Re: [PATCH v3 09/37] target/ppc: Move Vector Compare Equal/Not Equal/Greater Than to decodetree

[Richard Henderson]

On 2/10/22 23:34, matheus.ferst@eldorado.org.br wrote:
> +static void do_vcmp_rc(int vrt)
> +{
> +    TCGv_i64 t0, t1;
> +
> +    t0 = tcg_temp_new_i64();
> +    t1 = tcg_temp_new_i64();
> +
> +    get_avr64(t0, vrt, true);
> +    tcg_gen_ctpop_i64(t1, t0);
> +    get_avr64(t0, vrt, false);
> +    tcg_gen_ctpop_i64(t0, t0);
> +    tcg_gen_add_i64(t1, t0, t1);

I don't understand the ctpop here.  I would have expected:

    tcg_gen_and_i64(set, t0, t1);
    tcg_gen_or_i64(clr, t0, t1);
    tcg_gen_setcondi_i64(TCG_COND_EQ, set, set, -1); /* all bits set */
    tcg_gen_setcondi_i64(TCG_COND_EQ, clr, clr, 0);  /* all bits clear */


> +static bool do_vcmp(DisasContext *ctx, arg_VC *a, TCGCond cond, int vece)
> +{
> +    REQUIRE_VECTOR(ctx);
> +
> +    tcg_gen_gvec_cmp(cond, vece, avr_full_offset(a->vrt),
> +                     avr_full_offset(a->vra), avr_full_offset(a->vrb), 16, 16);
> +    tcg_gen_gvec_shli(vece, avr_full_offset(a->vrt), avr_full_offset(a->vrt),
> +                      (8 << vece) - 1, 16, 16);
> +    tcg_gen_gvec_sari(vece, avr_full_offset(a->vrt), avr_full_offset(a->vrt),
> +                      (8 << vece) - 1, 16, 16);

Vector compare already produces -1; no need for anything beyond the cmp.


r~