Re: [PATCH 47/55] target/arm: Implement MVE VQDMLADH and VQRDMLADH

[Richard Henderson]

On 6/7/21 9:58 AM, Peter Maydell wrote:
> +static int32_t do_vqdmladh_w(int32_t a, int32_t b, int32_t c, int32_t d,
> +                             int round, bool *sat)
> +{
> +    int64_t m1 = (int64_t)a * b;
> +    int64_t m2 = (int64_t)c * d;
> +    int64_t r;
> +    /*
> +     * Architecturally we should do the entire add, double, round
> +     * and then check for saturation. We do three saturating adds,
> +     * but we need to be careful about the order. If the first
> +     * m1 + m2 saturates then it's impossible for the *2+rc to
> +     * bring it back into the non-saturated range. However, if
> +     * m1 + m2 is negative then it's possible that doing the doubling
> +     * would take the intermediate result below INT64_MAX and the
> +     * addition of the rounding constant then brings it back in range.
> +     * So we add half the rounding constant before doubling rather
> +     * than adding the rounding constant after the doubling.
> +     */
> +    if (sadd64_overflow(m1, m2, &r) ||
> +        sadd64_overflow(r, (round << 30), &r) ||
> +        sadd64_overflow(r, r, &r)) {

Ooh, ahh, an operation that doesn't even exist in SVE2.
Nice use of the new interface, btw.

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

r~