[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [RFC v2 41/76] target/riscv: rvv-0.9: floating-point scalar move ins
From: |
Richard Henderson |
Subject: |
Re: [RFC v2 41/76] target/riscv: rvv-0.9: floating-point scalar move instructions |
Date: |
Thu, 30 Jul 2020 13:03:16 -0700 |
User-agent: |
Mozilla/5.0 (X11; Linux x86_64; rv:68.0) Gecko/20100101 Thunderbird/68.10.0 |
On 7/22/20 2:16 AM, frank.chang@sifive.com wrote:
> From: Frank Chang <frank.chang@sifive.com>
>
> NaN-boxed the scalar floating-point register based on RVV 0.9's rules.
>
> Signed-off-by: Frank Chang <frank.chang@sifive.com>
> ---
> target/riscv/insn32.decode | 4 +--
> target/riscv/insn_trans/trans_rvv.inc.c | 45 ++++++++++++++++---------
> 2 files changed, 31 insertions(+), 18 deletions(-)
>
> diff --git a/target/riscv/insn32.decode b/target/riscv/insn32.decode
> index ef53df7c73..4be1b88e2d 100644
> --- a/target/riscv/insn32.decode
> +++ b/target/riscv/insn32.decode
> @@ -583,8 +583,8 @@ vid_v 010100 . 00000 10001 010 ..... 1010111
> @r1_vm
> vmv_x_s 010000 1 ..... 00000 010 ..... 1010111 @r2rd
> vmv_s_x 010000 1 00000 ..... 110 ..... 1010111 @r2
> vext_x_v 001100 1 ..... ..... 010 ..... 1010111 @r
> -vfmv_f_s 001100 1 ..... 00000 001 ..... 1010111 @r2rd
> -vfmv_s_f 001101 1 00000 ..... 101 ..... 1010111 @r2
> +vfmv_f_s 010000 1 ..... 00000 001 ..... 1010111 @r2rd
> +vfmv_s_f 010000 1 00000 ..... 101 ..... 1010111 @r2
> vslideup_vx 001110 . ..... ..... 100 ..... 1010111 @r_vm
> vslideup_vi 001110 . ..... ..... 011 ..... 1010111 @r_vm
> vslide1up_vx 001110 . ..... ..... 110 ..... 1010111 @r_vm
> diff --git a/target/riscv/insn_trans/trans_rvv.inc.c
> b/target/riscv/insn_trans/trans_rvv.inc.c
> index 54c08ea1f8..56cd7444f2 100644
> --- a/target/riscv/insn_trans/trans_rvv.inc.c
> +++ b/target/riscv/insn_trans/trans_rvv.inc.c
> @@ -3326,14 +3326,22 @@ static bool trans_vmv_s_x(DisasContext *s,
> arg_vmv_s_x *a)
> /* Floating-Point Scalar Move Instructions */
> static bool trans_vfmv_f_s(DisasContext *s, arg_vfmv_f_s *a)
> {
> - if (!s->vill && has_ext(s, RVF) &&
> - (s->mstatus_fs != 0) && (s->sew != 0)) {
> - unsigned int len = 8 << s->sew;
> -
> - vec_element_loadi(s, cpu_fpr[a->rd], a->rs2, 0);
> - if (len < 64) {
> - tcg_gen_ori_i64(cpu_fpr[a->rd], cpu_fpr[a->rd],
> - MAKE_64BIT_MASK(len, 64 - len));
> + if (require_rvv(s) &&
> + vext_check_isa_ill(s) &&
> + has_ext(s, RVF) &&
> + (s->mstatus_fs != 0) &&
> + (s->sew != 0)) {
> + unsigned int ofs = (8 << s->sew);
> + unsigned int len = 64 - ofs;
> + TCGv_i64 t_nan;
> +
> + vec_element_loadi(s, cpu_fpr[a->rd], a->rs2, 0, false);
> + /* NaN-box f[rd] as necessary for SEW */
> + if (len) {
> + t_nan = tcg_const_i64(UINT64_MAX);
> + tcg_gen_deposit_i64(cpu_fpr[a->rd], cpu_fpr[a->rd],
> + t_nan, ofs, len);
> + tcg_temp_free_i64(t_nan);
> }
It would be better to use the gen_nanbox_[sh] functions here.
> t1 = tcg_temp_new_i64();
> if (s->sew == MO_64 && !has_ext(s, RVD)) {
> + /* SEW > FLEN, f[rs1] is NaN-boxed to SEW bits */
> tcg_gen_ori_i64(t1, cpu_fpr[a->rs1], MAKE_64BIT_MASK(32, 32));
With my nanbox patch set, RVF-only cpus will *always* have values nanboxed to
64-bits. So there should be no need to special case this, only provide a
comment noting that RVF and RVD can be treated equally.
> + } else if ((s->sew < MO_64 && has_ext(s, RVD)) ||
> + (s->sew < MO_32)) {
> + /* SEW < FLEN */
> + TCGv_i32 sew = tcg_const_i32(1 << (s->sew + 3));
> + gen_helper_narrower_nanbox_fpr(t1, cpu_fpr[a->rs1], sew,
> cpu_env);
Use gen_check_nanbox_[sh].
r~
- Re: [RFC v2 36/76] target/riscv: rvv-0.9: element index instruction, (continued)
- [RFC v2 37/76] target/riscv: rvv-0.9: allow load element with sign-extended, frank . chang, 2020/07/22
- [RFC v2 38/76] target/riscv: rvv-0.9: register gather instructions, frank . chang, 2020/07/22
- [RFC v2 39/76] target/riscv: rvv-0.9: integer scalar move instructions, frank . chang, 2020/07/22
- [RFC v2 40/76] target/riscv: rvv-0.9: floating-point move instruction, frank . chang, 2020/07/22
- [RFC v2 41/76] target/riscv: rvv-0.9: floating-point scalar move instructions, frank . chang, 2020/07/22
- Re: [RFC v2 41/76] target/riscv: rvv-0.9: floating-point scalar move instructions,
Richard Henderson <=
- [RFC v2 42/76] target/riscv: rvv-0.9: whole register move instructions, frank . chang, 2020/07/22
- [RFC v2 43/76] target/riscv: rvv-0.9: integer extension instructions, frank . chang, 2020/07/22
- [RFC v2 44/76] target/riscv: rvv-0.9: single-width averaging add and subtract instructions, frank . chang, 2020/07/22
- [RFC v2 45/76] target/riscv: rvv-0.9: single-width bit shift instructions, frank . chang, 2020/07/22