Re: heads and tails

[Florian Loitsch <florian@loitsch.net>]

On Tue, Apr 30, 2002 at 12:03:29AM -0400, Ondrej Lhotak wrote:
> We should define heads and tails differently. They should be all
> possible entry/exit points, respectively. For heads, we would have the
> first instruction (even if there is a goto to it), and the start of
> every exception handler (because they are like entry points, in that
> the flow into them is not explicit in the graph) (this is only in the
> brief unit graph). For tails, we would have all return statements, and,
> presumably, all statements that can throw an exception. I suspect this
> last one will cause problems...
another thought:
if we add the first unit of the graph into the header-list, we have
solved the problem for forward-analyses.
so now for backward-analyses:
i could imagine the following:
for the very first entry in an infinite loop take the tail-set (the
safe set) as "input"-set for the unit in the
flow-through-function. This not only eliminates the problem, but even
gives the best possible result.
We should probably rewrite BackwardAnalysis.java so we have something
like newInitialTailSet(), but beside of that there should not be much
work to do.
  florian
> 
> Ondrej
> 
> On Mon, Apr 29, 2002 at 11:35:42PM -0400, Florian Loitsch wrote:
> > hi,
> > When using the provided Analysis-classes, one just overwrites
> > the newInitialFlow() and customizeInitialFlowGraph() to assing the
> > desired flowsets to each unit. customizeInitialFlowGraph() is used, if
> > the heads/tails should have different sets.
> > Take as example a "defined-Locals"-analysis (similary to what Javac
> > does).
> > we assign the full set to each unit except the heads, where we put the
> > empty set.
> > if we did this using the current implementation of UnitGraph.java we
> > would, imo, run into troubles: heads, according to this class, are
> > units, that have no predecessor. but we could have a "goto" back to
> > this first unit. in this case we may not even have heads... Therefore
> > we wouldn't find any non-initialized variable.
> > 
> > while i haven't yet encountered this problem with heads, i've had it
> > with tails: if we have an infite loop, we will not have a tail
> > neither.
> > The problem is, that even fixing the UnitGraph.java so the first
> > element is always a head, will not fix the tails. and what is the tail
> > of an infite loop?
> > any suggestions?
> > mfg
> >  florian loitsch
> > -- 
> > People who think they know everything are very irritating to those of us who do.
> > - Anon.
> > 

-- 
People who think they know everything are very irritating to those of us who do.
- Anon.