1 //===- BottomUpClosure.cpp - Compute bottom-up interprocedural 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 BUDataStructures class, which represents the
11 // Bottom-Up Interprocedural closure of the data structure graph over the
12 // program. This is useful for applications like pool allocation, but **not**
13 // applications like alias analysis.
15 //===----------------------------------------------------------------------===//
17 #include "llvm/Analysis/DataStructure/DataStructure.h"
18 #include "llvm/Analysis/DataStructure/DSGraph.h"
19 #include "llvm/Module.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Support/Debug.h"
25 Statistic<> MaxSCC("budatastructure", "Maximum SCC Size in Call Graph");
26 Statistic<> NumBUInlines("budatastructures", "Number of graphs inlined");
27 Statistic<> NumCallEdges("budatastructures", "Number of 'actual' call edges");
29 RegisterAnalysis<BUDataStructures>
30 X("budatastructure", "Bottom-up Data Structure Analysis");
33 // run - Calculate the bottom up data structure graphs for each function in the
36 bool BUDataStructures::runOnModule(Module &M) {
37 LocalDataStructures &LocalDSA = getAnalysis<LocalDataStructures>();
38 GlobalECs = LocalDSA.getGlobalECs();
40 GlobalsGraph = new DSGraph(LocalDSA.getGlobalsGraph(), GlobalECs);
41 GlobalsGraph->setPrintAuxCalls();
43 IndCallGraphMap = new std::map<std::vector<Function*>,
44 std::pair<DSGraph*, std::vector<DSNodeHandle> > >();
46 std::vector<Function*> Stack;
47 hash_map<Function*, unsigned> ValMap;
50 Function *MainFunc = M.getMainFunction();
52 calculateGraphs(MainFunc, Stack, NextID, ValMap);
54 // Calculate the graphs for any functions that are unreachable from main...
55 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
56 if (!I->isExternal() && !DSInfo.count(I)) {
59 std::cerr << "*** Function unreachable from main: "
60 << I->getName() << "\n";
62 calculateGraphs(I, Stack, NextID, ValMap); // Calculate all graphs.
65 NumCallEdges += ActualCallees.size();
67 // If we computed any temporary indcallgraphs, free them now.
68 for (std::map<std::vector<Function*>,
69 std::pair<DSGraph*, std::vector<DSNodeHandle> > >::iterator I =
70 IndCallGraphMap->begin(), E = IndCallGraphMap->end(); I != E; ++I) {
71 I->second.second.clear(); // Drop arg refs into the graph.
72 delete I->second.first;
74 delete IndCallGraphMap;
76 // At the end of the bottom-up pass, the globals graph becomes complete.
77 // FIXME: This is not the right way to do this, but it is sorta better than
78 // nothing! In particular, externally visible globals and unresolvable call
79 // nodes at the end of the BU phase should make things that they point to
80 // incomplete in the globals graph.
82 GlobalsGraph->removeTriviallyDeadNodes();
83 GlobalsGraph->maskIncompleteMarkers();
85 // Mark external globals incomplete.
86 GlobalsGraph->markIncompleteNodes(DSGraph::IgnoreGlobals);
88 // Merge the globals variables (not the calls) from the globals graph back
89 // into the main function's graph so that the main function contains all of
90 // the information about global pools and GV usage in the program.
91 if (MainFunc && !MainFunc->isExternal()) {
92 DSGraph &MainGraph = getOrCreateGraph(MainFunc);
93 const DSGraph &GG = *MainGraph.getGlobalsGraph();
94 ReachabilityCloner RC(MainGraph, GG,
95 DSGraph::DontCloneCallNodes |
96 DSGraph::DontCloneAuxCallNodes);
98 // Clone the global nodes into this graph.
99 for (DSScalarMap::global_iterator I = GG.getScalarMap().global_begin(),
100 E = GG.getScalarMap().global_end(); I != E; ++I)
101 if (isa<GlobalVariable>(*I))
102 RC.getClonedNH(GG.getNodeForValue(*I));
104 MainGraph.maskIncompleteMarkers();
105 MainGraph.markIncompleteNodes(DSGraph::MarkFormalArgs |
106 DSGraph::IgnoreGlobals);
112 DSGraph &BUDataStructures::getOrCreateGraph(Function *F) {
113 // Has the graph already been created?
114 DSGraph *&Graph = DSInfo[F];
115 if (Graph) return *Graph;
117 // Copy the local version into DSInfo...
118 Graph = new DSGraph(getAnalysis<LocalDataStructures>().getDSGraph(*F),
121 Graph->setGlobalsGraph(GlobalsGraph);
122 Graph->setPrintAuxCalls();
124 // Start with a copy of the original call sites...
125 Graph->getAuxFunctionCalls() = Graph->getFunctionCalls();
129 static bool isVAHackFn(const Function *F) {
130 return F->getName() == "printf" || F->getName() == "sscanf" ||
131 F->getName() == "fprintf" || F->getName() == "open" ||
132 F->getName() == "sprintf" || F->getName() == "fputs" ||
133 F->getName() == "fscanf";
136 static bool isResolvableFunc(const Function* callee) {
137 return !callee->isExternal() || isVAHackFn(callee);
140 static void GetAllCallees(const DSCallSite &CS,
141 std::vector<Function*> &Callees) {
142 if (CS.isDirectCall()) {
143 if (isResolvableFunc(CS.getCalleeFunc()))
144 Callees.push_back(CS.getCalleeFunc());
145 } else if (!CS.getCalleeNode()->isIncomplete()) {
147 unsigned OldSize = Callees.size();
148 CS.getCalleeNode()->addFullFunctionList(Callees);
150 // If any of the callees are unresolvable, remove the whole batch!
151 for (unsigned i = OldSize, e = Callees.size(); i != e; ++i)
152 if (!isResolvableFunc(Callees[i])) {
153 Callees.erase(Callees.begin()+OldSize, Callees.end());
160 /// GetAllAuxCallees - Return a list containing all of the resolvable callees in
161 /// the aux list for the specified graph in the Callees vector.
162 static void GetAllAuxCallees(DSGraph &G, std::vector<Function*> &Callees) {
164 for (DSGraph::afc_iterator I = G.afc_begin(), E = G.afc_end(); I != E; ++I)
165 GetAllCallees(*I, Callees);
168 unsigned BUDataStructures::calculateGraphs(Function *F,
169 std::vector<Function*> &Stack,
171 hash_map<Function*, unsigned> &ValMap) {
172 assert(!ValMap.count(F) && "Shouldn't revisit functions!");
173 unsigned Min = NextID++, MyID = Min;
177 // FIXME! This test should be generalized to be any function that we have
178 // already processed, in the case when there isn't a main or there are
179 // unreachable functions!
180 if (F->isExternal()) { // sprintf, fprintf, sscanf, etc...
187 DSGraph &Graph = getOrCreateGraph(F);
189 // Find all callee functions.
190 std::vector<Function*> CalleeFunctions;
191 GetAllAuxCallees(Graph, CalleeFunctions);
193 // The edges out of the current node are the call site targets...
194 for (unsigned i = 0, e = CalleeFunctions.size(); i != e; ++i) {
195 Function *Callee = CalleeFunctions[i];
197 // Have we visited the destination function yet?
198 hash_map<Function*, unsigned>::iterator It = ValMap.find(Callee);
199 if (It == ValMap.end()) // No, visit it now.
200 M = calculateGraphs(Callee, Stack, NextID, ValMap);
201 else // Yes, get it's number.
203 if (M < Min) Min = M;
206 assert(ValMap[F] == MyID && "SCC construction assumption wrong!");
208 return Min; // This is part of a larger SCC!
210 // If this is a new SCC, process it now.
211 if (Stack.back() == F) { // Special case the single "SCC" case here.
212 DEBUG(std::cerr << "Visiting single node SCC #: " << MyID << " fn: "
213 << F->getName() << "\n");
215 DSGraph &G = getDSGraph(*F);
216 DEBUG(std::cerr << " [BU] Calculating graph for: " << F->getName()<< "\n");
218 DEBUG(std::cerr << " [BU] Done inlining: " << F->getName() << " ["
219 << G.getGraphSize() << "+" << G.getAuxFunctionCalls().size()
222 if (MaxSCC < 1) MaxSCC = 1;
224 // Should we revisit the graph? Only do it if there are now new resolvable
226 GetAllAuxCallees(Graph, CalleeFunctions);
227 if (!CalleeFunctions.empty()) {
229 return calculateGraphs(F, Stack, NextID, ValMap);
236 // SCCFunctions - Keep track of the functions in the current SCC
238 hash_set<DSGraph*> SCCGraphs;
241 std::vector<Function*>::iterator FirstInSCC = Stack.end();
242 DSGraph *SCCGraph = 0;
247 // Figure out which graph is the largest one, in order to speed things up
248 // a bit in situations where functions in the SCC have widely different
250 DSGraph &NFGraph = getDSGraph(*NF);
251 SCCGraphs.insert(&NFGraph);
252 // FIXME: If we used a better way of cloning graphs (ie, just splice all
253 // of the nodes into the new graph), this would be completely unneeded!
254 if (!SCCGraph || SCCGraph->getGraphSize() < NFGraph.getGraphSize())
258 std::cerr << "Calculating graph for SCC #: " << MyID << " of size: "
259 << SCCGraphs.size() << "\n";
261 // Compute the Max SCC Size...
262 if (MaxSCC < SCCGraphs.size())
263 MaxSCC = SCCGraphs.size();
265 // First thing first, collapse all of the DSGraphs into a single graph for
266 // the entire SCC. We computed the largest graph, so clone all of the other
267 // (smaller) graphs into it. Discard all of the old graphs.
269 for (hash_set<DSGraph*>::iterator I = SCCGraphs.begin(),
270 E = SCCGraphs.end(); I != E; ++I) {
272 if (&G != SCCGraph) {
273 SCCGraph->cloneInto(G);
275 // Update the DSInfo map and delete the old graph...
276 for (DSGraph::retnodes_iterator I = G.retnodes_begin(),
277 E = G.retnodes_end(); I != E; ++I)
278 DSInfo[I->first] = SCCGraph;
283 // Clean up the graph before we start inlining a bunch again...
284 SCCGraph->removeDeadNodes(DSGraph::KeepUnreachableGlobals);
286 // Now that we have one big happy family, resolve all of the call sites in
288 calculateGraph(*SCCGraph);
289 DEBUG(std::cerr << " [BU] Done inlining SCC [" << SCCGraph->getGraphSize()
290 << "+" << SCCGraph->getAuxFunctionCalls().size() << "]\n");
292 std::cerr << "DONE with SCC #: " << MyID << "\n";
294 // We never have to revisit "SCC" processed functions...
296 // Drop the stuff we don't need from the end of the stack
297 Stack.erase(FirstInSCC, Stack.end());
301 return MyID; // == Min
305 // releaseMemory - If the pass pipeline is done with this pass, we can release
306 // our memory... here...
308 void BUDataStructures::releaseMyMemory() {
309 for (hash_map<Function*, DSGraph*>::iterator I = DSInfo.begin(),
310 E = DSInfo.end(); I != E; ++I) {
311 I->second->getReturnNodes().erase(I->first);
312 if (I->second->getReturnNodes().empty())
316 // Empty map so next time memory is released, data structures are not
323 void BUDataStructures::calculateGraph(DSGraph &Graph) {
324 // Move our call site list into TempFCs so that inline call sites go into the
325 // new call site list and doesn't invalidate our iterators!
326 std::list<DSCallSite> TempFCs;
327 std::list<DSCallSite> &AuxCallsList = Graph.getAuxFunctionCalls();
328 TempFCs.swap(AuxCallsList);
330 DSGraph::ReturnNodesTy &ReturnNodes = Graph.getReturnNodes();
332 bool Printed = false;
333 std::vector<Function*> CalledFuncs;
334 while (!TempFCs.empty()) {
335 DSCallSite &CS = *TempFCs.begin();
339 // Fast path for noop calls. Note that we don't care about merging globals
340 // in the callee with nodes in the caller here.
341 if (CS.getRetVal().isNull() && CS.getNumPtrArgs() == 0) {
342 TempFCs.erase(TempFCs.begin());
344 } else if (CS.isDirectCall() && isVAHackFn(CS.getCalleeFunc())) {
345 TempFCs.erase(TempFCs.begin());
349 GetAllCallees(CS, CalledFuncs);
351 if (CalledFuncs.empty()) {
352 // Remember that we could not resolve this yet!
353 AuxCallsList.splice(AuxCallsList.end(), TempFCs, TempFCs.begin());
357 Instruction *TheCall = CS.getCallSite().getInstruction();
359 if (CalledFuncs.size() == 1) {
360 Function *Callee = CalledFuncs[0];
361 ActualCallees.insert(std::make_pair(TheCall, Callee));
363 // Get the data structure graph for the called function.
364 GI = &getDSGraph(*Callee); // Graph to inline
365 DEBUG(std::cerr << " Inlining graph for " << Callee->getName());
367 DEBUG(std::cerr << "[" << GI->getGraphSize() << "+"
368 << GI->getAuxFunctionCalls().size() << "] into '"
369 << Graph.getFunctionNames() << "' [" << Graph.getGraphSize() <<"+"
370 << Graph.getAuxFunctionCalls().size() << "]\n");
371 Graph.mergeInGraph(CS, *Callee, *GI,
372 DSGraph::StripAllocaBit|DSGraph::DontCloneCallNodes);
376 std::cerr << "In Fns: " << Graph.getFunctionNames() << "\n";
377 std::cerr << " calls " << CalledFuncs.size()
378 << " fns from site: " << CS.getCallSite().getInstruction()
379 << " " << *CS.getCallSite().getInstruction();
380 std::cerr << " Fns =";
381 unsigned NumPrinted = 0;
383 for (std::vector<Function*>::iterator I = CalledFuncs.begin(),
384 E = CalledFuncs.end(); I != E; ++I) {
385 if (NumPrinted++ < 8) std::cerr << " " << (*I)->getName();
387 // Add the call edges to the call graph.
388 ActualCallees.insert(std::make_pair(TheCall, *I));
392 // See if we already computed a graph for this set of callees.
393 std::sort(CalledFuncs.begin(), CalledFuncs.end());
394 std::pair<DSGraph*, std::vector<DSNodeHandle> > &IndCallGraph =
395 (*IndCallGraphMap)[CalledFuncs];
397 if (IndCallGraph.first == 0) {
398 std::vector<Function*>::iterator I = CalledFuncs.begin(),
399 E = CalledFuncs.end();
401 // Start with a copy of the first graph.
402 GI = IndCallGraph.first = new DSGraph(getDSGraph(**I), GlobalECs);
403 GI->setGlobalsGraph(Graph.getGlobalsGraph());
404 std::vector<DSNodeHandle> &Args = IndCallGraph.second;
406 // Get the argument nodes for the first callee. The return value is
407 // the 0th index in the vector.
408 GI->getFunctionArgumentsForCall(*I, Args);
410 // Merge all of the other callees into this graph.
411 for (++I; I != E; ++I) {
412 // If the graph already contains the nodes for the function, don't
413 // bother merging it in again.
414 if (!GI->containsFunction(*I)) {
415 GI->cloneInto(getDSGraph(**I));
419 std::vector<DSNodeHandle> NextArgs;
420 GI->getFunctionArgumentsForCall(*I, NextArgs);
421 unsigned i = 0, e = Args.size();
422 for (; i != e; ++i) {
423 if (i == NextArgs.size()) break;
424 Args[i].mergeWith(NextArgs[i]);
426 for (e = NextArgs.size(); i != e; ++i)
427 Args.push_back(NextArgs[i]);
430 // Clean up the final graph!
431 GI->removeDeadNodes(DSGraph::KeepUnreachableGlobals);
433 std::cerr << "***\n*** RECYCLED GRAPH ***\n***\n";
436 GI = IndCallGraph.first;
438 // Merge the unified graph into this graph now.
439 DEBUG(std::cerr << " Inlining multi callee graph "
440 << "[" << GI->getGraphSize() << "+"
441 << GI->getAuxFunctionCalls().size() << "] into '"
442 << Graph.getFunctionNames() << "' [" << Graph.getGraphSize() <<"+"
443 << Graph.getAuxFunctionCalls().size() << "]\n");
445 Graph.mergeInGraph(CS, IndCallGraph.second, *GI,
446 DSGraph::StripAllocaBit |
447 DSGraph::DontCloneCallNodes);
451 TempFCs.erase(TempFCs.begin());
454 // Recompute the Incomplete markers
455 Graph.maskIncompleteMarkers();
456 Graph.markIncompleteNodes(DSGraph::MarkFormalArgs);
458 // Delete dead nodes. Treat globals that are unreachable but that can
459 // reach live nodes as live.
460 Graph.removeDeadNodes(DSGraph::KeepUnreachableGlobals);
462 // When this graph is finalized, clone the globals in the graph into the
463 // globals graph to make sure it has everything, from all graphs.
464 DSScalarMap &MainSM = Graph.getScalarMap();
465 ReachabilityCloner RC(*GlobalsGraph, Graph, DSGraph::StripAllocaBit);
467 // Clone everything reachable from globals in the function graph into the
469 for (DSScalarMap::global_iterator I = MainSM.global_begin(),
470 E = MainSM.global_end(); I != E; ++I)
471 RC.getClonedNH(MainSM[*I]);
473 //Graph.writeGraphToFile(std::cerr, "bu_" + F.getName());
476 static const Function *getFnForValue(const Value *V) {
477 if (const Instruction *I = dyn_cast<Instruction>(V))
478 return I->getParent()->getParent();
479 else if (const Argument *A = dyn_cast<Argument>(V))
480 return A->getParent();
481 else if (const BasicBlock *BB = dyn_cast<BasicBlock>(V))
482 return BB->getParent();
486 /// deleteValue/copyValue - Interfaces to update the DSGraphs in the program.
487 /// These correspond to the interfaces defined in the AliasAnalysis class.
488 void BUDataStructures::deleteValue(Value *V) {
489 if (const Function *F = getFnForValue(V)) { // Function local value?
490 // If this is a function local value, just delete it from the scalar map!
491 getDSGraph(*F).getScalarMap().eraseIfExists(V);
495 if (Function *F = dyn_cast<Function>(V)) {
496 assert(getDSGraph(*F).getReturnNodes().size() == 1 &&
497 "cannot handle scc's");
503 assert(!isa<GlobalVariable>(V) && "Do not know how to delete GV's yet!");
506 void BUDataStructures::copyValue(Value *From, Value *To) {
507 if (From == To) return;
508 if (const Function *F = getFnForValue(From)) { // Function local value?
509 // If this is a function local value, just delete it from the scalar map!
510 getDSGraph(*F).getScalarMap().copyScalarIfExists(From, To);
514 if (Function *FromF = dyn_cast<Function>(From)) {
515 Function *ToF = cast<Function>(To);
516 assert(!DSInfo.count(ToF) && "New Function already exists!");
517 DSGraph *NG = new DSGraph(getDSGraph(*FromF), GlobalECs);
519 assert(NG->getReturnNodes().size() == 1 && "Cannot copy SCC's yet!");
521 // Change the Function* is the returnnodes map to the ToF.
522 DSNodeHandle Ret = NG->retnodes_begin()->second;
523 NG->getReturnNodes().clear();
524 NG->getReturnNodes()[ToF] = Ret;
528 if (const Function *F = getFnForValue(To)) {
529 DSGraph &G = getDSGraph(*F);
530 G.getScalarMap().copyScalarIfExists(From, To);
536 assert(0 && "Do not know how to copy this yet!");