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/Module.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Support/Debug.h"
21 #include "DSCallSiteIterator.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");
35 // run - Calculate the bottom up data structure graphs for each function in the
38 bool BUDataStructures::runOnModule(Module &M) {
39 LocalDataStructures &LocalDSA = getAnalysis<LocalDataStructures>();
40 GlobalsGraph = new DSGraph(LocalDSA.getGlobalsGraph());
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 // Merge the globals variables (not the calls) from the globals graph back
86 // into the main function's graph so that the main function contains all of
87 // the information about global pools and GV usage in the program.
89 DSGraph &MainGraph = getOrCreateGraph(MainFunc);
90 const DSGraph &GG = *MainGraph.getGlobalsGraph();
91 ReachabilityCloner RC(MainGraph, GG,
92 DSGraph::DontCloneCallNodes |
93 DSGraph::DontCloneAuxCallNodes);
95 // Clone the global nodes into this graph.
96 for (DSScalarMap::global_iterator I = GG.getScalarMap().global_begin(),
97 E = GG.getScalarMap().global_end(); I != E; ++I)
98 if (isa<GlobalVariable>(*I))
99 RC.getClonedNH(GG.getNodeForValue(*I));
101 MainGraph.maskIncompleteMarkers();
102 MainGraph.markIncompleteNodes(DSGraph::MarkFormalArgs |
103 DSGraph::IgnoreGlobals);
109 DSGraph &BUDataStructures::getOrCreateGraph(Function *F) {
110 // Has the graph already been created?
111 DSGraph *&Graph = DSInfo[F];
112 if (Graph) return *Graph;
114 // Copy the local version into DSInfo...
115 Graph = new DSGraph(getAnalysis<LocalDataStructures>().getDSGraph(*F));
117 Graph->setGlobalsGraph(GlobalsGraph);
118 Graph->setPrintAuxCalls();
120 // Start with a copy of the original call sites...
121 Graph->getAuxFunctionCalls() = Graph->getFunctionCalls();
125 unsigned BUDataStructures::calculateGraphs(Function *F,
126 std::vector<Function*> &Stack,
128 hash_map<Function*, unsigned> &ValMap) {
129 assert(!ValMap.count(F) && "Shouldn't revisit functions!");
130 unsigned Min = NextID++, MyID = Min;
134 // FIXME! This test should be generalized to be any function that we have
135 // already processed, in the case when there isn't a main or there are
136 // unreachable functions!
137 if (F->isExternal()) { // sprintf, fprintf, sscanf, etc...
144 DSGraph &Graph = getOrCreateGraph(F);
146 // The edges out of the current node are the call site targets...
147 for (DSCallSiteIterator I = DSCallSiteIterator::begin_aux(Graph),
148 E = DSCallSiteIterator::end_aux(Graph); I != E; ++I) {
149 Function *Callee = *I;
151 // Have we visited the destination function yet?
152 hash_map<Function*, unsigned>::iterator It = ValMap.find(Callee);
153 if (It == ValMap.end()) // No, visit it now.
154 M = calculateGraphs(Callee, Stack, NextID, ValMap);
155 else // Yes, get it's number.
157 if (M < Min) Min = M;
160 assert(ValMap[F] == MyID && "SCC construction assumption wrong!");
162 return Min; // This is part of a larger SCC!
164 // If this is a new SCC, process it now.
165 if (Stack.back() == F) { // Special case the single "SCC" case here.
166 DEBUG(std::cerr << "Visiting single node SCC #: " << MyID << " fn: "
167 << F->getName() << "\n");
169 DSGraph &G = getDSGraph(*F);
170 DEBUG(std::cerr << " [BU] Calculating graph for: " << F->getName()<< "\n");
172 DEBUG(std::cerr << " [BU] Done inlining: " << F->getName() << " ["
173 << G.getGraphSize() << "+" << G.getAuxFunctionCalls().size()
176 if (MaxSCC < 1) MaxSCC = 1;
178 // Should we revisit the graph?
179 if (DSCallSiteIterator::begin_aux(G) != DSCallSiteIterator::end_aux(G)) {
181 return calculateGraphs(F, Stack, NextID, ValMap);
188 // SCCFunctions - Keep track of the functions in the current SCC
190 hash_set<DSGraph*> SCCGraphs;
193 std::vector<Function*>::iterator FirstInSCC = Stack.end();
194 DSGraph *SCCGraph = 0;
199 // Figure out which graph is the largest one, in order to speed things up
200 // a bit in situations where functions in the SCC have widely different
202 DSGraph &NFGraph = getDSGraph(*NF);
203 SCCGraphs.insert(&NFGraph);
204 // FIXME: If we used a better way of cloning graphs (ie, just splice all
205 // of the nodes into the new graph), this would be completely unneeded!
206 if (!SCCGraph || SCCGraph->getGraphSize() < NFGraph.getGraphSize())
210 std::cerr << "Calculating graph for SCC #: " << MyID << " of size: "
211 << SCCGraphs.size() << "\n";
213 // Compute the Max SCC Size...
214 if (MaxSCC < SCCGraphs.size())
215 MaxSCC = SCCGraphs.size();
217 // First thing first, collapse all of the DSGraphs into a single graph for
218 // the entire SCC. We computed the largest graph, so clone all of the other
219 // (smaller) graphs into it. Discard all of the old graphs.
221 for (hash_set<DSGraph*>::iterator I = SCCGraphs.begin(),
222 E = SCCGraphs.end(); I != E; ++I) {
224 if (&G != SCCGraph) {
226 DSGraph::NodeMapTy NodeMap;
227 SCCGraph->cloneInto(G, SCCGraph->getScalarMap(),
228 SCCGraph->getReturnNodes(), NodeMap);
230 // Update the DSInfo map and delete the old graph...
231 for (DSGraph::retnodes_iterator I = G.retnodes_begin(),
232 E = G.retnodes_end(); I != E; ++I)
233 DSInfo[I->first] = SCCGraph;
238 // Clean up the graph before we start inlining a bunch again...
239 SCCGraph->removeDeadNodes(DSGraph::KeepUnreachableGlobals);
241 // Now that we have one big happy family, resolve all of the call sites in
243 calculateGraph(*SCCGraph);
244 DEBUG(std::cerr << " [BU] Done inlining SCC [" << SCCGraph->getGraphSize()
245 << "+" << SCCGraph->getAuxFunctionCalls().size() << "]\n");
247 std::cerr << "DONE with SCC #: " << MyID << "\n";
249 // We never have to revisit "SCC" processed functions...
251 // Drop the stuff we don't need from the end of the stack
252 Stack.erase(FirstInSCC, Stack.end());
256 return MyID; // == Min
260 // releaseMemory - If the pass pipeline is done with this pass, we can release
261 // our memory... here...
263 void BUDataStructures::releaseMemory() {
264 for (hash_map<Function*, DSGraph*>::iterator I = DSInfo.begin(),
265 E = DSInfo.end(); I != E; ++I) {
266 I->second->getReturnNodes().erase(I->first);
267 if (I->second->getReturnNodes().empty())
271 // Empty map so next time memory is released, data structures are not
278 static bool isVAHackFn(const Function *F) {
279 return F->getName() == "printf" || F->getName() == "sscanf" ||
280 F->getName() == "fprintf" || F->getName() == "open" ||
281 F->getName() == "sprintf" || F->getName() == "fputs" ||
282 F->getName() == "fscanf";
285 // isUnresolvableFunction - Return true if this is an unresolvable
286 // external function. A direct or indirect call to this cannot be resolved.
288 static bool isResolvableFunc(const Function* callee) {
289 return !callee->isExternal() || isVAHackFn(callee);
292 void BUDataStructures::calculateGraph(DSGraph &Graph) {
293 // Move our call site list into TempFCs so that inline call sites go into the
294 // new call site list and doesn't invalidate our iterators!
295 std::list<DSCallSite> TempFCs;
296 std::list<DSCallSite> &AuxCallsList = Graph.getAuxFunctionCalls();
297 TempFCs.swap(AuxCallsList);
299 DSGraph::ReturnNodesTy &ReturnNodes = Graph.getReturnNodes();
301 bool Printed = false;
302 std::vector<Function*> CalledFuncs;
303 while (!TempFCs.empty()) {
304 DSCallSite &CS = *TempFCs.begin();
308 if (CS.isDirectCall()) {
309 Function *F = CS.getCalleeFunc();
310 if (isResolvableFunc(F))
311 if (F->isExternal()) { // Call to fprintf, etc.
312 TempFCs.erase(TempFCs.begin());
315 CalledFuncs.push_back(F);
318 DSNode *Node = CS.getCalleeNode();
320 if (!Node->isIncomplete())
321 for (unsigned i = 0, e = Node->getGlobals().size(); i != e; ++i)
322 if (Function *CF = dyn_cast<Function>(Node->getGlobals()[i]))
323 if (isResolvableFunc(CF) && !CF->isExternal())
324 CalledFuncs.push_back(CF);
327 if (CalledFuncs.empty()) {
328 // Remember that we could not resolve this yet!
329 AuxCallsList.splice(AuxCallsList.end(), TempFCs, TempFCs.begin());
334 if (CalledFuncs.size() == 1) {
335 Function *Callee = CalledFuncs[0];
336 ActualCallees.insert(std::make_pair(CS.getCallSite().getInstruction(),
339 // Get the data structure graph for the called function.
340 GI = &getDSGraph(*Callee); // Graph to inline
341 DEBUG(std::cerr << " Inlining graph for " << Callee->getName());
343 DEBUG(std::cerr << "[" << GI->getGraphSize() << "+"
344 << GI->getAuxFunctionCalls().size() << "] into '"
345 << Graph.getFunctionNames() << "' [" << Graph.getGraphSize() <<"+"
346 << Graph.getAuxFunctionCalls().size() << "]\n");
347 Graph.mergeInGraph(CS, *Callee, *GI,
348 DSGraph::KeepModRefBits |
349 DSGraph::StripAllocaBit|DSGraph::DontCloneCallNodes);
353 std::cerr << "In Fns: " << Graph.getFunctionNames() << "\n";
354 std::cerr << " calls " << CalledFuncs.size()
355 << " fns from site: " << CS.getCallSite().getInstruction()
356 << " " << *CS.getCallSite().getInstruction();
357 unsigned NumToPrint = CalledFuncs.size();
358 if (NumToPrint > 8) NumToPrint = 8;
359 std::cerr << " Fns =";
360 for (std::vector<Function*>::iterator I = CalledFuncs.begin(),
361 E = CalledFuncs.end(); I != E && NumToPrint; ++I, --NumToPrint)
362 std::cerr << " " << (*I)->getName();
365 // See if we already computed a graph for this set of callees.
366 std::sort(CalledFuncs.begin(), CalledFuncs.end());
367 std::pair<DSGraph*, std::vector<DSNodeHandle> > &IndCallGraph =
368 (*IndCallGraphMap)[CalledFuncs];
370 if (IndCallGraph.first == 0) {
371 std::vector<Function*>::iterator I = CalledFuncs.begin(),
372 E = CalledFuncs.end();
374 // Start with a copy of the first graph.
375 GI = IndCallGraph.first = new DSGraph(getDSGraph(**I));
376 GI->setGlobalsGraph(Graph.getGlobalsGraph());
377 std::vector<DSNodeHandle> &Args = IndCallGraph.second;
379 // Get the argument nodes for the first callee. The return value is
380 // the 0th index in the vector.
381 GI->getFunctionArgumentsForCall(*I, Args);
383 // Merge all of the other callees into this graph.
384 for (++I; I != E; ++I) {
385 // If the graph already contains the nodes for the function, don't
386 // bother merging it in again.
387 if (!GI->containsFunction(*I)) {
388 DSGraph::NodeMapTy NodeMap;
389 GI->cloneInto(getDSGraph(**I), GI->getScalarMap(),
390 GI->getReturnNodes(), NodeMap);
394 std::vector<DSNodeHandle> NextArgs;
395 GI->getFunctionArgumentsForCall(*I, NextArgs);
396 unsigned i = 0, e = Args.size();
397 for (; i != e; ++i) {
398 if (i == NextArgs.size()) break;
399 Args[i].mergeWith(NextArgs[i]);
401 for (e = NextArgs.size(); i != e; ++i)
402 Args.push_back(NextArgs[i]);
405 // Clean up the final graph!
406 GI->removeDeadNodes(DSGraph::KeepUnreachableGlobals);
408 std::cerr << "***\n*** RECYCLED GRAPH ***\n***\n";
411 GI = IndCallGraph.first;
413 // Merge the unified graph into this graph now.
414 DEBUG(std::cerr << " Inlining multi callee graph "
415 << "[" << GI->getGraphSize() << "+"
416 << GI->getAuxFunctionCalls().size() << "] into '"
417 << Graph.getFunctionNames() << "' [" << Graph.getGraphSize() <<"+"
418 << Graph.getAuxFunctionCalls().size() << "]\n");
420 Graph.mergeInGraph(CS, IndCallGraph.second, *GI,
421 DSGraph::KeepModRefBits |
422 DSGraph::StripAllocaBit |
423 DSGraph::DontCloneCallNodes);
427 TempFCs.erase(TempFCs.begin());
430 // Recompute the Incomplete markers
431 assert(Graph.getInlinedGlobals().empty());
432 Graph.maskIncompleteMarkers();
433 Graph.markIncompleteNodes(DSGraph::MarkFormalArgs);
435 // Delete dead nodes. Treat globals that are unreachable but that can
436 // reach live nodes as live.
437 Graph.removeDeadNodes(DSGraph::KeepUnreachableGlobals);
439 // When this graph is finalized, clone the globals in the graph into the
440 // globals graph to make sure it has everything, from all graphs.
441 DSScalarMap &MainSM = Graph.getScalarMap();
442 ReachabilityCloner RC(*GlobalsGraph, Graph, DSGraph::StripAllocaBit);
444 // Clone everything reachable from globals in the function graph into the
446 for (DSScalarMap::global_iterator I = MainSM.global_begin(),
447 E = MainSM.global_end(); I != E; ++I)
448 RC.getClonedNH(MainSM[*I]);
450 //Graph.writeGraphToFile(std::cerr, "bu_" + F.getName());
453 static const Function *getFnForValue(const Value *V) {
454 if (const Instruction *I = dyn_cast<Instruction>(V))
455 return I->getParent()->getParent();
456 else if (const Argument *A = dyn_cast<Argument>(V))
457 return A->getParent();
458 else if (const BasicBlock *BB = dyn_cast<BasicBlock>(V))
459 return BB->getParent();
463 /// deleteValue/copyValue - Interfaces to update the DSGraphs in the program.
464 /// These correspond to the interfaces defined in the AliasAnalysis class.
465 void BUDataStructures::deleteValue(Value *V) {
466 if (const Function *F = getFnForValue(V)) { // Function local value?
467 // If this is a function local value, just delete it from the scalar map!
468 getDSGraph(*F).getScalarMap().eraseIfExists(V);
472 if (Function *F = dyn_cast<Function>(V)) {
473 assert(getDSGraph(*F).getReturnNodes().size() == 1 &&
474 "cannot handle scc's");
480 assert(!isa<GlobalVariable>(V) && "Do not know how to delete GV's yet!");
483 void BUDataStructures::copyValue(Value *From, Value *To) {
484 if (From == To) return;
485 if (const Function *F = getFnForValue(From)) { // Function local value?
486 // If this is a function local value, just delete it from the scalar map!
487 getDSGraph(*F).getScalarMap().copyScalarIfExists(From, To);
491 if (Function *FromF = dyn_cast<Function>(From)) {
492 Function *ToF = cast<Function>(To);
493 assert(!DSInfo.count(ToF) && "New Function already exists!");
494 DSGraph *NG = new DSGraph(getDSGraph(*FromF));
496 assert(NG->getReturnNodes().size() == 1 && "Cannot copy SCC's yet!");
498 // Change the Function* is the returnnodes map to the ToF.
499 DSNodeHandle Ret = NG->retnodes_begin()->second;
500 NG->getReturnNodes().clear();
501 NG->getReturnNodes()[ToF] = Ret;
505 assert(!isa<GlobalVariable>(From) && "Do not know how to copy GV's yet!");