Re: [PATCH v2 6/6] gdb/aarch64: Detect vector length changes when debugging remotely

[Thiago Jung Bauermann <thiago.bauermann@linaro.org>]

Simon Marchi <simark@simark.ca> writes:

> On 11/25/22 21:04, Thiago Jung Bauermann via Gdb-patches wrote:
>> If the remote target provides an expedited VG register, use it to update
>> the thread's gdbarch. This allows debugging programs that change their SVE
>> vector length during runtime.
>> 
>> This is accomplished by implementing the thread_architecture method in
>> remote_target, which returns the gdbarch corresponding to the expedited
>> registers provided by the last stop reply.
>> 
>> To allow changing the architecture based on the expedited registers, add a
>> new gdbarch method to allow arch-specific code to make the adjustment.
>
> I get this when applying, can you address that?
>
> Applying: gdb/aarch64: Detect vector length changes when debugging remotely
> .git/rebase-apply/patch:164: indent with spaces.
>                        "gdbarch_dump: update_architecture = <%s>\n",
> .git/rebase-apply/patch:165: indent with spaces.
>                        host_address_to_string (gdbarch->update_architecture));
> .git/rebase-apply/patch:186: indent with spaces.
>                                   gdbarch_update_architecture_ftype update_architecture)
> warning: 3 lines add whitespace errors.

This is because gdbarch.py generates these lines in gdbarch.c with
spaces-only indentation. I will send a separate patch to fix it.

> I think the idea makes sense.  Short of adding a packet for "get thread
> architecture" or extending the stop reply format to add a note about a
> thread having a special tdesc, I don't see another way to do this.  We

I guess we could define such a packet, but this approach is indeed
simpler.

> can't read the vq register using 'g', because to interpret the 'g'
> result, we need to know the full architecture.  We wouldn't know for
> sure the offset of vq in the reply.  By using expedited registers, we
> just need to make sure vq's register number is stable.  I guess that is
> the case?

The register number is determined by the function
aarch64_create_target_description. If SVE is supported, VG will always
have the register number 85. If not, then AFAIU the register number will
be unused (since the PAuth, MTE and TLS features only have a few
registers). But it could potentially be used by another feature in the
future...

Do we have to reserve the number 85 for the VG pseudo-register?

>> @@ -8068,6 +8078,21 @@ remote_target::process_stop_reply (struct stop_reply *stop_reply,
>>        /* Expedited registers.  */
>>        if (!stop_reply->regcache.empty ())
>>  	{
>> +	  /* If GDB already knows about this thread, we can give the
>> +	     architecture-specific code a chance to update the gdbarch based on
>> +	     the expedited registers.  */
>> +	  if (find_thread_ptid (this, ptid) != nullptr)
>> +	    {
>> +	      stop_reply->arch = gdbarch_update_architecture (stop_reply->arch,
>> +							      stop_reply->regcache);
>> +
>> +	      /* Save stop_reply->arch so that it can be returned by the
>> +		 thread_architecture method.  */
>> +	      remote_thread_info *remote_thr = get_remote_thread_info (this,
>> +								       ptid);
>> +	      remote_thr->expedited_arch = stop_reply->arch;
>> +	    }
>> +
>>  	  struct regcache *regcache
>>  	    = get_thread_arch_regcache (this, ptid, stop_reply->arch);
>
>
> What happens with threads that are not yet known?  Imagine the process
> starts with SVE == X, the main threads spawns a worker thread, the
> worker thread updates its SVE to Y, and the worker thread hits a
> breakpoint.  This is the first time we hear about that worker thread.

I tested this scenario and... it sort of works, but not quite. When the
unknown thread hits the breakpoint, gdbserver sends a T stop reply
packet with the new VG value in the expedited registers. It will be the
first time we hear about that worker thread, but at the same time we
will also learn about the value of its VG pseudo-register. So GDB
reports the breakpoint hit event and gets to the prompt as expected. You
can even print the new value of the VG pseudo-register, or any other
expedited register such as PC or SP.

It's only when you try to print the value of a non-expedited register
that you get an error indicating that GDB and gdbserver don't agree on
the size of the SVE registers (“Truncated register 51 in remote 'g'
packet”).

> If we don't do a "gdbarch_update_architecture" for it and save it
> somewhere, we'll never set the gdbarch with SVE == Y, will we?

We do that in remote_target::process_stop_reply, and I was under the
impression that it happened early enough for GDB to have the correct
gdbarch available when it needs it, but unfortunately that's not the
case.

I will add a testcase exercising this issue and see if I can find a
solution. Thank you for spotting it.

-- 
Thiago