On Wed, May 17, 2023 at 8:15 AM Kewen.Lin <li...@linux.ibm.com> wrote:
>
> Hi,
>
> Following Richi's suggestion in [1], I'm working on deferring
> cost evaluation next to the transformation, this patch is
> to enhance function vect_transform_slp_perm_load_1 which
> could under-cost for vector permutation, since the costing
> doesn't try to consider nvectors_per_build, it's inconsistent
> with the transformation part.
>
> Bootstrapped and regtested on x86_64-redhat-linux,
> aarch64-linux-gnu and powerpc64{,le}-linux-gnu.
>
> Is it ok for trunk?
>
> [1] https://gcc.gnu.org/pipermail/gcc-patches/2021-January/563624.html
>
> BR,
> Kewen
> -----
> gcc/ChangeLog:
>
> * tree-vect-slp.cc (vect_transform_slp_perm_load_1): Adjust the
> calculation on n_perms by considering nvectors_per_build.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.dg/vect/costmodel/ppc/costmodel-slp-perm.c: New test.
> ---
> .../vect/costmodel/ppc/costmodel-slp-perm.c | 23 +++++++
> gcc/tree-vect-slp.cc | 66 ++++++++++---------
> 2 files changed, 57 insertions(+), 32 deletions(-)
> create mode 100644
> gcc/testsuite/gcc.dg/vect/costmodel/ppc/costmodel-slp-perm.c
>
> diff --git a/gcc/testsuite/gcc.dg/vect/costmodel/ppc/costmodel-slp-perm.c
> b/gcc/testsuite/gcc.dg/vect/costmodel/ppc/costmodel-slp-perm.c
> new file mode 100644
> index 00000000000..e5c4dceddfb
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/vect/costmodel/ppc/costmodel-slp-perm.c
> @@ -0,0 +1,23 @@
> +/* { dg-do compile } */
> +/* { dg-require-effective-target vect_int } */
> +/* { dg-require-effective-target powerpc_p9vector_ok } */
> +/* Specify power9 to ensure the vectorization is profitable
> + and test point stands, otherwise it could be not profitable
> + to vectorize. */
> +/* { dg-additional-options "-mdejagnu-cpu=power9 -mpower9-vector" } */
> +
> +/* Verify we cost the exact count for required vec_perm. */
> +
> +int x[1024], y[1024];
> +
> +void
> +foo ()
> +{
> + for (int i = 0; i < 512; ++i)
> + {
> + x[2 * i] = y[1023 - (2 * i)];
> + x[2 * i + 1] = y[1023 - (2 * i + 1)];
> + }
> +}
> +
> +/* { dg-final { scan-tree-dump-times "2 times vec_perm" 1 "vect" } } */
> diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
> index e5c9d7e766e..af9a6dd4fa9 100644
> --- a/gcc/tree-vect-slp.cc
> +++ b/gcc/tree-vect-slp.cc
> @@ -8115,12 +8115,12 @@ vect_transform_slp_perm_load_1 (vec_info *vinfo,
> slp_tree node,
>
> mode = TYPE_MODE (vectype);
> poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
> + unsigned int nstmts = SLP_TREE_NUMBER_OF_VEC_STMTS (node);
>
> /* Initialize the vect stmts of NODE to properly insert the generated
> stmts later. */
> if (! analyze_only)
> - for (unsigned i = SLP_TREE_VEC_STMTS (node).length ();
> - i < SLP_TREE_NUMBER_OF_VEC_STMTS (node); i++)
> + for (unsigned i = SLP_TREE_VEC_STMTS (node).length (); i < nstmts; i++)
> SLP_TREE_VEC_STMTS (node).quick_push (NULL);
>
> /* Generate permutation masks for every NODE. Number of masks for each NODE
> @@ -8161,7 +8161,10 @@ vect_transform_slp_perm_load_1 (vec_info *vinfo,
> slp_tree node,
> (b) the permutes only need a single vector input. */
> mask.new_vector (nunits, group_size, 3);
> nelts_to_build = mask.encoded_nelts ();
> - nvectors_per_build = SLP_TREE_VEC_STMTS (node).length ();
> + /* It's possible to obtain zero nstmts during analyze_only, so make
> + it at least one to ensure the later computation for n_perms
> + proceed. */
> + nvectors_per_build = nstmts > 0 ? nstmts : 1;
> in_nlanes = DR_GROUP_SIZE (stmt_info) * 3;
> }
> else
> @@ -8252,40 +8255,39 @@ vect_transform_slp_perm_load_1 (vec_info *vinfo,
> slp_tree node,
> return false;
> }
>
> - ++*n_perms;
> -
> + tree mask_vec = NULL_TREE;
> if (!analyze_only)
> - {
> - tree mask_vec = vect_gen_perm_mask_checked (vectype,
> indices);
> + mask_vec = vect_gen_perm_mask_checked (vectype, indices);
>
> - if (second_vec_index == -1)
> - second_vec_index = first_vec_index;
> + if (second_vec_index == -1)
> + second_vec_index = first_vec_index;
>
> - for (unsigned int ri = 0; ri < nvectors_per_build; ++ri)
> + for (unsigned int ri = 0; ri < nvectors_per_build; ++ri)
> + {
> + ++*n_perms;
So the "real" change is doing
*n_perms += nvectors_per_build;
and *n_perms was unused when !analyze_only? And since at
analysis time we (sometimes?) have zero nvectors you have to
fixup above? Which cases are that?
In principle the patch looks good to me.
Richard.
> + if (analyze_only)
> + continue;
> + /* Generate the permute statement if necessary. */
> + tree first_vec = dr_chain[first_vec_index + ri];
> + tree second_vec = dr_chain[second_vec_index + ri];
> + gassign *stmt = as_a<gassign *> (stmt_info->stmt);
> + tree perm_dest
> + = vect_create_destination_var (gimple_assign_lhs (stmt),
> + vectype);
> + perm_dest = make_ssa_name (perm_dest);
> + gimple *perm_stmt
> + = gimple_build_assign (perm_dest, VEC_PERM_EXPR,
> first_vec,
> + second_vec, mask_vec);
> + vect_finish_stmt_generation (vinfo, stmt_info, perm_stmt,
> + gsi);
> + if (dce_chain)
> {
> - /* Generate the permute statement if necessary. */
> - tree first_vec = dr_chain[first_vec_index + ri];
> - tree second_vec = dr_chain[second_vec_index + ri];
> - gassign *stmt = as_a<gassign *> (stmt_info->stmt);
> - tree perm_dest
> - = vect_create_destination_var (gimple_assign_lhs
> (stmt),
> - vectype);
> - perm_dest = make_ssa_name (perm_dest);
> - gimple *perm_stmt
> - = gimple_build_assign (perm_dest, VEC_PERM_EXPR,
> - first_vec, second_vec,
> mask_vec);
> - vect_finish_stmt_generation (vinfo, stmt_info,
> perm_stmt,
> - gsi);
> - if (dce_chain)
> - {
> - bitmap_set_bit (used_defs, first_vec_index + ri);
> - bitmap_set_bit (used_defs, second_vec_index + ri);
> - }
> -
> - /* Store the vector statement in NODE. */
> - SLP_TREE_VEC_STMTS (node) [vect_stmts_counter++]
> - = perm_stmt;
> + bitmap_set_bit (used_defs, first_vec_index + ri);
> + bitmap_set_bit (used_defs, second_vec_index + ri);
> }
> +
> + /* Store the vector statement in NODE. */
> + SLP_TREE_VEC_STMTS (node)[vect_stmts_counter++] = perm_stmt;
> }
> }
> else if (!analyze_only)
> --
> 2.39.1
>
