https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103948
--- Comment #6 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to Richard Biener from comment #5)
> I guess that tree-vect-generic.c is not up-to-date with gimple-isel.cc. We
> should probably somehow factor out relevant pieces.
>
> Note vector lowering will see
>
> _1 = a_2(D) > b_3(D);
> _4 = VEC_COND_EXPR <_1, { -1, -1 }, { 0, 0 }>;
>
> and clearly a
>
> if (expand_vec_cond_expr_p (type, TREE_TYPE (a1), code))
>
> that's just type/code based cannot know expanding just a comparison will
> work fine. It does look like expand_vector_condition from vector lowering
> needs
> some work. Maybe the following works though:
>
> diff --git a/gcc/tree-vect-generic.c b/gcc/tree-vect-generic.c
> index 6afb6999cd7..5814a71a5bb 100644
> --- a/gcc/tree-vect-generic.c
> +++ b/gcc/tree-vect-generic.c
> @@ -1052,7 +1052,9 @@ expand_vector_condition (gimple_stmt_iterator *gsi,
> bitmap dce_ssa_names)
> }
> }
>
> - if (expand_vec_cond_expr_p (type, TREE_TYPE (a1), code))
> + if (expand_vec_cond_expr_p (type, TREE_TYPE (a1), code)
> + || (integer_all_onesp (b) && integer_zerop (c)
> + && expand_vec_cmp_expr_p (type, TREE_TYPE (a1), code)))
> {
> gcc_assert (TREE_CODE (a) == SSA_NAME || TREE_CODE (a) == VECTOR_CST);
> return true;
>
> obviously missing would be the case
> cmp ? { 0, 0 } : { -1, -1 } via code inversion, not sure if we canonicalize
> that way already and not sure if expansion later will pick that up correctly.
>
> A vector mode with just vec_cmp and not vcond will be good to have to cover
> all this with testcases ... ;)
>
> If the above works it's pre-approved.
Using the patch from Comment #2, the following testcase from Comment #0 can be
used:
--cut here--
typedef signed char vec __attribute__((vector_size(2)));
vec lt (vec a, vec b) { return a < b; }
--cut here--
vectorizes with -msse4 and fails to vectorize with -msse2.
I'll try your proposed patch from Comment #5 later today and report here.
