New implementation of data structure analysis. Only local analysis has been

implemented so far.

llvm-svn: 2871
This commit is contained in:
Chris Lattner 2002-07-10 22:38:08 +00:00
parent b2c5cb2525
commit c0dab43c31
3 changed files with 544 additions and 143 deletions

View File

@ -1,165 +1,127 @@
//===- DataStructure.cpp - Analysis for data structure identification -------=//
//===- DataStructure.cpp - Implement the core data structure analysis -----===//
//
// Implement the LLVM data structure analysis library.
// This file implements the core data structure functionality.
//
//===----------------------------------------------------------------------===//
#include "llvm/Analysis/DataStructure.h"
#include "llvm/Module.h"
#include <fstream>
#include "llvm/DerivedTypes.h"
#include <algorithm>
#include "Support/STLExtras.h"
AnalysisID LocalDataStructures::ID(AnalysisID::create<LocalDataStructures>());
//===----------------------------------------------------------------------===//
// DataStructure Class Implementation
//
// DSNode Implementation
//===----------------------------------------------------------------------===//
AnalysisID DataStructure::ID(AnalysisID::create<DataStructure>());
DSNode::DSNode(enum NodeTy NT, const Type *T) : Ty(T), NodeType(NT) {
// If this node has any fields, allocate them now, but leave them null.
switch (T->getPrimitiveID()) {
case Type::PointerTyID: Links.resize(1); break;
case Type::ArrayTyID: Links.resize(1); break;
case Type::StructTyID:
Links.resize(cast<StructType>(T)->getNumContainedTypes());
break;
default: break;
}
}
void DSNode::removeReferrer(DSNodeHandle *H) {
// Search backwards, because we depopulate the list from the back for
// efficiency (because it's a vector).
std::vector<DSNodeHandle*>::reverse_iterator I =
std::find(Referrers.rbegin(), Referrers.rend(), H);
assert(I != Referrers.rend() && "Referrer not pointing to node!");
Referrers.erase(I.base()-1);
}
// addEdgeTo - Add an edge from the current node to the specified node. This
// can cause merging of nodes in the graph.
//
void DSNode::addEdgeTo(unsigned LinkNo, DSNode *N) {
assert(LinkNo < Links.size() && "LinkNo out of range!");
if (N == 0 || Links[LinkNo] == N) return; // Nothing to do
if (Links[LinkNo] == 0) { // No merging to perform
Links[LinkNo] = N;
return;
}
// Merge the two nodes...
Links[LinkNo]->mergeWith(N);
}
// mergeWith - Merge this node into the specified node, moving all links to and
// from the argument node into the current node. The specified node may be a
// null pointer (in which case, nothing happens).
//
void DSNode::mergeWith(DSNode *N) {
if (N == 0 || N == this) return; // Noop
assert(N->Ty == Ty && N->Links.size() == Links.size() &&
"Cannot merge nodes of two different types!");
// Remove all edges pointing at N, causing them to point to 'this' instead.
while (!N->Referrers.empty())
*N->Referrers.back() = this;
// Make all of the outgoing links of N now be outgoing links of this. This
// can cause recursive merging!
//
for (unsigned i = 0, e = Links.size(); i != e; ++i) {
addEdgeTo(i, N->Links[i]);
N->Links[i] = 0; // Reduce unneccesary edges in graph. N is dead
}
NodeType |= N->NodeType;
N->NodeType = 0; // N is now a dead node.
}
//===----------------------------------------------------------------------===//
// DSGraph Implementation
//===----------------------------------------------------------------------===//
DSGraph::~DSGraph() {
FunctionCalls.clear();
ValueMap.clear();
RetNode = 0;
#ifndef NDEBUG
// Drop all intra-node references, so that assertions don't fail...
std::for_each(Nodes.begin(), Nodes.end(),
std::mem_fun(&DSNode::dropAllReferences));
#endif
// Delete all of the nodes themselves...
std::for_each(Nodes.begin(), Nodes.end(), deleter<DSNode>);
}
//===----------------------------------------------------------------------===//
// LocalDataStructures Implementation
//===----------------------------------------------------------------------===//
// releaseMemory - If the pass pipeline is done with this pass, we can release
// our memory... here...
void DataStructure::releaseMemory() {
for (InfoMap::iterator I = DSInfo.begin(), E = DSInfo.end(); I != E; ++I) {
delete I->second.first;
delete I->second.second;
}
//
void LocalDataStructures::releaseMemory() {
for (std::map<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();
}
// FIXME REMOVE
#include <sys/time.h>
#include "Support/CommandLine.h"
cl::Flag Time("t", "Print analysis time...");
// print - Print out the analysis results...
void DataStructure::print(std::ostream &O, Module *M) const {
if (Time) {
timeval TV1, TV2;
gettimeofday(&TV1, 0);
for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I)
if (!I->isExternal() && I->getName() == "main") {
//getDSGraph(*I);
getClosedDSGraph(I);
}
gettimeofday(&TV2, 0);
std::cerr << "Analysis took "
<< (TV2.tv_sec-TV1.tv_sec)*1000000+(TV2.tv_usec-TV1.tv_usec)
<< " microseconds.\n";
}
for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I)
bool LocalDataStructures::run(Module &M) {
// Calculate all of the graphs...
for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
if (!I->isExternal()) {
std::string Filename = "ds." + I->getName() + ".dot";
O << "Writing '" << Filename << "'...";
std::ofstream F(Filename.c_str());
if (F.good()) {
F << "digraph DataStructures {\n"
<< "\tnode [shape=Mrecord];\n"
<< "\tedge [arrowtail=\"dot\"];\n"
<< "\tsize=\"10,7.5\";\n"
<< "\trotate=\"90\";\n";
getDSGraph(I).printFunction(F, "Local");
getClosedDSGraph(I).printFunction(F, "Closed");
F << "}\n";
} else {
O << " error opening file for writing!\n";
}
if (Time)
O << " [" << getDSGraph(I).getGraphSize() << ", "
<< getClosedDSGraph(I).getGraphSize() << "]\n";
else
O << "\n";
std::map<Function*, DSGraph*>::iterator DI = DSInfo.find(I);
if (DI == DSInfo.end() || DI->second == 0)
DSInfo.insert(std::make_pair(&*I, new DSGraph(*I)));
}
return false;
}
//===----------------------------------------------------------------------===//
// PointerVal Class Implementation
//
void PointerVal::print(std::ostream &O) const {
if (Node) {
O << " Node: " << Node->getCaption() << "[" << Index << "]\n";
} else {
O << " NULL NODE\n";
}
}
//===----------------------------------------------------------------------===//
// PointerValSet Class Implementation
//
void PointerValSet::addRefs() {
for (unsigned i = 0, e = Vals.size(); i != e; ++i)
Vals[i].Node->addReferrer(this);
}
void PointerValSet::dropRefs() {
for (unsigned i = 0, e = Vals.size(); i != e; ++i)
Vals[i].Node->removeReferrer(this);
}
const PointerValSet &PointerValSet::operator=(const PointerValSet &PVS) {
dropRefs();
Vals.clear();
Vals = PVS.Vals;
addRefs();
return *this;
}
// operator< - Allow insertion into a map...
bool PointerValSet::operator<(const PointerValSet &PVS) const {
if (Vals.size() < PVS.Vals.size()) return true;
if (Vals.size() > PVS.Vals.size()) return false;
if (Vals.size() == 1) return Vals[0] < PVS.Vals[0]; // Most common case
std::vector<PointerVal> S1(Vals), S2(PVS.Vals);
sort(S1.begin(), S1.end());
sort(S2.begin(), S2.end());
return S1 < S2;
}
bool PointerValSet::operator==(const PointerValSet &PVS) const {
if (Vals.size() != PVS.Vals.size()) return false;
if (Vals.size() == 1) return Vals[0] == PVS.Vals[0]; // Most common case...
std::vector<PointerVal> S1(Vals), S2(PVS.Vals);
sort(S1.begin(), S1.end());
sort(S2.begin(), S2.end());
return S1 == S2;
}
bool PointerValSet::add(const PointerVal &PV, Value *Pointer) {
if (std::find(Vals.begin(), Vals.end(), PV) != Vals.end())
return false;
Vals.push_back(PV);
if (Pointer) PV.Node->addPointer(Pointer);
PV.Node->addReferrer(this);
return true;
}
// removePointerTo - Remove a single pointer val that points to the specified
// node...
void PointerValSet::removePointerTo(DSNode *Node) {
std::vector<PointerVal>::iterator I = std::find(Vals.begin(), Vals.end(), Node);
assert(I != Vals.end() && "Couldn't remove nonexistent edge!");
Vals.erase(I);
Node->removeReferrer(this);
}
void PointerValSet::print(std::ostream &O) const {
for (unsigned i = 0, e = Vals.size(); i != e; ++i)
Vals[i].print(O);
}

View File

@ -0,0 +1,283 @@
//===- ComputeLocal.cpp - Compute a local data structure graph for a fn ---===//
//
// Compute the local version of the data structure graph for a function. The
// external interface to this file is the DSGraph constructor.
//
//===----------------------------------------------------------------------===//
#include "llvm/Analysis/DataStructure.h"
#include "llvm/Function.h"
#include "llvm/iMemory.h"
#include "llvm/iTerminators.h"
#include "llvm/iPHINode.h"
#include "llvm/iOther.h"
#include "llvm/Constants.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Support/InstVisitor.h"
using std::map;
using std::vector;
//===----------------------------------------------------------------------===//
// GraphBuilder Class
//===----------------------------------------------------------------------===//
//
// This class is the builder class that constructs the local data structure
// graph by performing a single pass over the function in question.
//
namespace {
class GraphBuilder : InstVisitor<GraphBuilder> {
DSGraph &G;
vector<DSNode*> &Nodes;
DSNodeHandle &RetNode; // Node that gets returned...
map<Value*, DSNodeHandle> &ValueMap;
vector<vector<DSNodeHandle> > &FunctionCalls;
public:
GraphBuilder(DSGraph &g, vector<DSNode*> &nodes, DSNodeHandle &retNode,
map<Value*, DSNodeHandle> &vm,
vector<vector<DSNodeHandle> > &fc)
: G(g), Nodes(nodes), RetNode(retNode), ValueMap(vm), FunctionCalls(fc) {
visit(G.getFunction()); // Single pass over the function
removeDeadNodes();
}
private:
// Visitor functions, used to handle each instruction type we encounter...
friend class InstVisitor<GraphBuilder>;
void visitMallocInst(MallocInst &MI) { handleAlloc(MI, DSNode::NewNode); }
void visitAllocaInst(AllocaInst &AI) { handleAlloc(AI, DSNode::AllocaNode);}
void handleAlloc(AllocationInst &AI, DSNode::NodeTy NT);
void visitPHINode(PHINode &PN);
void visitGetElementPtrInst(GetElementPtrInst &GEP);
void visitReturnInst(ReturnInst &RI);
void visitLoadInst(LoadInst &LI);
void visitStoreInst(StoreInst &SI);
void visitCallInst(CallInst &CI);
void visitSetCondInst(SetCondInst &SCI) {} // SetEQ & friends are ignored
void visitFreeInst(FreeInst &FI) {} // Ignore free instructions
void visitInstruction(Instruction &I) {
#ifndef NDEBUG
bool bad = isa<PointerType>(I.getType());
for (Instruction::op_iterator i = I.op_begin(), E = I.op_end(); i!=E; ++i)
bad |= isa<PointerType>(i->get()->getType());
if (bad) {
std::cerr << "\n\n\nUNKNOWN PTR INSTRUCTION type: " << I << "\n\n\n";
assert(0 && "Cannot proceed");
}
#endif
}
private:
// Helper functions used to implement the visitation functions...
// createNode - Create a new DSNode, ensuring that it is properly added to
// the graph.
//
DSNode *createNode(DSNode::NodeTy NodeType, const Type *Ty);
// getValueNode - Return a DSNode that corresponds the the specified LLVM
// value. This either returns the already existing node, or creates a new
// one and adds it to the graph, if none exists.
//
DSNode *getValueNode(Value &V);
// getLink - This method is used to either return the specified link in the
// specified node if one exists. If a link does not already exist (it's
// null), then we create a new node, link it, then return it.
//
DSNode *getLink(DSNode *Node, unsigned Link);
// getSubscriptedNode - Perform the basic getelementptr functionality that
// must be factored out of gep, load and store while they are all MAI's.
//
DSNode *getSubscriptedNode(MemAccessInst &MAI, DSNode *Ptr);
// removeDeadNodes - After the graph has been constructed, this method
// removes all unreachable nodes that are created because they got merged
// with other nodes in the graph.
//
void removeDeadNodes();
};
}
//===----------------------------------------------------------------------===//
// DSGraph constructor - Simply use the GraphBuilder to construct the local
// graph.
DSGraph::DSGraph(Function &F) : Func(F), RetNode(0) {
// Use the graph builder to construct the local version of the graph
GraphBuilder B(*this, Nodes, RetNode, ValueMap, FunctionCalls);
}
//===----------------------------------------------------------------------===//
// Helper method implementations...
//
// createNode - Create a new DSNode, ensuring that it is properly added to the
// graph.
//
DSNode *GraphBuilder::createNode(DSNode::NodeTy NodeType, const Type *Ty) {
DSNode *N = new DSNode(NodeType, Ty);
Nodes.push_back(N);
return N;
}
// getValueNode - Return a DSNode that corresponds the the specified LLVM value.
// This either returns the already existing node, or creates a new one and adds
// it to the graph, if none exists.
//
DSNode *GraphBuilder::getValueNode(Value &V) {
assert(isa<PointerType>(V.getType()) && "Should only use pointer scalars!");
DSNodeHandle &N = ValueMap[&V];
if (N) return N; // Already have a node? Just return it...
// Otherwise we need to create a new scalar node...
N = createNode(DSNode::ScalarNode, V.getType());
if (isa<GlobalValue>(V)) {
// Traverse the global graph, adding nodes for them all, and marking them
// all globals. Be careful to mark functions global as well as the
// potential graph of global variables.
//
DSNode *G = getLink(N, 0);
G->NodeType |= DSNode::GlobalNode;
}
return N;
}
// getLink - This method is used to either return the specified link in the
// specified node if one exists. If a link does not already exist (it's
// null), then we create a new node, link it, then return it.
//
DSNode *GraphBuilder::getLink(DSNode *Node, unsigned Link) {
assert(Link < Node->getNumLinks() && "Link accessed out of range!");
if (Node->getLink(Link) == 0) {
DSNode::NodeTy NT;
const Type *Ty;
switch (Node->getType()->getPrimitiveID()) {
case Type::PointerTyID:
Ty = cast<PointerType>(Node->getType())->getElementType();
NT = DSNode::ShadowNode;
break;
case Type::ArrayTyID:
Ty = cast<ArrayType>(Node->getType())->getElementType();
NT = DSNode::SubElement;
break;
case Type::StructTyID:
Ty = cast<StructType>(Node->getType())->getContainedType(Link);
NT = DSNode::SubElement;
break;
default:
assert(0 && "Unexpected type to dereference!");
abort();
}
DSNode *New = createNode(NT, Ty);
Node->addEdgeTo(Link, New);
}
return Node->getLink(Link);
}
// getSubscriptedNode - Perform the basic getelementptr functionality that must
// be factored out of gep, load and store while they are all MAI's.
//
DSNode *GraphBuilder::getSubscriptedNode(MemAccessInst &MAI, DSNode *Ptr) {
for (unsigned i = MAI.getFirstIndexOperandNumber(), e = MAI.getNumOperands();
i != e; ++i)
if (MAI.getOperand(i)->getType() == Type::UIntTy)
Ptr = getLink(Ptr, 0);
else if (MAI.getOperand(i)->getType() == Type::UByteTy)
Ptr = getLink(Ptr, cast<ConstantUInt>(MAI.getOperand(i))->getValue());
if (MAI.getFirstIndexOperandNumber() == MAI.getNumOperands())
Ptr = getLink(Ptr, 0); // All MAI's have an implicit 0 if nothing else.
return Ptr;
}
// removeDeadNodes - After the graph has been constructed, this method removes
// all unreachable nodes that are created because they got merged with other
// nodes in the graph. These nodes will all be trivially unreachable, so we
// don't have to perform any non-trivial analysis here.
//
void GraphBuilder::removeDeadNodes() {
for (unsigned i = 0; i != Nodes.size(); )
if (!Nodes[i]->getReferrers().empty())
++i; // This node is alive!
else { // This node is dead!
delete Nodes[i]; // Free memory...
Nodes.erase(Nodes.begin()+i); // Remove from node list...
}
}
//===----------------------------------------------------------------------===//
// Specific instruction type handler implementations...
//
// Alloca & Malloc instruction implementation - Simply create a new memory
// object, pointing the scalar to it.
//
void GraphBuilder::handleAlloc(AllocationInst &AI, DSNode::NodeTy NodeType) {
DSNode *Scalar = getValueNode(AI);
DSNode *New = createNode(NodeType, AI.getAllocatedType());
Scalar->addEdgeTo(New); // Make the scalar point to the new node...
}
// PHINode - Make the scalar for the PHI node point to all of the things the
// incoming values point to... which effectively causes them to be merged.
//
void GraphBuilder::visitPHINode(PHINode &PN) {
if (!isa<PointerType>(PN.getType())) return; // Only pointer PHIs
DSNode *Scalar = getValueNode(PN);
for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
Scalar->mergeWith(getValueNode(*PN.getIncomingValue(i)));
}
void GraphBuilder::visitGetElementPtrInst(GetElementPtrInst &GEP) {
DSNode *Scalar = getValueNode(GEP);
DSNode *Ptr = getSubscriptedNode(GEP, getValueNode(*GEP.getOperand(0)));
Scalar->addEdgeTo(Ptr);
}
void GraphBuilder::visitLoadInst(LoadInst &LI) {
if (!isa<PointerType>(LI.getType())) return; // Only pointer PHIs
DSNode *Ptr = getSubscriptedNode(LI, getValueNode(*LI.getOperand(0)));
getValueNode(LI)->mergeWith(Ptr);
}
void GraphBuilder::visitStoreInst(StoreInst &SI) {
if (!isa<PointerType>(SI.getOperand(0)->getType())) return;
DSNode *Value = getValueNode(*SI.getOperand(0));
DSNode *DestPtr = getValueNode(*SI.getOperand(1));
Value->mergeWith(getSubscriptedNode(SI, DestPtr));
}
void GraphBuilder::visitReturnInst(ReturnInst &RI) {
if (RI.getNumOperands() && isa<PointerType>(RI.getOperand(0)->getType())) {
DSNode *Value = getValueNode(*RI.getOperand(0));
Value->mergeWith(RetNode);
RetNode = Value;
}
}
void GraphBuilder::visitCallInst(CallInst &CI) {
FunctionCalls.push_back(vector<DSNodeHandle>());
vector<DSNodeHandle> &Args = FunctionCalls.back();
for (unsigned i = 0, e = CI.getNumOperands(); i != e; ++i)
if (isa<PointerType>(CI.getOperand(i)->getType()))
Args.push_back(getValueNode(*CI.getOperand(i)));
}

View File

@ -0,0 +1,156 @@
//===- Printer.cpp - Code for printing data structure graphs nicely -------===//
//
// This file implements the 'dot' graph printer.
//
//===----------------------------------------------------------------------===//
#include "llvm/Analysis/DataStructure.h"
#include "llvm/Module.h"
#include "llvm/Assembly/Writer.h"
#include <fstream>
#include <sstream>
void DSNode::dump() const { print(std::cerr, 0); }
std::string DSNode::getCaption(Function *F) const {
std::stringstream OS;
WriteTypeSymbolic(OS, getType(), F ? F->getParent() : 0);
OS << " ";
if (NodeType & ScalarNode) OS << "S";
if (NodeType & AllocaNode) OS << "A";
if (NodeType & NewNode ) OS << "N";
if (NodeType & GlobalNode) OS << "G";
if (NodeType & SubElement) OS << "E";
if (NodeType & CastNode ) OS << "C";
return OS.str();
}
static std::string getValueName(Value *V, Function &F) {
std::stringstream OS;
WriteAsOperand(OS, V, true, true, F.getParent());
return OS.str();
}
static void replaceIn(std::string &S, char From, const std::string &To) {
for (unsigned i = 0; i < S.size(); )
if (S[i] == From) {
S.replace(S.begin()+i, S.begin()+i+1,
To.begin(), To.end());
i += To.size();
} else {
++i;
}
}
static string escapeLabel(const string &In) {
string Label(In);
replaceIn(Label, '\\', "\\\\"); // Escape caption...
replaceIn(Label, ' ', "\\ ");
replaceIn(Label, '{', "\\{");
replaceIn(Label, '}', "\\}");
return Label;
}
static void writeEdge(std::ostream &O, const void *SrcNode,
const char *SrcNodePortName, int SrcNodeIdx,
const DSNode *VS, const string &EdgeAttr = "") {
O << "\tNode" << SrcNode << SrcNodePortName;
if (SrcNodeIdx != -1) O << SrcNodeIdx;
O << " -> Node" << (void*)VS;
if (!EdgeAttr.empty())
O << "[" << EdgeAttr << "]";
O << ";\n";
}
void DSNode::print(std::ostream &O, Function *F) const {
string Caption = escapeLabel(getCaption(F));
O << "\tNode" << (void*)this << " [ label =\"{" << Caption;
if (!Links.empty()) {
O << "|{";
for (unsigned i = 0; i < Links.size(); ++i) {
if (i) O << "|";
O << "<g" << i << ">";
}
O << "}";
}
O << "}\"];\n";
for (unsigned i = 0; i < Links.size(); ++i)
if (Links[i])
writeEdge(O, this, ":g", i, Links[i]);
}
void DSGraph::print(std::ostream &O) const {
O << "digraph DataStructures {\n"
<< "\tnode [shape=Mrecord];\n"
<< "\tedge [arrowtail=\"dot\"];\n"
<< "\tsize=\"10,7.5\";\n"
<< "\trotate=\"90\";\n"
<< "\tlabel=\"Function\\ " << Func.getName() << "\";\n\n";
// Output all of the nodes...
for (unsigned i = 0, e = Nodes.size(); i != e; ++i)
Nodes[i]->print(O, &Func);
O << "\n";
// Output all of the nodes edges for scalar labels
for (std::map<Value*, DSNodeHandle>::const_iterator I = ValueMap.begin(),
E = ValueMap.end(); I != E; ++I) {
O << "\tNode" << (void*)I->first << "[ shape=circle, label =\""
<< escapeLabel(getValueName(I->first, Func)) << "\",style=dotted];\n";
writeEdge(O, I->first, "",-1, I->second.get(),"arrowtail=tee,style=dotted");
}
// Output the returned value pointer...
if (RetNode != 0) {
O << "\tNode0x1" << "[ shape=circle, label =\""
<< escapeLabel("Return") << "\"];\n";
writeEdge(O, (void*)1, "", -1, RetNode, "arrowtail=tee,style=dotted");
}
// Output all of the call nodes...
for (unsigned i = 0, e = FunctionCalls.size(); i != e; ++i) {
const std::vector<DSNodeHandle> &Call = FunctionCalls[i];
O << "\tNode" << (void*)&Call << " [shape=record,label=\"{call|{";
for (unsigned j = 0, e = Call.size(); j != e; ++j) {
if (j) O << "|";
O << "<g" << j << ">";
}
O << "}}\"];\n";
for (unsigned j = 0, e = Call.size(); j != e; ++j)
if (Call[j])
writeEdge(O, &Call, ":g", j, Call[j]);
}
O << "}\n";
}
// print - Print out the analysis results...
void LocalDataStructures::print(std::ostream &O, Module *M) const {
for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I)
if (!I->isExternal()) {
std::string Filename = "ds." + I->getName() + ".dot";
O << "Writing '" << Filename << "'...";
std::ofstream F(Filename.c_str());
if (F.good()) {
DSGraph &Graph = getDSGraph(*I);
Graph.print(F);
O << " [" << Graph.getGraphSize() << "]\n";
} else {
O << " error opening file for writing!\n";
}
}
}