Hi Simone,
setTopValue() is probably misnamed; it is used by DataflowSolver to determine the "top" value when computing a merge at a confluence point. It's not actually used anywhere else.
If you want to specially initialize your bitvector values, the DataflowValues class does have an InitializeValues() method that is called by solver at the beginning of the analysis:
/// InitializeValues - Invoked by the solver to initialize state needed for
/// dataflow analysis. This method is usually specialized by subclasses.
void InitializeValues(const CFG& cfg) {}
The LiveVariables analysis doesn't actually use this method since the default bitvector values are fine for initial value, but in InitializeValues you can iterate over the CFG and set the values as you like.
Would this work?
Hi again,
I just found out that using InitializeValue() to initialize the state of the dataflow solver is not enough.
Let's for example consider an easy dataflow analysis: dominator trees. I need to create a bitvector where each position refers to a block of the CFG. Before starting the solver I need to initialize the values in the following way:
n0 (root node) = {0, 0, 0, n0, 0, ...}
N-n0 (remaining nodes) = {1,1,1,1..,1,1}
and this is what I am doing in the InitializeValue function:
/// Initializes the CFGBlock values for Dominators analysis i.e.:
/// * Dom(root) = {root}
/// * for each n in N - {root}, Dom(n) = N
void Dominators::InitializeValues(const CFG& cfg) {
for(CFG::const_iterator I = cfg.begin(), E = cfg.end(); I != E; ++I) {
Dominators::ValTy& D = getBlockDataMap()[*I];
if(cfg.getEntry().getBlockID() == (*I)->getBlockID()) {
D.resetValues(getAnalysisData());
D((*I)->getBlockID(), getAnalysisData()) = IsDominator;
} else
D.setValues(getAnalysisData());
}
}
My transfer function looks like this:
class TransferFuncs : public CFGRecStmtVisitor<TransferFuncs> {
Dominators::AnalysisDataTy& AD;
Dominators::ValTy Dominance;
public:
TransferFuncs(Dominators::AnalysisDataTy& ad) : AD(ad) { Dominance.resetValues(AD); }
Dominators::ValTy& getVal() { return Dominance; }
CFG& getCFG() { return AD.getCFG(); }
void VisitTerminator(CFGBlock* B) {
Dominance(B->getBlockID(), AD) = IsDominator; // add this as dominator
}
void SetTopValue(Dominators::ValTy& V) { }
};
nothing difficult, but when I run the solver it gives me wrong results. 
I investigated the problem and it seems that the initial values I set in the initialize function are overwritten the first time the ProcessMerge() function is called by the DataflowSolver.
For what I understood from the code, the first time ProcessMerge() is called there is no edge data, therefore the last instruction in the function will always overwrite the existing block data with the one returned by the function TF.SetTopValue(), and in my case I don't have any TopValue to set.
To solve the problem I introduced a flag which track whether edge data exists. If not, the D.getBlockDataMap()[B] will not be overwritten.
What do you think?
cheers, Simone
