Re: [PATCH v2 2/3] target/ppc: Implement Vector Extract Mask

[Richard Henderson]

On 11/12/21 6:14 AM, matheus.ferst@eldorado.org.br wrote:
> From: Matheus Ferst <matheus.ferst@eldorado.org.br>
>
> Implement the following PowerISA v3.1 instructions:
> vextractbm: Vector Extract Byte Mask
> vextracthm: Vector Extract Halfword Mask
> vextractwm: Vector Extract Word Mask
> vextractdm: Vector Extract Doubleword Mask
> vextractqm: Vector Extract Quadword Mask
>
> Suggested-by: Richard Henderson <richard.henderson@linaro.org>
> Signed-off-by: Matheus Ferst <matheus.ferst@eldorado.org.br>
> ---
> v2:
> - Applied rth suggestion to do_vextractm
> ---
>   target/ppc/insn32.decode            |  6 +++
>   target/ppc/translate/vmx-impl.c.inc | 60 +++++++++++++++++++++++++++++
>   2 files changed, 66 insertions(+)
>
> diff --git a/target/ppc/insn32.decode b/target/ppc/insn32.decode
> index 9a28f1d266..639ac22bf0 100644
> --- a/target/ppc/insn32.decode
> +++ b/target/ppc/insn32.decode
> @@ -419,6 +419,12 @@ VEXPANDWM       000100 ..... 00010 ..... 11001000010    
> @VX_tb
>   VEXPANDDM       000100 ..... 00011 ..... 11001000010    @VX_tb
>   VEXPANDQM       000100 ..... 00100 ..... 11001000010    @VX_tb
>   
> +VEXTRACTBM      000100 ..... 01000 ..... 11001000010    @VX_tb
> +VEXTRACTHM      000100 ..... 01001 ..... 11001000010    @VX_tb
> +VEXTRACTWM      000100 ..... 01010 ..... 11001000010    @VX_tb
> +VEXTRACTDM      000100 ..... 01011 ..... 11001000010    @VX_tb
> +VEXTRACTQM      000100 ..... 01100 ..... 11001000010    @VX_tb
> +
>   # VSX Load/Store Instructions
>   
>   LXV             111101 ..... ..... ............ . 001   @DQ_TSX
> diff --git a/target/ppc/translate/vmx-impl.c.inc 
> b/target/ppc/translate/vmx-impl.c.inc
> index 58aca58f0f..dd7337c2f2 100644
> --- a/target/ppc/translate/vmx-impl.c.inc
> +++ b/target/ppc/translate/vmx-impl.c.inc
> @@ -1539,6 +1539,66 @@ static bool trans_VEXPANDQM(DisasContext *ctx, arg_VX_tb 
> *a)
>       return true;
>   }
>   
> +static bool do_vextractm(DisasContext *ctx, arg_VX_tb *a, unsigned vece)
> +{
> +    const uint64_t elem_width = 8 << vece, elem_count_half = 8 >> vece;
> +    TCGv_i64 t, b, tmp;
> +
> +    REQUIRE_INSNS_FLAGS2(ctx, ISA310);
> +    REQUIRE_VECTOR(ctx);
> +
> +    t = tcg_const_i64(0);
> +    b = tcg_temp_new_i64();
> +    tmp = tcg_temp_new_i64();
> +
> +    for (int w = 0; w < 2; w++) {
> +        get_avr64(b, a->vrb, w);
> +
> +        for (int i = 0; i < elem_count_half; i++) {
> +            int in_bit = (i + 1) * elem_width - 1;
> +            int out_bit = w * elem_count_half + i;
> +
> +            if (in_bit > out_bit) {
> +                tcg_gen_shri_i64(tmp, b, in_bit - out_bit);
> +            } else {
> +                tcg_gen_shli_i64(tmp, b, out_bit - in_bit);
> +            }
> +            tcg_gen_andi_i64(tmp, tmp, 1 << out_bit);
> +            tcg_gen_or_i64(t, t, tmp);
> +        }
> +    }
> +    tcg_gen_trunc_i64_tl(cpu_gpr[a->vrt], t);

Pardon me.  I realized after the fact that we can run the same algorithm as for mtvsrm (in 
the next patch) in reverse.

  & dup(1)
.......a.......b.......c.......d.......e.......f.......g.......h
  >> 32 - 4
...................................a.......b.......c.......d....
  |
.......a.......b.......c.......d...a...e...b...f...c...g...d...h
  >> 16 - 2
.....................a.......b.......c.......d...a...e...b...f..
  |
.......a.......b.....a.c.....b.d...a.c.e...b.d.f.a.c.e.g.b.d.f.h
  >> 8 - 1
..............a.......b.....a.c.....b.d...a.c.e...b.d.f.a.c.e.g.
  |
.......a......ab.....abc....abcd...abcde..abcdef.abcdefgabcdefgh
  & 0xff
........................................................abcdefgh

where one of the two final masks can be done via deposit:

    tcg_gen_andi_i64(hi, hi, 0xff);
    tcg_gen_deposit_i64(lo, lo, hi, 8, 56);

Which will reduce the instruction count of this implementation by half.


r~