Re: [PATCH v2 11/14] target/arm: default SVE length to 64 bytes for linux-user

[Alex Bennée]

Richard Henderson <address@hidden> writes:

> On 12/5/19 9:31 AM, Alex Bennée wrote:
>> 
>> Richard Henderson <address@hidden> writes:
>> 
>>> On 11/30/19 8:45 AM, Alex Bennée wrote:
>>>> The Linux kernel chooses the default of 64 bytes for SVE registers on
>>>> the basis that it is the largest size that won't grow the signal
>>>> frame. When debugging larger sizes are also unwieldy in gdb as each
>>>> zreg will take over a page of terminal to display.
>>>>
>>>> The user can of course always specify a larger size with the
>>>> sve-max-vq property on the command line:
>>>>
>>>>   -cpu max,sve-max-vq=16
>>>>
>>>> This should not make any difference to SVE enabled software as the SVE
>>>> is of course vector length agnostic.
>>>>
>>>> Signed-off-by: Alex Bennée <address@hidden>
>>>> ---
>>>>  target/arm/cpu64.c | 3 +++
>>>>  1 file changed, 3 insertions(+)
>>>
>>> 6 is the largest size that doesn't grow the signal frame.
>>> I imagine 4 was chosen because that's the only real hw atm.
>>>
>>>> +        /* Default sve-max-vq to a reasonable numer */
>>>> +        cpu->sve_max_vq = 4;
>>>
>>> I also agree that we should match the kernel, but this is not the right way.
>>> Changing max vq is not the same as changing the default vq.
>>>
>>> You should change the value of env->vfp.zcr_el[1] in arm_cpu_reset(), and 
>>> the
>>> user can increase the length with prctl(2) as they would be able to on real
>>> hardware that would have support for longer vector lengths.
>> 
>> No the intention is to default to a lower max VQ because...
>> 
>>> Also, I don't think you should mix this up with gdb stuff.
>> 
>> it is what we use for sizing the registers for the gdbstub. The other
>> option would be to use the effective zcr_el1 value at the time of the
>> gdbstub connecting but then things will go horribly wrong if the user
>> execute a prctl and widens their size.
>
> Why would you care about the size of the registers as passed by default?  You
> shouldn't need or want to change that default to make gdbstub work.
>
> The gdbstub should be passing along the vq value (via the "vg" 
> pseudo-register,
> iirc), and gdb should be working out what to display based on that.
>
> If that isn't happening, and you are only changing the default so that gdb
> quits displaying massive registers when they aren't in use, then you're doing
> something wrong with gdb and gdbstub.

Currently the upstream gdbserver sends the XML based on the VL at start-up. It
doesn't handle changes in the vector size.

-- 
Alex Bennée