1 //===- Local.cpp - Compute a local data structure graph for a function ----===//
3 // Compute the local version of the data structure graph for a function. The
4 // external interface to this file is the DSGraph constructor.
6 //===----------------------------------------------------------------------===//
8 #include "llvm/Analysis/DataStructure.h"
9 #include "llvm/Analysis/DSGraph.h"
10 #include "llvm/Constants.h"
11 #include "llvm/DerivedTypes.h"
12 #include "llvm/Function.h"
13 #include "llvm/GlobalVariable.h"
14 #include "llvm/Instructions.h"
15 #include "llvm/Support/InstVisitor.h"
16 #include "llvm/Target/TargetData.h"
17 #include "Support/CommandLine.h"
18 #include "Support/Debug.h"
19 #include "Support/Timer.h"
21 // FIXME: This should eventually be a FunctionPass that is automatically
22 // aggregated into a Pass.
24 #include "llvm/Module.h"
26 static RegisterAnalysis<LocalDataStructures>
27 X("datastructure", "Local Data Structure Analysis");
30 // FIXME: Do something smarter with target data!
31 TargetData TD("temp-td");
33 // isPointerType - Return true if this type is big enough to hold a pointer.
34 bool isPointerType(const Type *Ty) {
35 if (isa<PointerType>(Ty))
37 else if (Ty->isPrimitiveType() && Ty->isInteger())
38 return Ty->getPrimitiveSize() >= PointerSize;
47 DisableDirectCallOpt("disable-direct-call-dsopt", cl::Hidden,
48 cl::desc("Disable direct call optimization in "
49 "DSGraph construction"));
51 DisableFieldSensitivity("disable-ds-field-sensitivity", cl::Hidden,
52 cl::desc("Disable field sensitivity in DSGraphs"));
54 //===--------------------------------------------------------------------===//
56 //===--------------------------------------------------------------------===//
58 /// This class is the builder class that constructs the local data structure
59 /// graph by performing a single pass over the function in question.
61 class GraphBuilder : InstVisitor<GraphBuilder> {
63 DSNodeHandle *RetNode; // Node that gets returned...
64 DSGraph::ScalarMapTy &ScalarMap;
65 std::vector<DSCallSite> *FunctionCalls;
68 GraphBuilder(Function &f, DSGraph &g, DSNodeHandle &retNode,
69 std::vector<DSCallSite> &fc)
70 : G(g), RetNode(&retNode), ScalarMap(G.getScalarMap()),
73 // Create scalar nodes for all pointer arguments...
74 for (Function::aiterator I = f.abegin(), E = f.aend(); I != E; ++I)
75 if (isPointerType(I->getType()))
78 visit(f); // Single pass over the function
81 // GraphBuilder ctor for working on the globals graph
82 GraphBuilder(DSGraph &g)
83 : G(g), RetNode(0), ScalarMap(G.getScalarMap()), FunctionCalls(0) {
86 void mergeInGlobalInitializer(GlobalVariable *GV);
89 // Visitor functions, used to handle each instruction type we encounter...
90 friend class InstVisitor<GraphBuilder>;
91 void visitMallocInst(MallocInst &MI) { handleAlloc(MI, true); }
92 void visitAllocaInst(AllocaInst &AI) { handleAlloc(AI, false); }
93 void handleAlloc(AllocationInst &AI, bool isHeap);
95 void visitPHINode(PHINode &PN);
97 void visitGetElementPtrInst(User &GEP);
98 void visitReturnInst(ReturnInst &RI);
99 void visitLoadInst(LoadInst &LI);
100 void visitStoreInst(StoreInst &SI);
101 void visitCallInst(CallInst &CI);
102 void visitInvokeInst(InvokeInst &II);
103 void visitSetCondInst(SetCondInst &SCI) {} // SetEQ & friends are ignored
104 void visitFreeInst(FreeInst &FI);
105 void visitCastInst(CastInst &CI);
106 void visitInstruction(Instruction &I);
108 void visitCallSite(CallSite CS);
110 void MergeConstantInitIntoNode(DSNodeHandle &NH, Constant *C);
112 // Helper functions used to implement the visitation functions...
114 /// createNode - Create a new DSNode, ensuring that it is properly added to
117 DSNode *createNode(const Type *Ty = 0) {
118 DSNode *N = new DSNode(Ty, &G); // Create the node
119 if (DisableFieldSensitivity) {
120 N->foldNodeCompletely();
121 if (DSNode *FN = N->getForwardNode())
127 /// setDestTo - Set the ScalarMap entry for the specified value to point to
128 /// the specified destination. If the Value already points to a node, make
129 /// sure to merge the two destinations together.
131 void setDestTo(Value &V, const DSNodeHandle &NH);
133 /// getValueDest - Return the DSNode that the actual value points to.
135 DSNodeHandle getValueDest(Value &V);
137 /// getLink - This method is used to return the specified link in the
138 /// specified node if one exists. If a link does not already exist (it's
139 /// null), then we create a new node, link it, then return it.
141 DSNodeHandle &getLink(const DSNodeHandle &Node, unsigned Link = 0);
145 //===----------------------------------------------------------------------===//
146 // DSGraph constructor - Simply use the GraphBuilder to construct the local
148 DSGraph::DSGraph(Function &F, DSGraph *GG) : GlobalsGraph(GG) {
149 PrintAuxCalls = false;
151 DEBUG(std::cerr << " [Loc] Calculating graph for: " << F.getName() << "\n");
153 // Use the graph builder to construct the local version of the graph
154 GraphBuilder B(F, *this, ReturnNodes[&F], FunctionCalls);
156 Timer::addPeakMemoryMeasurement();
159 // Remove all integral constants from the scalarmap!
160 for (ScalarMapTy::iterator I = ScalarMap.begin(); I != ScalarMap.end();)
161 if (isa<ConstantIntegral>(I->first)) {
162 ScalarMapTy::iterator J = I++;
167 markIncompleteNodes(DSGraph::MarkFormalArgs);
169 // Remove any nodes made dead due to merging...
170 removeDeadNodes(DSGraph::KeepUnreachableGlobals);
174 //===----------------------------------------------------------------------===//
175 // Helper method implementations...
178 /// getValueDest - Return the DSNode that the actual value points to.
180 DSNodeHandle GraphBuilder::getValueDest(Value &Val) {
182 if (V == Constant::getNullValue(V->getType()))
183 return 0; // Null doesn't point to anything, don't add to ScalarMap!
185 DSNodeHandle &NH = ScalarMap[V];
187 return NH; // Already have a node? Just return it...
189 // Otherwise we need to create a new node to point to.
190 // Check first for constant expressions that must be traversed to
191 // extract the actual value.
192 if (Constant *C = dyn_cast<Constant>(V))
193 if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(C)) {
194 return NH = getValueDest(*CPR->getValue());
195 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
196 if (CE->getOpcode() == Instruction::Cast)
197 NH = getValueDest(*CE->getOperand(0));
198 else if (CE->getOpcode() == Instruction::GetElementPtr) {
199 visitGetElementPtrInst(*CE);
200 DSGraph::ScalarMapTy::iterator I = ScalarMap.find(CE);
201 assert(I != ScalarMap.end() && "GEP didn't get processed right?");
204 // This returns a conservative unknown node for any unhandled ConstExpr
205 return NH = createNode()->setUnknownNodeMarker();
207 if (NH.getNode() == 0) { // (getelementptr null, X) returns null
213 } else if (ConstantIntegral *CI = dyn_cast<ConstantIntegral>(C)) {
214 // Random constants are unknown mem
215 return NH = createNode()->setUnknownNodeMarker();
217 assert(0 && "Unknown constant type!");
220 // Otherwise we need to create a new node to point to...
222 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
223 // Create a new global node for this global variable...
224 N = createNode(GV->getType()->getElementType());
227 // Otherwise just create a shadow node
231 NH.setNode(N); // Remember that we are pointing to it...
237 /// getLink - This method is used to return the specified link in the
238 /// specified node if one exists. If a link does not already exist (it's
239 /// null), then we create a new node, link it, then return it. We must
240 /// specify the type of the Node field we are accessing so that we know what
241 /// type should be linked to if we need to create a new node.
243 DSNodeHandle &GraphBuilder::getLink(const DSNodeHandle &node, unsigned LinkNo) {
244 DSNodeHandle &Node = const_cast<DSNodeHandle&>(node);
245 DSNodeHandle &Link = Node.getLink(LinkNo);
246 if (!Link.getNode()) {
247 // If the link hasn't been created yet, make and return a new shadow node
254 /// setDestTo - Set the ScalarMap entry for the specified value to point to the
255 /// specified destination. If the Value already points to a node, make sure to
256 /// merge the two destinations together.
258 void GraphBuilder::setDestTo(Value &V, const DSNodeHandle &NH) {
259 DSNodeHandle &AINH = ScalarMap[&V];
260 if (AINH.getNode() == 0) // Not pointing to anything yet?
261 AINH = NH; // Just point directly to NH
267 //===----------------------------------------------------------------------===//
268 // Specific instruction type handler implementations...
271 /// Alloca & Malloc instruction implementation - Simply create a new memory
272 /// object, pointing the scalar to it.
274 void GraphBuilder::handleAlloc(AllocationInst &AI, bool isHeap) {
275 DSNode *N = createNode();
277 N->setHeapNodeMarker();
279 N->setAllocaNodeMarker();
283 // PHINode - Make the scalar for the PHI node point to all of the things the
284 // incoming values point to... which effectively causes them to be merged.
286 void GraphBuilder::visitPHINode(PHINode &PN) {
287 if (!isPointerType(PN.getType())) return; // Only pointer PHIs
289 DSNodeHandle &PNDest = ScalarMap[&PN];
290 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
291 PNDest.mergeWith(getValueDest(*PN.getIncomingValue(i)));
294 void GraphBuilder::visitGetElementPtrInst(User &GEP) {
295 DSNodeHandle Value = getValueDest(*GEP.getOperand(0));
296 if (Value.getNode() == 0) return;
298 const PointerType *PTy = cast<PointerType>(GEP.getOperand(0)->getType());
299 const Type *CurTy = PTy->getElementType();
301 if (Value.getNode()->mergeTypeInfo(CurTy, Value.getOffset())) {
302 // If the node had to be folded... exit quickly
303 setDestTo(GEP, Value); // GEP result points to folded node
308 // Handle the pointer index specially...
309 if (GEP.getNumOperands() > 1 &&
310 GEP.getOperand(1) != ConstantSInt::getNullValue(Type::LongTy)) {
312 // If we already know this is an array being accessed, don't do anything...
313 if (!TopTypeRec.isArray) {
314 TopTypeRec.isArray = true;
316 // If we are treating some inner field pointer as an array, fold the node
317 // up because we cannot handle it right. This can come because of
318 // something like this: &((&Pt->X)[1]) == &Pt->Y
320 if (Value.getOffset()) {
321 // Value is now the pointer we want to GEP to be...
322 Value.getNode()->foldNodeCompletely();
323 setDestTo(GEP, Value); // GEP result points to folded node
326 // This is a pointer to the first byte of the node. Make sure that we
327 // are pointing to the outter most type in the node.
328 // FIXME: We need to check one more case here...
334 // All of these subscripts are indexing INTO the elements we have...
336 for (unsigned i = 2, e = GEP.getNumOperands(); i < e; ++i)
337 if (GEP.getOperand(i)->getType() == Type::LongTy) {
338 // Get the type indexing into...
339 const SequentialType *STy = cast<SequentialType>(CurTy);
340 CurTy = STy->getElementType();
342 if (ConstantSInt *CS = dyn_cast<ConstantSInt>(GEP.getOperand(i))) {
343 Offset += CS->getValue()*TD.getTypeSize(CurTy);
345 // Variable index into a node. We must merge all of the elements of the
346 // sequential type here.
347 if (isa<PointerType>(STy))
348 std::cerr << "Pointer indexing not handled yet!\n";
350 const ArrayType *ATy = cast<ArrayType>(STy);
351 unsigned ElSize = TD.getTypeSize(CurTy);
352 DSNode *N = Value.getNode();
353 assert(N && "Value must have a node!");
354 unsigned RawOffset = Offset+Value.getOffset();
356 // Loop over all of the elements of the array, merging them into the
358 for (unsigned i = 1, e = ATy->getNumElements(); i != e; ++i)
359 // Merge all of the byte components of this array element
360 for (unsigned j = 0; j != ElSize; ++j)
361 N->mergeIndexes(RawOffset+j, RawOffset+i*ElSize+j);
365 } else if (GEP.getOperand(i)->getType() == Type::UByteTy) {
366 unsigned FieldNo = cast<ConstantUInt>(GEP.getOperand(i))->getValue();
367 const StructType *STy = cast<StructType>(CurTy);
368 Offset += TD.getStructLayout(STy)->MemberOffsets[FieldNo];
369 CurTy = STy->getContainedType(FieldNo);
372 // Add in the offset calculated...
373 Value.setOffset(Value.getOffset()+Offset);
375 // Value is now the pointer we want to GEP to be...
376 setDestTo(GEP, Value);
379 void GraphBuilder::visitLoadInst(LoadInst &LI) {
380 DSNodeHandle Ptr = getValueDest(*LI.getOperand(0));
381 if (Ptr.getNode() == 0) return;
383 // Make that the node is read from...
384 Ptr.getNode()->setReadMarker();
386 // Ensure a typerecord exists...
387 Ptr.getNode()->mergeTypeInfo(LI.getType(), Ptr.getOffset(), false);
389 if (isPointerType(LI.getType()))
390 setDestTo(LI, getLink(Ptr));
393 void GraphBuilder::visitStoreInst(StoreInst &SI) {
394 const Type *StoredTy = SI.getOperand(0)->getType();
395 DSNodeHandle Dest = getValueDest(*SI.getOperand(1));
396 if (Dest.getNode() == 0) return;
398 // Mark that the node is written to...
399 Dest.getNode()->setModifiedMarker();
401 // Ensure a type-record exists...
402 Dest.getNode()->mergeTypeInfo(StoredTy, Dest.getOffset());
404 // Avoid adding edges from null, or processing non-"pointer" stores
405 if (isPointerType(StoredTy))
406 Dest.addEdgeTo(getValueDest(*SI.getOperand(0)));
409 void GraphBuilder::visitReturnInst(ReturnInst &RI) {
410 if (RI.getNumOperands() && isPointerType(RI.getOperand(0)->getType()))
411 RetNode->mergeWith(getValueDest(*RI.getOperand(0)));
414 void GraphBuilder::visitCallInst(CallInst &CI) {
418 void GraphBuilder::visitInvokeInst(InvokeInst &II) {
422 void GraphBuilder::visitCallSite(CallSite CS) {
423 // Special case handling of certain libc allocation functions here.
424 if (Function *F = CS.getCalledFunction())
426 if (F->getName() == "calloc") {
427 setDestTo(*CS.getInstruction(),
428 createNode()->setHeapNodeMarker()->setModifiedMarker());
430 } else if (F->getName() == "realloc") {
431 DSNodeHandle RetNH = getValueDest(*CS.getInstruction());
432 RetNH.mergeWith(getValueDest(**CS.arg_begin()));
433 DSNode *N = RetNH.getNode();
434 if (N) N->setHeapNodeMarker()->setModifiedMarker()->setReadMarker();
439 // Set up the return value...
441 Instruction *I = CS.getInstruction();
442 if (isPointerType(I->getType()))
443 RetVal = getValueDest(*I);
446 if (DisableDirectCallOpt || !isa<Function>(CS.getCalledValue())) {
447 Callee = getValueDest(*CS.getCalledValue()).getNode();
449 std::cerr << "WARNING: Program is calling through a null pointer?\n"
451 return; // Calling a null pointer?
455 std::vector<DSNodeHandle> Args;
456 Args.reserve(CS.arg_end()-CS.arg_begin());
458 // Calculate the arguments vector...
459 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); I != E; ++I)
460 if (isPointerType((*I)->getType()))
461 Args.push_back(getValueDest(**I));
463 // Add a new function call entry...
465 FunctionCalls->push_back(DSCallSite(CS, RetVal, Callee, Args));
467 FunctionCalls->push_back(DSCallSite(CS, RetVal, CS.getCalledFunction(),
471 void GraphBuilder::visitFreeInst(FreeInst &FI) {
472 // Mark that the node is written to...
473 DSNode *N = getValueDest(*FI.getOperand(0)).getNode();
474 N->setModifiedMarker();
475 N->setHeapNodeMarker();
479 void GraphBuilder::visitCastInst(CastInst &CI) {
480 if (isPointerType(CI.getType()))
481 if (isPointerType(CI.getOperand(0)->getType())) {
482 // Cast one pointer to the other, just act like a copy instruction
483 setDestTo(CI, getValueDest(*CI.getOperand(0)));
485 // Cast something (floating point, small integer) to a pointer. We need
486 // to track the fact that the node points to SOMETHING, just something we
487 // don't know about. Make an "Unknown" node.
489 setDestTo(CI, createNode()->setUnknownNodeMarker());
494 // visitInstruction - For all other instruction types, if we have any arguments
495 // that are of pointer type, make them have unknown composition bits, and merge
496 // the nodes together.
497 void GraphBuilder::visitInstruction(Instruction &Inst) {
498 DSNodeHandle CurNode;
499 if (isPointerType(Inst.getType()))
500 CurNode = getValueDest(Inst);
501 for (User::op_iterator I = Inst.op_begin(), E = Inst.op_end(); I != E; ++I)
502 if (isPointerType((*I)->getType()))
503 CurNode.mergeWith(getValueDest(**I));
505 if (CurNode.getNode())
506 CurNode.getNode()->setUnknownNodeMarker();
511 //===----------------------------------------------------------------------===//
512 // LocalDataStructures Implementation
513 //===----------------------------------------------------------------------===//
515 // MergeConstantInitIntoNode - Merge the specified constant into the node
517 void GraphBuilder::MergeConstantInitIntoNode(DSNodeHandle &NH, Constant *C) {
518 // Ensure a type-record exists...
519 NH.getNode()->mergeTypeInfo(C->getType(), NH.getOffset());
521 if (C->getType()->isFirstClassType()) {
522 if (isPointerType(C->getType()))
523 // Avoid adding edges from null, or processing non-"pointer" stores
524 NH.addEdgeTo(getValueDest(*C));
528 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) {
529 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
530 // We don't currently do any indexing for arrays...
531 MergeConstantInitIntoNode(NH, cast<Constant>(CA->getOperand(i)));
532 } else if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
533 const StructLayout *SL = TD.getStructLayout(CS->getType());
534 for (unsigned i = 0, e = CS->getNumOperands(); i != e; ++i) {
535 DSNodeHandle NewNH(NH.getNode(), NH.getOffset()+SL->MemberOffsets[i]);
536 MergeConstantInitIntoNode(NewNH, cast<Constant>(CS->getOperand(i)));
539 assert(0 && "Unknown constant type!");
543 void GraphBuilder::mergeInGlobalInitializer(GlobalVariable *GV) {
544 assert(!GV->isExternal() && "Cannot merge in external global!");
545 // Get a node handle to the global node and merge the initializer into it.
546 DSNodeHandle NH = getValueDest(*GV);
547 MergeConstantInitIntoNode(NH, GV->getInitializer());
551 bool LocalDataStructures::run(Module &M) {
552 GlobalsGraph = new DSGraph();
554 // Calculate all of the graphs...
555 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
556 if (!I->isExternal())
557 DSInfo.insert(std::make_pair(I, new DSGraph(*I, GlobalsGraph)));
559 GraphBuilder GGB(*GlobalsGraph);
561 // Add initializers for all of the globals to the globals graph...
562 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I)
563 if (!I->isExternal())
564 GGB.mergeInGlobalInitializer(I);
566 GlobalsGraph->markIncompleteNodes(DSGraph::MarkFormalArgs);
567 GlobalsGraph->removeTriviallyDeadNodes();
571 // releaseMemory - If the pass pipeline is done with this pass, we can release
572 // our memory... here...
574 void LocalDataStructures::releaseMemory() {
575 for (hash_map<Function*, DSGraph*>::iterator I = DSInfo.begin(),
576 E = DSInfo.end(); I != E; ++I) {
577 I->second->getReturnNodes().erase(I->first);
578 if (I->second->getReturnNodes().empty())
582 // Empty map so next time memory is released, data structures are not