Re: [PATCH] Teach VRP to derive ranges from SWITCH_EXPR (fix PR21258)

["Richard Guenther" <richard dot guenther at gmail dot com>]

On 4/13/07, Jeffrey Law <law@redhat.com> wrote:
> On Fri, 2007-04-13 at 07:55 -0400, Diego Novillo wrote:
> > Richard Guenther wrote on 04/13/07 05:20:
> >
> > > No noticable slowdowns, I have applied the patch to mainline.  A patch
> > > for 14495 is in work.
> >
> > I forgot to ask, what was the net effect of the optimization in terms of
> > code size or performance of generated code?
> FWIW, I wouldn't have expected much of a code speed improvement, it's
> rare that a SWITCH_EXPR is on the critical path *and* there's
> opportunity for improvement via VRP derived ranges related to the
> switch statement.

Checking enabled GCC is an example where I would expect runtime
savings because the above cenario happens often (of course not in
critical loops, and the remaining branches are easy to predict).  Of course
VRP is currently limited in other ways, so the improvement is not as
big as possible (I'm working on that).

> It might be interesting to see if we can zap the code in
> tree-ssa-dom.c::record_edge_info -- in theory it should be
> subsumed by VRP.

OTOH DOM runs far more often than VRP so I would expect to see
fallouts as we'd have a different pass ordering.  Certainly worth investigating
though.

Richard.