Re: [PATCH] gdb: error out if architecture does not implement any "return_value" hook

[Simon Marchi <simark@simark.ca>]

On 2/28/23 15:20, Pedro Alves wrote:
> On 2023-02-28 2:50 p.m., Andrew Burgess via Gdb-patches wrote:
> 
>> Or maybe we shouldn't be throwing an error in some of these cases, maybe
>> in some cases we could warn and continue, just without the return value
>> fetching?
> 
> +1
> 
> It seems a bit too harsh for "finish" to error out, when it could just
> not print the return value.  get_return_value already returns NULL in
> a couple scenarios where we can't get at the return value, for instance.
> 
> For the "return" command, the patch is erroring out _after_ poping the frame.
> There are legitimate cases where GDB won't be able to retrieve the return
> value, like in the TYPE_NO_RETURN+query case handled in the function, which is
> checked _before_ frame_pop.  I'd think that if we are to error out when the arch
> can't write the return value, we should do it before frame_pop too.  Or we should
> warn that we can't write the return value, and proceed anyhow.
> 
> For ifuncs, calling the resolver and then erroring out is probably messing
> up run control.  It would probably be better to not try to resolve the ifunc
> at all.
> 
> For infcalls, I don't see a reason to error out if the user did "call func()"
> instead of "print func()".

I agree, I wrote this quickly to fix the problem, but didn't think all
this through.

>>
>> As I said above, I don't actually like GDB trying to handling this case
>> at all, I'd much rather we just force the port to stub these functions
>> during bring up.
> 
> There are legitimate cases when GDB isn't able to extract the return value.
> Say, the function does not follow normal ABI.  So we should already have
> code in place that gracefully handles not being able to get at the return value.
> If we code it right, then I guess in most cases an extra check for whether the
> arch implements the return value hook wouldn't end up complicating the code,
> it would just "fit right in".  I'm not sure it's worth the bother, though.
> I kind of tend to agree with you.

Yeah, that all makes sense.

If we en up with a way for gdbarch_return_value_as_value to return "I
don't know", then it will be easy for people who write ports to start with a
dummy implementation that always returns "I don't know".  That could
even be the default value for this gdbarch hook.  In either case, GDB
core indeed won't have to care about the case where the callback is
missing.

Simon

Simon