Polyhedron benchmark results on AMD64

[Diego Novillo]

Tobias Burnus wrote on 02/09/07 12:05:

> Does not look as if there was a big regression regarding aermod and
> mem-ssa, it moved too often slightly up and down to really pinpoint a
> mem-ssa effect. Protein seems to have worsend over the lasts three
> months, but also not in a big jump. And for ac we already did better a
> few weeks ago.
> 
OK, thanks.  With the two or three cases in the PRs, I think we have 
enough to work on for now.  From what I've seen, one of the main 
problems with Fortran and aliasing is the pass-by-reference rule in 
Fortran.  This often tricks alias analysis into overstating the amount 
of call-clobbered variables.

In the past, this was causing a lot of virtual operands in the IL, but 
with partitioning the problem is the loss of aliasing information 
precision because of how the addressable symbols get grouped.

Better partitioning heuristics will help.  Also, if we could somehow 
reduce the sheer number of addressables that would probably gives us a 
lot of benefit.  In the long term, I am trying to not rely so much into 
the static partitioning, but that requires solving other problems with 
the representation.