[Bug fortran/43829] Scalarization of reductions

[matz at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43829

Michael Matz <matz at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #23268|0                           |1
        is obsolete|                            |

--- Comment #35 from Michael Matz <matz at gcc dot gnu.org> 2011-10-04 13:17:00 UTC ---
Created attachment 25409
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=25409
Even more up-to-date patch

This patch applies to r179469.  The results on tonto, on a Bulldozer machine:

imf ./tonto.r179469.bdver.avx   3m22.045s
imf ./tonto.r179469.bdver.nosum 2m59.535s

That's with libimf preloaded.  And then it's actually faster then with ICC
(doesn't mean much, as on bulldozer we can use FMA4 and XOP).  Nevertheless
the speedup is now rougly 10%.  We really should work towards including
it for 4.7.