[4/9] Move power-of-two checks for interleaving

NEON has vld3 and vst3 instructions, which support an interleaving of
three vectors.  This patch therefore removes the blanket power-of-two
requirement for interleaving and enforces it on a per-operation
basis instead.

The patch also replaces:

  /* Check that the operation is supported.  */
  if (!vect_strided_store_supported (vectype))
    return false;

with:

  gcc_assert (vect_strided_store_supported (vectype, length));

because it was vectorizable_store's responsibility to check this upfront.
Likewise for loads.

Tested on x86_64-linux-gnu and arm-linux-gnueabi.  OK to install?

Richard


gcc/
	* tree-vectorizer.h (vect_strided_store_supported): Add a
	HOST_WIDE_INT argument.
	(vect_strided_load_supported): Likewise.
	(vect_permute_store_chain): Return void.
	(vect_transform_strided_load): Likewise.
	(vect_permute_load_chain): Delete.
	* tree-vect-data-refs.c (vect_strided_store_supported): Take a
	count argument.  Check that the count is a power of two.
	(vect_strided_load_supported): Likewise.
	(vect_permute_store_chain): Return void.  Update after above changes.
	Assert that the access is supported.
	(vect_permute_load_chain): Likewise.
	(vect_transform_strided_load): Return void.
	* tree-vect-stmts.c (vectorizable_store): Update calls after
	above interface changes.
	(vectorizable_load): Likewise.
	(vect_analyze_stmt): Don't check for strided powers of two here.

On Tue, Apr 12, 2011 at 3:44 PM, Richard Sandiford
<richard.sandiford@linaro.org> wrote:
> NEON has vld3 and vst3 instructions, which support an interleaving of
> three vectors.  This patch therefore removes the blanket power-of-two
> requirement for interleaving and enforces it on a per-operation
> basis instead.
>
> The patch also replaces:
>
>  /* Check that the operation is supported.  */
>  if (!vect_strided_store_supported (vectype))
>    return false;
>
> with:
>
>  gcc_assert (vect_strided_store_supported (vectype, length));
>
> because it was vectorizable_store's responsibility to check this upfront.
> Likewise for loads.
>
> Tested on x86_64-linux-gnu and arm-linux-gnueabi.  OK to install?

Ok.

Thanks,
Richard.

> Richard
>
>
> gcc/
>        * tree-vectorizer.h (vect_strided_store_supported): Add a
>        HOST_WIDE_INT argument.
>        (vect_strided_load_supported): Likewise.
>        (vect_permute_store_chain): Return void.
>        (vect_transform_strided_load): Likewise.
>        (vect_permute_load_chain): Delete.
>        * tree-vect-data-refs.c (vect_strided_store_supported): Take a
>        count argument.  Check that the count is a power of two.
>        (vect_strided_load_supported): Likewise.
>        (vect_permute_store_chain): Return void.  Update after above changes.
>        Assert that the access is supported.
>        (vect_permute_load_chain): Likewise.
>        (vect_transform_strided_load): Return void.
>        * tree-vect-stmts.c (vectorizable_store): Update calls after
>        above interface changes.
>        (vectorizable_load): Likewise.
>        (vect_analyze_stmt): Don't check for strided powers of two here.
>
> Index: gcc/tree-vectorizer.h
> ===================================================================
> --- gcc/tree-vectorizer.h       2011-04-12 11:55:07.000000000 +0100
> +++ gcc/tree-vectorizer.h       2011-04-12 11:55:09.000000000 +0100
> @@ -828,16 +828,14 @@ extern tree vect_create_data_ref_ptr (gi
>                                      gimple *, bool, bool *);
>  extern tree bump_vector_ptr (tree, gimple, gimple_stmt_iterator *, gimple, tree);
>  extern tree vect_create_destination_var (tree, tree);
> -extern bool vect_strided_store_supported (tree);
> -extern bool vect_strided_load_supported (tree);
> -extern bool vect_permute_store_chain (VEC(tree,heap) *,unsigned int, gimple,
> +extern bool vect_strided_store_supported (tree, unsigned HOST_WIDE_INT);
> +extern bool vect_strided_load_supported (tree, unsigned HOST_WIDE_INT);
> +extern void vect_permute_store_chain (VEC(tree,heap) *,unsigned int, gimple,
>                                     gimple_stmt_iterator *, VEC(tree,heap) **);
>  extern tree vect_setup_realignment (gimple, gimple_stmt_iterator *, tree *,
>                                     enum dr_alignment_support, tree,
>                                     struct loop **);
> -extern bool vect_permute_load_chain (VEC(tree,heap) *,unsigned int, gimple,
> -                                    gimple_stmt_iterator *, VEC(tree,heap) **);
> -extern bool vect_transform_strided_load (gimple, VEC(tree,heap) *, int,
> +extern void vect_transform_strided_load (gimple, VEC(tree,heap) *, int,
>                                          gimple_stmt_iterator *);
>  extern int vect_get_place_in_interleaving_chain (gimple, gimple);
>  extern tree vect_get_new_vect_var (tree, enum vect_var_kind, const char *);
> Index: gcc/tree-vect-data-refs.c
> ===================================================================
> --- gcc/tree-vect-data-refs.c   2011-04-12 11:55:07.000000000 +0100
> +++ gcc/tree-vect-data-refs.c   2011-04-12 11:55:09.000000000 +0100
> @@ -2196,19 +2196,6 @@ vect_analyze_group_access (struct data_r
>           return false;
>         }
>
> -      /* FORNOW: we handle only interleaving that is a power of 2.
> -         We don't fail here if it may be still possible to vectorize the
> -         group using SLP.  If not, the size of the group will be checked in
> -         vect_analyze_operations, and the vectorization will fail.  */
> -      if (exact_log2 (stride) == -1)
> -       {
> -         if (vect_print_dump_info (REPORT_DETAILS))
> -           fprintf (vect_dump, "interleaving is not a power of 2");
> -
> -         if (slp_impossible)
> -           return false;
> -       }
> -
>       if (stride == 0)
>         stride = count;
>
> @@ -3349,13 +3336,22 @@ vect_create_destination_var (tree scalar
>    and FALSE otherwise.  */
>
>  bool
> -vect_strided_store_supported (tree vectype)
> +vect_strided_store_supported (tree vectype, unsigned HOST_WIDE_INT count)
>  {
>   optab interleave_high_optab, interleave_low_optab;
>   enum machine_mode mode;
>
>   mode = TYPE_MODE (vectype);
>
> +  /* vect_permute_store_chain requires the group size to be a power of two.  */
> +  if (exact_log2 (count) == -1)
> +    {
> +      if (vect_print_dump_info (REPORT_DETAILS))
> +       fprintf (vect_dump, "the size of the group of strided accesses"
> +                " is not a power of 2");
> +      return false;
> +    }
> +
>   /* Check that the operation is supported.  */
>   interleave_high_optab = optab_for_tree_code (VEC_INTERLEAVE_HIGH_EXPR,
>                                               vectype, optab_default);
> @@ -3441,7 +3437,7 @@ vect_strided_store_supported (tree vecty
>    I3:  4 12 20 28  5 13 21 30
>    I4:  6 14 22 30  7 15 23 31.  */
>
> -bool
> +void
>  vect_permute_store_chain (VEC(tree,heap) *dr_chain,
>                          unsigned int length,
>                          gimple stmt,
> @@ -3455,9 +3451,7 @@ vect_permute_store_chain (VEC(tree,heap)
>   unsigned int j;
>   enum tree_code high_code, low_code;
>
> -  /* Check that the operation is supported.  */
> -  if (!vect_strided_store_supported (vectype))
> -    return false;
> +  gcc_assert (vect_strided_store_supported (vectype, length));
>
>   *result_chain = VEC_copy (tree, heap, dr_chain);
>
> @@ -3510,7 +3504,6 @@ vect_permute_store_chain (VEC(tree,heap)
>        }
>       dr_chain = VEC_copy (tree, heap, *result_chain);
>     }
> -  return true;
>  }
>
>  /* Function vect_setup_realignment
> @@ -3787,13 +3780,22 @@ vect_setup_realignment (gimple stmt, gim
>    and FALSE otherwise.  */
>
>  bool
> -vect_strided_load_supported (tree vectype)
> +vect_strided_load_supported (tree vectype, unsigned HOST_WIDE_INT count)
>  {
>   optab perm_even_optab, perm_odd_optab;
>   enum machine_mode mode;
>
>   mode = TYPE_MODE (vectype);
>
> +  /* vect_permute_load_chain requires the group size to be a power of two.  */
> +  if (exact_log2 (count) == -1)
> +    {
> +      if (vect_print_dump_info (REPORT_DETAILS))
> +       fprintf (vect_dump, "the size of the group of strided accesses"
> +                " is not a power of 2");
> +      return false;
> +    }
> +
>   perm_even_optab = optab_for_tree_code (VEC_EXTRACT_EVEN_EXPR, vectype,
>                                         optab_default);
>   if (!perm_even_optab)
> @@ -3905,7 +3907,7 @@ vect_strided_load_supported (tree vectyp
>    3rd vec (E2):  2 6 10 14 18 22 26 30
>    4th vec (E4):  3 7 11 15 19 23 27 31.  */
>
> -bool
> +static void
>  vect_permute_load_chain (VEC(tree,heap) *dr_chain,
>                         unsigned int length,
>                         gimple stmt,
> @@ -3918,9 +3920,7 @@ vect_permute_load_chain (VEC(tree,heap)
>   int i;
>   unsigned int j;
>
> -  /* Check that the operation is supported.  */
> -  if (!vect_strided_load_supported (vectype))
> -    return false;
> +  gcc_assert (vect_strided_load_supported (vectype, length));
>
>   *result_chain = VEC_copy (tree, heap, dr_chain);
>   for (i = 0; i < exact_log2 (length); i++)
> @@ -3963,7 +3963,6 @@ vect_permute_load_chain (VEC(tree,heap)
>        }
>       dr_chain = VEC_copy (tree, heap, *result_chain);
>     }
> -  return true;
>  }
>
>
> @@ -3974,7 +3973,7 @@ vect_permute_load_chain (VEC(tree,heap)
>    the scalar statements.
>  */
>
> -bool
> +void
>  vect_transform_strided_load (gimple stmt, VEC(tree,heap) *dr_chain, int size,
>                             gimple_stmt_iterator *gsi)
>  {
> @@ -3990,8 +3989,7 @@ vect_transform_strided_load (gimple stmt
>      vectors, that are ready for vector computation.  */
>   result_chain = VEC_alloc (tree, heap, size);
>   /* Permute.  */
> -  if (!vect_permute_load_chain (dr_chain, size, stmt, gsi, &result_chain))
> -    return false;
> +  vect_permute_load_chain (dr_chain, size, stmt, gsi, &result_chain);
>
>   /* Put a permuted data-ref in the VECTORIZED_STMT field.
>      Since we scan the chain starting from it's first node, their order
> @@ -4055,7 +4053,6 @@ vect_transform_strided_load (gimple stmt
>     }
>
>   VEC_free (tree, heap, result_chain);
> -  return true;
>  }
>
>  /* Function vect_force_dr_alignment_p.
> Index: gcc/tree-vect-stmts.c
> ===================================================================
> --- gcc/tree-vect-stmts.c       2011-04-12 11:55:09.000000000 +0100
> +++ gcc/tree-vect-stmts.c       2011-04-12 11:55:09.000000000 +0100
> @@ -3412,9 +3412,12 @@ vectorizable_store (gimple stmt, gimple_
>     {
>       strided_store = true;
>       first_stmt = DR_GROUP_FIRST_DR (stmt_info);
> -      if (!vect_strided_store_supported (vectype)
> -         && !PURE_SLP_STMT (stmt_info) && !slp)
> -       return false;
> +      if (!slp && !PURE_SLP_STMT (stmt_info))
> +       {
> +         group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt));
> +         if (!vect_strided_store_supported (vectype, group_size))
> +           return false;
> +       }
>
>       if (first_stmt == stmt)
>        {
> @@ -3617,9 +3620,8 @@ vectorizable_store (gimple stmt, gimple_
>            {
>              result_chain = VEC_alloc (tree, heap, group_size);
>              /* Permute.  */
> -             if (!vect_permute_store_chain (dr_chain, group_size, stmt, gsi,
> -                                            &result_chain))
> -               return false;
> +             vect_permute_store_chain (dr_chain, group_size, stmt, gsi,
> +                                       &result_chain);
>            }
>
>          next_stmt = first_stmt;
> @@ -3912,10 +3914,13 @@ vectorizable_load (gimple stmt, gimple_s
>       /* FORNOW */
>       gcc_assert (! nested_in_vect_loop);
>
> -      /* Check if interleaving is supported.  */
> -      if (!vect_strided_load_supported (vectype)
> -         && !PURE_SLP_STMT (stmt_info) && !slp)
> -       return false;
> +      first_stmt = DR_GROUP_FIRST_DR (stmt_info);
> +      if (!slp && !PURE_SLP_STMT (stmt_info))
> +       {
> +         group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt));
> +         if (!vect_strided_load_supported (vectype, group_size))
> +           return false;
> +       }
>     }
>
>   if (negative)
> @@ -4344,10 +4349,7 @@ vectorizable_load (gimple stmt, gimple_s
>         {
>           if (strided_load)
>            {
> -             if (!vect_transform_strided_load (stmt, dr_chain,
> -                                               group_size, gsi))
> -               return false;
> -
> +             vect_transform_strided_load (stmt, dr_chain, group_size, gsi);
>              *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
>            }
>           else
> @@ -4766,27 +4768,6 @@ vect_analyze_stmt (gimple stmt, bool *ne
>        return false;
>     }
>
> -  if (!PURE_SLP_STMT (stmt_info))
> -    {
> -      /* Groups of strided accesses whose size is not a power of 2 are not
> -         vectorizable yet using loop-vectorization.  Therefore, if this stmt
> -        feeds non-SLP-able stmts (i.e., this stmt has to be both SLPed and
> -        loop-based vectorized), the loop cannot be vectorized.  */
> -      if (STMT_VINFO_STRIDED_ACCESS (stmt_info)
> -          && exact_log2 (DR_GROUP_SIZE (vinfo_for_stmt (
> -                                        DR_GROUP_FIRST_DR (stmt_info)))) == -1)
> -        {
> -          if (vect_print_dump_info (REPORT_DETAILS))
> -            {
> -              fprintf (vect_dump, "not vectorized: the size of group "
> -                                  "of strided accesses is not a power of 2");
> -              print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
> -            }
> -
> -          return false;
> -        }
> -    }
> -
>   return true;
>  }
>
>