1 //===- EquivClassGraphs.cpp - Merge equiv-class graphs & inline bottom-up -===//
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 pass is the same as the complete bottom-up graphs, but
11 // with functions partitioned into equivalence classes and a single merged
12 // DS graph for all functions in an equivalence class. After this merging,
13 // graphs are inlined bottom-up on the SCCs of the final (CBU) call graph.
15 //===----------------------------------------------------------------------===//
17 #define DEBUG_TYPE "ECGraphs"
18 #include "llvm/Analysis/DataStructure/DataStructure.h"
19 #include "llvm/DerivedTypes.h"
20 #include "llvm/Module.h"
21 #include "llvm/Pass.h"
22 #include "llvm/Analysis/DataStructure/DSGraph.h"
23 #include "llvm/Support/CallSite.h"
24 #include "llvm/Support/Debug.h"
25 #include "llvm/ADT/SCCIterator.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/ADT/EquivalenceClasses.h"
28 #include "llvm/ADT/STLExtras.h"
33 RegisterAnalysis<EquivClassGraphs> X("eqdatastructure",
34 "Equivalence-class Bottom-up Data Structure Analysis");
35 Statistic<> NumEquivBUInlines("equivdatastructures",
36 "Number of graphs inlined");
37 Statistic<> NumFoldGraphInlines("Inline equiv-class graphs bottom up",
38 "Number of graphs inlined");
43 static void CheckAllGraphs(Module *M, GT &ECGraphs) {
44 DSGraph &GG = ECGraphs.getGlobalsGraph();
46 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I)
47 if (!I->isExternal()) {
48 DSGraph &G = ECGraphs.getDSGraph(*I);
49 if (G.retnodes_begin()->first != I)
50 continue; // Only check a graph once.
52 DSGraph::NodeMapTy GlobalsGraphNodeMapping;
53 G.computeGToGGMapping(GlobalsGraphNodeMapping);
58 // getSomeCalleeForCallSite - Return any one callee function at a call site.
60 Function *EquivClassGraphs::getSomeCalleeForCallSite(const CallSite &CS) const{
61 Function *thisFunc = CS.getCaller();
62 assert(thisFunc && "getSomeCalleeForCallSite(): Not a valid call site?");
63 DSGraph &DSG = getDSGraph(*thisFunc);
64 DSNode *calleeNode = DSG.getNodeForValue(CS.getCalledValue()).getNode();
65 std::map<DSNode*, Function *>::const_iterator I =
66 OneCalledFunction.find(calleeNode);
67 return (I == OneCalledFunction.end())? NULL : I->second;
70 // runOnModule - Calculate the bottom up data structure graphs for each function
73 bool EquivClassGraphs::runOnModule(Module &M) {
74 CBU = &getAnalysis<CompleteBUDataStructures>();
75 GlobalECs = CBU->getGlobalECs();
76 DEBUG(CheckAllGraphs(&M, *CBU));
78 GlobalsGraph = new DSGraph(CBU->getGlobalsGraph(), GlobalECs);
79 GlobalsGraph->setPrintAuxCalls();
81 ActualCallees = CBU->getActualCallees();
83 // Find equivalence classes of functions called from common call sites.
84 // Fold the CBU graphs for all functions in an equivalence class.
85 buildIndirectFunctionSets(M);
87 // Stack of functions used for Tarjan's SCC-finding algorithm.
88 std::vector<DSGraph*> Stack;
89 std::map<DSGraph*, unsigned> ValMap;
92 Function *MainFunc = M.getMainFunction();
93 if (MainFunc && !MainFunc->isExternal()) {
94 processSCC(getOrCreateGraph(*MainFunc), Stack, NextID, ValMap);
96 std::cerr << "Fold Graphs: No 'main' function found!\n";
99 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
100 if (!I->isExternal())
101 processSCC(getOrCreateGraph(*I), Stack, NextID, ValMap);
103 DEBUG(CheckAllGraphs(&M, *this));
105 getGlobalsGraph().removeTriviallyDeadNodes();
106 getGlobalsGraph().markIncompleteNodes(DSGraph::IgnoreGlobals);
108 // Merge the globals variables (not the calls) from the globals graph back
109 // into the main function's graph so that the main function contains all of
110 // the information about global pools and GV usage in the program.
111 if (MainFunc && !MainFunc->isExternal()) {
112 DSGraph &MainGraph = getOrCreateGraph(*MainFunc);
113 const DSGraph &GG = *MainGraph.getGlobalsGraph();
114 ReachabilityCloner RC(MainGraph, GG,
115 DSGraph::DontCloneCallNodes |
116 DSGraph::DontCloneAuxCallNodes);
118 // Clone the global nodes into this graph.
119 for (DSScalarMap::global_iterator I = GG.getScalarMap().global_begin(),
120 E = GG.getScalarMap().global_end(); I != E; ++I)
121 if (isa<GlobalVariable>(*I))
122 RC.getClonedNH(GG.getNodeForValue(*I));
124 MainGraph.maskIncompleteMarkers();
125 MainGraph.markIncompleteNodes(DSGraph::MarkFormalArgs |
126 DSGraph::IgnoreGlobals);
129 // Final processing. Note that dead node elimination may actually remove
130 // globals from a function graph that are immediately used. If there are no
131 // scalars pointing to the node (e.g. because the only use is a direct store
132 // to a scalar global) we have to make sure to rematerialize the globals back
133 // into the graphs here, or clients will break!
134 for (Module::global_iterator GI = M.global_begin(), E = M.global_end();
136 // This only happens to first class typed globals.
137 if (GI->getType()->getElementType()->isFirstClassType())
138 for (Value::use_iterator UI = GI->use_begin(), E = GI->use_end();
140 // This only happens to direct uses by instructions.
141 if (Instruction *User = dyn_cast<Instruction>(*UI)) {
142 DSGraph &DSG = getOrCreateGraph(*User->getParent()->getParent());
143 if (!DSG.getScalarMap().count(GI)) {
144 // If this global does not exist in the graph, but it is immediately
145 // used by an instruction in the graph, clone it over from the
147 ReachabilityCloner RC(DSG, *GlobalsGraph, 0);
148 RC.getClonedNH(GlobalsGraph->getNodeForValue(GI));
156 // buildIndirectFunctionSets - Iterate over the module looking for indirect
157 // calls to functions. If a call site can invoke any functions [F1, F2... FN],
158 // unify the N functions together in the FuncECs set.
160 void EquivClassGraphs::buildIndirectFunctionSets(Module &M) {
161 const ActualCalleesTy& AC = CBU->getActualCallees();
163 // Loop over all of the indirect calls in the program. If a call site can
164 // call multiple different functions, we need to unify all of the callees into
165 // the same equivalence class.
166 Instruction *LastInst = 0;
167 Function *FirstFunc = 0;
168 for (ActualCalleesTy::const_iterator I=AC.begin(), E=AC.end(); I != E; ++I) {
169 if (I->second->isExternal())
170 continue; // Ignore functions we cannot modify
172 CallSite CS = CallSite::get(I->first);
174 if (CS.getCalledFunction()) { // Direct call:
175 FuncECs.insert(I->second); // -- Make sure function has equiv class
176 FirstFunc = I->second; // -- First callee at this site
177 } else { // Else indirect call
178 // DEBUG(std::cerr << "CALLEE: " << I->second->getName()
179 // << " from : " << I->first);
180 if (I->first != LastInst) {
181 // This is the first callee from this call site.
183 FirstFunc = I->second;
184 // Instead of storing the lastInst For Indirection call Sites we store
185 // the DSNode for the function ptr arguemnt
186 Function *thisFunc = LastInst->getParent()->getParent();
187 DSGraph &TFG = CBU->getDSGraph(*thisFunc);
188 DSNode *calleeNode = TFG.getNodeForValue(CS.getCalledValue()).getNode();
189 OneCalledFunction[calleeNode] = FirstFunc;
190 FuncECs.insert(I->second);
192 // This is not the first possible callee from a particular call site.
193 // Union the callee in with the other functions.
194 FuncECs.unionSets(FirstFunc, I->second);
196 Function *thisFunc = LastInst->getParent()->getParent();
197 DSGraph &TFG = CBU->getDSGraph(*thisFunc);
198 DSNode *calleeNode = TFG.getNodeForValue(CS.getCalledValue()).getNode();
199 assert(OneCalledFunction.count(calleeNode) > 0 && "Missed a call?");
204 // Now include all functions that share a graph with any function in the
205 // equivalence class. More precisely, if F is in the class, and G(F) is
206 // its graph, then we include all other functions that are also in G(F).
207 // Currently, that is just the functions in the same call-graph-SCC as F.
209 DSGraph& funcDSGraph = CBU->getDSGraph(*I->second);
210 for (DSGraph::retnodes_iterator RI = funcDSGraph.retnodes_begin(),
211 RE = funcDSGraph.retnodes_end(); RI != RE; ++RI)
212 FuncECs.unionSets(FirstFunc, RI->first);
215 // Now that all of the equivalences have been built, merge the graphs for
216 // each equivalence class.
218 DEBUG(std::cerr << "\nIndirect Function Equivalence Sets:\n");
219 for (EquivalenceClasses<Function*>::iterator EQSI = FuncECs.begin(), E =
220 FuncECs.end(); EQSI != E; ++EQSI) {
221 if (!EQSI->isLeader()) continue;
223 EquivalenceClasses<Function*>::member_iterator SI =
224 FuncECs.member_begin(EQSI);
225 assert(SI != FuncECs.member_end() && "Empty equiv set??");
226 EquivalenceClasses<Function*>::member_iterator SN = SI;
228 if (SN == FuncECs.member_end())
229 continue; // Single function equivalence set, no merging to do.
234 DEBUG(std::cerr <<" Equivalence set for leader " << LF->getName() <<" = ");
235 for (SN = SI; SN != FuncECs.member_end(); ++SN)
236 DEBUG(std::cerr << " " << (*SN)->getName() << "," );
237 DEBUG(std::cerr << "\n");
240 // This equiv class has multiple functions: merge their graphs. First,
241 // clone the CBU graph for the leader and make it the common graph for the
242 // equivalence graph.
243 DSGraph &MergedG = getOrCreateGraph(*LF);
245 // Record the argument nodes for use in merging later below.
246 std::vector<DSNodeHandle> ArgNodes;
248 for (Function::arg_iterator AI = LF->arg_begin(), E = LF->arg_end();
250 if (DS::isPointerType(AI->getType()))
251 ArgNodes.push_back(MergedG.getNodeForValue(AI));
253 // Merge in the graphs of all other functions in this equiv. class. Note
254 // that two or more functions may have the same graph, and it only needs
255 // to be merged in once.
256 std::set<DSGraph*> GraphsMerged;
257 GraphsMerged.insert(&CBU->getDSGraph(*LF));
259 for (++SI; SI != FuncECs.member_end(); ++SI) {
261 DSGraph *&FG = DSInfo[F];
263 DSGraph &CBUGraph = CBU->getDSGraph(*F);
264 if (GraphsMerged.insert(&CBUGraph).second) {
265 // Record the "folded" graph for the function.
266 for (DSGraph::retnodes_iterator I = CBUGraph.retnodes_begin(),
267 E = CBUGraph.retnodes_end(); I != E; ++I) {
268 assert(DSInfo[I->first] == 0 && "Graph already exists for Fn!");
269 DSInfo[I->first] = &MergedG;
272 // Clone this member of the equivalence class into MergedG.
273 MergedG.cloneInto(CBUGraph);
276 // Merge the return nodes of all functions together.
277 MergedG.getReturnNodes()[LF].mergeWith(MergedG.getReturnNodes()[F]);
279 // Merge the function arguments with all argument nodes found so far.
280 // If there are extra function args, add them to the vector of argNodes
281 Function::arg_iterator AI2 = F->arg_begin(), AI2end = F->arg_end();
282 for (unsigned arg = 0, numArgs = ArgNodes.size();
283 arg != numArgs && AI2 != AI2end; ++AI2, ++arg)
284 if (DS::isPointerType(AI2->getType()))
285 ArgNodes[arg].mergeWith(MergedG.getNodeForValue(AI2));
287 for ( ; AI2 != AI2end; ++AI2)
288 if (DS::isPointerType(AI2->getType()))
289 ArgNodes.push_back(MergedG.getNodeForValue(AI2));
290 DEBUG(MergedG.AssertGraphOK());
293 DEBUG(std::cerr << "\n");
297 DSGraph &EquivClassGraphs::getOrCreateGraph(Function &F) {
298 // Has the graph already been created?
299 DSGraph *&Graph = DSInfo[&F];
300 if (Graph) return *Graph;
302 DSGraph &CBUGraph = CBU->getDSGraph(F);
304 // Copy the CBU graph...
305 Graph = new DSGraph(CBUGraph, GlobalECs); // updates the map via reference
306 Graph->setGlobalsGraph(&getGlobalsGraph());
307 Graph->setPrintAuxCalls();
309 // Make sure to update the DSInfo map for all functions in the graph!
310 for (DSGraph::retnodes_iterator I = Graph->retnodes_begin();
311 I != Graph->retnodes_end(); ++I)
312 if (I->first != &F) {
313 DSGraph *&FG = DSInfo[I->first];
314 assert(FG == 0 && "Merging function in SCC twice?");
322 unsigned EquivClassGraphs::
323 processSCC(DSGraph &FG, std::vector<DSGraph*> &Stack, unsigned &NextID,
324 std::map<DSGraph*, unsigned> &ValMap) {
325 std::map<DSGraph*, unsigned>::iterator It = ValMap.lower_bound(&FG);
326 if (It != ValMap.end() && It->first == &FG)
329 DEBUG(std::cerr << " ProcessSCC for function " << FG.getFunctionNames()
332 unsigned Min = NextID++, MyID = Min;
334 Stack.push_back(&FG);
336 // The edges out of the current node are the call site targets...
337 for (DSGraph::fc_iterator CI = FG.fc_begin(), CE = FG.fc_end();
339 Instruction *Call = CI->getCallSite().getInstruction();
341 // Loop over all of the actually called functions...
342 for (callee_iterator I = callee_begin(Call), E = callee_end(Call);
344 if (!I->second->isExternal()) {
345 // Process the callee as necessary.
346 unsigned M = processSCC(getOrCreateGraph(*I->second),
347 Stack, NextID, ValMap);
348 if (M < Min) Min = M;
352 assert(ValMap[&FG] == MyID && "SCC construction assumption wrong!");
354 return Min; // This is part of a larger SCC!
356 // If this is a new SCC, process it now.
357 bool MergedGraphs = false;
358 while (Stack.back() != &FG) {
359 DSGraph *NG = Stack.back();
362 // If the SCC found is not the same as those found in CBU, make sure to
363 // merge the graphs as appropriate.
366 // Update the DSInfo map and delete the old graph...
367 for (DSGraph::retnodes_iterator I = NG->retnodes_begin();
368 I != NG->retnodes_end(); ++I)
369 DSInfo[I->first] = &FG;
371 // Remove NG from the ValMap since the pointer may get recycled.
378 // Clean up the graph before we start inlining a bunch again.
380 FG.removeTriviallyDeadNodes();
390 /// processGraph - Process the CBU graphs for the program in bottom-up order on
391 /// the SCC of the __ACTUAL__ call graph. This builds final folded CBU graphs.
392 void EquivClassGraphs::processGraph(DSGraph &G) {
393 DEBUG(std::cerr << " ProcessGraph for function "
394 << G.getFunctionNames() << "\n");
396 hash_set<Instruction*> calls;
398 // Else we need to inline some callee graph. Visit all call sites.
399 // The edges out of the current node are the call site targets...
401 for (DSGraph::fc_iterator CI = G.fc_begin(), CE = G.fc_end(); CI != CE;
403 const DSCallSite &CS = *CI;
404 Instruction *TheCall = CS.getCallSite().getInstruction();
406 assert(calls.insert(TheCall).second &&
407 "Call instruction occurs multiple times in graph??");
409 if (CS.getRetVal().isNull() && CS.getNumPtrArgs() == 0)
412 // Inline the common callee graph into the current graph, if the callee
413 // graph has not changed. Note that all callees should have the same
414 // graph so we only need to do this once.
416 DSGraph* CalleeGraph = NULL;
417 callee_iterator I = callee_begin(TheCall), E = callee_end(TheCall);
420 // Loop over all potential callees to find the first non-external callee.
421 for (TNum = 0, Num = std::distance(I, E); I != E; ++I, ++TNum)
422 if (!I->second->isExternal())
425 // Now check if the graph has changed and if so, clone and inline it.
427 Function *CalleeFunc = I->second;
429 // Merge the callee's graph into this graph, if not already the same.
430 // Callees in the same equivalence class (which subsumes those
431 // in the same SCCs) have the same graph. Note that all recursion
432 // including self-recursion have been folded in the equiv classes.
434 CalleeGraph = &getOrCreateGraph(*CalleeFunc);
435 if (CalleeGraph != &G) {
436 ++NumFoldGraphInlines;
437 G.mergeInGraph(CS, *CalleeFunc, *CalleeGraph,
438 DSGraph::StripAllocaBit |
439 DSGraph::DontCloneCallNodes |
440 DSGraph::DontCloneAuxCallNodes);
441 DEBUG(std::cerr << " Inlining graph [" << i << "/"
442 << G.getFunctionCalls().size()-1
443 << ":" << TNum << "/" << Num-1 << "] for "
444 << CalleeFunc->getName() << "["
445 << CalleeGraph->getGraphSize() << "+"
446 << CalleeGraph->getAuxFunctionCalls().size()
447 << "] into '" /*<< G.getFunctionNames()*/ << "' ["
448 << G.getGraphSize() << "+" << G.getAuxFunctionCalls().size()
454 // Now loop over the rest of the callees and make sure they have the
455 // same graph as the one inlined above.
457 for (++I, ++TNum; I != E; ++I, ++TNum)
458 if (!I->second->isExternal())
459 assert(CalleeGraph == &getOrCreateGraph(*I->second) &&
460 "Callees at a call site have different graphs?");
464 // Recompute the Incomplete markers.
465 G.maskIncompleteMarkers();
466 G.markIncompleteNodes(DSGraph::MarkFormalArgs);
468 // Delete dead nodes. Treat globals that are unreachable but that can
469 // reach live nodes as live.
470 G.removeDeadNodes(DSGraph::KeepUnreachableGlobals);
472 // When this graph is finalized, clone the globals in the graph into the
473 // globals graph to make sure it has everything, from all graphs.
474 ReachabilityCloner RC(*G.getGlobalsGraph(), G, DSGraph::StripAllocaBit);
476 // Clone everything reachable from globals in the function graph into the
478 DSScalarMap &MainSM = G.getScalarMap();
479 for (DSScalarMap::global_iterator I = MainSM.global_begin(),
480 E = MainSM.global_end(); I != E; ++I)
481 RC.getClonedNH(MainSM[*I]);
483 DEBUG(std::cerr << " -- DONE ProcessGraph for function "
484 << G.getFunctionNames() << "\n");