Re: RFA: Recognize 'x' in response to 'p' packet

[Paul Schlie <schlie@comcast.net>]

> From: Paul Schlie <schlie@comcast.net>
>> From: Michael Snyder <msnyder@redhat.com>
>>> Paul Schlie wrote:
>>>> Jim Blandy
>>>> * remote.c (fetch_register_using_p): Recognize 'x's for the value
>>>>  of the register as indicating that the register's value is not
>>>>  available.
>>> 
>>> Out of curiosity, under what practical circumstances would the value of a
>>> register not be accessible? (and if not, shouldn't an error be returned, as
>>> opposed to an 'x' which is converted to a 0 anyway? Which I've noticed "g"
>>> packets also assume?)
>> 
>> Post-mortem debugging.  See tracepoint.c.  In a nutshell, you set up a
>> tracing experiment wherein a debugging agent on the target (like an
>> enhanced version of gdbserver) will collect data for you on the fly,
>> saving it to be retrieved later.  Data can include register values
>> and memory values.  During the post-mortem data examination stage,
>> the target agent enters a mode where, whenever GDB requests a register
>> or memory value, the agent supplies it from the collected cache rather
>> than from the 'live' target.  It lets GDB look at a saved machine state,
>> but you can only examine those registers that were saved.  Those that
>> weren't saved are not available, hence we need an idiom for reporting
>> this to gdb.  We can't use any legitimate integer value because a saved
>> register might actually have had that value.
> 
> Thanks, that makes a little more sense; but as xxxx values aren't represent-
> able within GDB, hence transformed upon receipt to 0; what's the point of
> transmit x's as opposed to 0's to begin with as an idiom until GDB is capable
> of differentiating valid from unknown values, which it's far from being
> capable of today, as it would imply that all logical values within GDB would
> need to be either maintained as text strings, or utilize structures containing
> it's value along with a tag indicating it's validity, and all expression
> corresponding evaluation routines would need to be updated.
> 
> (so although the proposal tries to solve a problem, it can't until GDB is
> capable of differentiation between valid and unknown values. If the objective
> is to determine for the purposes of information only which register value are
> valid independent of their value; therefore wonder if it may make sense to
> simply define a new server/target message function which requests the list of
> the validity of the register values, analogously it's request for their
> values, as today it's incapable of differentiation between valid and invalid
> ones if encoded into the value representation itself.)

Where upon a little more thought, if the objective is to enable gdb to
represent, track, and manipulate "unknown" values, it's likely most flexibly
done by representing all debug-values as text strings, expressed in whatever
radix required to precisely represent the value to the precision desired;
where then in addition to numerical digits, X's may represent
correspondingly unknown/indeterminate values, likely also useful when
interfacing with simulators analogously capable; but would require a fairly
large overhaul to accomplish, however until then if x's are treated as 0's
or whatever, might as well transmit the value desired to be assumed to begin
with, and not pretend anything would be accomplished otherwise, I would
think.)