http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50374
--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-09-19 15:59:06 UTC --- Created attachment 25320 --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=25320 gcc47-pr50374-wip.patch Here is the ported patch with rejects hopefully resolved and a few bugfixes for the 4.6 -> 4.7 changes. I've omitted the vcondc/vcondcu stuff because the trunk handles mixed float/integral vcond/u already, the rs6000 stuff (because IMHO the first step for rs6000 should be to add support for the mixed float/integral vcond/u for rs6000 and I'm not familiar with rs6000 enough), GTF/EQF etc. dropped (I think that it is not needed) and testcases separated out of it for now. I had to make changes beyond rejects as COND_EXPR assignments are now represented differently and the pattern stmts are no longer emitted into the basic blocks. I see no reason why this should be limited to floating point comparisons and integral indexes, IMHO e.g. finding location of smallest integer is common too. That said, I'm now stuck with it, on the fast-math-no-pre-minmax-loc-1.c (on x86_64 -O2 -ftree-vectorize -fno-tree-pre) testcase the compound pattern recognition seems to work, but still the vectorizer gives up: 17: Unsupported pattern. 17: not vectorized: unsupported use in stmt. 17: unexpected pattern. Ira, do you think you could have a quick look at this?
