[PATCH] Don't change DR_STMT if vect_pattern_recog_1 would fail (PR tree-optimization/51356)

[Ira Rosen]

On 1 December 2011 18:41, Jakub Jelinek <jakub@redhat.com> wrote:
> Hi!

Hi,

>
> As mentioned in the PR, vect_pattern_recog_1 attempts to find out
> if the computed type_in and type_out are already vector types or not,
> and uses VECTOR_MODE_P (TYPE_MODE (type_in)) as the test.  Unfortunately,
> get_vectype_for_scalar_type on some targets (e.g. PowerPC) returns a
> VECTOR_TYPE with TImode for a DImode integer/boolean scalar type.
> If that happens, vect_recog_bool_pattern assumes it will succeed and changes
> DR_STMT, but vect_mark_pattern_stmts isn't called and we ICE later on.
> Not sure what actually can be vectorized using scalar mode vectors,
> so either we adjust vect_recog_bool_pattern like this, or perhaps
> vect_pattern_recog_1 could use a different test (TREE_CODE (type_in) ==
> VECTOR_TYPE)?

But AFAIU in the later case we would fail to vectorize anyway, so I am
OK with your patch.

Thanks,
Ira

>
> This has been bootstrapped/regtested on x86_64-linux and i686-linux
> and fixes the testcase on PowerPC.
>
> 2011-12-01  Jakub Jelinek  <jakub@redhat.com>
>
>        PR tree-optimization/51356
>        * tree-vect-patterns.c (vect_recog_bool_pattern): Give up if
>        vectype doesn't have VECTOR_MODE_P.
>
> --- gcc/tree-vect-patterns.c.jj 2011-11-29 15:09:18.000000000 +0100
> +++ gcc/tree-vect-patterns.c    2011-11-30 17:57:42.183149742 +0100
> @@ -2078,6 +2078,8 @@ vect_recog_bool_pattern (VEC (gimple, he
>       stmt_vec_info pattern_stmt_info;
>       vectype = STMT_VINFO_VECTYPE (stmt_vinfo);
>       gcc_assert (vectype != NULL_TREE);
> +      if (!VECTOR_MODE_P (TYPE_MODE (vectype)))
> +       return NULL;
>       if (!check_bool_pattern (var, loop_vinfo))
>        return NULL;
>
>
>        Jakub
>