Re: PING fwprop and pr/19653

[Bernd Schmidt <bernds_cb1 at t-online dot de>]

Steven Bosscher wrote:

> > Benchmark results of baseline, 
> > (baseline + fwprop), and (baseline + fwprop - expensive-parts-of-cse),
> > across as many targets as possible, would be welcome.
>
> Why baseline+fwprop _with_ the expensive parts of cse?  The whole point
> is to get rid of those parts of CSE.  It would be a Really Bad Idea to
> put fwprop in the compiler as a new pass but not disable the expensive
> bits of CSE.

We need those tests to determine whether expensive-cse provides any 
benefit on top of fwprop.  If it does, then we'll have to think twice 
about removing it.  The alternative would be to improve its speed; Vlad 
had patches for that a few years ago.  But without numbers, we can't 
tell what should be done.

> Benchmarks for speed are tough to get for many targets, but we have at
> least x86 and x86_64 (done by me), POWER/PowerPC (by David Edelsohn and
> Paolo).  There really isn't any significant degradation, and there is a
> big win for compile time.

Numbers please, from the current set of patches, not an older version.

> For other targets, and for code size, hmm....  Paolo, ideas?  I can run
> regression tests on a few targets with simulators, if you like.   Bernd,
> have you tried fwprop with a code size benchmark (CSiBE, etc.)?

I haven't done any systematic testing yet, just some random experiments.


Bernd