Re: [PATCH] Maintain (mis-)alignment info in the first element of a group

[Richard Sandiford via Gcc-patches]

Richard Biener via Gcc-patches <gcc-patches@gcc.gnu.org> writes:
> This changes us to maintain and compute (mis-)alignment info for
> the first element of a group only rather than for each DR when
> doing interleaving and for the earliest, first, or first in the SLP
> node (or any pair or all three of those) when SLP vectorizing.
>
> For this to work out the easiest way I have changed the accessors
> DR_MISALIGNMENT and DR_TARGET_ALIGNMENT to do the indirection to
> the first element rather than adjusting all callers.
> dr_misalignment is moved out-of-line and I'm not too fond of the
> poly-int dances there (any hints?), but basically we are now
> adjusting the first elements misalignment based on the DR_INIT
> difference.
>
> Bootstrap & regtest running on x86_64-unknown-linux-gnu.
>
> Richard.
>
> 2021-09-13  Richard Biener  <rguent...@suse.de>
>
>       * tree-vectorizer.h (dr_misalignment): Move out of line.
>       (dr_target_alignment): New.
>       (DR_TARGET_ALIGNMENT): Wrap dr_target_alignment.
>       (set_dr_target_alignment): New.
>       (SET_DR_TARGET_ALIGNMENT): Wrap set_dr_target_alignment.
>       * tree-vect-data-refs.c (dr_misalignment): Compute and
>       return the group members misalignment.
>       (vect_compute_data_ref_alignment): Use SET_DR_TARGET_ALIGNMENT.
>       (vect_analyze_data_refs_alignment): Compute alignment only
>       for the first element of a DR group.
>       (vect_slp_analyze_node_alignment): Likewise.
> ---
>  gcc/tree-vect-data-refs.c | 65 ++++++++++++++++++++++++---------------
>  gcc/tree-vectorizer.h     | 24 ++++++++++-----
>  2 files changed, 57 insertions(+), 32 deletions(-)
>
> diff --git a/gcc/tree-vect-data-refs.c b/gcc/tree-vect-data-refs.c
> index 66e76132d14..b53d6a0b3f1 100644
> --- a/gcc/tree-vect-data-refs.c
> +++ b/gcc/tree-vect-data-refs.c
> @@ -887,6 +887,36 @@ vect_slp_analyze_instance_dependence (vec_info *vinfo, 
> slp_instance instance)
>    return res;
>  }
>  
> +/* Return the misalignment of DR_INFO.  */
> +
> +int
> +dr_misalignment (dr_vec_info *dr_info)
> +{
> +  if (STMT_VINFO_GROUPED_ACCESS (dr_info->stmt))
> +    {
> +      dr_vec_info *first_dr
> +     = STMT_VINFO_DR_INFO (DR_GROUP_FIRST_ELEMENT (dr_info->stmt));
> +      int misalign = first_dr->misalignment;
> +      gcc_assert (misalign != DR_MISALIGNMENT_UNINITIALIZED);
> +      if (misalign == DR_MISALIGNMENT_UNKNOWN)
> +     return misalign;
> +      poly_offset_int diff = (wi::to_poly_offset (DR_INIT (dr_info->dr))
> +                           - wi::to_poly_offset (DR_INIT (first_dr->dr)));
> +      poly_int64 mispoly = misalign + diff.to_constant ().to_shwi ();
> +      bool res = known_misalignment (mispoly,
> +                                  first_dr->target_alignment.to_constant (),
> +                                  &misalign);
> +      gcc_assert (res);
> +      return misalign;

Yeah, not too keen on the to_constants here.  The one on diff looks
redundant -- you could just use diff.force_shwi () instead, and
keep everything poly_int.

For the known_misalignment I think we should use:

   if (!can_div_trunc_p (mispoly, first_dr->target_alignment,
                         &quotient, &misalign))
     misalign = DR_MISALIGNMENT_UNKNOWN;
   return misalign;

There are then no to_constant assumptions.

Thanks,
Richard