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 // 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.
222 if (Constant *C = dyn_cast<Constant>(V))
223 if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(C)) {
224 return NH = getValueDest(*CPR->getValue());
225 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
226 if (CE->getOpcode() == Instruction::Cast)
227 NH = getValueDest(*CE->getOperand(0));
228 else if (CE->getOpcode() == Instruction::GetElementPtr) {
229 visitGetElementPtrInst(*CE);
230 DSScalarMap::iterator I = ScalarMap.find(CE);
231 assert(I != ScalarMap.end() && "GEP didn't get processed right?");
234 // This returns a conservative unknown node for any unhandled ConstExpr
235 return NH = createNode()->setUnknownNodeMarker();
237 if (NH.getNode() == 0) { // (getelementptr null, X) returns null
243 } else if (ConstantIntegral *CI = dyn_cast<ConstantIntegral>(C)) {
244 // Random constants are unknown mem
245 return NH = createNode()->setUnknownNodeMarker();
247 assert(0 && "Unknown constant type!");
250 // Otherwise we need to create a new node to point to...
252 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
253 // Create a new global node for this global variable...
254 N = createNode(GV->getType()->getElementType());
257 // Otherwise just create a shadow node
261 NH.setNode(N); // Remember that we are pointing to it...
267 /// getLink - This method is used to return the specified link in the
268 /// specified node if one exists. If a link does not already exist (it's
269 /// null), then we create a new node, link it, then return it. We must
270 /// specify the type of the Node field we are accessing so that we know what
271 /// type should be linked to if we need to create a new node.
273 DSNodeHandle &GraphBuilder::getLink(const DSNodeHandle &node, unsigned LinkNo) {
274 DSNodeHandle &Node = const_cast<DSNodeHandle&>(node);
275 DSNodeHandle &Link = Node.getLink(LinkNo);
276 if (!Link.getNode()) {
277 // If the link hasn't been created yet, make and return a new shadow node
284 /// setDestTo - Set the ScalarMap entry for the specified value to point to the
285 /// specified destination. If the Value already points to a node, make sure to
286 /// merge the two destinations together.
288 void GraphBuilder::setDestTo(Value &V, const DSNodeHandle &NH) {
289 ScalarMap[&V].mergeWith(NH);
293 //===----------------------------------------------------------------------===//
294 // Specific instruction type handler implementations...
297 /// Alloca & Malloc instruction implementation - Simply create a new memory
298 /// object, pointing the scalar to it.
300 void GraphBuilder::handleAlloc(AllocationInst &AI, bool isHeap) {
301 DSNode *N = createNode();
303 N->setHeapNodeMarker();
305 N->setAllocaNodeMarker();
309 // PHINode - Make the scalar for the PHI node point to all of the things the
310 // incoming values point to... which effectively causes them to be merged.
312 void GraphBuilder::visitPHINode(PHINode &PN) {
313 if (!isPointerType(PN.getType())) return; // Only pointer PHIs
315 DSNodeHandle &PNDest = ScalarMap[&PN];
316 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
317 PNDest.mergeWith(getValueDest(*PN.getIncomingValue(i)));
320 void GraphBuilder::visitGetElementPtrInst(User &GEP) {
321 DSNodeHandle Value = getValueDest(*GEP.getOperand(0));
322 if (Value.getNode() == 0) return;
324 // As a special case, if all of the index operands of GEP are constant zeros,
325 // handle this just like we handle casts (ie, don't do much).
326 bool AllZeros = true;
327 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i)
328 if (GEP.getOperand(i) !=
329 Constant::getNullValue(GEP.getOperand(i)->getType())) {
334 // If all of the indices are zero, the result points to the operand without
335 // applying the type.
337 setDestTo(GEP, Value);
342 const PointerType *PTy = cast<PointerType>(GEP.getOperand(0)->getType());
343 const Type *CurTy = PTy->getElementType();
345 if (Value.getNode()->mergeTypeInfo(CurTy, Value.getOffset())) {
346 // If the node had to be folded... exit quickly
347 setDestTo(GEP, Value); // GEP result points to folded node
351 const TargetData &TD = Value.getNode()->getTargetData();
354 // Handle the pointer index specially...
355 if (GEP.getNumOperands() > 1 &&
356 (!isa<Constant>(GEP.getOperand(1)) ||
357 !cast<Constant>(GEP.getOperand(1))->isNullValue())) {
359 // If we already know this is an array being accessed, don't do anything...
360 if (!TopTypeRec.isArray) {
361 TopTypeRec.isArray = true;
363 // If we are treating some inner field pointer as an array, fold the node
364 // up because we cannot handle it right. This can come because of
365 // something like this: &((&Pt->X)[1]) == &Pt->Y
367 if (Value.getOffset()) {
368 // Value is now the pointer we want to GEP to be...
369 Value.getNode()->foldNodeCompletely();
370 setDestTo(GEP, Value); // GEP result points to folded node
373 // This is a pointer to the first byte of the node. Make sure that we
374 // are pointing to the outter most type in the node.
375 // FIXME: We need to check one more case here...
381 // All of these subscripts are indexing INTO the elements we have...
383 for (gep_type_iterator I = gep_type_begin(GEP), E = gep_type_end(GEP);
385 if (const StructType *STy = dyn_cast<StructType>(*I)) {
386 unsigned FieldNo = cast<ConstantUInt>(I.getOperand())->getValue();
387 Offset += TD.getStructLayout(STy)->MemberOffsets[FieldNo];
388 } else if (const PointerType *PTy = dyn_cast<PointerType>(*I)) {
389 if (!isa<Constant>(I.getOperand()) ||
390 !cast<Constant>(I.getOperand())->isNullValue())
391 Value.getNode()->setArrayMarker();
396 if (const SequentialType *STy = cast<SequentialType>(*I)) {
397 CurTy = STy->getElementType();
398 if (ConstantSInt *CS = dyn_cast<ConstantSInt>(GEP.getOperand(i))) {
399 Offset += CS->getValue()*TD.getTypeSize(CurTy);
401 // Variable index into a node. We must merge all of the elements of the
402 // sequential type here.
403 if (isa<PointerType>(STy))
404 std::cerr << "Pointer indexing not handled yet!\n";
406 const ArrayType *ATy = cast<ArrayType>(STy);
407 unsigned ElSize = TD.getTypeSize(CurTy);
408 DSNode *N = Value.getNode();
409 assert(N && "Value must have a node!");
410 unsigned RawOffset = Offset+Value.getOffset();
412 // Loop over all of the elements of the array, merging them into the
414 for (unsigned i = 1, e = ATy->getNumElements(); i != e; ++i)
415 // Merge all of the byte components of this array element
416 for (unsigned j = 0; j != ElSize; ++j)
417 N->mergeIndexes(RawOffset+j, RawOffset+i*ElSize+j);
423 // Add in the offset calculated...
424 Value.setOffset(Value.getOffset()+Offset);
426 // Value is now the pointer we want to GEP to be...
427 setDestTo(GEP, Value);
430 void GraphBuilder::visitLoadInst(LoadInst &LI) {
431 DSNodeHandle Ptr = getValueDest(*LI.getOperand(0));
432 if (Ptr.getNode() == 0) return;
434 // Make that the node is read from...
435 Ptr.getNode()->setReadMarker();
437 // Ensure a typerecord exists...
438 Ptr.getNode()->mergeTypeInfo(LI.getType(), Ptr.getOffset(), false);
440 if (isPointerType(LI.getType()))
441 setDestTo(LI, getLink(Ptr));
444 void GraphBuilder::visitStoreInst(StoreInst &SI) {
445 const Type *StoredTy = SI.getOperand(0)->getType();
446 DSNodeHandle Dest = getValueDest(*SI.getOperand(1));
447 if (Dest.isNull()) return;
449 // Mark that the node is written to...
450 Dest.getNode()->setModifiedMarker();
452 // Ensure a type-record exists...
453 Dest.getNode()->mergeTypeInfo(StoredTy, Dest.getOffset());
455 // Avoid adding edges from null, or processing non-"pointer" stores
456 if (isPointerType(StoredTy))
457 Dest.addEdgeTo(getValueDest(*SI.getOperand(0)));
460 void GraphBuilder::visitReturnInst(ReturnInst &RI) {
461 if (RI.getNumOperands() && isPointerType(RI.getOperand(0)->getType()))
462 RetNode->mergeWith(getValueDest(*RI.getOperand(0)));
465 void GraphBuilder::visitVANextInst(VANextInst &I) {
466 getValueDest(*I.getOperand(0)).mergeWith(getValueDest(I));
469 void GraphBuilder::visitVAArgInst(VAArgInst &I) {
470 DSNodeHandle Ptr = getValueDest(*I.getOperand(0));
471 if (Ptr.isNull()) return;
473 // Make that the node is read from.
474 Ptr.getNode()->setReadMarker();
476 // Ensure a typerecord exists...
477 Ptr.getNode()->mergeTypeInfo(I.getType(), Ptr.getOffset(), false);
479 if (isPointerType(I.getType()))
480 setDestTo(I, getLink(Ptr));
484 void GraphBuilder::visitCallInst(CallInst &CI) {
488 void GraphBuilder::visitInvokeInst(InvokeInst &II) {
492 void GraphBuilder::visitCallSite(CallSite CS) {
493 Value *Callee = CS.getCalledValue();
494 if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(Callee))
495 Callee = CPR->getValue();
497 // Special case handling of certain libc allocation functions here.
498 if (Function *F = dyn_cast<Function>(Callee))
500 switch (F->getIntrinsicID()) {
501 case Intrinsic::vastart:
502 getValueDest(*CS.getInstruction()).getNode()->setAllocaNodeMarker();
504 case Intrinsic::vacopy:
505 getValueDest(*CS.getInstruction()).
506 mergeWith(getValueDest(**(CS.arg_begin())));
508 case Intrinsic::vaend:
510 case Intrinsic::memmove:
511 case Intrinsic::memcpy: {
512 // Merge the first & second arguments, and mark the memory read and
514 DSNodeHandle RetNH = getValueDest(**CS.arg_begin());
515 RetNH.mergeWith(getValueDest(**(CS.arg_begin()+1)));
516 if (DSNode *N = RetNH.getNode())
517 N->setModifiedMarker()->setReadMarker();
520 case Intrinsic::memset:
521 // Mark the memory modified.
522 if (DSNode *N = getValueDest(**CS.arg_begin()).getNode())
523 N->setModifiedMarker();
526 if (F->getName() == "calloc") {
527 setDestTo(*CS.getInstruction(),
528 createNode()->setHeapNodeMarker()->setModifiedMarker());
530 } else if (F->getName() == "realloc") {
531 DSNodeHandle RetNH = getValueDest(*CS.getInstruction());
532 RetNH.mergeWith(getValueDest(**CS.arg_begin()));
533 if (DSNode *N = RetNH.getNode())
534 N->setHeapNodeMarker()->setModifiedMarker()->setReadMarker();
536 } else if (F->getName() == "memmove") {
537 // Merge the first & second arguments, and mark the memory read and
539 DSNodeHandle RetNH = getValueDest(**CS.arg_begin());
540 RetNH.mergeWith(getValueDest(**(CS.arg_begin()+1)));
541 if (DSNode *N = RetNH.getNode())
542 N->setModifiedMarker()->setReadMarker();
545 } else if (F->getName() == "atoi" || F->getName() == "atof" ||
546 F->getName() == "atol" || F->getName() == "atoll" ||
547 F->getName() == "remove" || F->getName() == "unlink" ||
548 F->getName() == "rename" || F->getName() == "memcmp" ||
549 F->getName() == "strcmp" || F->getName() == "strncmp" ||
550 F->getName() == "execl" || F->getName() == "execlp" ||
551 F->getName() == "execle" || F->getName() == "execv" ||
552 F->getName() == "execvp" || F->getName() == "chmod" ||
553 F->getName() == "puts" || F->getName() == "write" ||
554 F->getName() == "open" || F->getName() == "create" ||
555 F->getName() == "truncate" || F->getName() == "chdir" ||
556 F->getName() == "mkdir" || F->getName() == "rmdir") {
557 // These functions read all of their pointer operands.
558 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
560 if (isPointerType((*AI)->getType()))
561 if (DSNode *N = getValueDest(**AI).getNode())
565 } else if (F->getName() == "read" || F->getName() == "pipe" ||
566 F->getName() == "wait" || F->getName() == "time") {
567 // These functions write all of their pointer operands.
568 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
570 if (isPointerType((*AI)->getType()))
571 if (DSNode *N = getValueDest(**AI).getNode())
572 N->setModifiedMarker();
575 } else if (F->getName() == "stat" || F->getName() == "fstat" ||
576 F->getName() == "lstat") {
577 // These functions read their first operand if its a pointer.
578 CallSite::arg_iterator AI = CS.arg_begin();
579 if (isPointerType((*AI)->getType())) {
580 DSNodeHandle Path = getValueDest(**AI);
581 if (DSNode *N = Path.getNode()) N->setReadMarker();
584 // Then they write into the stat buffer.
585 DSNodeHandle StatBuf = getValueDest(**++AI);
586 if (DSNode *N = StatBuf.getNode()) {
587 N->setModifiedMarker();
588 const Type *StatTy = F->getFunctionType()->getParamType(1);
589 if (const PointerType *PTy = dyn_cast<PointerType>(StatTy))
590 N->mergeTypeInfo(PTy->getElementType(), StatBuf.getOffset());
593 } else if (F->getName() == "strtod" || F->getName() == "strtof" ||
594 F->getName() == "strtold") {
595 // These functions read the first pointer
596 if (DSNode *Str = getValueDest(**CS.arg_begin()).getNode()) {
597 Str->setReadMarker();
598 // If the second parameter is passed, it will point to the first
600 const DSNodeHandle &EndPtrNH = getValueDest(**(CS.arg_begin()+1));
601 if (DSNode *End = EndPtrNH.getNode()) {
602 End->mergeTypeInfo(PointerType::get(Type::SByteTy),
603 EndPtrNH.getOffset(), false);
604 End->setModifiedMarker();
605 DSNodeHandle &Link = getLink(EndPtrNH);
606 Link.mergeWith(getValueDest(**CS.arg_begin()));
611 } else if (F->getName() == "fopen" || F->getName() == "fdopen" ||
612 F->getName() == "freopen") {
613 // These functions read all of their pointer operands.
614 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
616 if (isPointerType((*AI)->getType()))
617 if (DSNode *N = getValueDest(**AI).getNode())
620 // fopen allocates in an unknown way and writes to the file
621 // descriptor. Also, merge the allocated type into the node.
622 DSNodeHandle Result = getValueDest(*CS.getInstruction());
623 if (DSNode *N = Result.getNode()) {
624 N->setModifiedMarker()->setUnknownNodeMarker();
625 const Type *RetTy = F->getFunctionType()->getReturnType();
626 if (const PointerType *PTy = dyn_cast<PointerType>(RetTy))
627 N->mergeTypeInfo(PTy->getElementType(), Result.getOffset());
630 // If this is freopen, merge the file descriptor passed in with the
632 if (F->getName() == "freopen")
633 Result.mergeWith(getValueDest(**--CS.arg_end()));
636 } else if (F->getName() == "fclose" && CS.arg_end()-CS.arg_begin() ==1){
637 // fclose reads and deallocates the memory in an unknown way for the
638 // file descriptor. It merges the FILE type into the descriptor.
639 DSNodeHandle H = getValueDest(**CS.arg_begin());
640 if (DSNode *N = H.getNode()) {
641 N->setReadMarker()->setUnknownNodeMarker();
642 const Type *ArgTy = F->getFunctionType()->getParamType(0);
643 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
644 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
647 } else if (CS.arg_end()-CS.arg_begin() == 1 &&
648 (F->getName() == "fflush" || F->getName() == "feof" ||
649 F->getName() == "fileno" || F->getName() == "clearerr" ||
650 F->getName() == "rewind" || F->getName() == "ftell" ||
651 F->getName() == "ferror" || F->getName() == "fgetc" ||
652 F->getName() == "fgetc" || F->getName() == "_IO_getc")) {
653 // fflush reads and writes the memory for the file descriptor. It
654 // merges the FILE type into the descriptor.
655 DSNodeHandle H = getValueDest(**CS.arg_begin());
656 if (DSNode *N = H.getNode()) {
657 N->setReadMarker()->setModifiedMarker();
659 const Type *ArgTy = F->getFunctionType()->getParamType(0);
660 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
661 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
664 } else if (CS.arg_end()-CS.arg_begin() == 4 &&
665 (F->getName() == "fwrite" || F->getName() == "fread")) {
666 // fread writes the first operand, fwrite reads it. They both
667 // read/write the FILE descriptor, and merges the FILE type.
668 DSNodeHandle H = getValueDest(**--CS.arg_end());
669 if (DSNode *N = H.getNode()) {
670 N->setReadMarker()->setModifiedMarker();
671 const Type *ArgTy = F->getFunctionType()->getParamType(3);
672 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
673 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
676 H = getValueDest(**CS.arg_begin());
677 if (DSNode *N = H.getNode())
678 if (F->getName() == "fwrite")
681 N->setModifiedMarker();
683 } else if (F->getName() == "fgets" && CS.arg_end()-CS.arg_begin() == 3){
684 // fgets reads and writes the memory for the file descriptor. It
685 // merges the FILE type into the descriptor, and writes to the
686 // argument. It returns the argument as well.
687 CallSite::arg_iterator AI = CS.arg_begin();
688 DSNodeHandle H = getValueDest(**AI);
689 if (DSNode *N = H.getNode())
690 N->setModifiedMarker(); // Writes buffer
691 H.mergeWith(getValueDest(*CS.getInstruction())); // Returns buffer
694 // Reads and writes file descriptor, merge in FILE type.
695 H = getValueDest(**AI);
696 if (DSNode *N = H.getNode()) {
697 N->setReadMarker()->setModifiedMarker();
698 const Type *ArgTy = F->getFunctionType()->getParamType(2);
699 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
700 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
703 } else if (F->getName() == "ungetc" || F->getName() == "fputc" ||
704 F->getName() == "fputs" || F->getName() == "putc" ||
705 F->getName() == "ftell" || F->getName() == "rewind" ||
706 F->getName() == "_IO_putc") {
707 // These functions read and write the memory for the file descriptor,
708 // which is passes as the last argument.
709 DSNodeHandle H = getValueDest(**--CS.arg_end());
710 if (DSNode *N = H.getNode()) {
711 N->setReadMarker()->setModifiedMarker();
712 const Type *ArgTy = *--F->getFunctionType()->param_end();
713 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
714 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
717 // Any pointer arguments are read.
718 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
720 if (isPointerType((*AI)->getType()))
721 if (DSNode *N = getValueDest(**AI).getNode())
724 } else if (F->getName() == "fseek" || F->getName() == "fgetpos" ||
725 F->getName() == "fsetpos") {
726 // These functions read and write the memory for the file descriptor,
727 // and read/write all other arguments.
728 DSNodeHandle H = getValueDest(**CS.arg_begin());
729 if (DSNode *N = H.getNode()) {
730 const Type *ArgTy = *--F->getFunctionType()->param_end();
731 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
732 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
735 // Any pointer arguments are read.
736 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
738 if (isPointerType((*AI)->getType()))
739 if (DSNode *N = getValueDest(**AI).getNode())
740 N->setReadMarker()->setModifiedMarker();
742 } else if (F->getName() == "printf" || F->getName() == "fprintf" ||
743 F->getName() == "sprintf") {
744 CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
746 if (F->getName() == "fprintf") {
747 // fprintf reads and writes the FILE argument, and applies the type
749 DSNodeHandle H = getValueDest(**AI);
750 if (DSNode *N = H.getNode()) {
751 N->setModifiedMarker();
752 const Type *ArgTy = (*AI)->getType();
753 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
754 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
756 } else if (F->getName() == "sprintf") {
757 // sprintf writes the first string argument.
758 DSNodeHandle H = getValueDest(**AI++);
759 if (DSNode *N = H.getNode()) {
760 N->setModifiedMarker();
761 const Type *ArgTy = (*AI)->getType();
762 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
763 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
767 for (; AI != E; ++AI) {
768 // printf reads all pointer arguments.
769 if (isPointerType((*AI)->getType()))
770 if (DSNode *N = getValueDest(**AI).getNode())
774 } else if (F->getName() == "vprintf" || F->getName() == "vfprintf" ||
775 F->getName() == "vsprintf") {
776 CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
778 if (F->getName() == "vfprintf") {
779 // ffprintf reads and writes the FILE argument, and applies the type
781 DSNodeHandle H = getValueDest(**AI);
782 if (DSNode *N = H.getNode()) {
783 N->setModifiedMarker()->setReadMarker();
784 const Type *ArgTy = (*AI)->getType();
785 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
786 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
789 } else if (F->getName() == "vsprintf") {
790 // vsprintf writes the first string argument.
791 DSNodeHandle H = getValueDest(**AI++);
792 if (DSNode *N = H.getNode()) {
793 N->setModifiedMarker();
794 const Type *ArgTy = (*AI)->getType();
795 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
796 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
802 if (isPointerType((*AI)->getType()))
803 if (DSNode *N = getValueDest(**AI).getNode())
808 // Read the valist, and the pointed-to objects.
809 if (AI != E && isPointerType((*AI)->getType())) {
810 const DSNodeHandle &VAList = getValueDest(**AI);
811 if (DSNode *N = VAList.getNode()) {
813 N->mergeTypeInfo(PointerType::get(Type::SByteTy),
814 VAList.getOffset(), false);
816 DSNodeHandle &VAListObjs = getLink(VAList);
817 VAListObjs.getNode()->setReadMarker();
822 } else if (F->getName() == "scanf" || F->getName() == "fscanf" ||
823 F->getName() == "sscanf") {
824 CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
826 if (F->getName() == "fscanf") {
827 // fscanf reads and writes the FILE argument, and applies the type
829 DSNodeHandle H = getValueDest(**AI);
830 if (DSNode *N = H.getNode()) {
832 const Type *ArgTy = (*AI)->getType();
833 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
834 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
836 } else if (F->getName() == "sscanf") {
837 // sscanf reads the first string argument.
838 DSNodeHandle H = getValueDest(**AI++);
839 if (DSNode *N = H.getNode()) {
841 const Type *ArgTy = (*AI)->getType();
842 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
843 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
847 for (; AI != E; ++AI) {
848 // scanf writes all pointer arguments.
849 if (isPointerType((*AI)->getType()))
850 if (DSNode *N = getValueDest(**AI).getNode())
851 N->setModifiedMarker();
854 } else if (F->getName() == "strtok") {
855 // strtok reads and writes the first argument, returning it. It reads
856 // its second arg. FIXME: strtok also modifies some hidden static
857 // data. Someday this might matter.
858 CallSite::arg_iterator AI = CS.arg_begin();
859 DSNodeHandle H = getValueDest(**AI++);
860 if (DSNode *N = H.getNode()) {
861 N->setReadMarker()->setModifiedMarker(); // Reads/Writes buffer
862 const Type *ArgTy = F->getFunctionType()->getParamType(0);
863 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
864 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
866 H.mergeWith(getValueDest(*CS.getInstruction())); // Returns buffer
868 H = getValueDest(**AI); // Reads delimiter
869 if (DSNode *N = H.getNode()) {
871 const Type *ArgTy = F->getFunctionType()->getParamType(1);
872 if (const PointerType *PTy = dyn_cast<PointerType>(ArgTy))
873 N->mergeTypeInfo(PTy->getElementType(), H.getOffset());
876 } else if (F->getName() == "strchr" || F->getName() == "strrchr" ||
877 F->getName() == "strstr") {
878 // These read their arguments, and return the first one
879 DSNodeHandle H = getValueDest(**CS.arg_begin());
880 H.mergeWith(getValueDest(*CS.getInstruction())); // Returns buffer
882 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
884 if (isPointerType((*AI)->getType()))
885 if (DSNode *N = getValueDest(**AI).getNode())
888 if (DSNode *N = H.getNode())
891 } else if (F->getName() == "modf" && CS.arg_end()-CS.arg_begin() == 2) {
892 // This writes its second argument, and forces it to double.
893 DSNodeHandle H = getValueDest(**--CS.arg_end());
894 if (DSNode *N = H.getNode()) {
895 N->setModifiedMarker();
896 N->mergeTypeInfo(Type::DoubleTy, H.getOffset());
900 // Unknown function, warn if it returns a pointer type or takes a
902 bool Warn = isPointerType(CS.getInstruction()->getType());
904 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
906 if (isPointerType((*I)->getType())) {
911 std::cerr << "WARNING: Call to unknown external function '"
912 << F->getName() << "' will cause pessimistic results!\n";
917 // Set up the return value...
919 Instruction *I = CS.getInstruction();
920 if (isPointerType(I->getType()))
921 RetVal = getValueDest(*I);
923 DSNode *CalleeNode = 0;
924 if (DisableDirectCallOpt || !isa<Function>(Callee)) {
925 CalleeNode = getValueDest(*Callee).getNode();
926 if (CalleeNode == 0) {
927 std::cerr << "WARNING: Program is calling through a null pointer?\n"<< *I;
928 return; // Calling a null pointer?
932 std::vector<DSNodeHandle> Args;
933 Args.reserve(CS.arg_end()-CS.arg_begin());
935 // Calculate the arguments vector...
936 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); I != E; ++I)
937 if (isPointerType((*I)->getType()))
938 Args.push_back(getValueDest(**I));
940 // Add a new function call entry...
942 FunctionCalls->push_back(DSCallSite(CS, RetVal, CalleeNode, Args));
944 FunctionCalls->push_back(DSCallSite(CS, RetVal, cast<Function>(Callee),
948 void GraphBuilder::visitFreeInst(FreeInst &FI) {
949 // Mark that the node is written to...
950 if (DSNode *N = getValueDest(*FI.getOperand(0)).getNode())
951 N->setModifiedMarker()->setHeapNodeMarker();
955 void GraphBuilder::visitCastInst(CastInst &CI) {
956 if (isPointerType(CI.getType()))
957 if (isPointerType(CI.getOperand(0)->getType())) {
958 // Cast one pointer to the other, just act like a copy instruction
959 setDestTo(CI, getValueDest(*CI.getOperand(0)));
961 // Cast something (floating point, small integer) to a pointer. We need
962 // to track the fact that the node points to SOMETHING, just something we
963 // don't know about. Make an "Unknown" node.
965 setDestTo(CI, createNode()->setUnknownNodeMarker());
970 // visitInstruction - For all other instruction types, if we have any arguments
971 // that are of pointer type, make them have unknown composition bits, and merge
972 // the nodes together.
973 void GraphBuilder::visitInstruction(Instruction &Inst) {
974 DSNodeHandle CurNode;
975 if (isPointerType(Inst.getType()))
976 CurNode = getValueDest(Inst);
977 for (User::op_iterator I = Inst.op_begin(), E = Inst.op_end(); I != E; ++I)
978 if (isPointerType((*I)->getType()))
979 CurNode.mergeWith(getValueDest(**I));
981 if (CurNode.getNode())
982 CurNode.getNode()->setUnknownNodeMarker();
987 //===----------------------------------------------------------------------===//
988 // LocalDataStructures Implementation
989 //===----------------------------------------------------------------------===//
991 // MergeConstantInitIntoNode - Merge the specified constant into the node
993 void GraphBuilder::MergeConstantInitIntoNode(DSNodeHandle &NH, Constant *C) {
994 // Ensure a type-record exists...
995 NH.getNode()->mergeTypeInfo(C->getType(), NH.getOffset());
997 if (C->getType()->isFirstClassType()) {
998 if (isPointerType(C->getType()))
999 // Avoid adding edges from null, or processing non-"pointer" stores
1000 NH.addEdgeTo(getValueDest(*C));
1004 const TargetData &TD = NH.getNode()->getTargetData();
1006 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) {
1007 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
1008 // We don't currently do any indexing for arrays...
1009 MergeConstantInitIntoNode(NH, cast<Constant>(CA->getOperand(i)));
1010 } else if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
1011 const StructLayout *SL = TD.getStructLayout(CS->getType());
1012 for (unsigned i = 0, e = CS->getNumOperands(); i != e; ++i) {
1013 DSNodeHandle NewNH(NH.getNode(), NH.getOffset()+SL->MemberOffsets[i]);
1014 MergeConstantInitIntoNode(NewNH, cast<Constant>(CS->getOperand(i)));
1016 } else if (ConstantAggregateZero *CAZ = dyn_cast<ConstantAggregateZero>(C)) {
1019 assert(0 && "Unknown constant type!");
1023 void GraphBuilder::mergeInGlobalInitializer(GlobalVariable *GV) {
1024 assert(!GV->isExternal() && "Cannot merge in external global!");
1025 // Get a node handle to the global node and merge the initializer into it.
1026 DSNodeHandle NH = getValueDest(*GV);
1027 MergeConstantInitIntoNode(NH, GV->getInitializer());
1031 bool LocalDataStructures::run(Module &M) {
1032 GlobalsGraph = new DSGraph(getAnalysis<TargetData>());
1034 const TargetData &TD = getAnalysis<TargetData>();
1037 GraphBuilder GGB(*GlobalsGraph);
1039 // Add initializers for all of the globals to the globals graph...
1040 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I)
1041 if (!I->isExternal())
1042 GGB.mergeInGlobalInitializer(I);
1045 // Calculate all of the graphs...
1046 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
1047 if (!I->isExternal())
1048 DSInfo.insert(std::make_pair(I, new DSGraph(TD, *I, GlobalsGraph)));
1050 GlobalsGraph->removeTriviallyDeadNodes();
1051 GlobalsGraph->markIncompleteNodes(DSGraph::MarkFormalArgs);
1055 // releaseMemory - If the pass pipeline is done with this pass, we can release
1056 // our memory... here...
1058 void LocalDataStructures::releaseMemory() {
1059 for (hash_map<Function*, DSGraph*>::iterator I = DSInfo.begin(),
1060 E = DSInfo.end(); I != E; ++I) {
1061 I->second->getReturnNodes().erase(I->first);
1062 if (I->second->getReturnNodes().empty())
1066 // Empty map so next time memory is released, data structures are not
1069 delete GlobalsGraph;