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"
31 static RegisterAnalysis<LocalDataStructures>
32 X("datastructure", "Local Data Structure Analysis");
35 // isPointerType - Return true if this type is big enough to hold a pointer.
36 bool isPointerType(const Type *Ty) {
37 if (isa<PointerType>(Ty))
39 else if (Ty->isPrimitiveType() && Ty->isInteger())
40 return Ty->getPrimitiveSize() >= PointerSize;
49 DisableDirectCallOpt("disable-direct-call-dsopt", cl::Hidden,
50 cl::desc("Disable direct call optimization in "
51 "DSGraph construction"));
53 DisableFieldSensitivity("disable-ds-field-sensitivity", cl::Hidden,
54 cl::desc("Disable field sensitivity in DSGraphs"));
56 //===--------------------------------------------------------------------===//
58 //===--------------------------------------------------------------------===//
60 /// This class is the builder class that constructs the local data structure
61 /// graph by performing a single pass over the function in question.
63 class GraphBuilder : InstVisitor<GraphBuilder> {
65 DSNodeHandle *RetNode; // Node that gets returned...
66 DSGraph::ScalarMapTy &ScalarMap;
67 std::vector<DSCallSite> *FunctionCalls;
70 GraphBuilder(Function &f, DSGraph &g, DSNodeHandle &retNode,
71 std::vector<DSCallSite> &fc)
72 : G(g), RetNode(&retNode), ScalarMap(G.getScalarMap()),
75 // Create scalar nodes for all pointer arguments...
76 for (Function::aiterator I = f.abegin(), E = f.aend(); I != E; ++I)
77 if (isPointerType(I->getType()))
80 visit(f); // Single pass over the function
83 // GraphBuilder ctor for working on the globals graph
84 GraphBuilder(DSGraph &g)
85 : G(g), RetNode(0), ScalarMap(G.getScalarMap()), FunctionCalls(0) {
88 void mergeInGlobalInitializer(GlobalVariable *GV);
91 // Visitor functions, used to handle each instruction type we encounter...
92 friend class InstVisitor<GraphBuilder>;
93 void visitMallocInst(MallocInst &MI) { handleAlloc(MI, true); }
94 void visitAllocaInst(AllocaInst &AI) { handleAlloc(AI, false); }
95 void handleAlloc(AllocationInst &AI, bool isHeap);
97 void visitPHINode(PHINode &PN);
99 void visitGetElementPtrInst(User &GEP);
100 void visitReturnInst(ReturnInst &RI);
101 void visitLoadInst(LoadInst &LI);
102 void visitStoreInst(StoreInst &SI);
103 void visitCallInst(CallInst &CI);
104 void visitInvokeInst(InvokeInst &II);
105 void visitSetCondInst(SetCondInst &SCI) {} // SetEQ & friends are ignored
106 void visitFreeInst(FreeInst &FI);
107 void visitCastInst(CastInst &CI);
108 void visitInstruction(Instruction &I);
110 void visitCallSite(CallSite CS);
112 void MergeConstantInitIntoNode(DSNodeHandle &NH, Constant *C);
114 // Helper functions used to implement the visitation functions...
116 /// createNode - Create a new DSNode, ensuring that it is properly added to
119 DSNode *createNode(const Type *Ty = 0) {
120 DSNode *N = new DSNode(Ty, &G); // Create the node
121 if (DisableFieldSensitivity) {
122 N->foldNodeCompletely();
123 if (DSNode *FN = N->getForwardNode())
129 /// setDestTo - Set the ScalarMap entry for the specified value to point to
130 /// the specified destination. If the Value already points to a node, make
131 /// sure to merge the two destinations together.
133 void setDestTo(Value &V, const DSNodeHandle &NH);
135 /// getValueDest - Return the DSNode that the actual value points to.
137 DSNodeHandle getValueDest(Value &V);
139 /// getLink - This method is used to return the specified link in the
140 /// specified node if one exists. If a link does not already exist (it's
141 /// null), then we create a new node, link it, then return it.
143 DSNodeHandle &getLink(const DSNodeHandle &Node, unsigned Link = 0);
147 //===----------------------------------------------------------------------===//
148 // DSGraph constructor - Simply use the GraphBuilder to construct the local
150 DSGraph::DSGraph(const TargetData &td, Function &F, DSGraph *GG)
151 : GlobalsGraph(GG), TD(td) {
152 PrintAuxCalls = false;
154 DEBUG(std::cerr << " [Loc] Calculating graph for: " << F.getName() << "\n");
156 // Use the graph builder to construct the local version of the graph
157 GraphBuilder B(F, *this, ReturnNodes[&F], FunctionCalls);
159 Timer::addPeakMemoryMeasurement();
162 // Remove all integral constants from the scalarmap!
163 for (ScalarMapTy::iterator I = ScalarMap.begin(); I != ScalarMap.end();)
164 if (isa<ConstantIntegral>(I->first)) {
165 ScalarMapTy::iterator J = I++;
170 markIncompleteNodes(DSGraph::MarkFormalArgs);
172 // Remove any nodes made dead due to merging...
173 removeDeadNodes(DSGraph::KeepUnreachableGlobals);
177 //===----------------------------------------------------------------------===//
178 // Helper method implementations...
181 /// getValueDest - Return the DSNode that the actual value points to.
183 DSNodeHandle GraphBuilder::getValueDest(Value &Val) {
185 if (V == Constant::getNullValue(V->getType()))
186 return 0; // Null doesn't point to anything, don't add to ScalarMap!
188 DSNodeHandle &NH = ScalarMap[V];
190 return NH; // Already have a node? Just return it...
192 // Otherwise we need to create a new node to point to.
193 // Check first for constant expressions that must be traversed to
194 // extract the actual value.
195 if (Constant *C = dyn_cast<Constant>(V))
196 if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(C)) {
197 return NH = getValueDest(*CPR->getValue());
198 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
199 if (CE->getOpcode() == Instruction::Cast)
200 NH = getValueDest(*CE->getOperand(0));
201 else if (CE->getOpcode() == Instruction::GetElementPtr) {
202 visitGetElementPtrInst(*CE);
203 DSGraph::ScalarMapTy::iterator I = ScalarMap.find(CE);
204 assert(I != ScalarMap.end() && "GEP didn't get processed right?");
207 // This returns a conservative unknown node for any unhandled ConstExpr
208 return NH = createNode()->setUnknownNodeMarker();
210 if (NH.getNode() == 0) { // (getelementptr null, X) returns null
216 } else if (ConstantIntegral *CI = dyn_cast<ConstantIntegral>(C)) {
217 // Random constants are unknown mem
218 return NH = createNode()->setUnknownNodeMarker();
220 assert(0 && "Unknown constant type!");
223 // Otherwise we need to create a new node to point to...
225 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
226 // Create a new global node for this global variable...
227 N = createNode(GV->getType()->getElementType());
230 // Otherwise just create a shadow node
234 NH.setNode(N); // Remember that we are pointing to it...
240 /// getLink - This method is used to return the specified link in the
241 /// specified node if one exists. If a link does not already exist (it's
242 /// null), then we create a new node, link it, then return it. We must
243 /// specify the type of the Node field we are accessing so that we know what
244 /// type should be linked to if we need to create a new node.
246 DSNodeHandle &GraphBuilder::getLink(const DSNodeHandle &node, unsigned LinkNo) {
247 DSNodeHandle &Node = const_cast<DSNodeHandle&>(node);
248 DSNodeHandle &Link = Node.getLink(LinkNo);
249 if (!Link.getNode()) {
250 // If the link hasn't been created yet, make and return a new shadow node
257 /// setDestTo - Set the ScalarMap entry for the specified value to point to the
258 /// specified destination. If the Value already points to a node, make sure to
259 /// merge the two destinations together.
261 void GraphBuilder::setDestTo(Value &V, const DSNodeHandle &NH) {
262 DSNodeHandle &AINH = ScalarMap[&V];
263 if (AINH.getNode() == 0) // Not pointing to anything yet?
264 AINH = NH; // Just point directly to NH
270 //===----------------------------------------------------------------------===//
271 // Specific instruction type handler implementations...
274 /// Alloca & Malloc instruction implementation - Simply create a new memory
275 /// object, pointing the scalar to it.
277 void GraphBuilder::handleAlloc(AllocationInst &AI, bool isHeap) {
278 DSNode *N = createNode();
280 N->setHeapNodeMarker();
282 N->setAllocaNodeMarker();
286 // PHINode - Make the scalar for the PHI node point to all of the things the
287 // incoming values point to... which effectively causes them to be merged.
289 void GraphBuilder::visitPHINode(PHINode &PN) {
290 if (!isPointerType(PN.getType())) return; // Only pointer PHIs
292 DSNodeHandle &PNDest = ScalarMap[&PN];
293 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
294 PNDest.mergeWith(getValueDest(*PN.getIncomingValue(i)));
297 void GraphBuilder::visitGetElementPtrInst(User &GEP) {
298 DSNodeHandle Value = getValueDest(*GEP.getOperand(0));
299 if (Value.getNode() == 0) return;
301 const PointerType *PTy = cast<PointerType>(GEP.getOperand(0)->getType());
302 const Type *CurTy = PTy->getElementType();
304 if (Value.getNode()->mergeTypeInfo(CurTy, Value.getOffset())) {
305 // If the node had to be folded... exit quickly
306 setDestTo(GEP, Value); // GEP result points to folded node
310 const TargetData &TD = Value.getNode()->getTargetData();
313 // Handle the pointer index specially...
314 if (GEP.getNumOperands() > 1 &&
315 GEP.getOperand(1) != ConstantSInt::getNullValue(Type::LongTy)) {
317 // If we already know this is an array being accessed, don't do anything...
318 if (!TopTypeRec.isArray) {
319 TopTypeRec.isArray = true;
321 // If we are treating some inner field pointer as an array, fold the node
322 // up because we cannot handle it right. This can come because of
323 // something like this: &((&Pt->X)[1]) == &Pt->Y
325 if (Value.getOffset()) {
326 // Value is now the pointer we want to GEP to be...
327 Value.getNode()->foldNodeCompletely();
328 setDestTo(GEP, Value); // GEP result points to folded node
331 // This is a pointer to the first byte of the node. Make sure that we
332 // are pointing to the outter most type in the node.
333 // FIXME: We need to check one more case here...
339 // All of these subscripts are indexing INTO the elements we have...
341 for (unsigned i = 2, e = GEP.getNumOperands(); i < e; ++i)
342 if (GEP.getOperand(i)->getType() == Type::LongTy) {
343 // Get the type indexing into...
344 const SequentialType *STy = cast<SequentialType>(CurTy);
345 CurTy = STy->getElementType();
347 if (ConstantSInt *CS = dyn_cast<ConstantSInt>(GEP.getOperand(i))) {
348 Offset += CS->getValue()*TD.getTypeSize(CurTy);
350 // Variable index into a node. We must merge all of the elements of the
351 // sequential type here.
352 if (isa<PointerType>(STy))
353 std::cerr << "Pointer indexing not handled yet!\n";
355 const ArrayType *ATy = cast<ArrayType>(STy);
356 unsigned ElSize = TD.getTypeSize(CurTy);
357 DSNode *N = Value.getNode();
358 assert(N && "Value must have a node!");
359 unsigned RawOffset = Offset+Value.getOffset();
361 // Loop over all of the elements of the array, merging them into the
363 for (unsigned i = 1, e = ATy->getNumElements(); i != e; ++i)
364 // Merge all of the byte components of this array element
365 for (unsigned j = 0; j != ElSize; ++j)
366 N->mergeIndexes(RawOffset+j, RawOffset+i*ElSize+j);
370 } else if (GEP.getOperand(i)->getType() == Type::UByteTy) {
371 unsigned FieldNo = cast<ConstantUInt>(GEP.getOperand(i))->getValue();
372 const StructType *STy = cast<StructType>(CurTy);
373 Offset += TD.getStructLayout(STy)->MemberOffsets[FieldNo];
374 CurTy = STy->getContainedType(FieldNo);
377 // Add in the offset calculated...
378 Value.setOffset(Value.getOffset()+Offset);
380 // Value is now the pointer we want to GEP to be...
381 setDestTo(GEP, Value);
384 void GraphBuilder::visitLoadInst(LoadInst &LI) {
385 DSNodeHandle Ptr = getValueDest(*LI.getOperand(0));
386 if (Ptr.getNode() == 0) return;
388 // Make that the node is read from...
389 Ptr.getNode()->setReadMarker();
391 // Ensure a typerecord exists...
392 Ptr.getNode()->mergeTypeInfo(LI.getType(), Ptr.getOffset(), false);
394 if (isPointerType(LI.getType()))
395 setDestTo(LI, getLink(Ptr));
398 void GraphBuilder::visitStoreInst(StoreInst &SI) {
399 const Type *StoredTy = SI.getOperand(0)->getType();
400 DSNodeHandle Dest = getValueDest(*SI.getOperand(1));
401 if (Dest.getNode() == 0) return;
403 // Mark that the node is written to...
404 Dest.getNode()->setModifiedMarker();
406 // Ensure a type-record exists...
407 Dest.getNode()->mergeTypeInfo(StoredTy, Dest.getOffset());
409 // Avoid adding edges from null, or processing non-"pointer" stores
410 if (isPointerType(StoredTy))
411 Dest.addEdgeTo(getValueDest(*SI.getOperand(0)));
414 void GraphBuilder::visitReturnInst(ReturnInst &RI) {
415 if (RI.getNumOperands() && isPointerType(RI.getOperand(0)->getType()))
416 RetNode->mergeWith(getValueDest(*RI.getOperand(0)));
419 void GraphBuilder::visitCallInst(CallInst &CI) {
423 void GraphBuilder::visitInvokeInst(InvokeInst &II) {
427 void GraphBuilder::visitCallSite(CallSite CS) {
428 // Special case handling of certain libc allocation functions here.
429 if (Function *F = CS.getCalledFunction())
431 if (F->getName() == "calloc") {
432 setDestTo(*CS.getInstruction(),
433 createNode()->setHeapNodeMarker()->setModifiedMarker());
435 } else if (F->getName() == "realloc") {
436 DSNodeHandle RetNH = getValueDest(*CS.getInstruction());
437 RetNH.mergeWith(getValueDest(**CS.arg_begin()));
438 if (DSNode *N = RetNH.getNode())
439 N->setHeapNodeMarker()->setModifiedMarker()->setReadMarker();
441 } else if (F->getName() == "memset") {
442 // Merge the first argument with the return value, and mark the memory
444 DSNodeHandle RetNH = getValueDest(*CS.getInstruction());
445 RetNH.mergeWith(getValueDest(**CS.arg_begin()));
446 if (DSNode *N = RetNH.getNode())
447 N->setModifiedMarker();
449 } else if (F->getName() == "bzero") {
450 // Mark the memory modified.
451 DSNodeHandle H = getValueDest(**CS.arg_begin());
452 if (DSNode *N = H.getNode())
453 N->setModifiedMarker();
458 // Set up the return value...
460 Instruction *I = CS.getInstruction();
461 if (isPointerType(I->getType()))
462 RetVal = getValueDest(*I);
465 if (DisableDirectCallOpt || !isa<Function>(CS.getCalledValue())) {
466 Callee = getValueDest(*CS.getCalledValue()).getNode();
468 std::cerr << "WARNING: Program is calling through a null pointer?\n"
470 return; // Calling a null pointer?
474 std::vector<DSNodeHandle> Args;
475 Args.reserve(CS.arg_end()-CS.arg_begin());
477 // Calculate the arguments vector...
478 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); I != E; ++I)
479 if (isPointerType((*I)->getType()))
480 Args.push_back(getValueDest(**I));
482 // Add a new function call entry...
484 FunctionCalls->push_back(DSCallSite(CS, RetVal, Callee, Args));
486 FunctionCalls->push_back(DSCallSite(CS, RetVal, CS.getCalledFunction(),
490 void GraphBuilder::visitFreeInst(FreeInst &FI) {
491 // Mark that the node is written to...
492 DSNode *N = getValueDest(*FI.getOperand(0)).getNode();
493 N->setModifiedMarker();
494 N->setHeapNodeMarker();
498 void GraphBuilder::visitCastInst(CastInst &CI) {
499 if (isPointerType(CI.getType()))
500 if (isPointerType(CI.getOperand(0)->getType())) {
501 // Cast one pointer to the other, just act like a copy instruction
502 setDestTo(CI, getValueDest(*CI.getOperand(0)));
504 // Cast something (floating point, small integer) to a pointer. We need
505 // to track the fact that the node points to SOMETHING, just something we
506 // don't know about. Make an "Unknown" node.
508 setDestTo(CI, createNode()->setUnknownNodeMarker());
513 // visitInstruction - For all other instruction types, if we have any arguments
514 // that are of pointer type, make them have unknown composition bits, and merge
515 // the nodes together.
516 void GraphBuilder::visitInstruction(Instruction &Inst) {
517 DSNodeHandle CurNode;
518 if (isPointerType(Inst.getType()))
519 CurNode = getValueDest(Inst);
520 for (User::op_iterator I = Inst.op_begin(), E = Inst.op_end(); I != E; ++I)
521 if (isPointerType((*I)->getType()))
522 CurNode.mergeWith(getValueDest(**I));
524 if (CurNode.getNode())
525 CurNode.getNode()->setUnknownNodeMarker();
530 //===----------------------------------------------------------------------===//
531 // LocalDataStructures Implementation
532 //===----------------------------------------------------------------------===//
534 // MergeConstantInitIntoNode - Merge the specified constant into the node
536 void GraphBuilder::MergeConstantInitIntoNode(DSNodeHandle &NH, Constant *C) {
537 // Ensure a type-record exists...
538 NH.getNode()->mergeTypeInfo(C->getType(), NH.getOffset());
540 if (C->getType()->isFirstClassType()) {
541 if (isPointerType(C->getType()))
542 // Avoid adding edges from null, or processing non-"pointer" stores
543 NH.addEdgeTo(getValueDest(*C));
547 const TargetData &TD = NH.getNode()->getTargetData();
549 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) {
550 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
551 // We don't currently do any indexing for arrays...
552 MergeConstantInitIntoNode(NH, cast<Constant>(CA->getOperand(i)));
553 } else if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
554 const StructLayout *SL = TD.getStructLayout(CS->getType());
555 for (unsigned i = 0, e = CS->getNumOperands(); i != e; ++i) {
556 DSNodeHandle NewNH(NH.getNode(), NH.getOffset()+SL->MemberOffsets[i]);
557 MergeConstantInitIntoNode(NewNH, cast<Constant>(CS->getOperand(i)));
560 assert(0 && "Unknown constant type!");
564 void GraphBuilder::mergeInGlobalInitializer(GlobalVariable *GV) {
565 assert(!GV->isExternal() && "Cannot merge in external global!");
566 // Get a node handle to the global node and merge the initializer into it.
567 DSNodeHandle NH = getValueDest(*GV);
568 MergeConstantInitIntoNode(NH, GV->getInitializer());
572 bool LocalDataStructures::run(Module &M) {
573 GlobalsGraph = new DSGraph(getAnalysis<TargetData>());
575 const TargetData &TD = getAnalysis<TargetData>();
577 // Calculate all of the graphs...
578 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
579 if (!I->isExternal())
580 DSInfo.insert(std::make_pair(I, new DSGraph(TD, *I, GlobalsGraph)));
582 GraphBuilder GGB(*GlobalsGraph);
584 // Add initializers for all of the globals to the globals graph...
585 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I)
586 if (!I->isExternal())
587 GGB.mergeInGlobalInitializer(I);
589 GlobalsGraph->markIncompleteNodes(DSGraph::MarkFormalArgs);
590 GlobalsGraph->removeTriviallyDeadNodes();
594 // releaseMemory - If the pass pipeline is done with this pass, we can release
595 // our memory... here...
597 void LocalDataStructures::releaseMemory() {
598 for (hash_map<Function*, DSGraph*>::iterator I = DSInfo.begin(),
599 E = DSInfo.end(); I != E; ++I) {
600 I->second->getReturnNodes().erase(I->first);
601 if (I->second->getReturnNodes().empty())
605 // Empty map so next time memory is released, data structures are not