From: Richard Biener <richard.guenther@gmail.com>
To: "Kewen.Lin" <linkw@linux.ibm.com>
Cc: GCC Patches <gcc-patches@gcc.gnu.org>,
Richard Sandiford <richard.sandiford@arm.com>,
Segher Boessenkool <segher@kernel.crashing.org>,
Peter Bergner <bergner@linux.ibm.com>
Subject: Re: [PATCH 3/3] vect: Move VMAT_GATHER_SCATTER handlings from final loop nest
Date: Tue, 22 Aug 2023 15:28:09 +0200 [thread overview]
Message-ID: <CAFiYyc3vtEPB0_hM+M4jwSHw6WMFKHc6SPhbWo9hK0e-QNbvWg@mail.gmail.com> (raw)
In-Reply-To: <1c07d6a4-f322-6a1d-aaea-4d17733493fe@linux.ibm.com>
On Tue, Aug 22, 2023 at 10:52 AM Kewen.Lin <linkw@linux.ibm.com> wrote:
>
> Hi,
>
> Like r14-3317 which moves the handlings on memory access
> type VMAT_GATHER_SCATTER in vectorizable_load final loop
> nest, this one is to deal with vectorizable_store side.
>
> Bootstrapped and regtested on x86_64-redhat-linux,
> aarch64-linux-gnu and powerpc64{,le}-linux-gnu.
>
> Is it ok for trunk?
OK.
> BR,
> Kewen
> -----
>
> gcc/ChangeLog:
>
> * tree-vect-stmts.cc (vectorizable_store): Move the handlings on
> VMAT_GATHER_SCATTER in the final loop nest to its own loop,
> and update the final nest accordingly.
> ---
> gcc/tree-vect-stmts.cc | 258 +++++++++++++++++++++++++----------------
> 1 file changed, 159 insertions(+), 99 deletions(-)
>
> diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
> index 18f5ebcc09c..b959c1861ad 100644
> --- a/gcc/tree-vect-stmts.cc
> +++ b/gcc/tree-vect-stmts.cc
> @@ -8930,44 +8930,23 @@ vectorizable_store (vec_info *vinfo,
> return true;
> }
>
> - auto_vec<tree> result_chain (group_size);
> - auto_vec<tree> vec_offsets;
> - auto_vec<tree, 1> vec_oprnds;
> - for (j = 0; j < ncopies; j++)
> + if (memory_access_type == VMAT_GATHER_SCATTER)
> {
> - gimple *new_stmt;
> - if (j == 0)
> + gcc_assert (!slp && !grouped_store);
> + auto_vec<tree> vec_offsets;
> + for (j = 0; j < ncopies; j++)
> {
> - if (slp)
> - {
> - /* Get vectorized arguments for SLP_NODE. */
> - vect_get_vec_defs (vinfo, stmt_info, slp_node, 1, op,
> - &vec_oprnds);
> - vec_oprnd = vec_oprnds[0];
> - }
> - else
> + gimple *new_stmt;
> + if (j == 0)
> {
> - /* For interleaved stores we collect vectorized defs for all the
> - stores in the group in DR_CHAIN. DR_CHAIN is then used as an
> - input to vect_permute_store_chain().
> -
> - If the store is not grouped, DR_GROUP_SIZE is 1, and DR_CHAIN
> - is of size 1. */
> - stmt_vec_info next_stmt_info = first_stmt_info;
> - for (i = 0; i < group_size; i++)
> - {
> - /* Since gaps are not supported for interleaved stores,
> - DR_GROUP_SIZE is the exact number of stmts in the chain.
> - Therefore, NEXT_STMT_INFO can't be NULL_TREE. In case
> - that there is no interleaving, DR_GROUP_SIZE is 1,
> - and only one iteration of the loop will be executed. */
> - op = vect_get_store_rhs (next_stmt_info);
> - vect_get_vec_defs_for_operand (vinfo, next_stmt_info, ncopies,
> - op, gvec_oprnds[i]);
> - vec_oprnd = (*gvec_oprnds[i])[0];
> - dr_chain.quick_push (vec_oprnd);
> - next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info);
> - }
> + /* Since the store is not grouped, DR_GROUP_SIZE is 1, and
> + DR_CHAIN is of size 1. */
> + gcc_assert (group_size == 1);
> + op = vect_get_store_rhs (first_stmt_info);
> + vect_get_vec_defs_for_operand (vinfo, first_stmt_info, ncopies,
> + op, gvec_oprnds[0]);
> + vec_oprnd = (*gvec_oprnds[0])[0];
> + dr_chain.quick_push (vec_oprnd);
> if (mask)
> {
> vect_get_vec_defs_for_operand (vinfo, stmt_info, ncopies,
> @@ -8975,91 +8954,55 @@ vectorizable_store (vec_info *vinfo,
> mask_vectype);
> vec_mask = vec_masks[0];
> }
> - }
>
> - /* We should have catched mismatched types earlier. */
> - gcc_assert (useless_type_conversion_p (vectype,
> - TREE_TYPE (vec_oprnd)));
> - bool simd_lane_access_p
> - = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) != 0;
> - if (simd_lane_access_p
> - && !loop_masks
> - && TREE_CODE (DR_BASE_ADDRESS (first_dr_info->dr)) == ADDR_EXPR
> - && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr_info->dr), 0))
> - && integer_zerop (get_dr_vinfo_offset (vinfo, first_dr_info))
> - && integer_zerop (DR_INIT (first_dr_info->dr))
> - && alias_sets_conflict_p (get_alias_set (aggr_type),
> - get_alias_set (TREE_TYPE (ref_type))))
> - {
> - dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr_info->dr));
> - dataref_offset = build_int_cst (ref_type, 0);
> + /* We should have catched mismatched types earlier. */
> + gcc_assert (useless_type_conversion_p (vectype,
> + TREE_TYPE (vec_oprnd)));
> + if (STMT_VINFO_GATHER_SCATTER_P (stmt_info))
> + vect_get_gather_scatter_ops (loop_vinfo, loop, stmt_info,
> + slp_node, &gs_info, &dataref_ptr,
> + &vec_offsets);
> + else
> + dataref_ptr
> + = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
> + NULL, offset, &dummy, gsi,
> + &ptr_incr, false, bump);
> }
> - else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info))
> - vect_get_gather_scatter_ops (loop_vinfo, loop, stmt_info, slp_node,
> - &gs_info, &dataref_ptr, &vec_offsets);
> else
> - dataref_ptr
> - = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
> - simd_lane_access_p ? loop : NULL,
> - offset, &dummy, gsi, &ptr_incr,
> - simd_lane_access_p, bump);
> - }
> - else
> - {
> - gcc_assert (!LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo));
> - /* DR_CHAIN is then used as an input to vect_permute_store_chain().
> - If the store is not grouped, DR_GROUP_SIZE is 1, and DR_CHAIN is
> - of size 1. */
> - for (i = 0; i < group_size; i++)
> {
> - vec_oprnd = (*gvec_oprnds[i])[j];
> - dr_chain[i] = vec_oprnd;
> + gcc_assert (!LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo));
> + vec_oprnd = (*gvec_oprnds[0])[j];
> + dr_chain[0] = vec_oprnd;
> + if (mask)
> + vec_mask = vec_masks[j];
> + if (!STMT_VINFO_GATHER_SCATTER_P (stmt_info))
> + dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr,
> + gsi, stmt_info, bump);
> }
> - if (mask)
> - vec_mask = vec_masks[j];
> - if (dataref_offset)
> - dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, bump);
> - else if (!STMT_VINFO_GATHER_SCATTER_P (stmt_info))
> - dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi,
> - stmt_info, bump);
> - }
> -
> - new_stmt = NULL;
> - if (grouped_store)
> - /* Permute. */
> - vect_permute_store_chain (vinfo, dr_chain, group_size, stmt_info, gsi,
> - &result_chain);
>
> - stmt_vec_info next_stmt_info = first_stmt_info;
> - for (i = 0; i < vec_num; i++)
> - {
> - unsigned misalign;
> + new_stmt = NULL;
> unsigned HOST_WIDE_INT align;
> -
> tree final_mask = NULL_TREE;
> tree final_len = NULL_TREE;
> tree bias = NULL_TREE;
> if (loop_masks)
> final_mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks,
> - vec_num * ncopies, vectype,
> - vec_num * j + i);
> + ncopies, vectype, j);
> if (vec_mask)
> final_mask = prepare_vec_mask (loop_vinfo, mask_vectype, final_mask,
> vec_mask, gsi);
>
> - if (memory_access_type == VMAT_GATHER_SCATTER
> - && gs_info.ifn != IFN_LAST)
> + if (gs_info.ifn != IFN_LAST)
> {
> if (STMT_VINFO_GATHER_SCATTER_P (stmt_info))
> - vec_offset = vec_offsets[vec_num * j + i];
> + vec_offset = vec_offsets[j];
> tree scale = size_int (gs_info.scale);
>
> if (gs_info.ifn == IFN_MASK_LEN_SCATTER_STORE)
> {
> if (loop_lens)
> final_len = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
> - vec_num * ncopies, vectype,
> - vec_num * j + i, 1);
> + ncopies, vectype, j, 1);
> else
> final_len = build_int_cst (sizetype,
> TYPE_VECTOR_SUBPARTS (vectype));
> @@ -9091,9 +9034,8 @@ vectorizable_store (vec_info *vinfo,
> gimple_call_set_nothrow (call, true);
> vect_finish_stmt_generation (vinfo, stmt_info, call, gsi);
> new_stmt = call;
> - break;
> }
> - else if (memory_access_type == VMAT_GATHER_SCATTER)
> + else
> {
> /* Emulated scatter. */
> gcc_assert (!final_mask);
> @@ -9142,8 +9084,126 @@ vectorizable_store (vec_info *vinfo,
> new_stmt = gimple_build_assign (ref, elt);
> vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
> }
> - break;
> }
> + if (j == 0)
> + *vec_stmt = new_stmt;
> + STMT_VINFO_VEC_STMTS (stmt_info).safe_push (new_stmt);
> + }
> + return true;
> + }
> +
> + auto_vec<tree> result_chain (group_size);
> + auto_vec<tree, 1> vec_oprnds;
> + for (j = 0; j < ncopies; j++)
> + {
> + gimple *new_stmt;
> + if (j == 0)
> + {
> + if (slp)
> + {
> + /* Get vectorized arguments for SLP_NODE. */
> + vect_get_vec_defs (vinfo, stmt_info, slp_node, 1, op,
> + &vec_oprnds);
> + vec_oprnd = vec_oprnds[0];
> + }
> + else
> + {
> + /* For interleaved stores we collect vectorized defs for all the
> + stores in the group in DR_CHAIN. DR_CHAIN is then used as an
> + input to vect_permute_store_chain().
> +
> + If the store is not grouped, DR_GROUP_SIZE is 1, and DR_CHAIN
> + is of size 1. */
> + stmt_vec_info next_stmt_info = first_stmt_info;
> + for (i = 0; i < group_size; i++)
> + {
> + /* Since gaps are not supported for interleaved stores,
> + DR_GROUP_SIZE is the exact number of stmts in the chain.
> + Therefore, NEXT_STMT_INFO can't be NULL_TREE. In case
> + that there is no interleaving, DR_GROUP_SIZE is 1,
> + and only one iteration of the loop will be executed. */
> + op = vect_get_store_rhs (next_stmt_info);
> + vect_get_vec_defs_for_operand (vinfo, next_stmt_info, ncopies,
> + op, gvec_oprnds[i]);
> + vec_oprnd = (*gvec_oprnds[i])[0];
> + dr_chain.quick_push (vec_oprnd);
> + next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info);
> + }
> + if (mask)
> + {
> + vect_get_vec_defs_for_operand (vinfo, stmt_info, ncopies,
> + mask, &vec_masks,
> + mask_vectype);
> + vec_mask = vec_masks[0];
> + }
> + }
> +
> + /* We should have catched mismatched types earlier. */
> + gcc_assert (useless_type_conversion_p (vectype,
> + TREE_TYPE (vec_oprnd)));
> + bool simd_lane_access_p
> + = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) != 0;
> + if (simd_lane_access_p
> + && !loop_masks
> + && TREE_CODE (DR_BASE_ADDRESS (first_dr_info->dr)) == ADDR_EXPR
> + && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr_info->dr), 0))
> + && integer_zerop (get_dr_vinfo_offset (vinfo, first_dr_info))
> + && integer_zerop (DR_INIT (first_dr_info->dr))
> + && alias_sets_conflict_p (get_alias_set (aggr_type),
> + get_alias_set (TREE_TYPE (ref_type))))
> + {
> + dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr_info->dr));
> + dataref_offset = build_int_cst (ref_type, 0);
> + }
> + else
> + dataref_ptr
> + = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
> + simd_lane_access_p ? loop : NULL,
> + offset, &dummy, gsi, &ptr_incr,
> + simd_lane_access_p, bump);
> + }
> + else
> + {
> + gcc_assert (!LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo));
> + /* DR_CHAIN is then used as an input to vect_permute_store_chain().
> + If the store is not grouped, DR_GROUP_SIZE is 1, and DR_CHAIN is
> + of size 1. */
> + for (i = 0; i < group_size; i++)
> + {
> + vec_oprnd = (*gvec_oprnds[i])[j];
> + dr_chain[i] = vec_oprnd;
> + }
> + if (mask)
> + vec_mask = vec_masks[j];
> + if (dataref_offset)
> + dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, bump);
> + else
> + dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi,
> + stmt_info, bump);
> + }
> +
> + new_stmt = NULL;
> + if (grouped_store)
> + /* Permute. */
> + vect_permute_store_chain (vinfo, dr_chain, group_size, stmt_info, gsi,
> + &result_chain);
> +
> + stmt_vec_info next_stmt_info = first_stmt_info;
> + for (i = 0; i < vec_num; i++)
> + {
> + unsigned misalign;
> + unsigned HOST_WIDE_INT align;
> +
> + tree final_mask = NULL_TREE;
> + tree final_len = NULL_TREE;
> + tree bias = NULL_TREE;
> + if (loop_masks)
> + final_mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks,
> + vec_num * ncopies, vectype,
> + vec_num * j + i);
> + if (vec_mask)
> + final_mask = prepare_vec_mask (loop_vinfo, mask_vectype, final_mask,
> + vec_mask, gsi);
>
> if (i > 0)
> /* Bump the vector pointer. */
> --
> 2.31.1
next prev parent reply other threads:[~2023-08-22 13:29 UTC|newest]
Thread overview: 7+ messages / expand[flat|nested] mbox.gz Atom feed top
2023-08-22 8:45 [PATCH 1/3] vect: Remove some manual release in vectorizable_store Kewen.Lin
2023-08-22 8:49 ` [PATCH 2/3] vect: Move VMAT_LOAD_STORE_LANES handlings from final loop nest Kewen.Lin
2023-08-22 13:27 ` Richard Biener
2023-08-22 8:52 ` [PATCH 3/3] vect: Move VMAT_GATHER_SCATTER " Kewen.Lin
2023-08-22 13:28 ` Richard Biener [this message]
2023-08-22 12:32 ` [PATCH 1/3] vect: Remove some manual release in vectorizable_store Richard Biener
2023-08-23 5:12 ` Kewen.Lin
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=CAFiYyc3vtEPB0_hM+M4jwSHw6WMFKHc6SPhbWo9hK0e-QNbvWg@mail.gmail.com \
--to=richard.guenther@gmail.com \
--cc=bergner@linux.ibm.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=linkw@linux.ibm.com \
--cc=richard.sandiford@arm.com \
--cc=segher@kernel.crashing.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).