Re: [PATCH] Factor out division by squares and remove division around comparisons (1/2)

[Jackson Woodruff <jackson dot woodruff at foss dot arm dot com>]

Hi all,

Apologies again to those CC'ed, who (again) received this twice.

Joseph: Yes you are correct. I misread the original thread, now fixed.

Richard: I've moved the optimizations out of fold-const.c. One has been 
replicated in match.pd, and the other (x / C +- y / C -> (x +- y) / C) 
I've deleted as it only introduced a new optimization when running with 
the flags '-O0 -funsafe-math-optimizations'.

On O1 and up, the pattern that replaces 'x / C' with 'x * (1 / C)'
is enabled and then the pattern is covered by that and
(x * C +- y * C -> C * (x +- y)) (which is already present in match.pd)

I have also updated the testcase for those optimizations to use 'O1' to 
avoid that case.


On 08/24/2017 10:06 PM, Jeff Law wrote:
> On 08/17/2017 03:55 AM, Wilco Dijkstra wrote:
> > Richard Biener wrote:
> > > On Tue, Aug 15, 2017 at 4:11 PM, Wilco Dijkstra <Wilco.Dijkstra@arm.com> wrote:
> > > > Richard Biener wrote:
> > > > > > We also change the association of
> > > > > >
> > > > > >        x / (y * C) -> (x / C) / y
> > > > > >
> > > > > > If C is a constant.
> > > > >
> > > > > Why's that profitable?
> > > >
> > > > It enables (x * C1) / (y * C2) -> (x * C1/C2) / y for example.
> > > > Also 1/y is now available to the reciprocal optimization, see
> > > > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71026 for details.
> > >
> > > Sure, but on its own it's going to be slower.  So this isn't the
> > > correct way to enable those followup transforms.
> >
> > How can it be any slower? It's one division and one multiply in both cases.
> x / (y * C) -> (x / C) / y
>
> Goes from one division and one multiplication to two divisions.  I'm
> guessing that's what Richi is (reasonably) concerned about.
>
> So it may be the case that we need more complex pattern matching here
> for when to perform the first transformation to ultimately enable the
> second.

The only case where we don't remove the division but still execute this 
pattern is when run with'-O0 -freciprocal-math'.

As long as we have 'O1' or greater (and -freciprocal-math), that is 
enough to enable the removal of the reciprocal.


Jackson.

> Jeff

Attachment: patchfile-1
Description: Text document