[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Qemu-stable] [PATCH v2 05/14] target/arm: Implement FMOV (general)
From: |
Peter Maydell |
Subject: |
Re: [Qemu-stable] [PATCH v2 05/14] target/arm: Implement FMOV (general) for fp16 |
Date: |
Thu, 10 May 2018 17:21:58 +0100 |
On 2 May 2018 at 23:15, Richard Henderson <address@hidden> wrote:
> Adding the fp16 moves to/from general registers.
>
> Cc: address@hidden
> Signed-off-by: Richard Henderson <address@hidden>
> ---
> target/arm/translate-a64.c | 22 +++++++++++++++++++++-
> 1 file changed, 21 insertions(+), 1 deletion(-)
>
> diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
> index c64c3ed99d..247a4f0cce 100644
> --- a/target/arm/translate-a64.c
> +++ b/target/arm/translate-a64.c
> @@ -5463,6 +5463,15 @@ static void handle_fmov(DisasContext *s, int rd, int
> rn, int type, bool itof)
> tcg_gen_st_i64(tcg_rn, cpu_env, fp_reg_hi_offset(s, rd));
> clear_vec_high(s, true, rd);
> break;
> + case 3:
> + /* 16 bit */
> + tmp = tcg_temp_new_i64();
> + tcg_gen_ext16u_i64(tmp, tcg_rn);
> + write_fp_dreg(s, rd, tmp);
> + tcg_temp_free_i64(tmp);
> + break;
> + default:
> + g_assert_not_reached();
> }
> } else {
> TCGv_i64 tcg_rd = cpu_reg(s, rd);
> @@ -5480,6 +5489,12 @@ static void handle_fmov(DisasContext *s, int rd, int
> rn, int type, bool itof)
> /* 64 bits from top half */
> tcg_gen_ld_i64(tcg_rd, cpu_env, fp_reg_hi_offset(s, rn));
> break;
> + case 3:
> + /* 16 bit */
> + tcg_gen_ld16u_i64(tcg_rd, cpu_env, fp_reg_offset(s, rn, MO_16));
> + break;
> + default:
> + g_assert_not_reached();
> }
> }
> }
> @@ -5519,10 +5534,15 @@ static void disas_fp_int_conv(DisasContext *s,
> uint32_t insn)
> case 0xa: /* 64 bit */
> case 0xd: /* 64 bit to top half of quad */
> break;
> + case 0x6: /* 16-bit */
> + if (arm_dc_feature(s, ARM_FEATURE_V8_FP16)) {
> + break;
> + }
> + /* fallthru */
This is a switch on (sf << 3 | type << 1 | rmode), so this catches
the halfprec <-> 32 bit cases, but it misses the halfprec <-> 64 bit
ops, which have sf == 1. So I think you need a 'case 0xe' as well.
> default:
> /* all other sf/type/rmode combinations are invalid */
> unallocated_encoding(s);
> - break;
> + return;
This is a distinct bugfix, I think (though the only effect would
be generation of unnecessary dead code after the undef).
> }
>
> if (!fp_access_check(s)) {
> --
> 2.14.3
thanks
-- PMM