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);
121 void MergeConstantInitIntoNode(DSNodeHandle &NH, Constant *C);
123 // Helper functions used to implement the visitation functions...
125 /// createNode - Create a new DSNode, ensuring that it is properly added to
128 DSNode *createNode(const Type *Ty = 0) {
129 DSNode *N = new DSNode(Ty, &G); // Create the node
130 if (DisableFieldSensitivity) {
131 N->foldNodeCompletely();
132 if (DSNode *FN = N->getForwardNode())
138 /// setDestTo - Set the ScalarMap entry for the specified value to point to
139 /// the specified destination. If the Value already points to a node, make
140 /// sure to merge the two destinations together.
142 void setDestTo(Value &V, const DSNodeHandle &NH);
144 /// getValueDest - Return the DSNode that the actual value points to.
146 DSNodeHandle getValueDest(Value &V);
148 /// getLink - This method is used to return the specified link in the
149 /// specified node if one exists. If a link does not already exist (it's
150 /// null), then we create a new node, link it, then return it.
152 DSNodeHandle &getLink(const DSNodeHandle &Node, unsigned Link = 0);
158 //===----------------------------------------------------------------------===//
159 // DSGraph constructor - Simply use the GraphBuilder to construct the local
161 DSGraph::DSGraph(const TargetData &td, Function &F, DSGraph *GG)
162 : GlobalsGraph(GG), TD(td) {
163 PrintAuxCalls = false;
165 DEBUG(std::cerr << " [Loc] Calculating graph for: " << F.getName() << "\n");
167 // Use the graph builder to construct the local version of the graph
168 GraphBuilder B(F, *this, ReturnNodes[&F], FunctionCalls);
170 Timer::addPeakMemoryMeasurement();
173 // Remove all integral constants from the scalarmap!
174 for (DSScalarMap::iterator I = ScalarMap.begin(); I != ScalarMap.end();)
175 if (isa<ConstantIntegral>(I->first))
176 ScalarMap.erase(I++);
180 markIncompleteNodes(DSGraph::MarkFormalArgs);
182 // Remove any nodes made dead due to merging...
183 removeDeadNodes(DSGraph::KeepUnreachableGlobals);
187 //===----------------------------------------------------------------------===//
188 // Helper method implementations...
191 /// getValueDest - Return the DSNode that the actual value points to.
193 DSNodeHandle GraphBuilder::getValueDest(Value &Val) {
195 if (V == Constant::getNullValue(V->getType()))
196 return 0; // Null doesn't point to anything, don't add to ScalarMap!
198 DSNodeHandle &NH = ScalarMap[V];
200 return NH; // Already have a node? Just return it...
202 // Otherwise we need to create a new node to point to.
203 // Check first for constant expressions that must be traversed to
204 // extract the actual value.
205 if (Constant *C = dyn_cast<Constant>(V))
206 if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(C)) {
207 return NH = getValueDest(*CPR->getValue());
208 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
209 if (CE->getOpcode() == Instruction::Cast)
210 NH = getValueDest(*CE->getOperand(0));
211 else if (CE->getOpcode() == Instruction::GetElementPtr) {
212 visitGetElementPtrInst(*CE);
213 DSScalarMap::iterator I = ScalarMap.find(CE);
214 assert(I != ScalarMap.end() && "GEP didn't get processed right?");
217 // This returns a conservative unknown node for any unhandled ConstExpr
218 return NH = createNode()->setUnknownNodeMarker();
220 if (NH.getNode() == 0) { // (getelementptr null, X) returns null
226 } else if (ConstantIntegral *CI = dyn_cast<ConstantIntegral>(C)) {
227 // Random constants are unknown mem
228 return NH = createNode()->setUnknownNodeMarker();
230 assert(0 && "Unknown constant type!");
233 // Otherwise we need to create a new node to point to...
235 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
236 // Create a new global node for this global variable...
237 N = createNode(GV->getType()->getElementType());
240 // Otherwise just create a shadow node
244 NH.setNode(N); // Remember that we are pointing to it...
250 /// getLink - This method is used to return the specified link in the
251 /// specified node if one exists. If a link does not already exist (it's
252 /// null), then we create a new node, link it, then return it. We must
253 /// specify the type of the Node field we are accessing so that we know what
254 /// type should be linked to if we need to create a new node.
256 DSNodeHandle &GraphBuilder::getLink(const DSNodeHandle &node, unsigned LinkNo) {
257 DSNodeHandle &Node = const_cast<DSNodeHandle&>(node);
258 DSNodeHandle &Link = Node.getLink(LinkNo);
259 if (!Link.getNode()) {
260 // If the link hasn't been created yet, make and return a new shadow node
267 /// setDestTo - Set the ScalarMap entry for the specified value to point to the
268 /// specified destination. If the Value already points to a node, make sure to
269 /// merge the two destinations together.
271 void GraphBuilder::setDestTo(Value &V, const DSNodeHandle &NH) {
272 DSNodeHandle &AINH = ScalarMap[&V];
273 if (AINH.getNode() == 0) // Not pointing to anything yet?
274 AINH = NH; // Just point directly to NH
280 //===----------------------------------------------------------------------===//
281 // Specific instruction type handler implementations...
284 /// Alloca & Malloc instruction implementation - Simply create a new memory
285 /// object, pointing the scalar to it.
287 void GraphBuilder::handleAlloc(AllocationInst &AI, bool isHeap) {
288 DSNode *N = createNode();
290 N->setHeapNodeMarker();
292 N->setAllocaNodeMarker();
296 // PHINode - Make the scalar for the PHI node point to all of the things the
297 // incoming values point to... which effectively causes them to be merged.
299 void GraphBuilder::visitPHINode(PHINode &PN) {
300 if (!isPointerType(PN.getType())) return; // Only pointer PHIs
302 DSNodeHandle &PNDest = ScalarMap[&PN];
303 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
304 PNDest.mergeWith(getValueDest(*PN.getIncomingValue(i)));
307 void GraphBuilder::visitGetElementPtrInst(User &GEP) {
308 DSNodeHandle Value = getValueDest(*GEP.getOperand(0));
309 if (Value.getNode() == 0) return;
311 // As a special case, if all of the index operands of GEP are constant zeros,
312 // handle this just like we handle casts (ie, don't do much).
313 bool AllZeros = true;
314 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i)
315 if (GEP.getOperand(i) !=
316 Constant::getNullValue(GEP.getOperand(i)->getType())) {
321 // If all of the indices are zero, the result points to the operand without
322 // applying the type.
324 setDestTo(GEP, Value);
329 const PointerType *PTy = cast<PointerType>(GEP.getOperand(0)->getType());
330 const Type *CurTy = PTy->getElementType();
332 if (Value.getNode()->mergeTypeInfo(CurTy, Value.getOffset())) {
333 // If the node had to be folded... exit quickly
334 setDestTo(GEP, Value); // GEP result points to folded node
338 const TargetData &TD = Value.getNode()->getTargetData();
341 // Handle the pointer index specially...
342 if (GEP.getNumOperands() > 1 &&
343 GEP.getOperand(1) != ConstantSInt::getNullValue(Type::LongTy)) {
345 // If we already know this is an array being accessed, don't do anything...
346 if (!TopTypeRec.isArray) {
347 TopTypeRec.isArray = true;
349 // If we are treating some inner field pointer as an array, fold the node
350 // up because we cannot handle it right. This can come because of
351 // something like this: &((&Pt->X)[1]) == &Pt->Y
353 if (Value.getOffset()) {
354 // Value is now the pointer we want to GEP to be...
355 Value.getNode()->foldNodeCompletely();
356 setDestTo(GEP, Value); // GEP result points to folded node
359 // This is a pointer to the first byte of the node. Make sure that we
360 // are pointing to the outter most type in the node.
361 // FIXME: We need to check one more case here...
367 // All of these subscripts are indexing INTO the elements we have...
369 for (gep_type_iterator I = gep_type_begin(GEP), E = gep_type_end(GEP);
371 if (const StructType *STy = dyn_cast<StructType>(*I)) {
372 unsigned FieldNo = cast<ConstantUInt>(I.getOperand())->getValue();
373 Offset += TD.getStructLayout(STy)->MemberOffsets[FieldNo];
378 if (const SequentialType *STy = cast<SequentialType>(*I)) {
379 CurTy = STy->getElementType();
380 if (ConstantSInt *CS = dyn_cast<ConstantSInt>(GEP.getOperand(i))) {
381 Offset += CS->getValue()*TD.getTypeSize(CurTy);
383 // Variable index into a node. We must merge all of the elements of the
384 // sequential type here.
385 if (isa<PointerType>(STy))
386 std::cerr << "Pointer indexing not handled yet!\n";
388 const ArrayType *ATy = cast<ArrayType>(STy);
389 unsigned ElSize = TD.getTypeSize(CurTy);
390 DSNode *N = Value.getNode();
391 assert(N && "Value must have a node!");
392 unsigned RawOffset = Offset+Value.getOffset();
394 // Loop over all of the elements of the array, merging them into the
396 for (unsigned i = 1, e = ATy->getNumElements(); i != e; ++i)
397 // Merge all of the byte components of this array element
398 for (unsigned j = 0; j != ElSize; ++j)
399 N->mergeIndexes(RawOffset+j, RawOffset+i*ElSize+j);
405 // Add in the offset calculated...
406 Value.setOffset(Value.getOffset()+Offset);
408 // Value is now the pointer we want to GEP to be...
409 setDestTo(GEP, Value);
412 void GraphBuilder::visitLoadInst(LoadInst &LI) {
413 DSNodeHandle Ptr = getValueDest(*LI.getOperand(0));
414 if (Ptr.getNode() == 0) return;
416 // Make that the node is read from...
417 Ptr.getNode()->setReadMarker();
419 // Ensure a typerecord exists...
420 Ptr.getNode()->mergeTypeInfo(LI.getType(), Ptr.getOffset(), false);
422 if (isPointerType(LI.getType()))
423 setDestTo(LI, getLink(Ptr));
426 void GraphBuilder::visitStoreInst(StoreInst &SI) {
427 const Type *StoredTy = SI.getOperand(0)->getType();
428 DSNodeHandle Dest = getValueDest(*SI.getOperand(1));
429 if (Dest.getNode() == 0) return;
431 // Mark that the node is written to...
432 Dest.getNode()->setModifiedMarker();
434 // Ensure a type-record exists...
435 Dest.getNode()->mergeTypeInfo(StoredTy, Dest.getOffset());
437 // Avoid adding edges from null, or processing non-"pointer" stores
438 if (isPointerType(StoredTy))
439 Dest.addEdgeTo(getValueDest(*SI.getOperand(0)));
442 void GraphBuilder::visitReturnInst(ReturnInst &RI) {
443 if (RI.getNumOperands() && isPointerType(RI.getOperand(0)->getType()))
444 RetNode->mergeWith(getValueDest(*RI.getOperand(0)));
447 void GraphBuilder::visitCallInst(CallInst &CI) {
451 void GraphBuilder::visitInvokeInst(InvokeInst &II) {
455 void GraphBuilder::visitCallSite(CallSite CS) {
456 // Special case handling of certain libc allocation functions here.
457 if (Function *F = CS.getCalledFunction())
459 switch (F->getIntrinsicID()) {
460 case Intrinsic::memmove:
461 case Intrinsic::memcpy: {
462 // Merge the first & second arguments, and mark the memory read and
464 DSNodeHandle RetNH = getValueDest(**CS.arg_begin());
465 RetNH.mergeWith(getValueDest(**(CS.arg_begin()+1)));
466 if (DSNode *N = RetNH.getNode())
467 N->setModifiedMarker()->setReadMarker();
470 case Intrinsic::memset:
471 // Mark the memory modified.
472 if (DSNode *N = getValueDest(**CS.arg_begin()).getNode())
473 N->setModifiedMarker();
476 if (F->getName() == "calloc") {
477 setDestTo(*CS.getInstruction(),
478 createNode()->setHeapNodeMarker()->setModifiedMarker());
480 } else if (F->getName() == "realloc") {
481 DSNodeHandle RetNH = getValueDest(*CS.getInstruction());
482 RetNH.mergeWith(getValueDest(**CS.arg_begin()));
483 if (DSNode *N = RetNH.getNode())
484 N->setHeapNodeMarker()->setModifiedMarker()->setReadMarker();
486 } else if (F->getName() == "atoi" || F->getName() == "atof" ||
487 F->getName() == "atol" || F->getName() == "atoll" ||
488 F->getName() == "remove" || F->getName() == "unlink" ||
489 F->getName() == "rename" || F->getName() == "memcmp" ||
490 F->getName() == "strcmp" || F->getName() == "strncmp" ||
491 F->getName() == "execl" || F->getName() == "execlp" ||
492 F->getName() == "execle" || F->getName() == "execv" ||
493 F->getName() == "execvp" || F->getName() == "chmod" ||
494 F->getName() == "puts" || F->getName() == "write" ||
495 F->getName() == "open" || F->getName() == "create" ||
496 F->getName() == "truncate" || F->getName() == "chdir" ||
497 F->getName() == "mkdir" || F->getName() == "rmdir") {
498 // These functions read all of their pointer operands.
499 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
501 if (isPointerType((*AI)->getType()))
502 if (DSNode *N = getValueDest(**AI).getNode())
506 } else if (F->getName() == "read" || F->getName() == "pipe" ||
507 F->getName() == "wait") {
508 // These functions write all of their pointer operands.
509 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
511 if (isPointerType((*AI)->getType()))
512 if (DSNode *N = getValueDest(**AI).getNode())
513 N->setModifiedMarker();
516 } else if (F->getName() == "stat" || F->getName() == "fstat" ||
517 F->getName() == "lstat") {
518 // These functions read their first operand if its a pointer.
519 CallSite::arg_iterator AI = CS.arg_begin();
520 if (isPointerType((*AI)->getType())) {
521 DSNodeHandle Path = getValueDest(**AI);
522 if (DSNode *N = Path.getNode()) N->setReadMarker();
525 // Then they write into the stat buffer.
526 DSNodeHandle StatBuf = getValueDest(**++AI);
527 if (DSNode *N = StatBuf.getNode()) {
528 N->setModifiedMarker();
529 const Type *StatTy = F->getFunctionType()->getParamType(1);
530 if (const PointerType *PTy = dyn_cast<PointerType>(StatTy))
531 N->mergeTypeInfo(PTy->getElementType(), StatBuf.getOffset());
534 } else if (F->getName() == "fopen" || F->getName() == "fdopen") {
535 // fopen reads the mode argument strings.
536 CallSite::arg_iterator AI = CS.arg_begin();
537 if (isPointerType((*AI)->getType())) {
538 DSNodeHandle Path = getValueDest(**AI);
539 if (DSNode *N = Path.getNode()) N->setReadMarker();
542 DSNodeHandle Mode = getValueDest(**++AI);
543 if (DSNode *N = Mode.getNode()) N->setReadMarker();
545 // fopen allocates in an unknown way and writes to the file
546 // descriptor. Also, merge the allocated type into the node.
547 DSNodeHandle Result = getValueDest(*CS.getInstruction());
548 if (DSNode *N = Result.getNode()) {
549 N->setModifiedMarker()->setUnknownNodeMarker();
550 const Type *RetTy = F->getFunctionType()->getReturnType();
551 if (const PointerType *PTy = dyn_cast<PointerType>(RetTy))
552 N->mergeTypeInfo(PTy->getElementType(), Result.getOffset());
555 } else if (F->getName() == "fclose" && CS.arg_end()-CS.arg_begin() ==1){
556 // fclose reads and deallocates the memory in an unknown way for the
557 // file descriptor. It merges the FILE type into the descriptor.
558 DSNodeHandle H = getValueDest(**CS.arg_begin());
559 if (DSNode *N = H.getNode()) {
560 N->setReadMarker()->setUnknownNodeMarker();
561 const Type *ArgTy = F->getFunctionType()->getParamType(0);
562 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
563 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
566 } else if (CS.arg_end()-CS.arg_begin() == 1 &&
567 (F->getName() == "fflush" || F->getName() == "feof" ||
568 F->getName() == "fileno" || F->getName() == "clearerr" ||
569 F->getName() == "rewind" || F->getName() == "ftell" ||
570 F->getName() == "ferror")) {
571 // fflush reads and writes the memory for the file descriptor. It
572 // merges the FILE type into the descriptor.
573 DSNodeHandle H = getValueDest(**CS.arg_begin());
574 if (DSNode *N = H.getNode()) {
575 N->setReadMarker()->setModifiedMarker();
577 const Type *ArgTy = F->getFunctionType()->getParamType(0);
578 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
579 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
582 } else if (CS.arg_end()-CS.arg_begin() == 4 &&
583 (F->getName() == "fwrite" || F->getName() == "fread")) {
584 // fread writes the first operand, fwrite reads it. They both
585 // read/write the FILE descriptor, and merges the FILE type.
586 DSNodeHandle H = getValueDest(**--CS.arg_end());
587 if (DSNode *N = H.getNode()) {
588 N->setReadMarker()->setModifiedMarker();
589 const Type *ArgTy = F->getFunctionType()->getParamType(3);
590 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
591 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
594 H = getValueDest(**CS.arg_begin());
595 if (DSNode *N = H.getNode())
596 if (F->getName() == "fwrite")
599 N->setModifiedMarker();
601 } else if (F->getName() == "fgets" && CS.arg_end()-CS.arg_begin() == 3){
602 // fgets reads and writes the memory for the file descriptor. It
603 // merges the FILE type into the descriptor, and writes to the
604 // argument. It returns the argument as well.
605 CallSite::arg_iterator AI = CS.arg_begin();
606 DSNodeHandle H = getValueDest(**AI);
607 if (DSNode *N = H.getNode())
608 N->setModifiedMarker(); // Writes buffer
609 H.mergeWith(getValueDest(*CS.getInstruction())); // Returns buffer
612 // Reads and writes file descriptor, merge in FILE type.
613 H = getValueDest(**AI);
614 if (DSNode *N = H.getNode()) {
615 N->setReadMarker()->setModifiedMarker();
616 const Type *ArgTy = F->getFunctionType()->getParamType(2);
617 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
618 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
621 } else if (F->getName() == "ungetc" || F->getName() == "fputc" ||
622 F->getName() == "fputs" || F->getName() == "putc" ||
623 F->getName() == "ftell" || F->getName() == "rewind") {
624 // These functions read and write the memory for the file descriptor.
625 DSNodeHandle H = getValueDest(**--CS.arg_end());
626 if (DSNode *N = H.getNode()) {
627 N->setReadMarker()->setModifiedMarker();
628 const Type *ArgTy = *--F->getFunctionType()->param_end();
629 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
630 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
633 // Any pointer arguments are read.
634 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
636 if (isPointerType((*AI)->getType()))
637 if (DSNode *N = getValueDest(**AI).getNode())
640 } else if (F->getName() == "fseek" || F->getName() == "fgetpos" ||
641 F->getName() == "fsetpos") {
642 // These functions read and write the memory for the file descriptor,
643 // and read/write all other arguments.
644 DSNodeHandle H = getValueDest(**CS.arg_begin());
645 if (DSNode *N = H.getNode()) {
646 const Type *ArgTy = *--F->getFunctionType()->param_end();
647 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
648 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
651 // Any pointer arguments are read.
652 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
654 if (isPointerType((*AI)->getType()))
655 if (DSNode *N = getValueDest(**AI).getNode())
656 N->setReadMarker()->setModifiedMarker();
658 } else if (F->getName() == "printf" || F->getName() == "fprintf" ||
659 F->getName() == "sprintf") {
660 CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
662 if (F->getName() == "fprintf") {
663 // fprintf reads and writes the FILE argument, and applies the type
665 DSNodeHandle H = getValueDest(**AI);
666 if (DSNode *N = H.getNode()) {
667 N->setModifiedMarker();
668 const Type *ArgTy = (*AI)->getType();
669 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
670 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
672 } else if (F->getName() == "sprintf") {
673 // sprintf writes the first string argument.
674 DSNodeHandle H = getValueDest(**AI++);
675 if (DSNode *N = H.getNode()) {
676 N->setModifiedMarker();
677 const Type *ArgTy = (*AI)->getType();
678 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
679 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
683 for (; AI != E; ++AI) {
684 // printf reads all pointer arguments.
685 if (isPointerType((*AI)->getType()))
686 if (DSNode *N = getValueDest(**AI).getNode())
690 } else if (F->getName() == "scanf" || F->getName() == "fscanf" ||
691 F->getName() == "sscanf") {
692 CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
694 if (F->getName() == "fscanf") {
695 // fscanf reads and writes the FILE argument, and applies the type
697 DSNodeHandle H = getValueDest(**AI);
698 if (DSNode *N = H.getNode()) {
700 const Type *ArgTy = (*AI)->getType();
701 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
702 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
704 } else if (F->getName() == "sscanf") {
705 // sscanf reads the first string argument.
706 DSNodeHandle H = getValueDest(**AI++);
707 if (DSNode *N = H.getNode()) {
709 const Type *ArgTy = (*AI)->getType();
710 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
711 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
715 for (; AI != E; ++AI) {
716 // scanf writes all pointer arguments.
717 if (isPointerType((*AI)->getType()))
718 if (DSNode *N = getValueDest(**AI).getNode())
719 N->setModifiedMarker();
722 } else if (F->getName() == "strtok") {
723 // strtok reads and writes the first argument, returning it. It reads
724 // its second arg. FIXME: strtok also modifies some hidden static
725 // data. Someday this might matter.
726 CallSite::arg_iterator AI = CS.arg_begin();
727 DSNodeHandle H = getValueDest(**AI++);
728 if (DSNode *N = H.getNode()) {
729 N->setReadMarker()->setModifiedMarker(); // Reads/Writes buffer
730 const Type *ArgTy = F->getFunctionType()->getParamType(0);
731 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
732 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
734 H.mergeWith(getValueDest(*CS.getInstruction())); // Returns buffer
736 H = getValueDest(**AI); // Reads delimiter
737 if (DSNode *N = H.getNode()) {
739 const Type *ArgTy = F->getFunctionType()->getParamType(1);
740 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
741 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
744 } else if (F->getName() == "strchr" || F->getName() == "strrchr") {
745 // These read their first argument, and return it.
746 DSNodeHandle H = getValueDest(**CS.arg_begin());
747 if (DSNode *N = H.getNode())
749 H.mergeWith(getValueDest(*CS.getInstruction())); // Returns buffer
751 } else if (F->getName() == "modf" && CS.arg_end()-CS.arg_begin() == 2) {
752 // This writes its second argument, and forces it to double.
753 DSNodeHandle H = getValueDest(**--CS.arg_end());
754 if (DSNode *N = H.getNode()) {
755 N->setModifiedMarker();
756 N->mergeTypeInfo(Type::DoubleTy, H.getOffset());
760 // Unknown function, warn if it returns a pointer type or takes a
762 bool Warn = isPointerType(CS.getInstruction()->getType());
764 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
766 if (isPointerType((*I)->getType())) {
771 std::cerr << "WARNING: Call to unknown external function '"
772 << F->getName() << "' will cause pessimistic results!\n";
777 // Set up the return value...
779 Instruction *I = CS.getInstruction();
780 if (isPointerType(I->getType()))
781 RetVal = getValueDest(*I);
784 if (DisableDirectCallOpt || !isa<Function>(CS.getCalledValue())) {
785 Callee = getValueDest(*CS.getCalledValue()).getNode();
787 std::cerr << "WARNING: Program is calling through a null pointer?\n"
789 return; // Calling a null pointer?
793 std::vector<DSNodeHandle> Args;
794 Args.reserve(CS.arg_end()-CS.arg_begin());
796 // Calculate the arguments vector...
797 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); I != E; ++I)
798 if (isPointerType((*I)->getType()))
799 Args.push_back(getValueDest(**I));
801 // Add a new function call entry...
803 FunctionCalls->push_back(DSCallSite(CS, RetVal, Callee, Args));
805 FunctionCalls->push_back(DSCallSite(CS, RetVal, CS.getCalledFunction(),
809 void GraphBuilder::visitFreeInst(FreeInst &FI) {
810 // Mark that the node is written to...
811 if (DSNode *N = getValueDest(*FI.getOperand(0)).getNode())
812 N->setModifiedMarker()->setHeapNodeMarker();
816 void GraphBuilder::visitCastInst(CastInst &CI) {
817 if (isPointerType(CI.getType()))
818 if (isPointerType(CI.getOperand(0)->getType())) {
819 // Cast one pointer to the other, just act like a copy instruction
820 setDestTo(CI, getValueDest(*CI.getOperand(0)));
822 // Cast something (floating point, small integer) to a pointer. We need
823 // to track the fact that the node points to SOMETHING, just something we
824 // don't know about. Make an "Unknown" node.
826 setDestTo(CI, createNode()->setUnknownNodeMarker());
831 // visitInstruction - For all other instruction types, if we have any arguments
832 // that are of pointer type, make them have unknown composition bits, and merge
833 // the nodes together.
834 void GraphBuilder::visitInstruction(Instruction &Inst) {
835 DSNodeHandle CurNode;
836 if (isPointerType(Inst.getType()))
837 CurNode = getValueDest(Inst);
838 for (User::op_iterator I = Inst.op_begin(), E = Inst.op_end(); I != E; ++I)
839 if (isPointerType((*I)->getType()))
840 CurNode.mergeWith(getValueDest(**I));
842 if (CurNode.getNode())
843 CurNode.getNode()->setUnknownNodeMarker();
848 //===----------------------------------------------------------------------===//
849 // LocalDataStructures Implementation
850 //===----------------------------------------------------------------------===//
852 // MergeConstantInitIntoNode - Merge the specified constant into the node
854 void GraphBuilder::MergeConstantInitIntoNode(DSNodeHandle &NH, Constant *C) {
855 // Ensure a type-record exists...
856 NH.getNode()->mergeTypeInfo(C->getType(), NH.getOffset());
858 if (C->getType()->isFirstClassType()) {
859 if (isPointerType(C->getType()))
860 // Avoid adding edges from null, or processing non-"pointer" stores
861 NH.addEdgeTo(getValueDest(*C));
865 const TargetData &TD = NH.getNode()->getTargetData();
867 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) {
868 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
869 // We don't currently do any indexing for arrays...
870 MergeConstantInitIntoNode(NH, cast<Constant>(CA->getOperand(i)));
871 } else if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
872 const StructLayout *SL = TD.getStructLayout(CS->getType());
873 for (unsigned i = 0, e = CS->getNumOperands(); i != e; ++i) {
874 DSNodeHandle NewNH(NH.getNode(), NH.getOffset()+SL->MemberOffsets[i]);
875 MergeConstantInitIntoNode(NewNH, cast<Constant>(CS->getOperand(i)));
877 } else if (ConstantAggregateZero *CAZ = dyn_cast<ConstantAggregateZero>(C)) {
880 assert(0 && "Unknown constant type!");
884 void GraphBuilder::mergeInGlobalInitializer(GlobalVariable *GV) {
885 assert(!GV->isExternal() && "Cannot merge in external global!");
886 // Get a node handle to the global node and merge the initializer into it.
887 DSNodeHandle NH = getValueDest(*GV);
888 MergeConstantInitIntoNode(NH, GV->getInitializer());
892 bool LocalDataStructures::run(Module &M) {
893 GlobalsGraph = new DSGraph(getAnalysis<TargetData>());
895 const TargetData &TD = getAnalysis<TargetData>();
897 // Calculate all of the graphs...
898 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
899 if (!I->isExternal())
900 DSInfo.insert(std::make_pair(I, new DSGraph(TD, *I, GlobalsGraph)));
902 GraphBuilder GGB(*GlobalsGraph);
904 // Add initializers for all of the globals to the globals graph...
905 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I)
906 if (!I->isExternal())
907 GGB.mergeInGlobalInitializer(I);
909 GlobalsGraph->removeTriviallyDeadNodes();
910 GlobalsGraph->markIncompleteNodes(DSGraph::MarkFormalArgs);
914 // releaseMemory - If the pass pipeline is done with this pass, we can release
915 // our memory... here...
917 void LocalDataStructures::releaseMemory() {
918 for (hash_map<Function*, DSGraph*>::iterator I = DSInfo.begin(),
919 E = DSInfo.end(); I != E; ++I) {
920 I->second->getReturnNodes().erase(I->first);
921 if (I->second->getReturnNodes().empty())
925 // Empty map so next time memory is released, data structures are not