Re: [tree-ssa]: Points-To analysis update

[Daniel Berlin <dberlin at dberlin dot org>]

On Sun, 13 Oct 2002, Diego Novillo wrote:

> On Sun, 13 Oct 2002, Daniel Berlin wrote:
>
> > I'll submit a patch for approval in a minute that uses it in
> > compute_may_aliases.
> > I'll also explain the HAVE_BANSHEE stuff (andersen's analysis is built
> > on
> > a constraint engine generated by a program named banshee).
> Cool.  But before enabling this on the current analyses, I'd like
> to wait until we have at least a couple of passes that bootstrap
> fine and pass basic tests like SPEC.  Otherwise, we will not know
> whether we're fighting DFA or optimization bugs.
Of course.
> Anyway, we can always use the existing flag to predicate
> points-to analysis in the main alias function.
This is what i've done.
It doesn't use it at all unless you specify -ftree-points-to.
Even then, it only uses it after trying the other methods currently in
may_alias_p.

>  Any idea, how
> much reduction of false positive does this buy us?
I'm working on some numbers.

> We should
> balance that against the cost of running this analysis.
I've been doing timing tests, and the banshee generated analysis is
actually as fast or faster than steensgaard right now (though i'm sure 
in
whole program it wouldn't be), which is almost-linear (inverse ackerman)
in the size of the number of statements.

It's also currently the case that we can generate the sets faster
than compute_may_alias queries them (probably because it's an O(N^2)
loop), by a large margin.

I'm now trying out some huge testcases (100000 lines in a single 
function)
to see about scaling.
But it's taking forever to make the CFG on this testcase (It's one of
Brad's goto laden functions. 4 minutes so far and counting. Ahh. Done. 4
minutes, 20 seconds).

To give you some idea how large this function is, i blew out 8 meg of
stack using -fdump-tree-simple, and had to up it to 16 meg.
:)

> Diego.