {
836 const Type *ArgTy = (*AI)->getType();
837 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
838 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
842 for (; AI != E; ++AI) {
843 // scanf writes all pointer arguments.
844 if (isPointerType((*AI)->getType()))
845 if (DSNode *N = getValueDest(**AI).getNode())
846 N->setModifiedMarker();
849 } else if (F->getName() == "strtok") {
850 // strtok reads and writes the first argument, returning it. It reads
851 // its second arg. FIXME: strtok also modifies some hidden static
852 // data. Someday this might matter.
853 CallSite::arg_iterator AI = CS.arg_begin();
854 DSNodeHandle H = getValueDest(**AI++);
855 if (DSNode *N = H.getNode()) {
856 N->setReadMarker()->setModifiedMarker(); // Reads/Writes buffer
857 const Type *ArgTy = F->getFunctionType()->getParamType(0);
858 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
859 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
861 H.mergeWith(getValueDest(*CS.getInstruction())); // Returns buffer
863 H = getValueDest(**AI); // Reads delimiter
864 if (DSNode *N = H.getNode()) {
866 const Type *ArgTy = F->getFunctionType()->getParamType(1);
867 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
868 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
871 } else if (F->getName() == "strchr" || F->getName() == "strrchr" ||
872 F->getName() == "strstr") {
873 // These read their arguments, and return the first one
874 DSNodeHandle H = getValueDest(**CS.arg_begin());
875 H.mergeWith(getValueDest(*CS.getInstruction())); // Returns buffer
877 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
879 if (isPointerType((*AI)->getType()))
880 if (DSNode *N = getValueDest(**AI).getNode())
883 if (DSNode *N = H.getNode())
886 } else if (F->getName() == "modf" && CS.arg_end()-CS.arg_begin() == 2) {
887 // This writes its second argument, and forces it to double.
888 DSNodeHandle H = getValueDest(**--CS.arg_end());
889 if (DSNode *N = H.getNode()) {
890 N->setModifiedMarker();
891 N->mergeTypeInfo(Type::DoubleTy, H.getOffset());
895 // Unknown function, warn if it returns a pointer type or takes a
897 bool Warn = isPointerType(CS.getInstruction()->getType());
899 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
901 if (isPointerType((*I)->getType())) {
906 std::cerr << "WARNING: Call to unknown external function '"
907 << F->getName() << "' will cause pessimistic results!\n";
912 // Set up the return value...
914 Instruction *I = CS.getInstruction();
915 if (isPointerType(I->getType()))
916 RetVal = getValueDest(*I);
918 DSNode *CalleeNode = 0;
919 if (DisableDirectCallOpt || !isa<Function>(Callee)) {
920 CalleeNode = getValueDest(*Callee).getNode();
921 if (CalleeNode == 0) {
922 std::cerr << "WARNING: Program is calling through a null pointer?\n"<< *I;
923 return; // Calling a null pointer?
927 std::vector<DSNodeHandle> Args;
928 Args.reserve(CS.arg_end()-CS.arg_begin());
930 // Calculate the arguments vector...
931 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); I != E; ++I)
932 if (isPointerType((*I)->getType()))
933 Args.push_back(getValueDest(**I));
935 // Add a new function call entry...
937 FunctionCalls->push_back(DSCallSite(CS, RetVal, CalleeNode, Args));
939 FunctionCalls->push_back(DSCallSite(CS, RetVal, cast<Function>(Callee),
943 void GraphBuilder::visitFreeInst(FreeInst &FI) {
944 // Mark that the node is written to...
945 if (DSNode *N = getValueDest(*FI.getOperand(0)).getNode())
946 N->setModifiedMarker()->setHeapNodeMarker();
950 void GraphBuilder::visitCastInst(CastInst &CI) {
951 if (isPointerType(CI.getType()))
952 if (isPointerType(CI.getOperand(0)->getType())) {
953 // Cast one pointer to the other, just act like a copy instruction
954 setDestTo(CI, getValueDest(*CI.getOperand(0)));
956 // Cast something (floating point, small integer) to a pointer. We need
957 // to track the fact that the node points to SOMETHING, just something we
958 // don't know about. Make an "Unknown" node.
960 setDestTo(CI, createNode()->setUnknownNodeMarker());
965 // visitInstruction - For all other instruction types, if we have any arguments
966 // that are of pointer type, make them have unknown composition bits, and merge
967 // the nodes together.
968 void GraphBuilder::visitInstruction(Instruction &Inst) {
969 DSNodeHandle CurNode;
970 if (isPointerType(Inst.getType()))
971 CurNode = getValueDest(Inst);
972 for (User::op_iterator I = Inst.op_begin(), E = Inst.op_end(); I != E; ++I)
973 if (isPointerType((*I)->getType()))
974 CurNode.mergeWith(getValueDest(**I));
976 if (CurNode.getNode())
977 CurNode.getNode()->setUnknownNodeMarker();
982 //===----------------------------------------------------------------------===//
983 // LocalDataStructures Implementation
984 //===----------------------------------------------------------------------===//
986 // MergeConstantInitIntoNode - Merge the specified constant into the node
988 void GraphBuilder::MergeConstantInitIntoNode(DSNodeHandle &NH, Constant *C) {
989 // Ensure a type-record exists...
990 NH.getNode()->mergeTypeInfo(C->getType(), NH.getOffset());
992 if (C->getType()->isFirstClassType()) {
993 if (isPointerType(C->getType()))
994 // Avoid adding edges from null, or processing non-"pointer" stores
995 NH.addEdgeTo(getValueDest(*C));
999 const TargetData &TD = NH.getNode()->getTargetData();
1001 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) {
1002 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
1003 // We don't currently do any indexing for arrays...
1004 MergeConstantInitIntoNode(NH, cast<Constant>(CA->getOperand(i)));
1005 } else if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
1006 const StructLayout *SL = TD.getStructLayout(CS->getType());
1007 for (unsigned i = 0, e = CS->getNumOperands(); i != e; ++i) {
1008 DSNodeHandle NewNH(NH.getNode(), NH.getOffset()+SL->MemberOffsets[i]);
1009 MergeConstantInitIntoNode(NewNH, cast<Constant>(CS->getOperand(i)));
1011 } else if (ConstantAggregateZero *CAZ = dyn_cast<ConstantAggregateZero>(C)) {
1014 assert(0 && "Unknown constant type!");
1018 void GraphBuilder::mergeInGlobalInitializer(GlobalVariable *GV) {
1019 assert(!GV->isExternal() && "Cannot merge in external global!");
1020 // Get a node handle to the global node and merge the initializer into it.
1021 DSNodeHandle NH = getValueDest(*GV);
1022 MergeConstantInitIntoNode(NH, GV->getInitializer());
1026 bool LocalDataStructures::run(Module &M) {
1027 GlobalsGraph = new DSGraph(getAnalysis<TargetData>());
1029 const TargetData &TD = getAnalysis<TargetData>();
1032 GraphBuilder GGB(*GlobalsGraph);
1034 // Add initializers for all of the globals to the globals graph...
1035 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I)
1036 if (!I->isExternal())
1037 GGB.mergeInGlobalInitializer(I);
1040 // Calculate all of the graphs...
1041 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
1042 if (!I->isExternal())
1043 DSInfo.insert(std::make_pair(I, new DSGraph(TD, *I, GlobalsGraph)));
1045 GlobalsGraph->removeTriviallyDeadNodes();
1046 GlobalsGraph->markIncompleteNodes(DSGraph::MarkFormalArgs);
1050 // releaseMemory - If the pass pipeline is done with this pass, we can release
1051 // our memory... here...
1053 void LocalDataStructures::releaseMemory() {
1054 for (hash_map<Function*, DSGraph*>::iterator I = DSInfo.begin(),
1055 E = DSInfo.end(); I != E; ++I) {
1056 I->second->getReturnNodes().erase(I->first);
1057 if (I->second->getReturnNodes().empty())
1061 // Empty map so next time memory is released, data structures are not
1064 delete GlobalsGraph;