1
0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-10-30 15:32:52 +01:00
llvm-mirror/lib/Transforms/Instrumentation/ProfilePaths/ProfilePaths.cpp
2004-07-29 17:30:56 +00:00

250 lines
8.0 KiB
C++

//===-- ProfilePaths.cpp - interface to insert instrumentation --*- C++ -*-===//
//
// 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 inserts instrumentation for counting execution of paths though a given
// function Its implemented as a "Function" Pass, and called using opt
//
// This pass is implemented by using algorithms similar to
// 1."Efficient Path Profiling": Ball, T. and Larus, J. R.,
// Proceedings of Micro-29, Dec 1996, Paris, France.
// 2."Efficiently Counting Program events with support for on-line
// "queries": Ball T., ACM Transactions on Programming Languages
// and systems, Sep 1994.
//
// The algorithms work on a Graph constructed over the nodes made from Basic
// Blocks: The transformations then take place on the constructed graph
// (implementation in Graph.cpp and GraphAuxiliary.cpp) and finally, appropriate
// instrumentation is placed over suitable edges. (code inserted through
// EdgeCode.cpp).
//
// The algorithm inserts code such that every acyclic path in the CFG of a
// function is identified through a unique number. the code insertion is optimal
// in the sense that its inserted over a minimal set of edges. Also, the
// algorithm makes sure than initialization, path increment and counter update
// can be collapsed into minimum number of edges.
//
//===----------------------------------------------------------------------===//
#include "llvm/Transforms/Utils/UnifyFunctionExitNodes.h"
#include "llvm/Support/CFG.h"
#include "llvm/Constants.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Instructions.h"
#include "llvm/Module.h"
#include "Graph.h"
#include <fstream>
#include <cstdio>
namespace llvm {
struct ProfilePaths : public FunctionPass {
bool runOnFunction(Function &F);
// Before this pass, make sure that there is only one
// entry and only one exit node for the function in the CFG of the function
//
void ProfilePaths::getAnalysisUsage(AnalysisUsage &AU) const {
AU.addRequired<UnifyFunctionExitNodes>();
}
};
static RegisterOpt<ProfilePaths> X("paths", "Profile Paths");
static Node *findBB(std::vector<Node *> &st, BasicBlock *BB){
for(std::vector<Node *>::iterator si=st.begin(); si!=st.end(); ++si){
if(((*si)->getElement())==BB){
return *si;
}
}
return NULL;
}
//Per function pass for inserting counters and trigger code
bool ProfilePaths::runOnFunction(Function &F){
static int mn = -1;
static int CountCounter = 1;
if(F.isExternal()) {
return false;
}
//increment counter for instrumented functions. mn is now function#
mn++;
// Transform the cfg s.t. we have just one exit node
BasicBlock *ExitNode =
getAnalysis<UnifyFunctionExitNodes>().getReturnBlock();
//iterating over BBs and making graph
std::vector<Node *> nodes;
std::vector<Edge> edges;
Node *tmp;
Node *exitNode = 0, *startNode = 0;
// The nodes must be uniquely identified:
// That is, no two nodes must hav same BB*
for (Function::iterator BB = F.begin(), BE = F.end(); BB != BE; ++BB) {
Node *nd=new Node(BB);
nodes.push_back(nd);
if(&*BB == ExitNode)
exitNode=nd;
if(BB==F.begin())
startNode=nd;
}
// now do it again to insert edges
for (Function::iterator BB = F.begin(), BE = F.end(); BB != BE; ++BB){
Node *nd=findBB(nodes, BB);
assert(nd && "No node for this edge!");
for(succ_iterator s=succ_begin(BB), se=succ_end(BB); s!=se; ++s){
Node *nd2=findBB(nodes,*s);
assert(nd2 && "No node for this edge!");
Edge ed(nd,nd2,0);
edges.push_back(ed);
}
}
Graph g(nodes,edges, startNode, exitNode);
#ifdef DEBUG_PATH_PROFILES
std::cerr<<"Original graph\n";
printGraph(g);
#endif
BasicBlock *fr = &F.front();
// The graph is made acyclic: this is done
// by removing back edges for now, and adding them later on
std::vector<Edge> be;
std::map<Node *, int> nodePriority; //it ranks nodes in depth first order traversal
g.getBackEdges(be, nodePriority);
#ifdef DEBUG_PATH_PROFILES
std::cerr<<"BackEdges-------------\n";
for (std::vector<Edge>::iterator VI=be.begin(); VI!=be.end(); ++VI){
printEdge(*VI);
cerr<<"\n";
}
std::cerr<<"------\n";
#endif
#ifdef DEBUG_PATH_PROFILES
cerr<<"Backedges:"<<be.size()<<endl;
#endif
//Now we need to reflect the effect of back edges
//This is done by adding dummy edges
//If a->b is a back edge
//Then we add 2 back edges for it:
//1. from root->b (in vector stDummy)
//and 2. from a->exit (in vector exDummy)
std::vector<Edge> stDummy;
std::vector<Edge> exDummy;
addDummyEdges(stDummy, exDummy, g, be);
#ifdef DEBUG_PATH_PROFILES
std::cerr<<"After adding dummy edges\n";
printGraph(g);
#endif
// Now, every edge in the graph is assigned a weight
// This weight later adds on to assign path
// numbers to different paths in the graph
// All paths for now are acyclic,
// since no back edges in the graph now
// numPaths is the number of acyclic paths in the graph
int numPaths=valueAssignmentToEdges(g, nodePriority, be);
//if(numPaths<=1) return false;
static GlobalVariable *threshold = NULL;
static bool insertedThreshold = false;
if(!insertedThreshold){
threshold = new GlobalVariable(Type::IntTy, false,
GlobalValue::ExternalLinkage, 0,
"reopt_threshold");
F.getParent()->getGlobalList().push_back(threshold);
insertedThreshold = true;
}
assert(threshold && "GlobalVariable threshold not defined!");
if(fr->getParent()->getName() == "main"){
//initialize threshold
// FIXME: THIS IS HORRIBLY BROKEN. FUNCTION PASSES CANNOT DO THIS, EXCEPT
// IN THEIR INITIALIZE METHOD!!
Function *initialize =
F.getParent()->getOrInsertFunction("reoptimizerInitialize", Type::VoidTy,
PointerType::get(Type::IntTy), 0);
std::vector<Value *> trargs;
trargs.push_back(threshold);
new CallInst(initialize, trargs, "", fr->begin());
}
if(numPaths<=1 || numPaths >5000) return false;
#ifdef DEBUG_PATH_PROFILES
printGraph(g);
#endif
//create instruction allocation r and count
//r is the variable that'll act like an accumulator
//all along the path, we just add edge values to r
//and at the end, r reflects the path number
//count is an array: count[x] would store
//the number of executions of path numbered x
Instruction *rVar=new
AllocaInst(Type::IntTy,
ConstantUInt::get(Type::UIntTy,1),"R");
//Instruction *countVar=new
//AllocaInst(Type::IntTy,
// ConstantUInt::get(Type::UIntTy, numPaths), "Count");
//initialize counter array!
std::vector<Constant*> arrayInitialize;
for(int xi=0; xi<numPaths; xi++)
arrayInitialize.push_back(ConstantSInt::get(Type::IntTy, 0));
const ArrayType *ATy = ArrayType::get(Type::IntTy, numPaths);
Constant *initializer = ConstantArray::get(ATy, arrayInitialize);
char tempChar[20];
sprintf(tempChar, "Count%d", CountCounter);
CountCounter++;
std::string countStr = tempChar;
GlobalVariable *countVar = new GlobalVariable(ATy, false,
GlobalValue::InternalLinkage,
initializer, countStr,
F.getParent());
// insert initialization code in first (entry) BB
// this includes initializing r and count
insertInTopBB(&F.getEntryBlock(), numPaths, rVar, threshold);
//now process the graph: get path numbers,
//get increments along different paths,
//and assign "increments" and "updates" (to r and count)
//"optimally". Finally, insert llvm code along various edges
processGraph(g, rVar, countVar, be, stDummy, exDummy, numPaths, mn,
threshold);
return true; // Always modifies function
}
} // End llvm namespace