Re: [PATCHv2] gdb/x86: handle stap probe arguments in xmm registers

[Andrew Burgess <aburgess@redhat.com>]

Pedro, Tom,

Thanks for your feedback on this patch.  I've gone ahead and pushed
this.  For now I don't plan to priorities working on agent expression
support for this change, but I'm happy to do so if this becomes
important.

Thanks,
Andrew



Pedro Alves <pedro@palves.net> writes:

> On 2022-03-16 17:23, Tom Tromey wrote:
>> Andrew> This is because GDB doesn't currently support placing non-scalar types
>> Andrew> on the agent expression evaluation stack.  Solving this is clearly
>> Andrew> related to the original problem, but feels a bit like a second
>> Andrew> problem.  I'd like to get feedback on whether my approach to solving
>> Andrew> the original problem is acceptable or not before I start looking at
>> Andrew> how to handle xmm registers within agent expressions.
>> 
>> Note that there are many things that can't be represented in agent
>> expressions.  I recall filing a bug report about this -- there are some
>> DWARF expressions that can't be translated, and IIRC, floating point
>> isn't handled at all.  So, I wouldn't worry too much about this.  My
>> sense is that tracepoints aren't used a whole lot.
>
> Yeah.  They were definitely being used in production a few years back when we were
> revamping them, adding support to gdbserver, etc.  But it seems like users stopped using
> them since.  It's possible that the fact that more and more DWARF expressions
> can't be converted carries some weight.  Or users simply started using other
> tools for the job.
>
> In any case, if I were to tackle the whole problem set today, instead of extending agent 
> expressions bytecode to be able to do more of what DWARF can do, I'd be looking at expressing
> what should be collected via DWARF expressions/locations.  I mean, send DWARF expressions
> to the target side to be evaluated.