test

        const limb_t* __restrict__ bp = b.data.get();

        const limb_t* __restrict__ mp = mod.data.get();

        const size_t au = a.used;

        const size_t bu = b.used;

        for (size_t i = 0; i < n; ++i) {

            // Step 1: T += a[i] * b

            const dlimb_t ai = (i < au) ? (dlimb_t)ap[i] : 0;

            dlimb_t carry = 0;

            for (size_t j = 0; j < n; ++j) {

                dlimb_t cur = (dlimb_t)T[j] + ai * (dlimb_t)bp[j] + carry;

                const dlimb_t bj = (j < bu) ? (dlimb_t)bp[j] : 0;

                dlimb_t cur = (dlimb_t)T[j] + ai * bj + carry;

                T[j] = (limb_t)cur;

                carry = cur >> LIMB_BITS;

            }