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/InstVisitor.h"
21 #include "llvm/Target/TargetData.h"
22 #include "Support/CommandLine.h"
23 #include "Support/Debug.h"
24 #include "Support/Timer.h"
26 // FIXME: This should eventually be a FunctionPass that is automatically
27 // aggregated into a Pass.
29 #include "llvm/Module.h"
33 static RegisterAnalysis<LocalDataStructures>
34 X("datastructure", "Local Data Structure Analysis");
37 TrackIntegersAsPointers("dsa-track-integers",
38 cl::desc("If this is set, track integers as potential pointers"));
43 // isPointerType - Return true if this type is big enough to hold a pointer.
44 bool isPointerType(const Type *Ty) {
45 if (isa<PointerType>(Ty))
47 else if (TrackIntegersAsPointers && Ty->isPrimitiveType() &&Ty->isInteger())
48 return Ty->getPrimitiveSize() >= PointerSize;
57 DisableDirectCallOpt("disable-direct-call-dsopt", cl::Hidden,
58 cl::desc("Disable direct call optimization in "
59 "DSGraph construction"));
61 DisableFieldSensitivity("disable-ds-field-sensitivity", cl::Hidden,
62 cl::desc("Disable field sensitivity in DSGraphs"));
64 //===--------------------------------------------------------------------===//
66 //===--------------------------------------------------------------------===//
68 /// This class is the builder class that constructs the local data structure
69 /// graph by performing a single pass over the function in question.
71 class GraphBuilder : InstVisitor<GraphBuilder> {
73 DSNodeHandle *RetNode; // Node that gets returned...
74 DSGraph::ScalarMapTy &ScalarMap;
75 std::vector<DSCallSite> *FunctionCalls;
78 GraphBuilder(Function &f, DSGraph &g, DSNodeHandle &retNode,
79 std::vector<DSCallSite> &fc)
80 : G(g), RetNode(&retNode), ScalarMap(G.getScalarMap()),
83 // Create scalar nodes for all pointer arguments...
84 for (Function::aiterator I = f.abegin(), E = f.aend(); I != E; ++I)
85 if (isPointerType(I->getType()))
88 visit(f); // Single pass over the function
91 // GraphBuilder ctor for working on the globals graph
92 GraphBuilder(DSGraph &g)
93 : G(g), RetNode(0), ScalarMap(G.getScalarMap()), FunctionCalls(0) {
96 void mergeInGlobalInitializer(GlobalVariable *GV);
99 // Visitor functions, used to handle each instruction type we encounter...
100 friend class InstVisitor<GraphBuilder>;
101 void visitMallocInst(MallocInst &MI) { handleAlloc(MI, true); }
102 void visitAllocaInst(AllocaInst &AI) { handleAlloc(AI, false); }
103 void handleAlloc(AllocationInst &AI, bool isHeap);
105 void visitPHINode(PHINode &PN);
107 void visitGetElementPtrInst(User &GEP);
108 void visitReturnInst(ReturnInst &RI);
109 void visitLoadInst(LoadInst &LI);
110 void visitStoreInst(StoreInst &SI);
111 void visitCallInst(CallInst &CI);
112 void visitInvokeInst(InvokeInst &II);
113 void visitSetCondInst(SetCondInst &SCI) {} // SetEQ & friends are ignored
114 void visitFreeInst(FreeInst &FI);
115 void visitCastInst(CastInst &CI);
116 void visitInstruction(Instruction &I);
118 void visitCallSite(CallSite CS);
120 void MergeConstantInitIntoNode(DSNodeHandle &NH, Constant *C);
122 // Helper functions used to implement the visitation functions...
124 /// createNode - Create a new DSNode, ensuring that it is properly added to
127 DSNode *createNode(const Type *Ty = 0) {
128 DSNode *N = new DSNode(Ty, &G); // Create the node
129 if (DisableFieldSensitivity) {
130 N->foldNodeCompletely();
131 if (DSNode *FN = N->getForwardNode())
137 /// setDestTo - Set the ScalarMap entry for the specified value to point to
138 /// the specified destination. If the Value already points to a node, make
139 /// sure to merge the two destinations together.
141 void setDestTo(Value &V, const DSNodeHandle &NH);
143 /// getValueDest - Return the DSNode that the actual value points to.
145 DSNodeHandle getValueDest(Value &V);
147 /// getLink - This method is used to return the specified link in the
148 /// specified node if one exists. If a link does not already exist (it's
149 /// null), then we create a new node, link it, then return it.
151 DSNodeHandle &getLink(const DSNodeHandle &Node, unsigned Link = 0);
157 //===----------------------------------------------------------------------===//
158 // DSGraph constructor - Simply use the GraphBuilder to construct the local
160 DSGraph::DSGraph(const TargetData &td, Function &F, DSGraph *GG)
161 : GlobalsGraph(GG), TD(td) {
162 PrintAuxCalls = false;
164 DEBUG(std::cerr << " [Loc] Calculating graph for: " << F.getName() << "\n");
166 // Use the graph builder to construct the local version of the graph
167 GraphBuilder B(F, *this, ReturnNodes[&F], FunctionCalls);
169 Timer::addPeakMemoryMeasurement();
172 // Remove all integral constants from the scalarmap!
173 for (ScalarMapTy::iterator I = ScalarMap.begin(); I != ScalarMap.end();)
174 if (isa<ConstantIntegral>(I->first)) {
175 ScalarMapTy::iterator J = 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 DSGraph::ScalarMapTy::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 const PointerType *PTy = cast<PointerType>(GEP.getOperand(0)->getType());
312 const Type *CurTy = PTy->getElementType();
314 if (Value.getNode()->mergeTypeInfo(CurTy, Value.getOffset())) {
315 // If the node had to be folded... exit quickly
316 setDestTo(GEP, Value); // GEP result points to folded node
320 const TargetData &TD = Value.getNode()->getTargetData();
323 // Handle the pointer index specially...
324 if (GEP.getNumOperands() > 1 &&
325 GEP.getOperand(1) != ConstantSInt::getNullValue(Type::LongTy)) {
327 // If we already know this is an array being accessed, don't do anything...
328 if (!TopTypeRec.isArray) {
329 TopTypeRec.isArray = true;
331 // If we are treating some inner field pointer as an array, fold the node
332 // up because we cannot handle it right. This can come because of
333 // something like this: &((&Pt->X)[1]) == &Pt->Y
335 if (Value.getOffset()) {
336 // Value is now the pointer we want to GEP to be...
337 Value.getNode()->foldNodeCompletely();
338 setDestTo(GEP, Value); // GEP result points to folded node
341 // This is a pointer to the first byte of the node. Make sure that we
342 // are pointing to the outter most type in the node.
343 // FIXME: We need to check one more case here...
349 // All of these subscripts are indexing INTO the elements we have...
351 for (unsigned i = 2, e = GEP.getNumOperands(); i < e; ++i)
352 if (GEP.getOperand(i)->getType() == Type::LongTy) {
353 // Get the type indexing into...
354 const SequentialType *STy = cast<SequentialType>(CurTy);
355 CurTy = STy->getElementType();
357 if (ConstantSInt *CS = dyn_cast<ConstantSInt>(GEP.getOperand(i))) {
358 Offset += CS->getValue()*TD.getTypeSize(CurTy);
360 // Variable index into a node. We must merge all of the elements of the
361 // sequential type here.
362 if (isa<PointerType>(STy))
363 std::cerr << "Pointer indexing not handled yet!\n";
365 const ArrayType *ATy = cast<ArrayType>(STy);
366 unsigned ElSize = TD.getTypeSize(CurTy);
367 DSNode *N = Value.getNode();
368 assert(N && "Value must have a node!");
369 unsigned RawOffset = Offset+Value.getOffset();
371 // Loop over all of the elements of the array, merging them into the
373 for (unsigned i = 1, e = ATy->getNumElements(); i != e; ++i)
374 // Merge all of the byte components of this array element
375 for (unsigned j = 0; j != ElSize; ++j)
376 N->mergeIndexes(RawOffset+j, RawOffset+i*ElSize+j);
380 } else if (GEP.getOperand(i)->getType() == Type::UByteTy) {
381 unsigned FieldNo = cast<ConstantUInt>(GEP.getOperand(i))->getValue();
382 const StructType *STy = cast<StructType>(CurTy);
383 Offset += TD.getStructLayout(STy)->MemberOffsets[FieldNo];
384 CurTy = STy->getContainedType(FieldNo);
387 // Add in the offset calculated...
388 Value.setOffset(Value.getOffset()+Offset);
390 // Value is now the pointer we want to GEP to be...
391 setDestTo(GEP, Value);
394 void GraphBuilder::visitLoadInst(LoadInst &LI) {
395 DSNodeHandle Ptr = getValueDest(*LI.getOperand(0));
396 if (Ptr.getNode() == 0) return;
398 // Make that the node is read from...
399 Ptr.getNode()->setReadMarker();
401 // Ensure a typerecord exists...
402 Ptr.getNode()->mergeTypeInfo(LI.getType(), Ptr.getOffset(), false);
404 if (isPointerType(LI.getType()))
405 setDestTo(LI, getLink(Ptr));
408 void GraphBuilder::visitStoreInst(StoreInst &SI) {
409 const Type *StoredTy = SI.getOperand(0)->getType();
410 DSNodeHandle Dest = getValueDest(*SI.getOperand(1));
411 if (Dest.getNode() == 0) return;
413 // Mark that the node is written to...
414 Dest.getNode()->setModifiedMarker();
416 // Ensure a type-record exists...
417 Dest.getNode()->mergeTypeInfo(StoredTy, Dest.getOffset());
419 // Avoid adding edges from null, or processing non-"pointer" stores
420 if (isPointerType(StoredTy))
421 Dest.addEdgeTo(getValueDest(*SI.getOperand(0)));
424 void GraphBuilder::visitReturnInst(ReturnInst &RI) {
425 if (RI.getNumOperands() && isPointerType(RI.getOperand(0)->getType()))
426 RetNode->mergeWith(getValueDest(*RI.getOperand(0)));
429 void GraphBuilder::visitCallInst(CallInst &CI) {
433 void GraphBuilder::visitInvokeInst(InvokeInst &II) {
437 void GraphBuilder::visitCallSite(CallSite CS) {
438 // Special case handling of certain libc allocation functions here.
439 if (Function *F = CS.getCalledFunction())
441 if (F->getName() == "calloc") {
442 setDestTo(*CS.getInstruction(),
443 createNode()->setHeapNodeMarker()->setModifiedMarker());
445 } else if (F->getName() == "realloc") {
446 DSNodeHandle RetNH = getValueDest(*CS.getInstruction());
447 RetNH.mergeWith(getValueDest(**CS.arg_begin()));
448 if (DSNode *N = RetNH.getNode())
449 N->setHeapNodeMarker()->setModifiedMarker()->setReadMarker();
451 } else if (F->getName() == "memset") {
452 // Merge the first argument with the return value, and mark the memory
454 DSNodeHandle RetNH = getValueDest(*CS.getInstruction());
455 RetNH.mergeWith(getValueDest(**CS.arg_begin()));
456 if (DSNode *N = RetNH.getNode())
457 N->setModifiedMarker();
459 } else if (F->getName() == "memmove") {
460 // Merge the first & second arguments with the result, and mark the
461 // memory read and modified.
462 DSNodeHandle RetNH = getValueDest(*CS.getInstruction());
463 RetNH.mergeWith(getValueDest(**CS.arg_begin()));
464 RetNH.mergeWith(getValueDest(**(CS.arg_begin()+1)));
465 if (DSNode *N = RetNH.getNode())
466 N->setModifiedMarker()->setReadMarker();
468 } else if (F->getName() == "bzero") {
469 // Mark the memory modified.
470 DSNodeHandle H = getValueDest(**CS.arg_begin());
471 if (DSNode *N = H.getNode())
472 N->setModifiedMarker();
477 // Set up the return value...
479 Instruction *I = CS.getInstruction();
480 if (isPointerType(I->getType()))
481 RetVal = getValueDest(*I);
484 if (DisableDirectCallOpt || !isa<Function>(CS.getCalledValue())) {
485 Callee = getValueDest(*CS.getCalledValue()).getNode();
487 std::cerr << "WARNING: Program is calling through a null pointer?\n"
489 return; // Calling a null pointer?
493 std::vector<DSNodeHandle> Args;
494 Args.reserve(CS.arg_end()-CS.arg_begin());
496 // Calculate the arguments vector...
497 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); I != E; ++I)
498 if (isPointerType((*I)->getType()))
499 Args.push_back(getValueDest(**I));
501 // Add a new function call entry...
503 FunctionCalls->push_back(DSCallSite(CS, RetVal, Callee, Args));
505 FunctionCalls->push_back(DSCallSite(CS, RetVal, CS.getCalledFunction(),
509 void GraphBuilder::visitFreeInst(FreeInst &FI) {
510 // Mark that the node is written to...
511 DSNode *N = getValueDest(*FI.getOperand(0)).getNode();
512 N->setModifiedMarker();
513 N->setHeapNodeMarker();
517 void GraphBuilder::visitCastInst(CastInst &CI) {
518 if (isPointerType(CI.getType()))
519 if (isPointerType(CI.getOperand(0)->getType())) {
520 // Cast one pointer to the other, just act like a copy instruction
521 setDestTo(CI, getValueDest(*CI.getOperand(0)));
523 // Cast something (floating point, small integer) to a pointer. We need
524 // to track the fact that the node points to SOMETHING, just something we
525 // don't know about. Make an "Unknown" node.
527 setDestTo(CI, createNode()->setUnknownNodeMarker());
532 // visitInstruction - For all other instruction types, if we have any arguments
533 // that are of pointer type, make them have unknown composition bits, and merge
534 // the nodes together.
535 void GraphBuilder::visitInstruction(Instruction &Inst) {
536 DSNodeHandle CurNode;
537 if (isPointerType(Inst.getType()))
538 CurNode = getValueDest(Inst);
539 for (User::op_iterator I = Inst.op_begin(), E = Inst.op_end(); I != E; ++I)
540 if (isPointerType((*I)->getType()))
541 CurNode.mergeWith(getValueDest(**I));
543 if (CurNode.getNode())
544 CurNode.getNode()->setUnknownNodeMarker();
549 //===----------------------------------------------------------------------===//
550 // LocalDataStructures Implementation
551 //===----------------------------------------------------------------------===//
553 // MergeConstantInitIntoNode - Merge the specified constant into the node
555 void GraphBuilder::MergeConstantInitIntoNode(DSNodeHandle &NH, Constant *C) {
556 // Ensure a type-record exists...
557 NH.getNode()->mergeTypeInfo(C->getType(), NH.getOffset());
559 if (C->getType()->isFirstClassType()) {
560 if (isPointerType(C->getType()))
561 // Avoid adding edges from null, or processing non-"pointer" stores
562 NH.addEdgeTo(getValueDest(*C));
566 const TargetData &TD = NH.getNode()->getTargetData();
568 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) {
569 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
570 // We don't currently do any indexing for arrays...
571 MergeConstantInitIntoNode(NH, cast<Constant>(CA->getOperand(i)));
572 } else if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
573 const StructLayout *SL = TD.getStructLayout(CS->getType());
574 for (unsigned i = 0, e = CS->getNumOperands(); i != e; ++i) {
575 DSNodeHandle NewNH(NH.getNode(), NH.getOffset()+SL->MemberOffsets[i]);
576 MergeConstantInitIntoNode(NewNH, cast<Constant>(CS->getOperand(i)));
579 assert(0 && "Unknown constant type!");
583 void GraphBuilder::mergeInGlobalInitializer(GlobalVariable *GV) {
584 assert(!GV->isExternal() && "Cannot merge in external global!");
585 // Get a node handle to the global node and merge the initializer into it.
586 DSNodeHandle NH = getValueDest(*GV);
587 MergeConstantInitIntoNode(NH, GV->getInitializer());
591 bool LocalDataStructures::run(Module &M) {
592 GlobalsGraph = new DSGraph(getAnalysis<TargetData>());
594 const TargetData &TD = getAnalysis<TargetData>();
596 // Calculate all of the graphs...
597 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
598 if (!I->isExternal())
599 DSInfo.insert(std::make_pair(I, new DSGraph(TD, *I, GlobalsGraph)));
601 GraphBuilder GGB(*GlobalsGraph);
603 // Add initializers for all of the globals to the globals graph...
604 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I)
605 if (!I->isExternal())
606 GGB.mergeInGlobalInitializer(I);
608 GlobalsGraph->markIncompleteNodes(DSGraph::MarkFormalArgs);
609 GlobalsGraph->removeTriviallyDeadNodes();
613 // releaseMemory - If the pass pipeline is done with this pass, we can release
614 // our memory... here...
616 void LocalDataStructures::releaseMemory() {
617 for (hash_map<Function*, DSGraph*>::iterator I = DSInfo.begin(),
618 E = DSInfo.end(); I != E; ++I) {
619 I->second->getReturnNodes().erase(I->first);
620 if (I->second->getReturnNodes().empty())
624 // Empty map so next time memory is released, data structures are not