On March 22, 2016 4:55:13 PM GMT+01:00, Marc Glisse <marc.gli...@inria.fr> wrote: >On Tue, 22 Mar 2016, Richard Biener wrote: > >> >> This adjusts the PR70251 fix as discussed in the PR audit trail >> and fixes a bug in genmatch required (bah, stupid GENERIC comparisons >in >> GIMPLE operands...). > >Thanks ! > >Hmm, the transformation is still disabled on AVX512: > >> ! /* A + (B vcmp C ? 1 : 0) -> A - (B vcmp C ? -1 : 0), since vector >comparisons >> ! return all -1 or all 0 results. */ >> /* ??? We could instead convert all instances of the vec_cond to >negate, >> but that isn't necessarily a win on its own. */ >> (simplify >> ! (plus:c @3 (view_convert? (vec_cond:s @0 integer_each_onep@1 >integer_zerop@2))) >> (if (VECTOR_TYPE_P (type) >> && TYPE_VECTOR_SUBPARTS (type) == TYPE_VECTOR_SUBPARTS >(TREE_TYPE (@0)) >> && (TYPE_MODE (TREE_TYPE (type)) >> == TYPE_MODE (TREE_TYPE (TREE_TYPE (@0))))) >> ! (minus @3 (view_convert (vec_cond @0 (negate @1) @2))))) > >It seems that the references to @0 in the "if" should use @1 instead >(at >least the last one). I assume this test is to make sure that A has as >many >integer elements of the same size as the result of the vec_cond_expr.
It looks like that is always guaranteed by the input form and instead these are now useless checks which were guarding the original view-convert transform. I'll remove them in a followup. Richard. >Sorry for giving you an incomplete change in the PR.
