https://gcc.gnu.org/bugzilla/show_bug.cgi?id=56766
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |NEW
Last reconfirmed| |2015-05-12
CC| |uros at gcc dot gnu.org
Ever confirmed|0 |1
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
The vectorizer patch is now in (I added a testcase for addsubps because for
that
it happens to work).
We still need to fixup the sse3_addsubv2df3 pattern or fix combine to try
multiple "canonical" forms of vec_merge vs. (vec_select (vec_concat ...)).
Or decide which one is canonical (I'd prefer simply dropping vec_merge - with
AVX512 we've run out of bits for the CONST_INT selector....)
One issue with the vec_select form is that it requires an intermediate mode of
double-size while vec_merge avoids this.
Eventually one can also teach combine that certain (vec_select (vec_concat ..))
can be treated as (vec_merge ...).
void testd (double * __restrict__ p, double * __restrict q)
{
p[0] = p[0] - q[0];
p[1] = p[1] + q[1];
}
should vectorize to addsubpd with -msse3.
