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/Support/GetElementPtrTypeIterator.h"
21 #include "llvm/Support/InstVisitor.h"
22 #include "llvm/Target/TargetData.h"
23 #include "Support/CommandLine.h"
24 #include "Support/Debug.h"
25 #include "Support/Timer.h"
27 // FIXME: This should eventually be a FunctionPass that is automatically
28 // aggregated into a Pass.
30 #include "llvm/Module.h"
34 static RegisterAnalysis<LocalDataStructures>
35 X("datastructure", "Local Data Structure Analysis");
38 TrackIntegersAsPointers("dsa-track-integers",
39 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 if (F->getName() == "calloc") {
460 setDestTo(*CS.getInstruction(),
461 createNode()->setHeapNodeMarker()->setModifiedMarker());
463 } else if (F->getName() == "realloc") {
464 DSNodeHandle RetNH = getValueDest(*CS.getInstruction());
465 RetNH.mergeWith(getValueDest(**CS.arg_begin()));
466 if (DSNode *N = RetNH.getNode())
467 N->setHeapNodeMarker()->setModifiedMarker()->setReadMarker();
469 } else if (F->getName() == "memset") {
470 // Merge the first argument with the return value, and mark the memory
472 DSNodeHandle RetNH = getValueDest(*CS.getInstruction());
473 RetNH.mergeWith(getValueDest(**CS.arg_begin()));
474 if (DSNode *N = RetNH.getNode())
475 N->setModifiedMarker();
477 } else if (F->getName() == "memmove") {
478 // Merge the first & second arguments with the result, and mark the
479 // memory read and modified.
480 DSNodeHandle RetNH = getValueDest(*CS.getInstruction());
481 RetNH.mergeWith(getValueDest(**CS.arg_begin()));
482 RetNH.mergeWith(getValueDest(**(CS.arg_begin()+1)));
483 if (DSNode *N = RetNH.getNode())
484 N->setModifiedMarker()->setReadMarker();
486 } else if (F->getName() == "bzero") {
487 // Mark the memory modified.
488 DSNodeHandle H = getValueDest(**CS.arg_begin());
489 if (DSNode *N = H.getNode())
490 N->setModifiedMarker();
495 // Set up the return value...
497 Instruction *I = CS.getInstruction();
498 if (isPointerType(I->getType()))
499 RetVal = getValueDest(*I);
502 if (DisableDirectCallOpt || !isa<Function>(CS.getCalledValue())) {
503 Callee = getValueDest(*CS.getCalledValue()).getNode();
505 std::cerr << "WARNING: Program is calling through a null pointer?\n"
507 return; // Calling a null pointer?
511 std::vector<DSNodeHandle> Args;
512 Args.reserve(CS.arg_end()-CS.arg_begin());
514 // Calculate the arguments vector...
515 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); I != E; ++I)
516 if (isPointerType((*I)->getType()))
517 Args.push_back(getValueDest(**I));
519 // Add a new function call entry...
521 FunctionCalls->push_back(DSCallSite(CS, RetVal, Callee, Args));
523 FunctionCalls->push_back(DSCallSite(CS, RetVal, CS.getCalledFunction(),
527 void GraphBuilder::visitFreeInst(FreeInst &FI) {
528 // Mark that the node is written to...
529 DSNode *N = getValueDest(*FI.getOperand(0)).getNode();
530 N->setModifiedMarker();
531 N->setHeapNodeMarker();
535 void GraphBuilder::visitCastInst(CastInst &CI) {
536 if (isPointerType(CI.getType()))
537 if (isPointerType(CI.getOperand(0)->getType())) {
538 // Cast one pointer to the other, just act like a copy instruction
539 setDestTo(CI, getValueDest(*CI.getOperand(0)));
541 // Cast something (floating point, small integer) to a pointer. We need
542 // to track the fact that the node points to SOMETHING, just something we
543 // don't know about. Make an "Unknown" node.
545 setDestTo(CI, createNode()->setUnknownNodeMarker());
550 // visitInstruction - For all other instruction types, if we have any arguments
551 // that are of pointer type, make them have unknown composition bits, and merge
552 // the nodes together.
553 void GraphBuilder::visitInstruction(Instruction &Inst) {
554 DSNodeHandle CurNode;
555 if (isPointerType(Inst.getType()))
556 CurNode = getValueDest(Inst);
557 for (User::op_iterator I = Inst.op_begin(), E = Inst.op_end(); I != E; ++I)
558 if (isPointerType((*I)->getType()))
559 CurNode.mergeWith(getValueDest(**I));
561 if (CurNode.getNode())
562 CurNode.getNode()->setUnknownNodeMarker();
567 //===----------------------------------------------------------------------===//
568 // LocalDataStructures Implementation
569 //===----------------------------------------------------------------------===//
571 // MergeConstantInitIntoNode - Merge the specified constant into the node
573 void GraphBuilder::MergeConstantInitIntoNode(DSNodeHandle &NH, Constant *C) {
574 // Ensure a type-record exists...
575 NH.getNode()->mergeTypeInfo(C->getType(), NH.getOffset());
577 if (C->getType()->isFirstClassType()) {
578 if (isPointerType(C->getType()))
579 // Avoid adding edges from null, or processing non-"pointer" stores
580 NH.addEdgeTo(getValueDest(*C));
584 const TargetData &TD = NH.getNode()->getTargetData();
586 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) {
587 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
588 // We don't currently do any indexing for arrays...
589 MergeConstantInitIntoNode(NH, cast<Constant>(CA->getOperand(i)));
590 } else if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
591 const StructLayout *SL = TD.getStructLayout(CS->getType());
592 for (unsigned i = 0, e = CS->getNumOperands(); i != e; ++i) {
593 DSNodeHandle NewNH(NH.getNode(), NH.getOffset()+SL->MemberOffsets[i]);
594 MergeConstantInitIntoNode(NewNH, cast<Constant>(CS->getOperand(i)));
597 assert(0 && "Unknown constant type!");
601 void GraphBuilder::mergeInGlobalInitializer(GlobalVariable *GV) {
602 assert(!GV->isExternal() && "Cannot merge in external global!");
603 // Get a node handle to the global node and merge the initializer into it.
604 DSNodeHandle NH = getValueDest(*GV);
605 MergeConstantInitIntoNode(NH, GV->getInitializer());
609 bool LocalDataStructures::run(Module &M) {
610 GlobalsGraph = new DSGraph(getAnalysis<TargetData>());
612 const TargetData &TD = getAnalysis<TargetData>();
614 // Calculate all of the graphs...
615 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
616 if (!I->isExternal())
617 DSInfo.insert(std::make_pair(I, new DSGraph(TD, *I, GlobalsGraph)));
619 GraphBuilder GGB(*GlobalsGraph);
621 // Add initializers for all of the globals to the globals graph...
622 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I)
623 if (!I->isExternal())
624 GGB.mergeInGlobalInitializer(I);
626 GlobalsGraph->markIncompleteNodes(DSGraph::MarkFormalArgs);
627 GlobalsGraph->removeTriviallyDeadNodes();
631 // releaseMemory - If the pass pipeline is done with this pass, we can release
632 // our memory... here...
634 void LocalDataStructures::releaseMemory() {
635 for (hash_map<Function*, DSGraph*>::iterator I = DSInfo.begin(),
636 E = DSInfo.end(); I != E; ++I) {
637 I->second->getReturnNodes().erase(I->first);
638 if (I->second->getReturnNodes().empty())
642 // Empty map so next time memory is released, data structures are not