Re: [qemu-s390x] [Qemu-devel] [PATCH v1 06/33] s390x/tcg: Implement VECTOR GENERATE BYTE MASK

[Richard Henderson]

On 2/26/19 3:38 AM, David Hildenbrand wrote:
> +static DisasJumpType op_vgbm(DisasContext *s, DisasOps *o)
> +{
> +    const uint16_t i2 = get_field(s->fields, i2);
> +    TCGv_i32 ones = tcg_const_i32(-1u);
> +    TCGv_i32 zeroes = tcg_const_i32(0);
> +    int i;
> +
> +    for (i = 0; i < 16; i++) {
> +        if (extract32(i2, 15 - i, 1)) {
> +            write_vec_element_i32(ones, get_field(s->fields, v1), i, MO_8);
> +        } else {
> +            write_vec_element_i32(zeroes, get_field(s->fields, v1), i, MO_8);
> +        }
> +    }
> +    tcg_temp_free_i32(ones);
> +    tcg_temp_free_i32(zeroes);
> +    return DISAS_NEXT;
> +}

While this works, it's not in the spirit of

> Programming Note: VECTOR GENERATE BYTE
> MASK is the preferred method for setting a vector
> register to all zeroes or ones.

Better, I think, with

uint64_t generate_byte_mask(uint8_t mask)
{
    uint64_t r = 0;
    int i;
    for (i = 0; i < 8; i++) {
        if ((mask >> i) & 1) {
            r |= 0xffull << (i * 8);
        }
    }
    return r;
}

    if (i2 == (i2 & 0xff) * 0x0101) {
        /* masks for both halves of the vector are the same.
           trust tcg to produce a good constant loading.  */
        tcg_gen_gvec_dup64i(vec_full_reg_offset(s, v1), 16, 16,
                            generate_byte_mask(i2 & 0xff));
    } else {
        TCGv_i64 t = tcg_temp_new_i64();
        tcg_gen_movi_i64(t, generate_byte_mask(i2 >> 8));
        write_vec_element_i64(t, v1, 0, MO_64);
        tcg_gen_movi_i64(t, generate_byte_mask(i2 & 0xff));
        write_vec_element_i64(t, v1, 1, MO_64);
        tcg_temp_free_i64();
    }

Somewhere behind tcg_gen_gvec_dup64i, I check to see if the constant can be
decomposed further, which will eventually bottom out at

        vpxor   %xmm0,%xmm0,%xmm0               // all zeros
        vpcmpeq %xmm0,%xmm0,%xmm0               // all ones

and even more interesting combinations for tcg/aarch64.



r~