llvm-6502/lib/Analysis/DataStructure/TopDownClosure.cpp
Chris Lattner c3f5f7701f Instead of callign removeTriviallyDeadNodes on the global graph every time
removeDeadNodes is called, only call it at the end of the pass being run.
This saves 1.3 seconds running DSA on 177.mesa (5.3->4.0s), which is
pretty big.  This is only possible because of the automatic garbage
collection done on forwarding nodes.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11178 91177308-0d34-0410-b5e6-96231b3b80d8
2004-02-08 01:51:48 +00:00

292 lines
12 KiB
C++

//===- TopDownClosure.cpp - Compute the top-down interprocedure closure ---===//
//
// The LLVM Compiler Infrastructure
//
// This file was developed by the LLVM research group and is distributed under
// the University of Illinois Open Source License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file implements the TDDataStructures class, which represents the
// Top-down Interprocedural closure of the data structure graph over the
// program. This is useful (but not strictly necessary?) for applications
// like pointer analysis.
//
//===----------------------------------------------------------------------===//
#include "llvm/Analysis/DataStructure.h"
#include "llvm/Module.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Analysis/DSGraph.h"
#include "Support/Debug.h"
#include "Support/Statistic.h"
using namespace llvm;
namespace {
RegisterAnalysis<TDDataStructures> // Register the pass
Y("tddatastructure", "Top-down Data Structure Analysis");
Statistic<> NumTDInlines("tddatastructures", "Number of graphs inlined");
}
void TDDataStructures::markReachableFunctionsExternallyAccessible(DSNode *N,
hash_set<DSNode*> &Visited) {
if (!N || Visited.count(N)) return;
Visited.insert(N);
for (unsigned i = 0, e = N->getNumLinks(); i != e; ++i) {
DSNodeHandle &NH = N->getLink(i*N->getPointerSize());
if (DSNode *NN = NH.getNode()) {
const std::vector<GlobalValue*> &Globals = NN->getGlobals();
for (unsigned G = 0, e = Globals.size(); G != e; ++G)
if (Function *F = dyn_cast<Function>(Globals[G]))
ArgsRemainIncomplete.insert(F);
markReachableFunctionsExternallyAccessible(NN, Visited);
}
}
}
// run - Calculate the top down data structure graphs for each function in the
// program.
//
bool TDDataStructures::run(Module &M) {
BUDataStructures &BU = getAnalysis<BUDataStructures>();
GlobalsGraph = new DSGraph(BU.getGlobalsGraph());
GlobalsGraph->setPrintAuxCalls();
// Figure out which functions must not mark their arguments complete because
// they are accessible outside this compilation unit. Currently, these
// arguments are functions which are reachable by global variables in the
// globals graph.
const DSScalarMap &GGSM = GlobalsGraph->getScalarMap();
hash_set<DSNode*> Visited;
for (DSScalarMap::global_iterator I = GGSM.global_begin(), E = GGSM.global_end();
I != E; ++I)
markReachableFunctionsExternallyAccessible(GGSM.find(*I)->second.getNode(), Visited);
// Loop over unresolved call nodes. Any functions passed into (but not
// returned!) from unresolvable call nodes may be invoked outside of the
// current module.
const std::vector<DSCallSite> &Calls = GlobalsGraph->getAuxFunctionCalls();
for (unsigned i = 0, e = Calls.size(); i != e; ++i) {
const DSCallSite &CS = Calls[i];
for (unsigned arg = 0, e = CS.getNumPtrArgs(); arg != e; ++arg)
markReachableFunctionsExternallyAccessible(CS.getPtrArg(arg).getNode(),
Visited);
}
Visited.clear();
// Functions without internal linkage also have unknown incoming arguments!
for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
if (!I->isExternal() && !I->hasInternalLinkage())
ArgsRemainIncomplete.insert(I);
// We want to traverse the call graph in reverse post-order. To do this, we
// calculate a post-order traversal, then reverse it.
hash_set<DSGraph*> VisitedGraph;
std::vector<DSGraph*> PostOrder;
const BUDataStructures::ActualCalleesTy &ActualCallees =
getAnalysis<BUDataStructures>().getActualCallees();
// Calculate top-down from main...
if (Function *F = M.getMainFunction())
ComputePostOrder(*F, VisitedGraph, PostOrder, ActualCallees);
// Next calculate the graphs for each unreachable function...
for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
ComputePostOrder(*I, VisitedGraph, PostOrder, ActualCallees);
VisitedGraph.clear(); // Release memory!
// Visit each of the graphs in reverse post-order now!
while (!PostOrder.empty()) {
inlineGraphIntoCallees(*PostOrder.back());
PostOrder.pop_back();
}
ArgsRemainIncomplete.clear();
GlobalsGraph->removeTriviallyDeadNodes();
return false;
}
DSGraph &TDDataStructures::getOrCreateDSGraph(Function &F) {
DSGraph *&G = DSInfo[&F];
if (G == 0) { // Not created yet? Clone BU graph...
G = new DSGraph(getAnalysis<BUDataStructures>().getDSGraph(F));
G->getAuxFunctionCalls().clear();
G->setPrintAuxCalls();
G->setGlobalsGraph(GlobalsGraph);
}
return *G;
}
void TDDataStructures::ComputePostOrder(Function &F,hash_set<DSGraph*> &Visited,
std::vector<DSGraph*> &PostOrder,
const BUDataStructures::ActualCalleesTy &ActualCallees) {
if (F.isExternal()) return;
DSGraph &G = getOrCreateDSGraph(F);
if (Visited.count(&G)) return;
Visited.insert(&G);
// Recursively traverse all of the callee graphs.
const std::vector<DSCallSite> &FunctionCalls = G.getFunctionCalls();
for (unsigned i = 0, e = FunctionCalls.size(); i != e; ++i) {
Instruction *CallI = FunctionCalls[i].getCallSite().getInstruction();
std::pair<BUDataStructures::ActualCalleesTy::const_iterator,
BUDataStructures::ActualCalleesTy::const_iterator>
IP = ActualCallees.equal_range(CallI);
for (BUDataStructures::ActualCalleesTy::const_iterator I = IP.first;
I != IP.second; ++I)
ComputePostOrder(*I->second, Visited, PostOrder, ActualCallees);
}
PostOrder.push_back(&G);
}
// releaseMemory - If the pass pipeline is done with this pass, we can release
// our memory... here...
//
// FIXME: This should be releaseMemory and will work fine, except that LoadVN
// has no way to extend the lifetime of the pass, which screws up ds-aa.
//
void TDDataStructures::releaseMyMemory() {
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 TDDataStructures::inlineGraphIntoCallees(DSGraph &Graph) {
// Recompute the Incomplete markers and eliminate unreachable nodes.
Graph.maskIncompleteMarkers();
// If any of the functions has incomplete incoming arguments, don't mark any
// of them as complete.
bool HasIncompleteArgs = false;
const DSGraph::ReturnNodesTy &GraphReturnNodes = Graph.getReturnNodes();
for (DSGraph::ReturnNodesTy::const_iterator I = GraphReturnNodes.begin(),
E = GraphReturnNodes.end(); I != E; ++I)
if (ArgsRemainIncomplete.count(I->first)) {
HasIncompleteArgs = true;
break;
}
// Now fold in the necessary globals from the GlobalsGraph. A global G
// must be folded in if it exists in the current graph (i.e., is not dead)
// and it was not inlined from any of my callers. If it was inlined from
// a caller, it would have been fully consistent with the GlobalsGraph
// in the caller so folding in is not necessary. Otherwise, this node came
// solely from this function's BU graph and so has to be made consistent.
//
Graph.updateFromGlobalGraph();
// Recompute the Incomplete markers. Depends on whether args are complete
unsigned Flags
= HasIncompleteArgs ? DSGraph::MarkFormalArgs : DSGraph::IgnoreFormalArgs;
Graph.markIncompleteNodes(Flags | DSGraph::IgnoreGlobals);
// Delete dead nodes. Treat globals that are unreachable as dead also.
Graph.removeDeadNodes(DSGraph::RemoveUnreachableGlobals);
// We are done with computing the current TD Graph! Now move on to
// inlining the current graph into the graphs for its callees, if any.
//
const std::vector<DSCallSite> &FunctionCalls = Graph.getFunctionCalls();
if (FunctionCalls.empty()) {
DEBUG(std::cerr << " [TD] No callees for: " << Graph.getFunctionNames()
<< "\n");
return;
}
// Now that we have information about all of the callees, propagate the
// current graph into the callees. Clone only the reachable subgraph at
// each call-site, not the entire graph (even though the entire graph
// would be cloned only once, this should still be better on average).
//
DEBUG(std::cerr << " [TD] Inlining '" << Graph.getFunctionNames() <<"' into "
<< FunctionCalls.size() << " call nodes.\n");
const BUDataStructures::ActualCalleesTy &ActualCallees =
getAnalysis<BUDataStructures>().getActualCallees();
// Loop over all the call sites and all the callees at each call site. Build
// a mapping from called DSGraph's to the call sites in this function that
// invoke them. This is useful because we can be more efficient if there are
// multiple call sites to the callees in the graph from this caller.
std::multimap<DSGraph*, std::pair<Function*, const DSCallSite*> > CallSites;
for (unsigned i = 0, e = FunctionCalls.size(); i != e; ++i) {
Instruction *CallI = FunctionCalls[i].getCallSite().getInstruction();
// For each function in the invoked function list at this call site...
std::pair<BUDataStructures::ActualCalleesTy::const_iterator,
BUDataStructures::ActualCalleesTy::const_iterator>
IP = ActualCallees.equal_range(CallI);
// Loop over each actual callee at this call site
for (BUDataStructures::ActualCalleesTy::const_iterator I = IP.first;
I != IP.second; ++I) {
DSGraph& CalleeGraph = getDSGraph(*I->second);
assert(&CalleeGraph != &Graph && "TD need not inline graph into self!");
CallSites.insert(std::make_pair(&CalleeGraph,
std::make_pair(I->second, &FunctionCalls[i])));
}
}
// Now that we built the mapping, actually perform the inlining a callee graph
// at a time.
std::multimap<DSGraph*,std::pair<Function*,const DSCallSite*> >::iterator CSI;
for (CSI = CallSites.begin(); CSI != CallSites.end(); ) {
DSGraph &CalleeGraph = *CSI->first;
// Iterate through all of the call sites of this graph, cloning and merging
// any nodes required by the call.
ReachabilityCloner RC(CalleeGraph, Graph, DSGraph::StripModRefBits);
// Clone over any global nodes that appear in both graphs.
for (DSScalarMap::global_iterator
SI = CalleeGraph.getScalarMap().global_begin(),
SE = CalleeGraph.getScalarMap().global_end(); SI != SE; ++SI) {
DSScalarMap::const_iterator GI = Graph.getScalarMap().find(*SI);
if (GI != Graph.getScalarMap().end())
RC.merge(CalleeGraph.getNodeForValue(*SI), GI->second);
}
// Loop over all of the distinct call sites in the caller of the callee.
for (; CSI != CallSites.end() && CSI->first == &CalleeGraph; ++CSI) {
Function &CF = *CSI->second.first;
const DSCallSite &CS = *CSI->second.second;
DEBUG(std::cerr << " [TD] Resolving arguments for callee graph '"
<< CalleeGraph.getFunctionNames()
<< "': " << CF.getFunctionType()->getNumParams()
<< " args\n at call site (DSCallSite*) 0x" << &CS << "\n");
// Get the formal argument and return nodes for the called function and
// merge them with the cloned subgraph.
RC.mergeCallSite(CalleeGraph.getCallSiteForArguments(CF), CS);
++NumTDInlines;
}
}
DEBUG(std::cerr << " [TD] Done inlining into callees for: "
<< Graph.getFunctionNames() << " [" << Graph.getGraphSize() << "+"
<< Graph.getFunctionCalls().size() << "]\n");
}