[Bug target/110776] [14 Regression] powerpc-darwin bootstrap broken after r14-2490 with ICE rs6000.cc:5069 building libgfortran

["rguenther at suse dot de" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110776

--- Comment #5 from rguenther at suse dot de <rguenther at suse dot de> ---
On Tue, 25 Jul 2023, linkw at gcc dot gnu.org wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110776
> 
> Kewen Lin <linkw at gcc dot gnu.org> changed:
> 
>            What    |Removed                     |Added
> ----------------------------------------------------------------------------
>                  CC|                            |rguenth at gcc dot gnu.org
> 
> --- Comment #3 from Kewen Lin <linkw at gcc dot gnu.org> ---
> The failure exposed one special case: there is one stmt 
> 
> # VUSE <.MEM_404>
> _15 = *_14;
> 
> its STMT_VINFO_STRIDED_P (stmt_info) is set, memory_access_type is
> VMAT_ELEMENTWISE and alignment_support_scheme is dr_unaligned_supported, its
> vector type is "vector(1) long int", so in the handling it's taken as:
> 
>       /* Load vector(1) scalar_type if it's 1 element-wise vectype.  */
>       else if (nloads == 1)
>         ltype = vectype;
> 
> as its ltype is vector type, we cost it with 
> 
>                   if (VECTOR_TYPE_P (ltype))
>                     vect_get_load_cost (vinfo, stmt_info, 1,
>                                         alignment_support_scheme, misalignment,
>                                         false, &inside_cost, nullptr, cost_vec,
>                                         cost_vec, true);
> 
> as it's dr_unaligned_supported, it's costed as unaligned_load then causes the
> ICE. One idea is to teach rs6000_builtin_vectorization_cost about one single
> lane vector unaligned load as scalar_load. But I think it's simple to treat
> single lane vector load as scalar_load, as I expect veclower will lower it
> later.
> 
> diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
> index 4622d6a04ef..bdf4c12cd03 100644
> --- a/gcc/tree-vect-stmts.cc
> +++ b/gcc/tree-vect-stmts.cc
> @@ -9985,7 +9985,9 @@ vectorizable_load (vec_info *vinfo,
>              {
>                if (costing_p)
>                  {
> -                  if (VECTOR_TYPE_P (ltype))
> +                  /* For a single lane vector type, we should cost it as
> +                     scalar_load to avoid ICE, see PR110776.  */
> +                  if (VECTOR_TYPE_P (ltype) && lnel > 1)
>                      vect_get_load_cost (vinfo, stmt_info, 1,
>                                          alignment_support_scheme,
> misalignment,
>                                          false, &inside_cost, nullptr,
> cost_vec,
> 
> Hi Richi, what do you think of this?

I think apart from the consideration what a single element vector
is compared to a scalar, a more to-the-point fix is

  if (VECTOR_TYPE_P (ltype)
      && memory_access_type != VMAT_ELEMENTWISE)