//
// Method: calculateGraphs()
//
// Description:
// Perform recursive bottom-up inlining of DSGraphs from callee to caller.
//
// Inputs:
// F - The function which should have its callees' DSGraphs merged into its
// own DSGraph.
// Stack - The stack used for Tarjan's SCC-finding algorithm.
// NextID - The nextID value used for Tarjan's SCC-finding algorithm.
// ValMap - The map used for Tarjan's SCC-finding algorithm.
//
// Return value:
//
unsigned
BUDataStructures::calculateGraphs (const Function *F,
TarjanStack & Stack,
unsigned & NextID,
TarjanMap & ValMap) {
assert(!ValMap.count(F) && "Shouldn't revisit functions!");
unsigned Min = NextID++, MyID = Min;
ValMap[F] = Min;
Stack.push_back(F);
//
// FIXME: This test should be generalized to be any function that we have
// already processed in the case when there isn't a main() or there are
// unreachable functions!
//
if (F->isDeclaration()) { // sprintf, fprintf, sscanf, etc...
// No callees!
Stack.pop_back();
ValMap[F] = ~0;
return Min;
}
//
// Get the DSGraph of the current function. Make one if one doesn't exist.
//
DSGraph* Graph = getOrCreateGraph(F);
//
// Find all callee functions. Use the DSGraph for this (do not use the call
// graph (DSCallgraph) as we're still in the process of constructing it).
//
FuncSet CalleeFunctions;
getAllAuxCallees(Graph, CalleeFunctions);
//
// Iterate through each call target (these are the edges out of the current
// node (i.e., the current function) in Tarjan graph parlance). Find the
// minimum assigned ID.
//
for (FuncSet::iterator I = CalleeFunctions.begin(), E = CalleeFunctions.end();
I != E; ++I) {
const Function *Callee = *I;
unsigned M;
//
// If we have not visited this callee before, visit it now (this is the
// post-order component of the Bottom-Up algorithm). Otherwise, look up
// the assigned ID value from the Tarjan Value Map.
//
TarjanMap::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 we've found a function with a smaller ID than this funtion, record
// that ID as the minimum ID.
//
if (M < Min) Min = M;
}
assert(ValMap[F] == MyID && "SCC construction assumption wrong!");
//
// If the minimum ID found is not this function's ID, then this function is
// part of a larger SCC.
//
if (Min != MyID)
return Min;
//
// If this is a new SCC, process it now.
//
if (Stack.back() == F) { // Special case the single "SCC" case here.
DEBUG(errs() << "Visiting single node SCC #: " << MyID << " fn: "
<< F->getName() << "\n");
Stack.pop_back();
DEBUG(errs() << " [BU] Calculating graph for: " << F->getName()<< "\n");
DSGraph* G = getOrCreateGraph(F);
calculateGraph(G);
DEBUG(errs() << " [BU] Done inlining: " << F->getName() << " ["
<< G->getGraphSize() << "+" << G->getAuxFunctionCalls().size()
<< "]\n");
if (MaxSCC < 1) MaxSCC = 1;
//
// Should we revisit the graph? Only do it if there are now new resolvable
// callees.
FuncSet NewCallees;
getAllAuxCallees(G, NewCallees);
if (!NewCallees.empty()) {
if (hasNewCallees(NewCallees, CalleeFunctions)) {
DEBUG(errs() << "Recalculating " << F->getName() << " due to new knowledge\n");
ValMap.erase(F);
++NumRecalculations;
return calculateGraphs(F, Stack, NextID, ValMap);
}
++NumRecalculationsSkipped;
}
ValMap[F] = ~0U;
return MyID;
} else {
unsigned SCCSize = 1;
const Function *NF = Stack.back();
if(NF != F)
ValMap[NF] = ~0U;
DSGraph* SCCGraph = getDSGraph(*NF);
//
// First thing first: collapse all of the DSGraphs into a single graph for
// the entire SCC. Splice all of the graphs into one and discard all of
// the old graphs.
//
while (NF != F) {
Stack.pop_back();
NF = Stack.back();
if(NF != F)
ValMap[NF] = ~0U;
DSGraph* NFG = getDSGraph(*NF);
if (NFG != SCCGraph) {
// Update the Function -> DSG map.
for (DSGraph::retnodes_iterator I = NFG->retnodes_begin(),
E = NFG->retnodes_end(); I != E; ++I)
setDSGraph(*I->first, SCCGraph);
SCCGraph->spliceFrom(NFG);
delete NFG;
++SCCSize;
}
}
Stack.pop_back();
DEBUG(errs() << "Calculating graph for SCC #: " << MyID << " of size: "
<< SCCSize << "\n");
// Compute the Max SCC Size.
if (MaxSCC < SCCSize)
MaxSCC = SCCSize;
// Clean up the graph before we start inlining a bunch again...
SCCGraph->removeDeadNodes(DSGraph::KeepUnreachableGlobals);
// Now that we have one big happy family, resolve all of the call sites in
// the graph...
calculateGraph(SCCGraph);
DEBUG(errs() << " [BU] Done inlining SCC [" << SCCGraph->getGraphSize()
<< "+" << SCCGraph->getAuxFunctionCalls().size() << "]\n"
<< "DONE with SCC #: " << MyID << "\n");
FuncSet NewCallees;
getAllAuxCallees(SCCGraph, NewCallees);
if (!NewCallees.empty()) {
if (hasNewCallees(NewCallees, CalleeFunctions)) {
DEBUG(errs() << "Recalculating SCC Graph " << F->getName() << " due to new knowledge\n");
ValMap.erase(F);
++NumRecalculations;
return calculateGraphs(F, Stack, NextID, ValMap);
}
++NumRecalculationsSkipped;
}
ValMap[F] = ~0U;
return MyID;
}
}
/// visitGraph - Visit the functions in the specified graph, updating the
/// specified lattice values for all of their uses.
///
void StructureFieldVisitorBase::
visitGraph(DSGraph &DSG, std::multimap<DSNode*, LatticeValue*> &NodeLVs) {
assert(!NodeLVs.empty() && "No lattice values to compute!");
// To visit a graph, first step, we visit the instruction making up each
// function in the graph, but ignore calls when processing them. We handle
// call nodes explicitly by looking at call nodes in the graph if needed. We
// handle instructions before calls to avoid interprocedural analysis if we
// can drive lattice values to bottom early.
//
SFVInstVisitor IV(DSG, Callbacks, NodeLVs);
for (DSGraph::retnodes_iterator FI = DSG.retnodes_begin(),
E = DSG.retnodes_end(); FI != E; ++FI)
for (Function::iterator BB = FI->first->begin(), E = FI->first->end();
BB != E; ++BB)
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
if (IV.visit(*I) && NodeLVs.empty())
return; // Nothing left to analyze.
// Keep track of which actual direct callees are handled.
std::set<Function*> CalleesHandled;
// Once we have visited all of the instructions in the function bodies, if
// there are lattice values that have not been driven to bottom, see if any of
// the nodes involved are passed into function calls. If so, we potentially
// have to recursively traverse the call graph.
for (DSGraph::fc_iterator CS = DSG.fc_begin(), E = DSG.fc_end();
CS != E; ++CS) {
// Figure out the mapping from a node in the caller (potentially several)
// nodes in the callee.
DSGraph::NodeMapTy CallNodeMap;
Instruction *TheCall = CS->getCallSite().getInstruction();
// If this is an indirect function call, assume nothing gets passed through
// it. FIXME: THIS IS BROKEN! Just get the ECG for the fn ptr if it's not
// direct.
if (CS->isIndirectCall())
continue;
// If this is an external function call, it cannot be involved with this
// node, because otherwise the node would be marked incomplete!
if (CS->getCalleeFunc()->isExternal())
continue;
// If we can handle this function call, remove it from the set of direct
// calls found by the visitor.
CalleesHandled.insert(CS->getCalleeFunc());
std::vector<DSNodeHandle> Args;
DSGraph *CG = &ECG.getDSGraph(*CS->getCalleeFunc());
CG->getFunctionArgumentsForCall(CS->getCalleeFunc(), Args);
if (!CS->getRetVal().isNull())
DSGraph::computeNodeMapping(Args[0], CS->getRetVal(), CallNodeMap);
for (unsigned i = 0, e = CS->getNumPtrArgs(); i != e; ++i) {
if (i == Args.size()-1) break;
DSGraph::computeNodeMapping(Args[i+1], CS->getPtrArg(i), CallNodeMap);
}
Args.clear();
// The mapping we just computed maps from nodes in the callee to nodes in
// the caller, so we can't query it efficiently. Instead of going through
// the trouble of inverting the map to do this (linear time with the size of
// the mapping), we just do a linear search to see if any affected nodes are
// passed into this call.
bool CallCanModifyDataFlow = false;
for (DSGraph::NodeMapTy::iterator MI = CallNodeMap.begin(),
E = CallNodeMap.end(); MI != E; ++MI)
if (NodeLVs.count(MI->second.getNode()))
// Okay, the node is passed in, check to see if the call might do
// something interesting to it (i.e. if analyzing the call can produce
// anything other than "top").
if ((CallCanModifyDataFlow = NodeCanPossiblyBeInteresting(MI->first,
Callbacks)))
break;
// If this function call cannot impact the analysis (either because the
// nodes we are tracking are not passed into the call, or the DSGraph for
// the callee tells us that analysis of the callee can't provide interesting
// information), ignore it.
if (!CallCanModifyDataFlow)
continue;
// Okay, either compute analysis results for the callee function, or reuse
// results previously computed.
std::multimap<DSNode*, LatticeValue*> &CalleeFacts = getCalleeFacts(*CG);
// Merge all of the facts for the callee into the facts for the caller. If
// this reduces anything in the caller to 'bottom', remove them.
for (DSGraph::NodeMapTy::iterator MI = CallNodeMap.begin(),
E = CallNodeMap.end(); MI != E; ++MI) {
// If we have Lattice facts in the caller for this node in the callee,
// merge any information from the callee into the caller.
// If the node is not accessed in the callee at all, don't update.
if (MI->first->getType() == Type::VoidTy)
continue;
// If there are no data-flow facts live in the caller for this node, don't
// both processing it.
std::multimap<DSNode*, LatticeValue*>::iterator NLVI =
NodeLVs.find(MI->second.getNode());
if (NLVI == NodeLVs.end()) continue;
// Iterate over all of the lattice values that have corresponding fields
// in the callee, merging in information as we go. Be careful about the
// fact that the callee may get passed the address of a substructure and
// other funny games.
//if (CalleeFacts.count(const_cast<DSNode*>(MI->first)) == 0) {
DSNode *CalleeNode = const_cast<DSNode*>(MI->first);
unsigned CalleeNodeOffset = MI->second.getOffset();
while (NLVI->first == MI->second.getNode()) {
// Figure out what offset in the callee this field would land.
unsigned FieldOff = NLVI->second->getFieldOffset()+CalleeNodeOffset;
// If the field is not within the callee node, ignore it.
if (FieldOff >= CalleeNode->getSize()) {
++NLVI;
continue;
}
// Okay, check to see if we have a lattice value for the field at offset
// FieldOff in the callee node.
const LatticeValue *CalleeLV = 0;
std::multimap<DSNode*, LatticeValue*>::iterator CFI =
CalleeFacts.lower_bound(CalleeNode);
for (; CFI != CalleeFacts.end() && CFI->first == CalleeNode; ++CFI)
if (CFI->second->getFieldOffset() == FieldOff) {
CalleeLV = CFI->second; // Found it!
break;
}
// If we don't, the lattice value hit bottom and we should remove the
// lattice value in the caller.
if (!CalleeLV) {
delete NLVI->second; // The lattice value hit bottom.
NodeLVs.erase(NLVI++);
continue;
}
// Finally, if we did find a corresponding entry, merge the information
// into the caller's lattice value and keep going.
if (NLVI->second->mergeInValue(CalleeLV)) {
// Okay, merging these two caused the caller value to hit bottom.
// Remove it.
delete NLVI->second; // The lattice value hit bottom.
NodeLVs.erase(NLVI++);
}
++NLVI; // We successfully merged in some information!
}
// If we ran out of facts to prove, just exit.
if (NodeLVs.empty()) return;
}
}
// The local analysis pass inconveniently discards many local function calls
// from the graph if they are to known functions. Loop over direct function
// calls not handled above and visit them as appropriate.
while (!IV.DirectCallSites.empty()) {
Instruction *Call = *IV.DirectCallSites.begin();
IV.DirectCallSites.erase(IV.DirectCallSites.begin());
// Is this one actually handled by DSA?
if (CalleesHandled.count(cast<Function>(Call->getOperand(0))))
continue;
// Collect the pointers involved in this call.
std::vector<Value*> Pointers;
if (isa<PointerType>(Call->getType()))
Pointers.push_back(Call);
for (unsigned i = 1, e = Call->getNumOperands(); i != e; ++i)
if (isa<PointerType>(Call->getOperand(i)->getType()))
Pointers.push_back(Call->getOperand(i));
// If this is an intrinsic function call, figure out which one.
unsigned IID = cast<Function>(Call->getOperand(0))->getIntrinsicID();
for (unsigned i = 0, e = Pointers.size(); i != e; ++i) {
// If any of our lattice values are passed into this call, which is
// specially handled by the local analyzer, inform the lattice function.
DSNode *N = DSG.getNodeForValue(Pointers[i]).getNode();
for (std::multimap<DSNode*, LatticeValue*>::iterator LVI =
NodeLVs.lower_bound(N); LVI != NodeLVs.end() && LVI->first == N;) {
bool AtBottom = false;
switch (IID) {
default:
AtBottom = LVI->second->visitRecognizedCall(*Call);
break;
case Intrinsic::memset:
if (Callbacks & Visit::Stores)
AtBottom = LVI->second->visitMemSet(*cast<CallInst>(Call));
break;
}
if (AtBottom) {
delete LVI->second;
NodeLVs.erase(LVI++);
} else {
++LVI;
}
}
}
}
}
//
// Method: postOrderInline()
//
// Description:
// This methods does a post order traversal of the call graph and performs
// bottom-up inlining of the DSGraphs.
//
void
BUDataStructures::postOrderInline (Module & M) {
// Variables used for Tarjan SCC-finding algorithm. These are passed into
// the recursive function used to find SCCs.
std::vector<const Function*> Stack;
std::map<const Function*, unsigned> ValMap;
unsigned NextID = 1;
// Do post order traversal on the global ctors. Use this information to update
// the globals graph.
const char *Name = "llvm.global_ctors";
GlobalVariable *GV = M.getNamedGlobal(Name);
if (GV && !(GV->isDeclaration()) && !(GV->hasLocalLinkage())) {
// Should be an array of '{ int, void ()* }' structs. The first value is
// the init priority, which we ignore.
ConstantArray *InitList = dyn_cast<ConstantArray>(GV->getInitializer());
if (InitList) {
for (unsigned i = 0, e = InitList->getNumOperands(); i != e; ++i)
if (ConstantStruct *CS = dyn_cast<ConstantStruct>(InitList->getOperand(i))) {
if (CS->getNumOperands() != 2)
break; // Not array of 2-element structs.
Constant *FP = CS->getOperand(1);
if (FP->isNullValue())
break; // Found a null terminator, exit.
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(FP))
if (CE->isCast())
FP = CE->getOperand(0);
Function *F = dyn_cast<Function>(FP);
if (F && !F->isDeclaration() && !ValMap.count(F)) {
calculateGraphs(F, Stack, NextID, ValMap);
CloneAuxIntoGlobal(getDSGraph(*F));
}
}
GlobalsGraph->removeTriviallyDeadNodes();
GlobalsGraph->maskIncompleteMarkers();
// Mark external globals incomplete.
GlobalsGraph->markIncompleteNodes(DSGraph::IgnoreGlobals);
GlobalsGraph->computeExternalFlags(DSGraph::DontMarkFormalsExternal);
GlobalsGraph->computeIntPtrFlags();
//
// Create equivalence classes for aliasing globals so that we only need to
// record one global per DSNode.
//
formGlobalECs();
// propogte information calculated
// from the globals graph to the other graphs.
for (Module::iterator F = M.begin(); F != M.end(); ++F) {
if (!(F->isDeclaration())){
DSGraph *Graph = getDSGraph(*F);
cloneGlobalsInto(Graph, DSGraph::DontCloneCallNodes |
DSGraph::DontCloneAuxCallNodes);
Graph->buildCallGraph(callgraph, GlobalFunctionList, filterCallees);
Graph->maskIncompleteMarkers();
Graph->markIncompleteNodes(DSGraph::MarkFormalArgs |
DSGraph::IgnoreGlobals);
Graph->computeExternalFlags(DSGraph::DontMarkFormalsExternal);
Graph->computeIntPtrFlags();
}
}
}
}
//
// Start the post order traversal with the main() function. If there is no
// main() function, don't worry; we'll have a separate traversal for inlining
// graphs for functions not reachable from main().
//
Function *MainFunc = M.getFunction ("main");
if (MainFunc && !MainFunc->isDeclaration()) {
calculateGraphs(MainFunc, Stack, NextID, ValMap);
CloneAuxIntoGlobal(getDSGraph(*MainFunc));
}
//
// Calculate the graphs for any functions that are unreachable from main...
//
for (Function &F : M)
if (!F.isDeclaration() && !ValMap.count(&F)) {
if (MainFunc)
DEBUG(errs() << debugname << ": Function unreachable from main: "
<< F.getName() << "\n");
calculateGraphs(&F, Stack, NextID, ValMap); // Calculate all graphs.
CloneAuxIntoGlobal(getDSGraph(F));
// Mark this graph as processed. Do this by finding all functions
// in the graph that map to it, and mark them visited.
// Note that this really should be handled neatly by calculateGraphs
// itself, not here. However this catches the worst offenders.
DSGraph *G = getDSGraph(F);
for(DSGraph::retnodes_iterator RI = G->retnodes_begin(),
RE = G->retnodes_end(); RI != RE; ++RI) {
if (getDSGraph(*RI->first) == G) {
if (!ValMap.count(RI->first))
ValMap[RI->first] = ~0U;
else
assert(ValMap[RI->first] == ~0U);
}
}
}
return;
}
unsigned BUDataStructures::calculateGraphs(const Function *F,
std::vector<const Function*> &Stack,
unsigned &NextID,
hash_map<const Function*, unsigned> &ValMap) {
assert(!ValMap.count(F) && "Shouldn't revisit functions!");
unsigned Min = NextID++, MyID = Min;
ValMap[F] = Min;
Stack.push_back(F);
// FIXME! This test should be generalized to be any function that we have
// already processed, in the case when there isn't a main or there are
// unreachable functions!
if (F->isDeclaration()) { // sprintf, fprintf, sscanf, etc...
// No callees!
Stack.pop_back();
ValMap[F] = ~0;
return Min;
}
DSGraph* Graph = getOrFetchDSGraph(F);
// Find all callee functions.
std::vector<const Function*> CalleeFunctions;
GetAllAuxCallees(Graph, CalleeFunctions);
std::sort(CalleeFunctions.begin(), CalleeFunctions.end());
std::vector<const Function*>::iterator uid = std::unique(CalleeFunctions.begin(), CalleeFunctions.end());
CalleeFunctions.resize(uid - CalleeFunctions.begin());
// The edges out of the current node are the call site targets...
for (unsigned i = 0, e = CalleeFunctions.size(); i != e; ++i) {
const Function *Callee = CalleeFunctions[i];
unsigned M;
// Have we visited the destination function yet?
hash_map<const 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(errs() << "Visiting single node SCC #: " << MyID << " fn: "
<< F->getName() << "\n");
Stack.pop_back();
DEBUG(errs() << " [BU] Calculating graph for: " << F->getName()<< "\n");
calculateGraph(Graph);
DEBUG(errs() << " [BU] Done inlining: " << F->getName() << " ["
<< Graph->getGraphSize() << "+" << Graph->getAuxFunctionCalls().size()
<< "]\n");
if (MaxSCC < 1) MaxSCC = 1;
// Should we revisit the graph? Only do it if there are now new resolvable
// callees or new callees
GetAllAuxCallees(Graph, CalleeFunctions);
if (CalleeFunctions.size()) {
DEBUG(errs() << "Recalculating " << F->getName() << " due to new knowledge\n");
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
//
std::vector<DSGraph*> SCCGraphs;
unsigned SCCSize = 1;
const Function *NF = Stack.back();
ValMap[NF] = ~0U;
DSGraph* SCCGraph = getDSGraph(NF);
// First thing first, collapse all of the DSGraphs into a single graph for
// the entire SCC. Splice all of the graphs into one and discard all of the
// old graphs.
//
while (NF != F) {
Stack.pop_back();
NF = Stack.back();
ValMap[NF] = ~0U;
DSGraph* NFG = getDSGraph(NF);
if (NFG != SCCGraph) {
// Update the Function -> DSG map.
for (DSGraph::retnodes_iterator I = NFG->retnodes_begin(),
E = NFG->retnodes_end(); I != E; ++I)
setDSGraph(I->first, SCCGraph);
SCCGraph->spliceFrom(NFG);
delete NFG;
++SCCSize;
}
}
Stack.pop_back();
DEBUG(errs() << "Calculating graph for SCC #: " << MyID << " of size: "
<< SCCSize << "\n");
// Compute the Max SCC Size.
if (MaxSCC < SCCSize)
MaxSCC = SCCSize;
// Clean up the graph before we start inlining a bunch again...
SCCGraph->removeDeadNodes(DSGraph::KeepUnreachableGlobals);
// Now that we have one big happy family, resolve all of the call sites in
// the graph...
calculateGraph(SCCGraph);
DEBUG(errs() << " [BU] Done inlining SCC [" << SCCGraph->getGraphSize()
<< "+" << SCCGraph->getAuxFunctionCalls().size() << "]\n"
<< "DONE with SCC #: " << MyID << "\n");
// We never have to revisit "SCC" processed functions...
return MyID;
}
return MyID; // == Min
}
