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.h"
16 #include "llvm/Analysis/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",
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 N->foldNodeCompletely();
134 if (DSNode *FN = N->getForwardNode())
140 /// setDestTo - Set the ScalarMap entry for the specified value to point to
141 /// the specified destination. If the Value already points to a node, make
142 /// sure to merge the two destinations together.
144 void setDestTo(Value &V, const DSNodeHandle &NH);
146 /// getValueDest - Return the DSNode that the actual value points to.
148 DSNodeHandle getValueDest(Value &V);
150 /// getLink - This method is used to return the specified link in the
151 /// specified node if one exists. If a link does not already exist (it's
152 /// null), then we create a new node, link it, then return it.
154 DSNodeHandle &getLink(const DSNodeHandle &Node, unsigned Link = 0);
160 //===----------------------------------------------------------------------===//
161 // DSGraph constructor - Simply use the GraphBuilder to construct the local
163 DSGraph::DSGraph(const TargetData &td, Function &F, DSGraph *GG)
164 : GlobalsGraph(GG), TD(td) {
165 PrintAuxCalls = false;
167 DEBUG(std::cerr << " [Loc] Calculating graph for: " << F.getName() << "\n");
169 // Use the graph builder to construct the local version of the graph
170 GraphBuilder B(F, *this, ReturnNodes[&F], FunctionCalls);
172 Timer::addPeakMemoryMeasurement();
175 // Remove all integral constants from the scalarmap!
176 for (DSScalarMap::iterator I = ScalarMap.begin(); I != ScalarMap.end();)
177 if (isa<ConstantIntegral>(I->first))
178 ScalarMap.erase(I++);
182 // If there are any constant globals referenced in this function, merge their
183 // initializers into the local graph from the globals graph.
184 if (ScalarMap.global_begin() != ScalarMap.global_end()) {
185 ReachabilityCloner RC(*this, *GG, 0);
187 for (DSScalarMap::global_iterator I = ScalarMap.global_begin();
188 I != ScalarMap.global_end(); ++I)
189 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(*I))
190 if (!GV->isExternal() && GV->isConstant())
191 RC.merge(ScalarMap[GV], GG->ScalarMap[GV]);
194 markIncompleteNodes(DSGraph::MarkFormalArgs);
196 // Remove any nodes made dead due to merging...
197 removeDeadNodes(DSGraph::KeepUnreachableGlobals);
201 //===----------------------------------------------------------------------===//
202 // Helper method implementations...
205 /// getValueDest - Return the DSNode that the actual value points to.
207 DSNodeHandle GraphBuilder::getValueDest(Value &Val) {
209 if (V == Constant::getNullValue(V->getType()))
210 return 0; // Null doesn't point to anything, don't add to ScalarMap!
212 DSNodeHandle &NH = ScalarMap[V];
214 return NH; // Already have a node? Just return it...
216 // Otherwise we need to create a new node to point to.
217 // Check first for constant expressions that must be traversed to
218 // extract the actual value.
219 if (Constant *C = dyn_cast<Constant>(V))
220 if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(C)) {
221 return NH = getValueDest(*CPR->getValue());
222 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
223 if (CE->getOpcode() == Instruction::Cast)
224 NH = getValueDest(*CE->getOperand(0));
225 else if (CE->getOpcode() == Instruction::GetElementPtr) {
226 visitGetElementPtrInst(*CE);
227 DSScalarMap::iterator I = ScalarMap.find(CE);
228 assert(I != ScalarMap.end() && "GEP didn't get processed right?");
231 // This returns a conservative unknown node for any unhandled ConstExpr
232 return NH = createNode()->setUnknownNodeMarker();
234 if (NH.getNode() == 0) { // (getelementptr null, X) returns null
240 } else if (ConstantIntegral *CI = dyn_cast<ConstantIntegral>(C)) {
241 // Random constants are unknown mem
242 return NH = createNode()->setUnknownNodeMarker();
244 assert(0 && "Unknown constant type!");
247 // Otherwise we need to create a new node to point to...
249 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
250 // Create a new global node for this global variable...
251 N = createNode(GV->getType()->getElementType());
254 // Otherwise just create a shadow node
258 NH.setNode(N); // Remember that we are pointing to it...
264 /// getLink - This method is used to return the specified link in the
265 /// specified node if one exists. If a link does not already exist (it's
266 /// null), then we create a new node, link it, then return it. We must
267 /// specify the type of the Node field we are accessing so that we know what
268 /// type should be linked to if we need to create a new node.
270 DSNodeHandle &GraphBuilder::getLink(const DSNodeHandle &node, unsigned LinkNo) {
271 DSNodeHandle &Node = const_cast<DSNodeHandle&>(node);
272 DSNodeHandle &Link = Node.getLink(LinkNo);
273 if (!Link.getNode()) {
274 // If the link hasn't been created yet, make and return a new shadow node
281 /// setDestTo - Set the ScalarMap entry for the specified value to point to the
282 /// specified destination. If the Value already points to a node, make sure to
283 /// merge the two destinations together.
285 void GraphBuilder::setDestTo(Value &V, const DSNodeHandle &NH) {
286 ScalarMap[&V].mergeWith(NH);
290 //===----------------------------------------------------------------------===//
291 // Specific instruction type handler implementations...
294 /// Alloca & Malloc instruction implementation - Simply create a new memory
295 /// object, pointing the scalar to it.
297 void GraphBuilder::handleAlloc(AllocationInst &AI, bool isHeap) {
298 DSNode *N = createNode();
300 N->setHeapNodeMarker();
302 N->setAllocaNodeMarker();
306 // PHINode - Make the scalar for the PHI node point to all of the things the
307 // incoming values point to... which effectively causes them to be merged.
309 void GraphBuilder::visitPHINode(PHINode &PN) {
310 if (!isPointerType(PN.getType())) return; // Only pointer PHIs
312 DSNodeHandle &PNDest = ScalarMap[&PN];
313 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
314 PNDest.mergeWith(getValueDest(*PN.getIncomingValue(i)));
317 void GraphBuilder::visitGetElementPtrInst(User &GEP) {
318 DSNodeHandle Value = getValueDest(*GEP.getOperand(0));
319 if (Value.getNode() == 0) return;
321 // As a special case, if all of the index operands of GEP are constant zeros,
322 // handle this just like we handle casts (ie, don't do much).
323 bool AllZeros = true;
324 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i)
325 if (GEP.getOperand(i) !=
326 Constant::getNullValue(GEP.getOperand(i)->getType())) {
331 // If all of the indices are zero, the result points to the operand without
332 // applying the type.
334 setDestTo(GEP, Value);
339 const PointerType *PTy = cast<PointerType>(GEP.getOperand(0)->getType());
340 const Type *CurTy = PTy->getElementType();
342 if (Value.getNode()->mergeTypeInfo(CurTy, Value.getOffset())) {
343 // If the node had to be folded... exit quickly
344 setDestTo(GEP, Value); // GEP result points to folded node
348 const TargetData &TD = Value.getNode()->getTargetData();
351 // Handle the pointer index specially...
352 if (GEP.getNumOperands() > 1 &&
353 GEP.getOperand(1) != ConstantSInt::getNullValue(Type::LongTy)) {
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();
490 if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(Callee))
491 Callee = CPR->getValue();
493 // Special case handling of certain libc allocation functions here.
494 if (Function *F = dyn_cast<Function>(Callee))
496 switch (F->getIntrinsicID()) {
497 case Intrinsic::va_start:
498 getValueDest(*CS.getInstruction()).getNode()->setAllocaNodeMarker();
500 case Intrinsic::va_copy:
501 getValueDest(*CS.getInstruction()).
502 mergeWith(getValueDest(**(CS.arg_begin())));
504 // FIXME: the #undef is a quick fix for compilation on Sparc
506 case Intrinsic::va_end:
508 case Intrinsic::memmove:
509 case Intrinsic::memcpy: {
510 // Merge the first & second arguments, and mark the memory read and
512 DSNodeHandle RetNH = getValueDest(**CS.arg_begin());
513 RetNH.mergeWith(getValueDest(**(CS.arg_begin()+1)));
514 if (DSNode *N = RetNH.getNode())
515 N->setModifiedMarker()->setReadMarker();
518 case Intrinsic::memset:
519 // Mark the memory modified.
520 if (DSNode *N = getValueDest(**CS.arg_begin()).getNode())
521 N->setModifiedMarker();
524 if (F->getName() == "calloc") {
525 setDestTo(*CS.getInstruction(),
526 createNode()->setHeapNodeMarker()->setModifiedMarker());
528 } else if (F->getName() == "realloc") {
529 DSNodeHandle RetNH = getValueDest(*CS.getInstruction());
530 RetNH.mergeWith(getValueDest(**CS.arg_begin()));
531 if (DSNode *N = RetNH.getNode())
532 N->setHeapNodeMarker()->setModifiedMarker()->setReadMarker();
534 } else if (F->getName() == "memmove") {
535 // Merge the first & second arguments, and mark the memory read and
537 DSNodeHandle RetNH = getValueDest(**CS.arg_begin());
538 RetNH.mergeWith(getValueDest(**(CS.arg_begin()+1)));
539 if (DSNode *N = RetNH.getNode())
540 N->setModifiedMarker()->setReadMarker();
543 } else if (F->getName() == "atoi" || F->getName() == "atof" ||
544 F->getName() == "atol" || F->getName() == "atoll" ||
545 F->getName() == "remove" || F->getName() == "unlink" ||
546 F->getName() == "rename" || F->getName() == "memcmp" ||
547 F->getName() == "strcmp" || F->getName() == "strncmp" ||
548 F->getName() == "execl" || F->getName() == "execlp" ||
549 F->getName() == "execle" || F->getName() == "execv" ||
550 F->getName() == "execvp" || F->getName() == "chmod" ||
551 F->getName() == "puts" || F->getName() == "write" ||
552 F->getName() == "open" || F->getName() == "create" ||
553 F->getName() == "truncate" || F->getName() == "chdir" ||
554 F->getName() == "mkdir" || F->getName() == "rmdir") {
555 // These functions read all of their pointer operands.
556 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
558 if (isPointerType((*AI)->getType()))
559 if (DSNode *N = getValueDest(**AI).getNode())
563 } else if (F->getName() == "read" || F->getName() == "pipe" ||
564 F->getName() == "wait" || F->getName() == "time") {
565 // These functions write all of their pointer operands.
566 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
568 if (isPointerType((*AI)->getType()))
569 if (DSNode *N = getValueDest(**AI).getNode())
570 N->setModifiedMarker();
573 } else if (F->getName() == "stat" || F->getName() == "fstat" ||
574 F->getName() == "lstat") {
575 // These functions read their first operand if its a pointer.
576 CallSite::arg_iterator AI = CS.arg_begin();
577 if (isPointerType((*AI)->getType())) {
578 DSNodeHandle Path = getValueDest(**AI);
579 if (DSNode *N = Path.getNode()) N->setReadMarker();
582 // Then they write into the stat buffer.
583 DSNodeHandle StatBuf = getValueDest(**++AI);
584 if (DSNode *N = StatBuf.getNode()) {
585 N->setModifiedMarker();
586 const Type *StatTy = F->getFunctionType()->getParamType(1);
587 if (const PointerType *PTy = dyn_cast<PointerType>(StatTy))
588 N->mergeTypeInfo(PTy->getElementType(), StatBuf.getOffset());
591 } else if (F->getName() == "strtod" || F->getName() == "strtof" ||
592 F->getName() == "strtold") {
593 // These functions read the first pointer
594 if (DSNode *Str = getValueDest(**CS.arg_begin()).getNode()) {
595 Str->setReadMarker();
596 // If the second parameter is passed, it will point to the first
598 const DSNodeHandle &EndPtrNH = getValueDest(**(CS.arg_begin()+1));
599 if (DSNode *End = EndPtrNH.getNode()) {
600 End->mergeTypeInfo(PointerType::get(Type::SByteTy),
601 EndPtrNH.getOffset(), false);
602 End->setModifiedMarker();
603 DSNodeHandle &Link = getLink(EndPtrNH);
604 Link.mergeWith(getValueDest(**CS.arg_begin()));
609 } else if (F->getName() == "fopen" || F->getName() == "fdopen" ||
610 F->getName() == "freopen") {
611 // These functions read all of their pointer operands.
612 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
614 if (isPointerType((*AI)->getType()))
615 if (DSNode *N = getValueDest(**AI).getNode())
618 // fopen allocates in an unknown way and writes to the file
619 // descriptor. Also, merge the allocated type into the node.
620 DSNodeHandle Result = getValueDest(*CS.getInstruction());
621 if (DSNode *N = Result.getNode()) {
622 N->setModifiedMarker()->setUnknownNodeMarker();
623 const Type *RetTy = F->getFunctionType()->getReturnType();
624 if (const PointerType *PTy = dyn_cast<PointerType>(RetTy))
625 N->mergeTypeInfo(PTy->getElementType(), Result.getOffset());
628 // If this is freopen, merge the file descriptor passed in with the
630 if (F->getName() == "freopen")
631 Result.mergeWith(getValueDest(**--CS.arg_end()));
634 } else if (F->getName() == "fclose" && CS.arg_end()-CS.arg_begin() ==1){
635 // fclose reads and deallocates the memory in an unknown way for the
636 // file descriptor. It merges the FILE type into the descriptor.
637 DSNodeHandle H = getValueDest(**CS.arg_begin());
638 if (DSNode *N = H.getNode()) {
639 N->setReadMarker()->setUnknownNodeMarker();
640 const Type *ArgTy = F->getFunctionType()->getParamType(0);
641 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
642 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
645 } else if (CS.arg_end()-CS.arg_begin() == 1 &&
646 (F->getName() == "fflush" || F->getName() == "feof" ||
647 F->getName() == "fileno" || F->getName() == "clearerr" ||
648 F->getName() == "rewind" || F->getName() == "ftell" ||
649 F->getName() == "ferror" || F->getName() == "fgetc" ||
650 F->getName() == "fgetc" || F->getName() == "_IO_getc")) {
651 // fflush reads and writes the memory for the file descriptor. It
652 // merges the FILE type into the descriptor.
653 DSNodeHandle H = getValueDest(**CS.arg_begin());
654 if (DSNode *N = H.getNode()) {
655 N->setReadMarker()->setModifiedMarker();
657 const Type *ArgTy = F->getFunctionType()->getParamType(0);
658 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
659 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
662 } else if (CS.arg_end()-CS.arg_begin() == 4 &&
663 (F->getName() == "fwrite" || F->getName() == "fread")) {
664 // fread writes the first operand, fwrite reads it. They both
665 // read/write the FILE descriptor, and merges the FILE type.
666 DSNodeHandle H = getValueDest(**--CS.arg_end());
667 if (DSNode *N = H.getNode()) {
668 N->setReadMarker()->setModifiedMarker();
669 const Type *ArgTy = F->getFunctionType()->getParamType(3);
670 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
671 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
674 H = getValueDest(**CS.arg_begin());
675 if (DSNode *N = H.getNode())
676 if (F->getName() == "fwrite")
679 N->setModifiedMarker();
681 } else if (F->getName() == "fgets" && CS.arg_end()-CS.arg_begin() == 3){
682 // fgets reads and writes the memory for the file descriptor. It
683 // merges the FILE type into the descriptor, and writes to the
684 // argument. It returns the argument as well.
685 CallSite::arg_iterator AI = CS.arg_begin();
686 DSNodeHandle H = getValueDest(**AI);
687 if (DSNode *N = H.getNode())
688 N->setModifiedMarker(); // Writes buffer
689 H.mergeWith(getValueDest(*CS.getInstruction())); // Returns buffer
692 // Reads and writes file descriptor, merge in FILE type.
693 H = getValueDest(**AI);
694 if (DSNode *N = H.getNode()) {
695 N->setReadMarker()->setModifiedMarker();
696 const Type *ArgTy = F->getFunctionType()->getParamType(2);
697 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
698 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
701 } else if (F->getName() == "ungetc" || F->getName() == "fputc" ||
702 F->getName() == "fputs" || F->getName() == "putc" ||
703 F->getName() == "ftell" || F->getName() == "rewind" ||
704 F->getName() == "_IO_putc") {
705 // These functions read and write the memory for the file descriptor,
706 // which is passes as the last argument.
707 DSNodeHandle H = getValueDest(**--CS.arg_end());
708 if (DSNode *N = H.getNode()) {
709 N->setReadMarker()->setModifiedMarker();
710 const Type *ArgTy = *--F->getFunctionType()->param_end();
711 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
712 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
715 // Any pointer arguments are read.
716 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
718 if (isPointerType((*AI)->getType()))
719 if (DSNode *N = getValueDest(**AI).getNode())
722 } else if (F->getName() == "fseek" || F->getName() == "fgetpos" ||
723 F->getName() == "fsetpos") {
724 // These functions read and write the memory for the file descriptor,
725 // and read/write all other arguments.
726 DSNodeHandle H = getValueDest(**CS.arg_begin());
727 if (DSNode *N = H.getNode()) {
728 const Type *ArgTy = *--F->getFunctionType()->param_end();
729 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
730 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
733 // Any pointer arguments are read.
734 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
736 if (isPointerType((*AI)->getType()))
737 if (DSNode *N = getValueDest(**AI).getNode())
738 N->setReadMarker()->setModifiedMarker();
740 } else if (F->getName() == "printf" || F->getName() == "fprintf" ||
741 F->getName() == "sprintf") {
742 CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
744 if (F->getName() == "fprintf") {
745 // fprintf reads and writes the FILE argument, and applies the type
747 DSNodeHandle H = getValueDest(**AI);
748 if (DSNode *N = H.getNode()) {
749 N->setModifiedMarker();
750 const Type *ArgTy = (*AI)->getType();
751 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
752 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
754 } else if (F->getName() == "sprintf") {
755 // sprintf writes the first string argument.
756 DSNodeHandle H = getValueDest(**AI++);
757 if (DSNode *N = H.getNode()) {
758 N->setModifiedMarker();
759 const Type *ArgTy = (*AI)->getType();
760 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
761 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
765 for (; AI != E; ++AI) {
766 // printf reads all pointer arguments.
767 if (isPointerType((*AI)->getType()))
768 if (DSNode *N = getValueDest(**AI).getNode())
772 } else if (F->getName() == "vprintf" || F->getName() == "vfprintf" ||
773 F->getName() == "vsprintf") {
774 CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
776 if (F->getName() == "vfprintf") {
777 // ffprintf reads and writes the FILE argument, and applies the type
779 DSNodeHandle H = getValueDest(**AI);
780 if (DSNode *N = H.getNode()) {
781 N->setModifiedMarker()->setReadMarker();
782 const Type *ArgTy = (*AI)->getType();
783 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
784 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
787 } else if (F->getName() == "vsprintf") {
788 // vsprintf writes the first string argument.
789 DSNodeHandle H = getValueDest(**AI++);
790 if (DSNode *N = H.getNode()) {
791 N->setModifiedMarker();
792 const Type *ArgTy = (*AI)->getType();
793 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
794 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
800 if (isPointerType((*AI)->getType()))
801 if (DSNode *N = getValueDest(**AI).getNode())
806 // Read the valist, and the pointed-to objects.
807 if (AI != E && isPointerType((*AI)->getType())) {
808 const DSNodeHandle &VAList = getValueDest(**AI);
809 if (DSNode *N = VAList.getNode()) {
811 N->mergeTypeInfo(PointerType::get(Type::SByteTy),
812 VAList.getOffset(), false);
814 DSNodeHandle &VAListObjs = getLink(VAList);
815 VAListObjs.getNode()->setReadMarker();
820 } else if (F->getName() == "scanf" || F->getName() == "fscanf" ||
821 F->getName() == "sscanf") {
822 CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
824 if (F->getName() == "fscanf") {
825 // fscanf reads and writes the FILE argument, and applies the type
827 DSNodeHandle H = getValueDest(**AI);
828 if (DSNode *N = H.getNode()) {
830 const Type *ArgTy = (*AI)->getType();
831 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
832 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
834 } else if (F->getName() == "sscanf") {
835 // sscanf reads the first string argument.
836 DSNodeHandle H = getValueDest(**AI++);
837 if (DSNode *N = H.getNode()) {
839 const Type *ArgTy = (*AI)->getType();
840 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
841 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
845 for (; AI != E; ++AI) {
846 // scanf writes all pointer arguments.
847 if (isPointerType((*AI)->getType()))
848 if (DSNode *N = getValueDest(**AI).getNode())
849 N->setModifiedMarker();
852 } else if (F->getName() == "strtok") {
853 // strtok reads and writes the first argument, returning it. It reads
854 // its second arg. FIXME: strtok also modifies some hidden static
855 // data. Someday this might matter.
856 CallSite::arg_iterator AI = CS.arg_begin();
857 DSNodeHandle H = getValueDest(**AI++);
858 if (DSNode *N = H.getNode()) {
859 N->setReadMarker()->setModifiedMarker(); // Reads/Writes buffer
860 const Type *ArgTy = F->getFunctionType()->getParamType(0);
861 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
862 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
864 H.mergeWith(getValueDest(*CS.getInstruction())); // Returns buffer
866 H = getValueDest(**AI); // Reads delimiter
867 if (DSNode *N = H.getNode()) {
869 const Type *ArgTy = F->getFunctionType()->getParamType(1);
870 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
871 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
874 } else if (F->getName() == "strchr" || F->getName() == "strrchr" ||
875 F->getName() == "strstr") {
876 // These read their arguments, and return the first one
877 DSNodeHandle H = getValueDest(**CS.arg_begin());
878 H.mergeWith(getValueDest(*CS.getInstruction())); // Returns buffer
880 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
882 if (isPointerType((*AI)->getType()))
883 if (DSNode *N = getValueDest(**AI).getNode())
886 if (DSNode *N = H.getNode())
889 } else if (F->getName() == "modf" && CS.arg_end()-CS.arg_begin() == 2) {
890 // This writes its second argument, and forces it to double.
891 DSNodeHandle H = getValueDest(**--CS.arg_end());
892 if (DSNode *N = H.getNode()) {
893 N->setModifiedMarker();
894 N->mergeTypeInfo(Type::DoubleTy, H.getOffset());
898 // Unknown function, warn if it returns a pointer type or takes a
900 bool Warn = isPointerType(CS.getInstruction()->getType());
902 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
904 if (isPointerType((*I)->getType())) {
909 std::cerr << "WARNING: Call to unknown external function '"
910 << F->getName() << "' will cause pessimistic results!\n";
915 // Set up the return value...
917 Instruction *I = CS.getInstruction();
918 if (isPointerType(I->getType()))
919 RetVal = getValueDest(*I);
921 DSNode *CalleeNode = 0;
922 if (DisableDirectCallOpt || !isa<Function>(Callee)) {
923 CalleeNode = getValueDest(*Callee).getNode();
924 if (CalleeNode == 0) {
925 std::cerr << "WARNING: Program is calling through a null pointer?\n"<< *I;
926 return; // Calling a null pointer?
930 std::vector<DSNodeHandle> Args;
931 Args.reserve(CS.arg_end()-CS.arg_begin());
933 // Calculate the arguments vector...
934 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); I != E; ++I)
935 if (isPointerType((*I)->getType()))
936 Args.push_back(getValueDest(**I));
938 // Add a new function call entry...
940 FunctionCalls->push_back(DSCallSite(CS, RetVal, CalleeNode, Args));
942 FunctionCalls->push_back(DSCallSite(CS, RetVal, cast<Function>(Callee),
946 void GraphBuilder::visitFreeInst(FreeInst &FI) {
947 // Mark that the node is written to...
948 if (DSNode *N = getValueDest(*FI.getOperand(0)).getNode())
949 N->setModifiedMarker()->setHeapNodeMarker();
953 void GraphBuilder::visitCastInst(CastInst &CI) {
954 if (isPointerType(CI.getType()))
955 if (isPointerType(CI.getOperand(0)->getType())) {
956 // Cast one pointer to the other, just act like a copy instruction
957 setDestTo(CI, getValueDest(*CI.getOperand(0)));
959 // Cast something (floating point, small integer) to a pointer. We need
960 // to track the fact that the node points to SOMETHING, just something we
961 // don't know about. Make an "Unknown" node.
963 setDestTo(CI, createNode()->setUnknownNodeMarker());
968 // visitInstruction - For all other instruction types, if we have any arguments
969 // that are of pointer type, make them have unknown composition bits, and merge
970 // the nodes together.
971 void GraphBuilder::visitInstruction(Instruction &Inst) {
972 DSNodeHandle CurNode;
973 if (isPointerType(Inst.getType()))
974 CurNode = getValueDest(Inst);
975 for (User::op_iterator I = Inst.op_begin(), E = Inst.op_end(); I != E; ++I)
976 if (isPointerType((*I)->getType()))
977 CurNode.mergeWith(getValueDest(**I));
979 if (CurNode.getNode())
980 CurNode.getNode()->setUnknownNodeMarker();
985 //===----------------------------------------------------------------------===//
986 // LocalDataStructures Implementation
987 //===----------------------------------------------------------------------===//
989 // MergeConstantInitIntoNode - Merge the specified constant into the node
991 void GraphBuilder::MergeConstantInitIntoNode(DSNodeHandle &NH, Constant *C) {
992 // Ensure a type-record exists...
993 NH.getNode()->mergeTypeInfo(C->getType(), NH.getOffset());
995 if (C->getType()->isFirstClassType()) {
996 if (isPointerType(C->getType()))
997 // Avoid adding edges from null, or processing non-"pointer" stores
998 NH.addEdgeTo(getValueDest(*C));
1002 const TargetData &TD = NH.getNode()->getTargetData();
1004 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) {
1005 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
1006 // We don't currently do any indexing for arrays...
1007 MergeConstantInitIntoNode(NH, cast<Constant>(CA->getOperand(i)));
1008 } else if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
1009 const StructLayout *SL = TD.getStructLayout(CS->getType());
1010 for (unsigned i = 0, e = CS->getNumOperands(); i != e; ++i) {
1011 DSNodeHandle NewNH(NH.getNode(), NH.getOffset()+SL->MemberOffsets[i]);
1012 MergeConstantInitIntoNode(NewNH, cast<Constant>(CS->getOperand(i)));
1014 } else if (ConstantAggregateZero *CAZ = dyn_cast<ConstantAggregateZero>(C)) {
1017 assert(0 && "Unknown constant type!");
1021 void GraphBuilder::mergeInGlobalInitializer(GlobalVariable *GV) {
1022 assert(!GV->isExternal() && "Cannot merge in external global!");
1023 // Get a node handle to the global node and merge the initializer into it.
1024 DSNodeHandle NH = getValueDest(*GV);
1025 MergeConstantInitIntoNode(NH, GV->getInitializer());
1029 bool LocalDataStructures::run(Module &M) {
1030 GlobalsGraph = new DSGraph(getAnalysis<TargetData>());
1032 const TargetData &TD = getAnalysis<TargetData>();
1035 GraphBuilder GGB(*GlobalsGraph);
1037 // Add initializers for all of the globals to the globals graph...
1038 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I)
1039 if (!I->isExternal())
1040 GGB.mergeInGlobalInitializer(I);
1043 // Calculate all of the graphs...
1044 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
1045 if (!I->isExternal())
1046 DSInfo.insert(std::make_pair(I, new DSGraph(TD, *I, GlobalsGraph)));
1048 GlobalsGraph->removeTriviallyDeadNodes();
1049 GlobalsGraph->markIncompleteNodes(DSGraph::MarkFormalArgs);
1053 // releaseMemory - If the pass pipeline is done with this pass, we can release
1054 // our memory... here...
1056 void LocalDataStructures::releaseMemory() {
1057 for (hash_map<Function*, DSGraph*>::iterator I = DSInfo.begin(),
1058 E = DSInfo.end(); I != E; ++I) {
1059 I->second->getReturnNodes().erase(I->first);
1060 if (I->second->getReturnNodes().empty())
1064 // Empty map so next time memory is released, data structures are not
1067 delete GlobalsGraph;