On Wed, Aug 6, 2025 at 7:04 AM Edwin Lu <e...@rivosinc.com> wrote:
>
> This patch tries to add support for a variant of SAT_TRUNC where
> negative numbers are clipped to 0 instead of NARROW_TYPE_MAX_VALUE.
> This form is seen in x264, aka
>
> UT clip (T a)
> {
> return a & (UT)(-1) ? (-a) >> 31 : a;
the patch below says
> + /* SAT_U_TRUNC = (UT)X & (NT)(-1) ? (-X) >> 31 : X
so I'm not sure what this is about? Is an actual example
unsigned char clip (unsigned int a)
{
return a & (unsigned char)-1 ? -a >> 31 : a;
}
? But then I'm confused about the hard-coded 31, -X >> 31 should
be the inverted sign-bit of X?
The pattern restricts this to signed X it seems and the result is also
signed. So what is this supposed to be? A form of MAX(0, x) for
signed x?
> }
>
> Where sizeof(UT) < sizeof(T)
>
> I'm unable to get the SAT_TRUNC pattern to appear on x86_64, however it
> does appear when building for riscv as seen below:
>
> Before this patch:
> <bb 3> [local count: 764504183]:
> # i_21 = PHI <i_14(8), 0(15)>
> # vectp_x.10_54 = PHI <vectp_x.10_55(8), x_10(D)(15)>
> # vectp_res.20_66 = PHI <vectp_res.20_67(8), res_11(D)(15)>
> # ivtmp_70 = PHI <ivtmp_71(8), _69(15)>
> _72 = .SELECT_VL (ivtmp_70, POLY_INT_CST [4, 4]);
> _1 = (long unsigned int) i_21;
> _2 = _1 * 4;
> _3 = x_10(D) + _2;
> ivtmp_53 = _72 * 4;
> vect__4.12_57 = .MASK_LEN_LOAD (vectp_x.10_54, 32B, { -1, ... }, _56(D),
> _72, 0);
> vect_x.13_58 = VIEW_CONVERT_EXPR<vector([4,4]) unsigned int>(vect__4.12_57);
> vect__38.15_60 = -vect_x.13_58;
> vect__15.16_61 = VIEW_CONVERT_EXPR<vector([4,4]) int>(vect__38.15_60);
> vect__16.17_62 = vect__15.16_61 >> 31;
> mask__29.14_59 = vect_x.13_58 > { 255, ... };
> vect__17.18_63 = VEC_COND_EXPR <mask__29.14_59, vect__16.17_62,
> vect__4.12_57>;
> vect__18.19_64 = (vector([4,4]) unsigned char) vect__17.18_63;
> _4 = *_3;
> _5 = res_11(D) + _1;
> x.0_12 = (unsigned int) _4;
> _38 = -x.0_12;
> _15 = (int) _38;
> _16 = _15 >> 31;
> _29 = x.0_12 > 255;
> _17 = _29 ? _16 : _4;
> _18 = (unsigned char) _17;
> .MASK_LEN_STORE (vectp_res.20_66, 8B, { -1, ... }, _72, 0, vect__18.19_64);
> i_14 = i_21 + 1;
> vectp_x.10_55 = vectp_x.10_54 + ivtmp_53;
> vectp_res.20_67 = vectp_res.20_66 + _72;
> ivtmp_71 = ivtmp_70 - _72;
> if (ivtmp_71 != 0)
> goto <bb 8>; [89.00%]
> else
> goto <bb 17>; [11.00%]
>
> After this patch:
>
> <bb 3> [local count: 764504183]:
> # i_21 = PHI <i_14(8), 0(15)>
> # vectp_x.10_68 = PHI <vectp_x.10_69(8), x_10(D)(15)>
> # vectp_res.15_75 = PHI <vectp_res.15_76(8), res_11(D)(15)>
> # ivtmp_79 = PHI <ivtmp_80(8), _78(15)>
> _81 = .SELECT_VL (ivtmp_79, POLY_INT_CST [4, 4]);
> _1 = (long unsigned int) i_21;
> _2 = _1 * 4;
> _3 = x_10(D) + _2;
> ivtmp_67 = _81 * 4;
> vect__4.12_71 = .MASK_LEN_LOAD (vectp_x.10_68, 32B, { -1, ... }, _70(D),
> _81, 0);
> vect_patt_37.13_72 = MAX_EXPR <{ 0, ... }, vect__4.12_71>;
> vect_patt_39.14_73 = .SAT_TRUNC (vect_patt_37.13_72);
> _4 = *_3;
> _5 = res_11(D) + _1;
> x.0_12 = (unsigned int) _4;
> _38 = -x.0_12;
> _15 = (int) _38;
> _16 = _15 >> 31;
> _29 = x.0_12 > 255;
> _17 = _29 ? _16 : _4;
> _18 = (unsigned char) _17;
> .MASK_LEN_STORE (vectp_res.15_75, 8B, { -1, ... }, _81, 0,
> vect_patt_39.14_73);
> i_14 = i_21 + 1;
> vectp_x.10_69 = vectp_x.10_68 + ivtmp_67;
> vectp_res.15_76 = vectp_res.15_75 + _81;
> ivtmp_80 = ivtmp_79 - _81;
> if (ivtmp_80 != 0)
> goto <bb 8>; [89.00%]
> else
> goto <bb 17>; [11.00%]
>
> gcc/ChangeLog:
>
> * match.pd: New NARROW_CLIP variant for SAT_TRUNC.
> * tree-vect-patterns.cc (gimple_unsigned_integer_narrow_clip):
> Add new decl for NARROW_CLIP.
> (vect_recog_sat_trunc_pattern): Add NARROW_CLIP check.
>
> Signed-off-by: Edwin Lu <e...@rivosinc.com>
> ---
> gcc/match.pd | 23 +++++++++++++++++++++++
> gcc/tree-vect-patterns.cc | 32 ++++++++++++++++++++++++++++++++
> 2 files changed, 55 insertions(+)
>
> diff --git a/gcc/match.pd b/gcc/match.pd
> index 82e6e291ae1..546e8fb3685 100644
> --- a/gcc/match.pd
> +++ b/gcc/match.pd
> @@ -3360,6 +3360,29 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> }
> (if (wi::eq_p (sum, wi::uhwi (0, precision))))))))
>
> +(if (INTEGRAL_TYPE_P (type) && TYPE_UNSIGNED (type))
> + (match (unsigned_integer_narrow_clip @0)
> + /* SAT_U_TRUNC = (UT)X & (NT)(-1) ? (-X) >> 31 : X
> +
> + The gimple representation uses X > (NT)(-1) instead of
> + using & so match on gt instead of bit_and. */
> + (convert (cond^ (gt (nop_convert? @0) INTEGER_CST@1)
> + (rshift:s (nop_convert? (negate (nop_convert? @0))) INTEGER_CST@2)
> + @0))
> + (if (! TYPE_UNSIGNED (TREE_TYPE (@0)))
> + (with
> + {
> + unsigned itype_precision = TYPE_PRECISION (TREE_TYPE (@0));
> + unsigned otype_precision = TYPE_PRECISION (type);
> + wide_int trunc_max = wi::mask (otype_precision, false, itype_precision);
> + wide_int int_cst_1 = wi::to_wide (@1, itype_precision);
> + wide_int int_cst_2 = wi::to_wide (@2, itype_precision);
> + wide_int shift_amount = wi::uhwi ((HOST_WIDE_INT_1U << 5) - 1,
> + itype_precision); // Aka 31
> + }
> + (if (otype_precision < itype_precision && wi::eq_p (trunc_max,
> + int_cst_1) && wi::eq_p(int_cst_2, shift_amount)))))))
> +
> /* Saturation truncate for unsigned integer. */
> (if (INTEGRAL_TYPE_P (type) && TYPE_UNSIGNED (type))
> (match (unsigned_integer_sat_trunc @0)
> diff --git a/gcc/tree-vect-patterns.cc b/gcc/tree-vect-patterns.cc
> index ffb320fbf23..a6588950be4 100644
> --- a/gcc/tree-vect-patterns.cc
> +++ b/gcc/tree-vect-patterns.cc
> @@ -4505,6 +4505,8 @@ extern bool gimple_unsigned_integer_sat_add (tree,
> tree*, tree (*)(tree));
> extern bool gimple_unsigned_integer_sat_sub (tree, tree*, tree (*)(tree));
> extern bool gimple_unsigned_integer_sat_trunc (tree, tree*, tree (*)(tree));
>
> +extern bool gimple_unsigned_integer_narrow_clip (tree, tree*, tree
> (*)(tree));
> +
> extern bool gimple_signed_integer_sat_add (tree, tree*, tree (*)(tree));
> extern bool gimple_signed_integer_sat_sub (tree, tree*, tree (*)(tree));
> extern bool gimple_signed_integer_sat_trunc (tree, tree*, tree (*)(tree));
> @@ -4739,6 +4741,36 @@ vect_recog_sat_trunc_pattern (vec_info *vinfo,
> stmt_vec_info stmt_vinfo,
> tree lhs = gimple_assign_lhs (last_stmt);
> tree otype = TREE_TYPE (lhs);
>
> + if ((gimple_unsigned_integer_narrow_clip (lhs, ops, NULL))
> + && type_has_mode_precision_p (otype))
> + {
> + tree itype = TREE_TYPE (ops[0]);
> + tree v_itype = get_vectype_for_scalar_type (vinfo, itype);
> + tree v_otype = get_vectype_for_scalar_type (vinfo, otype);
> + internal_fn fn = IFN_SAT_TRUNC;
> +
> + if (v_itype != NULL_TREE && v_otype != NULL_TREE
> + && direct_internal_fn_supported_p (fn, tree_pair (v_otype, v_itype),
> + OPTIMIZE_FOR_BOTH))
> + {
> + tree temp = vect_recog_temp_ssa_var (itype, NULL);
> + gimple * max_stmt = gimple_build_assign (temp, build2 (MAX_EXPR,
> itype, build_zero_cst(itype), ops[0]));
> + append_pattern_def_seq (vinfo, stmt_vinfo, max_stmt, v_itype);
> +
> + gcall *call = gimple_build_call_internal (fn, 1, temp);
> + tree out_ssa = vect_recog_temp_ssa_var (otype, NULL);
> +
> + gimple_call_set_lhs (call, out_ssa);
> + gimple_call_set_nothrow (call, /* nothrow_p */ false);
> + gimple_set_location (call, gimple_location (last_stmt));
> +
> + *type_out = v_otype;
> +
> + return call;
> + }
> +
> + }
> +
> if ((gimple_unsigned_integer_sat_trunc (lhs, ops, NULL)
> || gimple_signed_integer_sat_trunc (lhs, ops, NULL))
> && type_has_mode_precision_p (otype))
> --
> 2.43.0
>
