Re: [RFC] third liveness pass

[Jan Hubicka <jh at suse dot cz>]

>  > Would be OK to make patch that moves PROP_SCAN_DEAD_STORES out of
>  > PROP_FINAL and makes the PROP_FINAL pass to include it conditionally?  I
>  > still guess we can include dead code removal as it pays back by avoiding
>  > reload/global to work too hard.
> Seems reasonable.
> 
> Another idea for doing DSE in a much cleaner, potentially less expensive
> way would be to look at dependency information.
> 
> Look at the dependencies generated when scheduling if we find that a store Y
> is a direct descendant of store X in the dependency tree and X and Y are the
> same memory location, then Y _may_ be a dead store.
> 
> The problem boils down to determining if there is some valid schedule which
> could result in X firing resolving the final dependency for Y.  If there is
> such a schedule, then Y is a dead store.
> 
> So this means that we'd need to prove that none of other descendants of X
> reach Y in the dependency tree -- and do so in a reasonable amount of
> computation.
> 
> So conceptually you could hook in a memory optimizer to run immediately
> after we build the dependency graph for scheduling. 
> 
> Alternately you can hook right into the scheduler code and detect some of these
> as we actually resolve dependencies -- it's fast, but less clean and may
> miss some redundant memory operations.
> 
> These schemes may or may not be less expensive than the DSE code currently
> in flow.c, but it might be worth an investigation.

Yes, I was playing with this idea as well.  It would be nice to break
out the dependency graph generation out of scheduler, as for instance
reg-stack can use it as well, but unforutnately I am not sure I can do
this at the moment, but I will keep it in the mind.

Honza
> 
> jeff
>