Re: [PATCH v1 03/20] s390x/tcg: Implement VECTOR MULTIPLY SUM LOGICAL

[Richard Henderson]

On 9/30/20 9:55 AM, David Hildenbrand wrote:
> +    /* Multipy both even elements from v2 and v3 */
> +    read_vec_element_i64(l1, get_field(s, v2), 0, ES_64);
> +    read_vec_element_i64(h1, get_field(s, v3), 0, ES_64);
> +    tcg_gen_mulu2_i64(l1, h1, l1, h1);
> +    /* Shift result left by one bit if requested */
> +    if (extract32(get_field(s, m6), 3, 1)) {
> +        tcg_gen_extract2_i64(h1, l1, h1, 63);
> +        tcg_gen_shli_i64(l1, l1, 1);
> +    }

Not a bug, but some hosts require 3 insns for extract2 (so 4 total for this
sequence).

This doubling can also be had via add2:

    tcg_gen_add2_i64(l1, h1, l1, h1, l1, h1);

At which point most hosts will require only 2 insns for this sequence.  The two
hosts that don't have a carry bit (mips, riscv), will still be able to perform
the add in 3 insns.

So add is never more expensive and sometimes half as expensive.

Regardless,
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>


r~