Re: [Qemu-devel] [PATCH 6/9] target/arm: Implement FCVT (scalar, integer) for fp16

["Alex Bennée" <alex.bennee@linaro.org>]

Richard Henderson <richard.henderson@linaro.org> writes:

> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>

Reviewed-by: Alex Bennée <alex.bennee@linaro.org>

> ---
>  target/arm/helper.h        |  6 +++
>  target/arm/helper.c        | 38 +++++++++++++++++-
>  target/arm/translate-a64.c | 96 ++++++++++++++++++++++++++++++++++++++--------
>  3 files changed, 122 insertions(+), 18 deletions(-)
>
> diff --git a/target/arm/helper.h b/target/arm/helper.h
> index eafd5d746b..f494b10f1b 100644
> --- a/target/arm/helper.h
> +++ b/target/arm/helper.h
> @@ -151,6 +151,10 @@ DEF_HELPER_3(vfp_touhd_round_to_zero, i64, f64, i32, ptr)
>  DEF_HELPER_3(vfp_tould_round_to_zero, i64, f64, i32, ptr)
>  DEF_HELPER_3(vfp_touhh, i32, f16, i32, ptr)
>  DEF_HELPER_3(vfp_toshh, i32, f16, i32, ptr)
> +DEF_HELPER_3(vfp_toulh, i32, f16, i32, ptr)
> +DEF_HELPER_3(vfp_toslh, i32, f16, i32, ptr)
> +DEF_HELPER_3(vfp_touqh, i64, f16, i32, ptr)
> +DEF_HELPER_3(vfp_tosqh, i64, f16, i32, ptr)
>  DEF_HELPER_3(vfp_toshs, i32, f32, i32, ptr)
>  DEF_HELPER_3(vfp_tosls, i32, f32, i32, ptr)
>  DEF_HELPER_3(vfp_tosqs, i64, f32, i32, ptr)
> @@ -177,6 +181,8 @@ DEF_HELPER_3(vfp_ultod, f64, i64, i32, ptr)
>  DEF_HELPER_3(vfp_uqtod, f64, i64, i32, ptr)
>  DEF_HELPER_3(vfp_sltoh, f16, i32, i32, ptr)
>  DEF_HELPER_3(vfp_ultoh, f16, i32, i32, ptr)
> +DEF_HELPER_3(vfp_sqtoh, f16, i64, i32, ptr)
> +DEF_HELPER_3(vfp_uqtoh, f16, i64, i32, ptr)
>
>  DEF_HELPER_FLAGS_2(set_rmode, TCG_CALL_NO_RWG, i32, i32, ptr)
>  DEF_HELPER_FLAGS_2(set_neon_rmode, TCG_CALL_NO_RWG, i32, i32, env)
> diff --git a/target/arm/helper.c b/target/arm/helper.c
> index 743f34bd0a..dbc10b454a 100644
> --- a/target/arm/helper.c
> +++ b/target/arm/helper.c
> @@ -11416,8 +11416,12 @@ VFP_CONV_FIX_A64(uq, s, 32, 64, uint64)
>  #undef VFP_CONV_FIX_A64
>
>  /* Conversion to/from f16 can overflow to infinity before/after scaling.
> - * Therefore we convert to f64 (which does not round), scale,
> - * and then convert f64 to f16 (which may round).
> + * Therefore we convert to f64, scale, and then convert f64 to f16; or
> + * vice versa for conversion to integer.
> + *
> + * For 16- and 32-bit integers, the conversion to f64 never rounds.
> + * For 64-bit integers, any integer that would cause rounding will also
> + * overflow to f16 infinity, so there is no double rounding problem.
>   */
>
>  static float16 do_postscale_fp16(float64 f, int shift, float_status *fpst)
> @@ -11435,6 +11439,16 @@ float16 HELPER(vfp_ultoh)(uint32_t x, uint32_t shift, void *fpst)
>      return do_postscale_fp16(uint32_to_float64(x, fpst), shift, fpst);
>  }
>
> +float16 HELPER(vfp_sqtoh)(uint64_t x, uint32_t shift, void *fpst)
> +{
> +    return do_postscale_fp16(int64_to_float64(x, fpst), shift, fpst);
> +}
> +
> +float16 HELPER(vfp_uqtoh)(uint64_t x, uint32_t shift, void *fpst)
> +{
> +    return do_postscale_fp16(uint64_to_float64(x, fpst), shift, fpst);
> +}
> +
>  static float64 do_prescale_fp16(float16 f, int shift, float_status *fpst)
>  {
>      if (unlikely(float16_is_any_nan(f))) {
> @@ -11464,6 +11478,26 @@ uint32_t HELPER(vfp_touhh)(float16 x, uint32_t shift, void *fpst)
>      return float64_to_uint16(do_prescale_fp16(x, shift, fpst), fpst);
>  }
>
> +uint32_t HELPER(vfp_toslh)(float16 x, uint32_t shift, void *fpst)
> +{
> +    return float64_to_int32(do_prescale_fp16(x, shift, fpst), fpst);
> +}
> +
> +uint32_t HELPER(vfp_toulh)(float16 x, uint32_t shift, void *fpst)
> +{
> +    return float64_to_uint32(do_prescale_fp16(x, shift, fpst), fpst);
> +}
> +
> +uint64_t HELPER(vfp_tosqh)(float16 x, uint32_t shift, void *fpst)
> +{
> +    return float64_to_int64(do_prescale_fp16(x, shift, fpst), fpst);
> +}
> +
> +uint64_t HELPER(vfp_touqh)(float16 x, uint32_t shift, void *fpst)
> +{
> +    return float64_to_uint64(do_prescale_fp16(x, shift, fpst), fpst);
> +}
> +
>  /* Set the current fp rounding mode and return the old one.
>   * The argument is a softfloat float_round_ value.
>   */
> diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
> index 36bb5f6f08..4f6317aa0f 100644
> --- a/target/arm/translate-a64.c
> +++ b/target/arm/translate-a64.c
> @@ -5186,11 +5186,11 @@ static void handle_fpfpcvt(DisasContext *s, int rd, int rn, int opcode,
>                             bool itof, int rmode, int scale, int sf, int type)
>  {
>      bool is_signed = !(opcode & 1);
> -    bool is_double = type;
>      TCGv_ptr tcg_fpstatus;
> -    TCGv_i32 tcg_shift;
> +    TCGv_i32 tcg_shift, tcg_single;
> +    TCGv_i64 tcg_double;
>
> -    tcg_fpstatus = get_fpstatus_ptr(false);
> +    tcg_fpstatus = get_fpstatus_ptr(type == 3);
>
>      tcg_shift = tcg_const_i32(64 - scale);
>
> @@ -5208,8 +5208,9 @@ static void handle_fpfpcvt(DisasContext *s, int rd, int rn, int opcode,
>              tcg_int = tcg_extend;
>          }
>
> -        if (is_double) {
> -            TCGv_i64 tcg_double = tcg_temp_new_i64();
> +        switch (type) {
> +        case 1: /* float64 */
> +            tcg_double = tcg_temp_new_i64();
>              if (is_signed) {
>                  gen_helper_vfp_sqtod(tcg_double, tcg_int,
>                                       tcg_shift, tcg_fpstatus);
> @@ -5219,8 +5220,10 @@ static void handle_fpfpcvt(DisasContext *s, int rd, int rn, int opcode,
>              }
>              write_fp_dreg(s, rd, tcg_double);
>              tcg_temp_free_i64(tcg_double);
> -        } else {
> -            TCGv_i32 tcg_single = tcg_temp_new_i32();
> +            break;
> +
> +        case 0: /* float32 */
> +            tcg_single = tcg_temp_new_i32();
>              if (is_signed) {
>                  gen_helper_vfp_sqtos(tcg_single, tcg_int,
>                                       tcg_shift, tcg_fpstatus);
> @@ -5230,6 +5233,23 @@ static void handle_fpfpcvt(DisasContext *s, int rd, int rn, int opcode,
>              }
>              write_fp_sreg(s, rd, tcg_single);
>              tcg_temp_free_i32(tcg_single);
> +            break;
> +
> +        case 3: /* float16 */
> +            tcg_single = tcg_temp_new_i32();
> +            if (is_signed) {
> +                gen_helper_vfp_sqtoh(tcg_single, tcg_int,
> +                                     tcg_shift, tcg_fpstatus);
> +            } else {
> +                gen_helper_vfp_uqtoh(tcg_single, tcg_int,
> +                                     tcg_shift, tcg_fpstatus);
> +            }
> +            write_fp_sreg(s, rd, tcg_single);
> +            tcg_temp_free_i32(tcg_single);
> +            break;
> +
> +        default:
> +            g_assert_not_reached();
>          }
>      } else {
>          TCGv_i64 tcg_int = cpu_reg(s, rd);
> @@ -5246,8 +5266,9 @@ static void handle_fpfpcvt(DisasContext *s, int rd, int rn, int opcode,
>
>          gen_helper_set_rmode(tcg_rmode, tcg_rmode, tcg_fpstatus);
>
> -        if (is_double) {
> -            TCGv_i64 tcg_double = read_fp_dreg(s, rn);
> +        switch (type) {
> +        case 1: /* float64 */
> +            tcg_double = read_fp_dreg(s, rn);
>              if (is_signed) {
>                  if (!sf) {
>                      gen_helper_vfp_tosld(tcg_int, tcg_double,
> @@ -5265,9 +5286,14 @@ static void handle_fpfpcvt(DisasContext *s, int rd, int rn, int opcode,
>                                           tcg_shift, tcg_fpstatus);
>                  }
>              }
> +            if (!sf) {
> +                tcg_gen_ext32u_i64(tcg_int, tcg_int);
> +            }
>              tcg_temp_free_i64(tcg_double);
> -        } else {
> -            TCGv_i32 tcg_single = read_fp_sreg(s, rn);
> +            break;
> +
> +        case 0: /* float32 */
> +            tcg_single = read_fp_sreg(s, rn);
>              if (sf) {
>                  if (is_signed) {
>                      gen_helper_vfp_tosqs(tcg_int, tcg_single,
> @@ -5289,14 +5315,39 @@ static void handle_fpfpcvt(DisasContext *s, int rd, int rn, int opcode,
>                  tcg_temp_free_i32(tcg_dest);
>              }
>              tcg_temp_free_i32(tcg_single);
> +            break;
> +
> +        case 3: /* float16 */
> +            tcg_single = read_fp_sreg(s, rn);
> +            if (sf) {
> +                if (is_signed) {
> +                    gen_helper_vfp_tosqh(tcg_int, tcg_single,
> +                                         tcg_shift, tcg_fpstatus);
> +                } else {
> +                    gen_helper_vfp_touqh(tcg_int, tcg_single,
> +                                         tcg_shift, tcg_fpstatus);
> +                }
> +            } else {
> +                TCGv_i32 tcg_dest = tcg_temp_new_i32();
> +                if (is_signed) {
> +                    gen_helper_vfp_toslh(tcg_dest, tcg_single,
> +                                         tcg_shift, tcg_fpstatus);
> +                } else {
> +                    gen_helper_vfp_toulh(tcg_dest, tcg_single,
> +                                         tcg_shift, tcg_fpstatus);
> +                }
> +                tcg_gen_extu_i32_i64(tcg_int, tcg_dest);
> +                tcg_temp_free_i32(tcg_dest);
> +            }
> +            tcg_temp_free_i32(tcg_single);
> +            break;
> +
> +        default:
> +            g_assert_not_reached();
>          }
>
>          gen_helper_set_rmode(tcg_rmode, tcg_rmode, tcg_fpstatus);
>          tcg_temp_free_i32(tcg_rmode);
> -
> -        if (!sf) {
> -            tcg_gen_ext32u_i64(tcg_int, tcg_int);
> -        }
>      }
>
>      tcg_temp_free_ptr(tcg_fpstatus);
> @@ -5465,7 +5516,20 @@ static void disas_fp_int_conv(DisasContext *s, uint32_t insn)
>          /* actual FP conversions */
>          bool itof = extract32(opcode, 1, 1);
>
> -        if (type > 1 || (rmode != 0 && opcode > 1)) {
> +        if (rmode != 0 && opcode > 1) {
> +            unallocated_encoding(s);
> +            return;
> +        }
> +        switch (type) {
> +        case 0: /* float32 */
> +        case 1: /* float64 */
> +            break;
> +        case 3: /* float16 */
> +            if (arm_dc_feature(s, ARM_FEATURE_V8_FP16)) {
> +                break;
> +            }
> +            /* fallthru */
> +        default:
>              unallocated_encoding(s);
>              return;
>          }


--
Alex Bennée