//===----------------------------------------------------------------------===//
#include "llvm/Analysis/DataStructure.h"
-#include "llvm/Analysis/DSGraph.h"
#include "llvm/Module.h"
#include "Support/Statistic.h"
-using std::map;
-
-static RegisterAnalysis<BUDataStructures>
-X("budatastructure", "Bottom-up Data Structure Analysis Closure");
-
-// TODO: FIXME
-namespace DataStructureAnalysis {
- // isPointerType - Return true if this first class type is big enough to hold
- // a pointer.
- //
- bool isPointerType(const Type *Ty);
+#include "Support/Debug.h"
+#include "DSCallSiteIterator.h"
+
+namespace {
+ Statistic<> MaxSCC("budatastructure", "Maximum SCC Size in Call Graph");
+ Statistic<> NumBUInlines("budatastructures", "Number of graphs inlined");
+ Statistic<> NumCallEdges("budatastructures", "Number of 'actual' call edges");
+
+ RegisterAnalysis<BUDataStructures>
+ X("budatastructure", "Bottom-up Data Structure Analysis");
}
-using namespace DataStructureAnalysis;
-
-
-// releaseMemory - If the pass pipeline is done with this pass, we can release
-// our memory... here...
-//
-void BUDataStructures::releaseMemory() {
- // Delete all call site information
- CallSites.clear();
- for (map<const Function*, DSGraph*>::iterator I = DSInfo.begin(),
- E = DSInfo.end(); I != E; ++I)
- delete I->second;
-
- // Empty map so next time memory is released, data structures are not
- // re-deleted.
- DSInfo.clear();
-}
+using namespace DS;
// run - Calculate the bottom up data structure graphs for each function in the
// program.
//
bool BUDataStructures::run(Module &M) {
- // Simply calculate the graphs for each function...
+ LocalDataStructures &LocalDSA = getAnalysis<LocalDataStructures>();
+ GlobalsGraph = new DSGraph(LocalDSA.getGlobalsGraph());
+ GlobalsGraph->setPrintAuxCalls();
+
+ Function *MainFunc = M.getMainFunction();
+ if (MainFunc)
+ calculateReachableGraphs(MainFunc);
+
+ // Calculate the graphs for any functions that are unreachable from main...
for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
- if (!I->isExternal())
- calculateGraph(*I);
+ if (!I->isExternal() && !DSInfo.count(I)) {
+#ifndef NDEBUG
+ if (MainFunc)
+ std::cerr << "*** Function unreachable from main: "
+ << I->getName() << "\n";
+#endif
+ calculateReachableGraphs(I); // Calculate all graphs...
+ }
+
+ NumCallEdges += ActualCallees.size();
return false;
}
-// ResolveArguments - Resolve the formal and actual arguments for a function
-// call.
-//
-static void ResolveArguments(DSCallSite &Call, Function &F,
- map<Value*, DSNodeHandle> &ValueMap) {
- // Resolve all of the function arguments...
- Function::aiterator AI = F.abegin();
- for (unsigned i = 2, e = Call.size(); i != e; ++i) {
- // Advance the argument iterator to the first pointer argument...
- while (!isPointerType(AI->getType())) ++AI;
-
- // Add the link from the argument scalar to the provided value
- DSNodeHandle &NN = ValueMap[AI];
- NN.addEdgeTo(Call[i]);
- ++AI;
- }
+void BUDataStructures::calculateReachableGraphs(Function *F) {
+ std::vector<Function*> Stack;
+ hash_map<Function*, unsigned> ValMap;
+ unsigned NextID = 1;
+ calculateGraphs(F, Stack, NextID, ValMap);
}
-DSGraph &BUDataStructures::calculateGraph(Function &F) {
- // Make sure this graph has not already been calculated, or that we don't get
- // into an infinite loop with mutually recursive functions.
- //
- DSGraph *&Graph = DSInfo[&F];
+DSGraph &BUDataStructures::getOrCreateGraph(Function *F) {
+ // Has the graph already been created?
+ DSGraph *&Graph = DSInfo[F];
if (Graph) return *Graph;
// Copy the local version into DSInfo...
- Graph = new DSGraph(getAnalysis<LocalDataStructures>().getDSGraph(F));
+ Graph = new DSGraph(getAnalysis<LocalDataStructures>().getDSGraph(*F));
-#if 0
- // Populate the GlobalsGraph with globals from this one.
- Graph->GlobalsGraph->cloneGlobals(*Graph, /*cloneCalls*/ false);
-#endif
+ Graph->setGlobalsGraph(GlobalsGraph);
+ Graph->setPrintAuxCalls();
- // Start resolving calls...
- std::vector<DSCallSite> &FCs = Graph->getFunctionCalls();
-
- DEBUG(std::cerr << " [BU] Inlining: " << F.getName() << "\n");
-
- bool Inlined;
- do {
- Inlined = false;
-
- for (unsigned i = 0; i != FCs.size(); ++i) {
- // Copy the call, because inlining graphs may invalidate the FCs vector.
- DSCallSite Call = FCs[i];
-
- // If the function list is complete...
- if ((Call.getCalleeNode().getNode()->NodeType & DSNode::Incomplete)==0) {
- // Start inlining all of the functions we can... some may not be
- // inlinable if they are external...
- //
- std::vector<GlobalValue*> Callees(Call.getCalleeNode().getNode()->getGlobals());
-
- // Loop over the functions, inlining whatever we can...
- for (unsigned c = 0; c != Callees.size(); ++c) {
- // Must be a function type, so this cast MUST succeed.
- Function &FI = cast<Function>(*Callees[c]);
-
- // Record that this is a call site of FI.
- assert(&Call.getCaller() == &F && "Invalid caller in DSCallSite?");
- CallSites[&FI].push_back(DSCallSite(Call));
-
- if (&FI == &F) {
- // Self recursion... simply link up the formal arguments with the
- // actual arguments...
-
- DEBUG(std::cerr << "\t[BU] Self Inlining: " << F.getName() << "\n");
-
- if (Call.getReturnValueNode().getNode()) // Handle the return value if present...
- Graph->getRetNode().mergeWith(Call.getReturnValueNode());
-
- // Resolve the arguments in the call to the actual values...
- ResolveArguments(Call, F, Graph->getValueMap());
-
- // Erase the entry in the callees vector
- Callees.erase(Callees.begin()+c--);
-
- } else if (!FI.isExternal()) {
- DEBUG(std::cerr << "\t[BU] In " << F.getName() << " inlining: "
- << FI.getName() << "\n");
-
- // Get the data structure graph for the called function, closing it
- // if possible (which is only impossible in the case of mutual
- // recursion...
- //
- DSGraph &GI = calculateGraph(FI); // Graph to inline
-
- DEBUG(std::cerr << "\t\t[BU] Got graph for " << FI.getName()
- << " in: " << F.getName() << "\n");
-
- // Clone the callee's graph into the current graph, keeping
- // track of where scalars in the old graph _used_ to point,
- // and of the new nodes matching nodes of the old graph.
- map<Value*, DSNodeHandle> OldValMap;
- map<const DSNode*, DSNode*> OldNodeMap;
-
- // The clone call may invalidate any of the vectors in the data
- // structure graph. Strip locals and don't copy the list of callers
- DSNodeHandle RetVal = Graph->cloneInto(GI, OldValMap, OldNodeMap,
- /*StripScalars*/ true,
- /*StripAllocas*/ true,
- /*CopyCallers*/ false,
- /*CopyOrigCalls*/ false);
-
- // Resolve the arguments in the call to the actual values...
- ResolveArguments(Call, FI, OldValMap);
-
- if (Call.getReturnValueNode().getNode()) // Handle the return value if present
- RetVal.mergeWith(Call.getReturnValueNode());
-
- // Erase the entry in the Callees vector
- Callees.erase(Callees.begin()+c--);
-
- } else if (FI.getName() == "printf" || FI.getName() == "sscanf" ||
- FI.getName() == "fprintf" || FI.getName() == "open" ||
- FI.getName() == "sprintf") {
-
- // Erase the entry in the globals vector
- Callees.erase(Callees.begin()+c--);
- }
- }
-
- if (Callees.empty()) { // Inlined all of the function calls?
- // Erase the call if it is resolvable...
- FCs.erase(FCs.begin()+i--); // Don't skip a the next call...
- Inlined = true;
- } else if (Callees.size() != Call.getCalleeNode().getNode()->getGlobals().size()) {
- // Was able to inline SOME, but not all of the functions. Construct a
- // new global node here.
- //
- assert(0 && "Unimpl!");
- Inlined = true;
- }
+ // Start with a copy of the original call sites...
+ Graph->getAuxFunctionCalls() = Graph->getFunctionCalls();
+ return *Graph;
+}
+
+unsigned BUDataStructures::calculateGraphs(Function *F,
+ std::vector<Function*> &Stack,
+ unsigned &NextID,
+ hash_map<Function*, unsigned> &ValMap) {
+ assert(ValMap.find(F) == ValMap.end() && "Shouldn't revisit functions!");
+ unsigned Min = NextID++, MyID = Min;
+ ValMap[F] = Min;
+ Stack.push_back(F);
+
+ if (F->isExternal()) { // sprintf, fprintf, sscanf, etc...
+ // No callees!
+ Stack.pop_back();
+ ValMap[F] = ~0;
+ return Min;
+ }
+
+ DSGraph &Graph = getOrCreateGraph(F);
+
+ // The edges out of the current node are the call site targets...
+ for (DSCallSiteIterator I = DSCallSiteIterator::begin_aux(Graph),
+ E = DSCallSiteIterator::end_aux(Graph); I != E; ++I) {
+ Function *Callee = *I;
+ unsigned M;
+ // Have we visited the destination function yet?
+ hash_map<Function*, unsigned>::iterator It = ValMap.find(Callee);
+ if (It == ValMap.end()) // No, visit it now.
+ M = calculateGraphs(Callee, Stack, NextID, ValMap);
+ else // Yes, get it's number.
+ M = It->second;
+ if (M < Min) Min = M;
+ }
+
+ assert(ValMap[F] == MyID && "SCC construction assumption wrong!");
+ if (Min != MyID)
+ return Min; // This is part of a larger SCC!
+
+ // If this is a new SCC, process it now.
+ if (Stack.back() == F) { // Special case the single "SCC" case here.
+ DEBUG(std::cerr << "Visiting single node SCC #: " << MyID << " fn: "
+ << F->getName() << "\n");
+ Stack.pop_back();
+ DSGraph &G = getDSGraph(*F);
+ DEBUG(std::cerr << " [BU] Calculating graph for: " << F->getName()<< "\n");
+ calculateGraph(G);
+ DEBUG(std::cerr << " [BU] Done inlining: " << F->getName() << " ["
+ << G.getGraphSize() << "+" << G.getAuxFunctionCalls().size()
+ << "]\n");
+
+ if (MaxSCC < 1) MaxSCC = 1;
+
+ // Should we revisit the graph?
+ if (DSCallSiteIterator::begin_aux(G) != DSCallSiteIterator::end_aux(G)) {
+ ValMap.erase(F);
+ return calculateGraphs(F, Stack, NextID, ValMap);
+ } else {
+ ValMap[F] = ~0U;
+ }
+ return MyID;
+
+ } else {
+ // SCCFunctions - Keep track of the functions in the current SCC
+ //
+ hash_set<Function*> SCCFunctions;
+
+ Function *NF;
+ std::vector<Function*>::iterator FirstInSCC = Stack.end();
+ DSGraph *SCCGraph = 0;
+ do {
+ NF = *--FirstInSCC;
+ ValMap[NF] = ~0U;
+ SCCFunctions.insert(NF);
+
+ // Figure out which graph is the largest one, in order to speed things up
+ // a bit in situations where functions in the SCC have widely different
+ // graph sizes.
+ DSGraph &NFGraph = getDSGraph(*NF);
+ if (!SCCGraph || SCCGraph->getGraphSize() < NFGraph.getGraphSize())
+ SCCGraph = &NFGraph;
+ } while (NF != F);
+
+ std::cerr << "Calculating graph for SCC #: " << MyID << " of size: "
+ << SCCFunctions.size() << "\n";
+
+ // Compute the Max SCC Size...
+ if (MaxSCC < SCCFunctions.size())
+ MaxSCC = SCCFunctions.size();
+
+ // First thing first, collapse all of the DSGraphs into a single graph for
+ // the entire SCC. We computed the largest graph, so clone all of the other
+ // (smaller) graphs into it. Discard all of the old graphs.
+ //
+ for (hash_set<Function*>::iterator I = SCCFunctions.begin(),
+ E = SCCFunctions.end(); I != E; ++I) {
+ DSGraph &G = getDSGraph(**I);
+ if (&G != SCCGraph) {
+ DSGraph::NodeMapTy NodeMap;
+ SCCGraph->cloneInto(G, SCCGraph->getScalarMap(),
+ SCCGraph->getReturnNodes(), NodeMap, 0);
+ // Update the DSInfo map and delete the old graph...
+ DSInfo[*I] = SCCGraph;
+ delete &G;
}
}
- // Recompute the Incomplete markers. If there are any function calls left
- // now that are complete, we must loop!
- if (Inlined) {
- Graph->maskIncompleteMarkers();
- Graph->markIncompleteNodes();
- Graph->removeDeadNodes(/*KeepAllGlobals*/ true, /*KeepCalls*/ true);
+ // Clean up the graph before we start inlining a bunch again...
+ SCCGraph->removeTriviallyDeadNodes();
+
+ // Now that we have one big happy family, resolve all of the call sites in
+ // the graph...
+ calculateGraph(*SCCGraph);
+ DEBUG(std::cerr << " [BU] Done inlining SCC [" << SCCGraph->getGraphSize()
+ << "+" << SCCGraph->getAuxFunctionCalls().size() << "]\n");
+
+ std::cerr << "DONE with SCC #: " << MyID << "\n";
+
+ // We never have to revisit "SCC" processed functions...
+
+ // Drop the stuff we don't need from the end of the stack
+ Stack.erase(FirstInSCC, Stack.end());
+ return MyID;
+ }
+
+ return MyID; // == Min
+}
+
+
+// releaseMemory - If the pass pipeline is done with this pass, we can release
+// our memory... here...
+//
+void BUDataStructures::releaseMemory() {
+ for (hash_map<Function*, DSGraph*>::iterator I = DSInfo.begin(),
+ E = DSInfo.end(); I != E; ++I) {
+ I->second->getReturnNodes().erase(I->first);
+ if (I->second->getReturnNodes().empty())
+ delete I->second;
+ }
+
+ // Empty map so next time memory is released, data structures are not
+ // re-deleted.
+ DSInfo.clear();
+ delete GlobalsGraph;
+ GlobalsGraph = 0;
+}
+
+void BUDataStructures::calculateGraph(DSGraph &Graph) {
+ // Move our call site list into TempFCs so that inline call sites go into the
+ // new call site list and doesn't invalidate our iterators!
+ std::vector<DSCallSite> TempFCs;
+ std::vector<DSCallSite> &AuxCallsList = Graph.getAuxFunctionCalls();
+ TempFCs.swap(AuxCallsList);
+
+ DSGraph::ReturnNodesTy &ReturnNodes = Graph.getReturnNodes();
+
+ // Loop over all of the resolvable call sites
+ unsigned LastCallSiteIdx = ~0U;
+ for (DSCallSiteIterator I = DSCallSiteIterator::begin(TempFCs),
+ E = DSCallSiteIterator::end(TempFCs); I != E; ++I) {
+ // If we skipped over any call sites, they must be unresolvable, copy them
+ // to the real call site list.
+ LastCallSiteIdx++;
+ for (; LastCallSiteIdx < I.getCallSiteIdx(); ++LastCallSiteIdx)
+ AuxCallsList.push_back(TempFCs[LastCallSiteIdx]);
+ LastCallSiteIdx = I.getCallSiteIdx();
+
+ // Resolve the current call...
+ Function *Callee = *I;
+ DSCallSite CS = I.getCallSite();
+
+ if (Callee->isExternal()) {
+ // Ignore this case, simple varargs functions we cannot stub out!
+ } else if (ReturnNodes.find(Callee) != ReturnNodes.end()) {
+ // Self recursion... simply link up the formal arguments with the
+ // actual arguments...
+ DEBUG(std::cerr << " Self Inlining: " << Callee->getName() << "\n");
+
+ // Handle self recursion by resolving the arguments and return value
+ Graph.mergeInGraph(CS, *Callee, Graph, 0);
+
+ } else {
+ ActualCallees.insert(std::make_pair(&CS.getCallInst(), Callee));
+
+ // Get the data structure graph for the called function.
+ //
+ DSGraph &GI = getDSGraph(*Callee); // Graph to inline
+
+ DEBUG(std::cerr << " Inlining graph for " << Callee->getName()
+ << "[" << GI.getGraphSize() << "+"
+ << GI.getAuxFunctionCalls().size() << "] into '"
+ << Graph.getFunctionNames() << "' [" << Graph.getGraphSize() << "+"
+ << Graph.getAuxFunctionCalls().size() << "]\n");
+
+ // Handle self recursion by resolving the arguments and return value
+ Graph.mergeInGraph(CS, *Callee, GI,
+ DSGraph::KeepModRefBits |
+ DSGraph::StripAllocaBit | DSGraph::DontCloneCallNodes);
+ ++NumBUInlines;
+
+#if 0
+ Graph.writeGraphToFile(std::cerr, "bu_" + F.getName() + "_after_" +
+ Callee->getName());
+#endif
}
- } while (Inlined && !FCs.empty());
+ }
- Graph->maskIncompleteMarkers();
- Graph->markIncompleteNodes();
- Graph->removeTriviallyDeadNodes(false);
- Graph->removeDeadNodes(/*KeepAllGlobals*/ true, /*KeepCalls*/ true);
+ // Make sure to catch any leftover unresolvable calls...
+ for (++LastCallSiteIdx; LastCallSiteIdx < TempFCs.size(); ++LastCallSiteIdx)
+ AuxCallsList.push_back(TempFCs[LastCallSiteIdx]);
- DEBUG(std::cerr << " [BU] Done inlining: " << F.getName() << " ["
- << Graph->getGraphSize() << "+" << Graph->getFunctionCalls().size()
- << "]\n");
+ TempFCs.clear();
- return *Graph;
+ // Re-materialize nodes from the globals graph.
+ // Do not ignore globals inlined from callees -- they are not up-to-date!
+ Graph.getInlinedGlobals().clear();
+ Graph.updateFromGlobalGraph();
+
+ // Recompute the Incomplete markers
+ Graph.maskIncompleteMarkers();
+ Graph.markIncompleteNodes(DSGraph::MarkFormalArgs);
+
+ // Delete dead nodes. Treat globals that are unreachable but that can
+ // reach live nodes as live.
+ Graph.removeDeadNodes(DSGraph::KeepUnreachableGlobals);
+
+ //Graph.writeGraphToFile(std::cerr, "bu_" + F.getName());
}
+