Re: [patch] Loop distribution for single nested loops

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

On Dec 5, 2007 9:27 PM, Sebastian Pop <sebpop@gmail.com> wrote:
> On Oct 3, 2007 6:49 AM, Diego Novillo <dnovillo@google.com> wrote:
> > I like this pass.  Would it make sense to enable this pass by
> > default?  Or enable it only with the vectorizer/parallelizer?  I
> > think this is something we could at least enable at -O3.
> >
>
> I've also implemented a cost function for the loop distribution: it
> aggregates the components containing read or write accesses to
> similar memory, that otherwise would end in different loops.  I don't
> know yet the performance impact of the loop distribution on real code,
> but I'll give a run on spec2006, and then send out the numbers.
> During the build of gcc there are 44 loops distributed.  The usual
> pattern is the distribution of different arrays initialization.
>
> For counting the number of distributed loops, I've left in the attached
> patch a if (0 && ...) that allows you to print to stdout the distributed
> cases if you remove the 0. I will not include that part in a commit.
>
> > Before comitting the pass, I think it would be very useful to
> > enable it by default at -O2 and do a full bootstrap/test cycle,
> > that will help you shake the most obvious bugs.
> >
>
> I've bootstrap/test the attached patch with the loop distribution
> enabled by default at -O2 on {i686,amd64}-linux.  There are a huge
> number of tests that fail in the vectorization part, as the loop
> distribution modifies the patterns that the vectorizer expects to see.
> Some nice examples include the vectorization of unaligned stores,
> something that we expected to see ;-).  See below for the complete
> list of autovect fails.
>
> The attached patch passed bootstrap and test on both i686-linux and
> amd64-linux.  Okay for trunk?

I don't think this is appropriate for mainline at this point.  Absolutely not
if enabled by default.

Richard.