1 //===- TopDownClosure.cpp - Compute the top-down interprocedure closure ---===//
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 // This file implements the TDDataStructures class, which represents the
11 // Top-down Interprocedural closure of the data structure graph over the
12 // program. This is useful (but not strictly necessary?) for applications
13 // like pointer analysis.
15 //===----------------------------------------------------------------------===//
17 #include "llvm/Analysis/DataStructure/DataStructure.h"
18 #include "llvm/Module.h"
19 #include "llvm/DerivedTypes.h"
20 #include "llvm/Analysis/DataStructure/DSGraph.h"
21 #include "llvm/Support/Debug.h"
22 #include "llvm/ADT/Statistic.h"
26 RegisterAnalysis<TDDataStructures> // Register the pass
27 Y("tddatastructure", "Top-down Data Structure Analysis");
29 Statistic<> NumTDInlines("tddatastructures", "Number of graphs inlined");
32 void TDDataStructures::markReachableFunctionsExternallyAccessible(DSNode *N,
33 hash_set<DSNode*> &Visited) {
34 if (!N || Visited.count(N)) return;
37 for (unsigned i = 0, e = N->getNumLinks(); i != e; ++i) {
38 DSNodeHandle &NH = N->getLink(i*N->getPointerSize());
39 if (DSNode *NN = NH.getNode()) {
40 const std::vector<GlobalValue*> &Globals = NN->getGlobals();
41 for (unsigned G = 0, e = Globals.size(); G != e; ++G)
42 if (Function *F = dyn_cast<Function>(Globals[G]))
43 ArgsRemainIncomplete.insert(F);
45 markReachableFunctionsExternallyAccessible(NN, Visited);
51 // run - Calculate the top down data structure graphs for each function in the
54 bool TDDataStructures::runOnModule(Module &M) {
55 BUDataStructures &BU = getAnalysis<BUDataStructures>();
56 GlobalsGraph = new DSGraph(BU.getGlobalsGraph());
57 GlobalsGraph->setPrintAuxCalls();
59 // Figure out which functions must not mark their arguments complete because
60 // they are accessible outside this compilation unit. Currently, these
61 // arguments are functions which are reachable by global variables in the
63 const DSScalarMap &GGSM = GlobalsGraph->getScalarMap();
64 hash_set<DSNode*> Visited;
65 for (DSScalarMap::global_iterator I=GGSM.global_begin(), E=GGSM.global_end();
67 markReachableFunctionsExternallyAccessible(GGSM.find(*I)->second.getNode(),
70 // Loop over unresolved call nodes. Any functions passed into (but not
71 // returned!) from unresolvable call nodes may be invoked outside of the
73 for (DSGraph::afc_iterator I = GlobalsGraph->afc_begin(),
74 E = GlobalsGraph->afc_end(); I != E; ++I)
75 for (unsigned arg = 0, e = I->getNumPtrArgs(); arg != e; ++arg)
76 markReachableFunctionsExternallyAccessible(I->getPtrArg(arg).getNode(),
80 // Functions without internal linkage also have unknown incoming arguments!
81 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
82 if (!I->isExternal() && !I->hasInternalLinkage())
83 ArgsRemainIncomplete.insert(I);
85 // We want to traverse the call graph in reverse post-order. To do this, we
86 // calculate a post-order traversal, then reverse it.
87 hash_set<DSGraph*> VisitedGraph;
88 std::vector<DSGraph*> PostOrder;
89 const BUDataStructures::ActualCalleesTy &ActualCallees =
90 getAnalysis<BUDataStructures>().getActualCallees();
92 // Calculate top-down from main...
93 if (Function *F = M.getMainFunction())
94 ComputePostOrder(*F, VisitedGraph, PostOrder, ActualCallees);
96 // Next calculate the graphs for each unreachable function...
97 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
98 ComputePostOrder(*I, VisitedGraph, PostOrder, ActualCallees);
100 VisitedGraph.clear(); // Release memory!
102 // Visit each of the graphs in reverse post-order now!
103 while (!PostOrder.empty()) {
104 inlineGraphIntoCallees(*PostOrder.back());
105 PostOrder.pop_back();
108 ArgsRemainIncomplete.clear();
109 GlobalsGraph->removeTriviallyDeadNodes();
115 DSGraph &TDDataStructures::getOrCreateDSGraph(Function &F) {
116 DSGraph *&G = DSInfo[&F];
117 if (G == 0) { // Not created yet? Clone BU graph...
118 G = new DSGraph(getAnalysis<BUDataStructures>().getDSGraph(F));
119 G->getAuxFunctionCalls().clear();
120 G->setPrintAuxCalls();
121 G->setGlobalsGraph(GlobalsGraph);
127 void TDDataStructures::ComputePostOrder(Function &F,hash_set<DSGraph*> &Visited,
128 std::vector<DSGraph*> &PostOrder,
129 const BUDataStructures::ActualCalleesTy &ActualCallees) {
130 if (F.isExternal()) return;
131 DSGraph &G = getOrCreateDSGraph(F);
132 if (Visited.count(&G)) return;
135 // Recursively traverse all of the callee graphs.
136 for (DSGraph::fc_iterator CI = G.fc_begin(), E = G.fc_end(); CI != E; ++CI) {
137 Instruction *CallI = CI->getCallSite().getInstruction();
138 std::pair<BUDataStructures::ActualCalleesTy::const_iterator,
139 BUDataStructures::ActualCalleesTy::const_iterator>
140 IP = ActualCallees.equal_range(CallI);
142 for (BUDataStructures::ActualCalleesTy::const_iterator I = IP.first;
144 ComputePostOrder(*I->second, Visited, PostOrder, ActualCallees);
147 PostOrder.push_back(&G);
154 // releaseMemory - If the pass pipeline is done with this pass, we can release
155 // our memory... here...
157 // FIXME: This should be releaseMemory and will work fine, except that LoadVN
158 // has no way to extend the lifetime of the pass, which screws up ds-aa.
160 void TDDataStructures::releaseMyMemory() {
161 for (hash_map<Function*, DSGraph*>::iterator I = DSInfo.begin(),
162 E = DSInfo.end(); I != E; ++I) {
163 I->second->getReturnNodes().erase(I->first);
164 if (I->second->getReturnNodes().empty())
168 // Empty map so next time memory is released, data structures are not
175 void TDDataStructures::inlineGraphIntoCallees(DSGraph &Graph) {
176 // Recompute the Incomplete markers and eliminate unreachable nodes.
177 Graph.maskIncompleteMarkers();
179 // If any of the functions has incomplete incoming arguments, don't mark any
180 // of them as complete.
181 bool HasIncompleteArgs = false;
182 const DSGraph::ReturnNodesTy &GraphReturnNodes = Graph.getReturnNodes();
183 for (DSGraph::ReturnNodesTy::const_iterator I = GraphReturnNodes.begin(),
184 E = GraphReturnNodes.end(); I != E; ++I)
185 if (ArgsRemainIncomplete.count(I->first)) {
186 HasIncompleteArgs = true;
190 // Now fold in the necessary globals from the GlobalsGraph. A global G
191 // must be folded in if it exists in the current graph (i.e., is not dead)
192 // and it was not inlined from any of my callers. If it was inlined from
193 // a caller, it would have been fully consistent with the GlobalsGraph
194 // in the caller so folding in is not necessary. Otherwise, this node came
195 // solely from this function's BU graph and so has to be made consistent.
197 Graph.updateFromGlobalGraph();
199 // Recompute the Incomplete markers. Depends on whether args are complete
201 = HasIncompleteArgs ? DSGraph::MarkFormalArgs : DSGraph::IgnoreFormalArgs;
202 Graph.markIncompleteNodes(Flags | DSGraph::IgnoreGlobals);
204 // Delete dead nodes. Treat globals that are unreachable as dead also.
205 Graph.removeDeadNodes(DSGraph::RemoveUnreachableGlobals);
207 // We are done with computing the current TD Graph! Now move on to
208 // inlining the current graph into the graphs for its callees, if any.
210 if (Graph.fc_begin() == Graph.fc_end()) {
211 DEBUG(std::cerr << " [TD] No callees for: " << Graph.getFunctionNames()
216 // Now that we have information about all of the callees, propagate the
217 // current graph into the callees. Clone only the reachable subgraph at
218 // each call-site, not the entire graph (even though the entire graph
219 // would be cloned only once, this should still be better on average).
221 DEBUG(std::cerr << " [TD] Inlining '" << Graph.getFunctionNames() <<"' into "
222 << Graph.getFunctionCalls().size() << " call nodes.\n");
224 const BUDataStructures::ActualCalleesTy &ActualCallees =
225 getAnalysis<BUDataStructures>().getActualCallees();
227 // Loop over all the call sites and all the callees at each call site. Build
228 // a mapping from called DSGraph's to the call sites in this function that
229 // invoke them. This is useful because we can be more efficient if there are
230 // multiple call sites to the callees in the graph from this caller.
231 std::multimap<DSGraph*, std::pair<Function*, const DSCallSite*> > CallSites;
233 for (DSGraph::fc_iterator CI = Graph.fc_begin(), E = Graph.fc_end();
235 Instruction *CallI = CI->getCallSite().getInstruction();
236 // For each function in the invoked function list at this call site...
237 std::pair<BUDataStructures::ActualCalleesTy::const_iterator,
238 BUDataStructures::ActualCalleesTy::const_iterator>
239 IP = ActualCallees.equal_range(CallI);
240 // Loop over each actual callee at this call site
241 for (BUDataStructures::ActualCalleesTy::const_iterator I = IP.first;
242 I != IP.second; ++I) {
243 DSGraph& CalleeGraph = getDSGraph(*I->second);
244 if (&CalleeGraph != &Graph)
245 CallSites.insert(std::make_pair(&CalleeGraph,
246 std::make_pair(I->second, &*CI)));
250 // Now that we built the mapping, actually perform the inlining a callee graph
252 std::multimap<DSGraph*,std::pair<Function*,const DSCallSite*> >::iterator CSI;
253 for (CSI = CallSites.begin(); CSI != CallSites.end(); ) {
254 DSGraph &CalleeGraph = *CSI->first;
255 // Iterate through all of the call sites of this graph, cloning and merging
256 // any nodes required by the call.
257 ReachabilityCloner RC(CalleeGraph, Graph, DSGraph::StripModRefBits);
259 // Clone over any global nodes that appear in both graphs.
260 for (DSScalarMap::global_iterator
261 SI = CalleeGraph.getScalarMap().global_begin(),
262 SE = CalleeGraph.getScalarMap().global_end(); SI != SE; ++SI) {
263 DSScalarMap::const_iterator GI = Graph.getScalarMap().find(*SI);
264 if (GI != Graph.getScalarMap().end())
265 RC.merge(CalleeGraph.getNodeForValue(*SI), GI->second);
268 // Loop over all of the distinct call sites in the caller of the callee.
269 for (; CSI != CallSites.end() && CSI->first == &CalleeGraph; ++CSI) {
270 Function &CF = *CSI->second.first;
271 const DSCallSite &CS = *CSI->second.second;
272 DEBUG(std::cerr << " [TD] Resolving arguments for callee graph '"
273 << CalleeGraph.getFunctionNames()
274 << "': " << CF.getFunctionType()->getNumParams()
275 << " args\n at call site (DSCallSite*) 0x" << &CS << "\n");
277 // Get the formal argument and return nodes for the called function and
278 // merge them with the cloned subgraph.
279 RC.mergeCallSite(CalleeGraph.getCallSiteForArguments(CF), CS);
284 DEBUG(std::cerr << " [TD] Done inlining into callees for: "
285 << Graph.getFunctionNames() << " [" << Graph.getGraphSize() << "+"
286 << Graph.getFunctionCalls().size() << "]\n");
289 static const Function *getFnForValue(const Value *V) {
290 if (const Instruction *I = dyn_cast<Instruction>(V))
291 return I->getParent()->getParent();
292 else if (const Argument *A = dyn_cast<Argument>(V))
293 return A->getParent();
294 else if (const BasicBlock *BB = dyn_cast<BasicBlock>(V))
295 return BB->getParent();
299 void TDDataStructures::deleteValue(Value *V) {
300 if (const Function *F = getFnForValue(V)) { // Function local value?
301 // If this is a function local value, just delete it from the scalar map!
302 getDSGraph(*F).getScalarMap().eraseIfExists(V);
306 if (Function *F = dyn_cast<Function>(V)) {
307 assert(getDSGraph(*F).getReturnNodes().size() == 1 &&
308 "cannot handle scc's");
314 assert(!isa<GlobalVariable>(V) && "Do not know how to delete GV's yet!");
317 void TDDataStructures::copyValue(Value *From, Value *To) {
318 if (From == To) return;
319 if (const Function *F = getFnForValue(From)) { // Function local value?
320 // If this is a function local value, just delete it from the scalar map!
321 getDSGraph(*F).getScalarMap().copyScalarIfExists(From, To);
325 if (Function *FromF = dyn_cast<Function>(From)) {
326 Function *ToF = cast<Function>(To);
327 assert(!DSInfo.count(ToF) && "New Function already exists!");
328 DSGraph *NG = new DSGraph(getDSGraph(*FromF));
330 assert(NG->getReturnNodes().size() == 1 && "Cannot copy SCC's yet!");
332 // Change the Function* is the returnnodes map to the ToF.
333 DSNodeHandle Ret = NG->getReturnNodes().begin()->second;
334 NG->getReturnNodes().clear();
335 NG->getReturnNodes()[ToF] = Ret;
339 assert(!isa<GlobalVariable>(From) && "Do not know how to copy GV's yet!");