Re: [RFC PATCH v2 35/44] target/loongarch: Implement LSX fpu fcvt instructions

[Richard Henderson]

On 3/27/23 20:06, Song Gao wrote:
> This patch includes:
> - VFCVT{L/H}.{S.H/D.S};
> - VFCVT.{H.S/S.D};
> - VFRINT[{RNE/RZ/RP/RM}].{S/D};
> - VFTINT[{RNE/RZ/RP/RM}].{W.S/L.D};
> - VFTINT[RZ].{WU.S/LU.D};
> - VFTINT[{RNE/RZ/RP/RM}].W.D;
> - VFTINT[{RNE/RZ/RP/RM}]{L/H}.L.S;
> - VFFINT.{S.W/D.L}[U];
> - VFFINT.S.L, VFFINT{L/H}.D.W.
>
> Signed-off-by: Song Gao <gaosong@loongson.cn>
> ---
>   fpu/softfloat.c                             |  55 +++
>   include/fpu/softfloat.h                     |  27 ++
>   target/loongarch/disas.c                    |  56 +++
>   target/loongarch/helper.h                   |  56 +++
>   target/loongarch/insn_trans/trans_lsx.c.inc |  56 +++
>   target/loongarch/insns.decode               |  56 +++
>   target/loongarch/lsx_helper.c               | 369 ++++++++++++++++++++
>   7 files changed, 675 insertions(+)
>
> diff --git a/fpu/softfloat.c b/fpu/softfloat.c
> index c7454c3eb1..79975c6b01 100644
> --- a/fpu/softfloat.c
> +++ b/fpu/softfloat.c
> @@ -2988,6 +2988,25 @@ float64 float64_round_to_int(float64 a, float_status *s)
>       return float64_round_pack_canonical(&p, s);
>   }
>   
> +#define FRINT_RM(rm, rmode, bits)                             \
> +float ## bits float ## bits ## _round_to_int_ ## rm(          \
> +                         float ## bits a, float_status *s)    \
> +{                                                             \
> +    FloatParts64 pa;   \
> +    float ## bits ## _unpack_canonical(&pa, a, s); \
> +    parts_round_to_int(&pa, rmode, 0, s, &float64_params);    \
> +    return float ## bits ## _round_pack_canonical(&pa, s);    \
> +}
> +FRINT_RM(rne, float_round_nearest_even, 32)
> +FRINT_RM(rm,  float_round_down,         32)
> +FRINT_RM(rp,  float_round_up,           32)
> +FRINT_RM(rz,  float_round_to_zero,      32)
> +FRINT_RM(rne, float_round_nearest_even, 64)
> +FRINT_RM(rm,  float_round_down,         64)
> +FRINT_RM(rp,  float_round_up,           64)
> +FRINT_RM(rz,  float_round_to_zero,      64)
> +#undef FRINT_RM


No, you should simply swap your float_status rounding mode around the operation.
See the arm/tcg gen_set_rmode function.


r~