On Wed, 28 Jun 2023, Tamar Christina wrote:
> Hi All,
>
> The bitfield vectorization support does not currently recognize bitfields
> inside
> gconds. This means they can't be used as conditions for early break
> vectorization which is a functionality we require.
>
> This adds support for them by explicitly matching and handling gcond as a
> source.
>
> Testcases are added in the testsuite update patch as the only way to get there
> is with the early break vectorization. See tests:
>
> - vect-early-break_20.c
> - vect-early-break_21.c
>
> Bootstrapped Regtested on aarch64-none-linux-gnu and no issues.
>
> Ok for master?
>
> Thanks,
> Tamar
>
> gcc/ChangeLog:
>
> * tree-vect-patterns.cc (vect_init_pattern_stmt): Copy STMT_VINFO_TYPE
> from original statement.
> (vect_recog_bitfield_ref_pattern): Support bitfields in gcond.
>
> Co-Authored-By: Andre Vieira <[email protected]>
>
> --- inline copy of patch --
> diff --git a/gcc/tree-vect-patterns.cc b/gcc/tree-vect-patterns.cc
> index
> 60bc9be6819af9bd28a81430869417965ba9d82d..c221b1d64449ce3b6c8864bbec4b17ddf938c2d6
> 100644
> --- a/gcc/tree-vect-patterns.cc
> +++ b/gcc/tree-vect-patterns.cc
> @@ -128,6 +128,7 @@ vect_init_pattern_stmt (vec_info *vinfo, gimple
> *pattern_stmt,
> STMT_VINFO_RELATED_STMT (pattern_stmt_info) = orig_stmt_info;
> STMT_VINFO_DEF_TYPE (pattern_stmt_info)
> = STMT_VINFO_DEF_TYPE (orig_stmt_info);
> + STMT_VINFO_TYPE (pattern_stmt_info) = STMT_VINFO_TYPE (orig_stmt_info);
> if (!STMT_VINFO_VECTYPE (pattern_stmt_info))
> {
> gcc_assert (!vectype
> @@ -2488,27 +2489,37 @@ static gimple *
there's a comment above this mentioning what we look for - please
update it with the gcond case.
> vect_recog_bitfield_ref_pattern (vec_info *vinfo, stmt_vec_info stmt_info,
> tree *type_out)
> {
> - gassign *first_stmt = dyn_cast <gassign *> (stmt_info->stmt);
> + gassign *conv_stmt = dyn_cast <gassign *> (stmt_info->stmt);
> + gcond *cond_stmt = dyn_cast <gcond *> (stmt_info->stmt);
>
> - if (!first_stmt)
> - return NULL;
> -
> - gassign *bf_stmt;
> - if (CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (first_stmt))
> - && TREE_CODE (gimple_assign_rhs1 (first_stmt)) == SSA_NAME)
> + gimple *bf_stmt = NULL;
> + tree cond_cst = NULL_TREE;
> + if (cond_stmt)
please make that
if (gcond *cond_stmt = dyn_cast <gcond *> (stmt_info->stmt))
> {
> - gimple *second_stmt
> - = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (first_stmt));
> - bf_stmt = dyn_cast <gassign *> (second_stmt);
> - if (!bf_stmt
> - || gimple_assign_rhs_code (bf_stmt) != BIT_FIELD_REF)
> + tree op = gimple_cond_lhs (cond_stmt);
> + if (TREE_CODE (op) != SSA_NAME)
> + return NULL;
> + bf_stmt = dyn_cast <gassign *> (SSA_NAME_DEF_STMT (op));
> + cond_cst = gimple_cond_rhs (cond_stmt);
> + if (TREE_CODE (cond_cst) != INTEGER_CST)
> return NULL;
> }
> - else
> + else if (conv_stmt
similar
> + && CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (conv_stmt))
> + && TREE_CODE (gimple_assign_rhs1 (conv_stmt)) == SSA_NAME)
> + {
> + gimple *second_stmt = SSA_NAME_DEF_STMT (gimple_assign_rhs1
> (conv_stmt));
> + bf_stmt = dyn_cast <gassign *> (second_stmt);
> + }
> +
> + if (!bf_stmt
> + || gimple_assign_rhs_code (bf_stmt) != BIT_FIELD_REF)
> return NULL;
>
> tree bf_ref = gimple_assign_rhs1 (bf_stmt);
> tree container = TREE_OPERAND (bf_ref, 0);
> + tree ret_type = cond_cst ? TREE_TYPE (container)
> + : TREE_TYPE (gimple_assign_lhs (conv_stmt));
>
> if (!bit_field_offset (bf_ref).is_constant ()
> || !bit_field_size (bf_ref).is_constant ()
> @@ -2522,8 +2533,6 @@ vect_recog_bitfield_ref_pattern (vec_info *vinfo,
> stmt_vec_info stmt_info,
>
> gimple *use_stmt, *pattern_stmt;
> use_operand_p use_p;
> - tree ret = gimple_assign_lhs (first_stmt);
> - tree ret_type = TREE_TYPE (ret);
> bool shift_first = true;
> tree container_type = TREE_TYPE (container);
> tree vectype = get_vectype_for_scalar_type (vinfo, container_type);
> @@ -2560,7 +2569,8 @@ vect_recog_bitfield_ref_pattern (vec_info *vinfo,
> stmt_vec_info stmt_info,
> /* If the only use of the result of this BIT_FIELD_REF + CONVERT is a
> PLUS_EXPR then do the shift last as some targets can combine the shift
> and
> add into a single instruction. */
> - if (single_imm_use (gimple_assign_lhs (first_stmt), &use_p, &use_stmt))
> + if (conv_stmt
> + && single_imm_use (gimple_assign_lhs (conv_stmt), &use_p, &use_stmt))
> {
> if (gimple_code (use_stmt) == GIMPLE_ASSIGN
> && gimple_assign_rhs_code (use_stmt) == PLUS_EXPR)
> @@ -2620,7 +2630,21 @@ vect_recog_bitfield_ref_pattern (vec_info *vinfo,
> stmt_vec_info stmt_info,
> NOP_EXPR, result);
> }
>
> - *type_out = STMT_VINFO_VECTYPE (stmt_info);
> + if (cond_cst)
> + {
> + append_pattern_def_seq (vinfo, stmt_info, pattern_stmt, vectype);
> + pattern_stmt
> + = gimple_build_cond (gimple_cond_code (cond_stmt),
> + gimple_get_lhs (pattern_stmt),
> + fold_convert (ret_type, cond_cst),
> + gimple_cond_true_label (cond_stmt),
> + gimple_cond_false_label (cond_stmt));
> + *type_out = STMT_VINFO_VECTYPE (stmt_info);
is there any vectype set for a gcond?
I must say the flow of the function is a bit convoluted now. Is it
possible to factor out a helper so we can fully separate the
gassign vs. gcond handling in this function?
Thanks,
Richard.
> + }
> + else
> + *type_out
> + = get_vectype_for_scalar_type (vinfo,
> + TREE_TYPE (gimple_get_lhs (pattern_stmt)));
> vect_pattern_detected ("bitfield_ref pattern", stmt_info->stmt);
>
> return pattern_stmt;
>
