Re: [Qemu-devel] [PATCH 06/10] target-arm: A64: Add SIMD ZIP/UZP/TRN

[Alex Bennée]

address@hidden writes:

> On 01/10/2014 09:12 AM, Peter Maydell wrote:
>> +    for (i = 0; i < elements; i++) {
>> +        switch (opcode) {
>> +        case 1: /* UZP1/2 */
>> +        {
>> +            int midpoint = elements / 2;
>> +            if (i < midpoint) {
>> +                read_vec_element(s, tcg_res, rn, 2 * i + part, size);
>> +            } else {
>> +                read_vec_element(s, tcg_res, rm,
>> +                                 2 * (i - midpoint) + part, size);
>> +            }
>> +            break;
>> +        }
>
> You're generating up to 16 * 3 + 2 = 50 opcodes here.  I do wonder if it
> wouldn't be better to implement these as helpers.  But,

What's the hit in terms of opcodes for calling a helper function? I
would have thought you spend 10s of opcodes with messing around with
stack frames and the like before you get to the helper.

I was wondering if there is a demand for more SIMD like TCG opcodes but
I suspect all the arches are subtly different enough to make it a pain
to abstract.

>
> Reviewed-by: Richard Henderson <address@hidden>
>
>
> r~

-- 
Sent with my mu4e