On Mon, Jan 24, 2022 at 2:01 PM liuhongt via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> The pattern converts (cond (cmp a b) (convert c) (convert d))
> to (convert (cond (cmp a b) c d)) when
> 1) types_match (c, d)
> 2) single_use for (convert c) and (convert d)
> 3) TYPE_PRECISION (TREE_TYPE (c)) == TYPE_PRECISION (TREE_TYPE (a))
> 4) INTEGERAL_TYPE_P (TREE_TYPE (c))
>
> The pattern can save packing of mask and data(partial for data, 2 vs
> 1).
>
> Bootstrapped and regtested for x86_64-pc-linux-gnu{-m32,}
> and x86_64-pc-linux-gnu{-m32\ -march=native,\ -march=native} on CLX.
> Ok for trunk?
Sorry for the slow response, see for some comments below
> gcc/ChangeLog:
>
> PR target/103771
> * match.pd (cond_expr_convert_p): New match.
> * tree-vect-patterns.cc (gimple_cond_expr_convert_p): Declare.
> (vect_recog_cond_expr_convert_pattern): New.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/i386/pr103771-2.c: New test.
> ---
> gcc/match.pd | 8 +++
> gcc/testsuite/gcc.target/i386/pr103771-2.c | 8 +++
> gcc/tree-vect-patterns.cc | 83 ++++++++++++++++++++++
> 3 files changed, 99 insertions(+)
> create mode 100644 gcc/testsuite/gcc.target/i386/pr103771-2.c
>
> diff --git a/gcc/match.pd b/gcc/match.pd
> index c68eed70a26..5808c4561ee 100644
> --- a/gcc/match.pd
> +++ b/gcc/match.pd
> @@ -7647,3 +7647,11 @@ and,
> to the number of trailing zeroes. */
> (match (ctz_table_index @1 @2 @3)
> (rshift (mult (bit_and:c (negate @1) @1) INTEGER_CST@2) INTEGER_CST@3))
> +
> +(match (cond_expr_convert_p @0 @2 @3 @6)
> + (cond (simple_comparison@6 @0 @1) (convert@4 @2) (convert@5 @3))
> + (if (types_match (TREE_TYPE (@2), TREE_TYPE (@3))
But in principle @2 or @3 could safely differ in sign, you'd then need to ensure
to insert sign conversions to @2/@3 to the signedness of @4/@5.
> + && INTEGRAL_TYPE_P (type)
> + && INTEGRAL_TYPE_P (TREE_TYPE (@2))
> + && single_use (@4)
> + && single_use (@5))))
> diff --git a/gcc/testsuite/gcc.target/i386/pr103771-2.c
> b/gcc/testsuite/gcc.target/i386/pr103771-2.c
> new file mode 100644
> index 00000000000..962a3a74ecf
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr103771-2.c
> @@ -0,0 +1,8 @@
> +/* { dg-do compile } */
> +/* { dg-options "-march=cascadelake -O3" } */
> +/* { dg-final { scan-assembler-not "kunpck" } } */
> +/* { dg-final { scan-assembler-not "kand" } } */
> +/* { dg-final { scan-assembler-not "kor" } } */
> +/* { dg-final { scan-assembler-not "kshift" } } */
> +
> +#include "pr103771.c"
> diff --git a/gcc/tree-vect-patterns.cc b/gcc/tree-vect-patterns.cc
> index bea04992160..cbdfa96789c 100644
> --- a/gcc/tree-vect-patterns.cc
> +++ b/gcc/tree-vect-patterns.cc
> @@ -924,6 +924,88 @@ vect_reassociating_reduction_p (vec_info *vinfo,
> return true;
> }
>
> +/* match.pd function to match
> + (cond (cmp@3 a b) (convert@1 c) (convert@2 d))
> + with conditions:
> + 1) there's single_use for both @1 and @2.
> + 2) c and d have same type.
> + record a and c and d and @3. */
> +
> +extern bool gimple_cond_expr_convert_p (tree, tree*, tree (*)(tree));
> +
> +/* Function vect_recog_cond_expr_convert
> +
> + Try to find the following pattern:
> +
> + TYPE1 A, B;
> + TYPE2 C,D;
> + TYPE3 E;
> + TYPE3 op_true = (TYPE3)A;
> + TYPE4 op_false = (TYPE3)B;
> +
> + E = C cmp D ? op_true : op_false;
> +
> + where
> + TYPE_PRECISION (TYPE1) != TYPE_PRECISION (TYPE3);
you are not testing for this anywhere?
Btw, matching up the comments with the types is somewhat difficult,
maybe using TYPE_AB, TYPE_CD, TYPE_E instead of 1,2,3 will
make that easier ;)
> + TYPE_PRECISION (TYPE1) == TYPE_PRECISION (TYPE2);
> + single_use of op_true and op_false.
> +
> + Input:
> +
> + * STMT_VINFO: The stmt from which the pattern search begins.
> + here it starts with E = c cmp D ? op_true : op_false;
> +
> + Output:
> +
> + TYPE1 E' = C cmp D ? A : B;
> + TYPE3 E = (TYPE3) E';
> +
> + * TYPE_OUT: The vector type of the output of this pattern.
> +
> + * Return value: A new stmt that will be used to replace the sequence of
> + stmts that constitute the pattern. In this case it will be:
> + E = (TYPE3)E';
> + E' = C cmp D ? A : B; is recorded in pattern definition statements; */
> +
> +static gimple *
> +vect_recog_cond_expr_convert_pattern (vec_info *vinfo,
> + stmt_vec_info stmt_vinfo, tree
> *type_out)
> +{
> + gassign *last_stmt = dyn_cast <gassign *> (stmt_vinfo->stmt);
> + tree lhs, match[4], temp, type, new_lhs;
> + gimple *cond_stmt;
> + gimple *pattern_stmt;
> +
> + if (!last_stmt)
> + return NULL;
> +
> + lhs = gimple_assign_lhs (last_stmt);
> +
> + /* Find E = C cmp D ? (TYPE3) A ? (TYPE3) B;
> + TYPE_PRECISION (A) == TYPE_PRECISION (C). */
> + if (!gimple_cond_expr_convert_p (lhs, &match[0], NULL)
> + || (TYPE_PRECISION (TREE_TYPE (match[0]))
> + != TYPE_PRECISION (TREE_TYPE (match[1]))))
I think the precision check should be part of the match.pd pattern. You
do not check that the comparison operands are integral - I think float
comparisons would be OK in principle but the precision check will not
work there.
Thanks,
Richard.
> + return NULL;
> +
> + vect_pattern_detected ("vect_recog_cond_expr_convert_pattern", last_stmt);
> +
> + type = TREE_TYPE (match[1]);
> + temp = vect_recog_temp_ssa_var (type, NULL);
> + cond_stmt = gimple_build_assign (temp, build3 (COND_EXPR, type, match[3],
> + match[1], match[2]));
> + append_pattern_def_seq (vinfo, stmt_vinfo, cond_stmt,
> + get_vectype_for_scalar_type (vinfo, type));
> + new_lhs = vect_recog_temp_ssa_var (TREE_TYPE (lhs), NULL);
> + pattern_stmt = gimple_build_assign (new_lhs, CONVERT_EXPR, temp);
> + *type_out = STMT_VINFO_VECTYPE (stmt_vinfo);
> +
> + if (dump_enabled_p ())
> + dump_printf_loc (MSG_NOTE, vect_location,
> + "created pattern stmt: %G", pattern_stmt);
> + return pattern_stmt;
> +}
> +
> /* Function vect_recog_dot_prod_pattern
>
> Try to find the following pattern:
> @@ -5492,6 +5574,7 @@ static vect_recog_func vect_vect_recog_func_ptrs[] = {
> /* Must come after over_widening, which narrows the shift as much as
> possible beforehand. */
> { vect_recog_average_pattern, "average" },
> + { vect_recog_cond_expr_convert_pattern, "cond_expr_convert" },
> { vect_recog_mulhs_pattern, "mult_high" },
> { vect_recog_cast_forwprop_pattern, "cast_forwprop" },
> { vect_recog_widen_mult_pattern, "widen_mult" },
> --
> 2.18.1
>
