Re: [tree-ssa] Returning to alias analysis

[Daniel Berlin <dberlin at dberlin dot org>]

> In this test we spend very little time in alias analysis before or 
> after
> my changes (.28 and .24 seconds respectively).  However, after my 
> changes
> the time for PHI node insertion explodes (.01 seconds vs 100 seconds).
> Why?

> This is due to my changes allowing the alias analysis code to 
> disambiguate
> all the memory references in that test (something like 32k of them).  
> This
> in turn means that we have thousands of objects for which we'll need to
> create PHI nodes (instead of just 3). Most of the PHI node insertion 
> time
> is spent computing global live information those 32k objects.

I think I mentioned this problem as to why i haven't pushed for PTA to 
be the default.
It's too good at disambiguating, so PHI insertions blows up.

> But note that the test still completes without blowing the machine up 
> --
> meaning that the semi-pruned vs fully-pruned PHI node insertion 
> heuristics
> are working properly.
>
> I've got one additional tweak to investigate which could improve things
> further..
>
> Jeff