Re: [patch rfa:doco rfc:NEWS] mi1 -> mi2; rm mi0

[Jim Ingham <jingham@apple.com>]

Andrew,

On Friday, October 25, 2002, at 05:57  PM, 
gdb-patches-digest-help@sources.redhat.com wrote:

> From: Andrew Cagney <ac131313@redhat.com>
> Date: Fri Oct 25, 2002  2:48:09  PM US/Pacific
> To: Jim Ingham <jingham@apple.com>
> Cc: Keith Seitz <keiths@redhat.com>, gdb-patches@sources.redhat.com
> Subject:
>
>
> [Did anyone mention that Apple's mail tool is officially weird?]
> [was a good holiday]
>
>
>>> On Wednesday, October 2, 2002, at 08:01  AM, Keith Seitz wrote:
>>>> On Tue, 1 Oct 2002, Jim Ingham wrote:
>>>> I don't want a tight coupling in TIME, or in the output sequence.  I
>>>> want meta-info which tells me with 100% certainty that this result 
>>>> is
>>>> the result of command foo, that I sent some time in the past, 
>>>> precisely
>>>> so I DON'T have to rely on anything about the sequencing of input or
>>>> output to make this connection.
>>> Ok, but this is exactly what I don't understand: what does it matter 
>>> where
>>> the breakpoint comes from? All that matters is that you have issued a
>>> command, and you don't do anything out of the normaluntil an event 
>>> from
>>> GDB telling you that it has set a breakpoint. You just sit in your 
>>> event
>>> loop waiting for any event (GUI, paint, GDB, blah blah blah).
>
>> In large part, it is just that if I can track cause & effect 
>> unambiguously, everything is much cleaner.  From PB, I issue a 
>> command, I want to tick off when that command succeeded, and gather 
>> all the data from the result I need.  If I get a note back from gdb 
>> saying, "The third command down in the command stream that you 
>> issued, here is the result to it." then I can do this without any 
>> effort.  However, in your proposed implementation, I have to peer at 
>> the command stream and wait both for an anonymous breakpoint created 
>> message, and a for the done message (to know things are formally 
>> done, and because it would be very dangerous to just toss aside bits 
>> of the output stream, since it reduces your chances of catching & 
>> handling errors intelligently.)  Then I have to use some smarts to 
>> tie the breakpoint created message to the particular message I sent 
>> to gdb.  This is harder to do, and fragile.
>
>> In my mind, one of the major advantages of the MI was that it made 
>> this tie between the command you sent and the result that came from 
>> it explicit and unambiguous.  This was in response to a lot of the 
>> pain that we suffered in insight where any random spew from the 
>> command line would get itself inserted into the "results" of the 
>> command you thought you were executing.  You are breaking this tie by 
>> moving the results of a command from the command result to an 
>> untagged asynchronous event.  For breakpoints, this is probably just 
>> an inconvenience.  But if we follow this as a principle, then in more 
>> complex cases, where it is not so easy to tell whether an 
>> asynchronous notification arose from a command initiated by the MI, 
>> or from some other action we will likely start getting confused 
>> again.  For instance, we might have something happening in the 
>> target, or some console command, or something more gnarly like the 
>> results of a breakpoint command which also continued the target while 
>> you thought it was just running...
>
> You mean like 
> http://sources.redhat.com/gdb/papers/libgdb2/wp/gdbtk2.jpg
>

Cute...

> There are two things here:
>
> - events
> - results
>
> And a question of what, if anything, each should be bound two.
>
> For the case in question, I think you're greatest concern is with 
> ensuring that the result of the request is clearly bound to the 
> command completion indication.  With that in mind, would something 
> like:
>
> 	-breakpoint-create
> 	=breakpoint-created,....
> 	=breakpoint-created,....
> 	^done,breakpoint=1
>
> be reasonable?  It clearly binds the created breakpoint with the 
> request while at the same time doesn't treat the creation of this 
> breakpoint any different to the creation of any other breakpoint.

Right.  I think the principle should be that any command that has a 
direct result, that result should be reported in the "done" message of 
the command.  I don't mind that it also get reported as an event.  For 
instance, this means that if something like your listing above occurs, 
I know that the first breakpoint is the one I intended to make, and the 
second one, dunno about.  This will be important, for instance, if we 
ever teach gdb that breaking on template source code generated a whole 
bunch of breakpoints.  The GUI will want to gather up these breakpoints 
and treat them as a single bundle (since if the source line breakpoint 
is deleted in the UI, it will need to delete all these breakpoints).  
So knowing that one of the breakpoint-created events is some odd 
accident would be useful...

>
> I don't know how feasible it is to do this, mind.
>

Shouldn't be hard for the breakpoints, we do it already...

>
> >
> > In short, you are breaking one of the core parts of the MI model, 
> for no very apparent reason.  In this case, it might be benign, but it 
> is a very useful part of the design, and the reasons for doing it are 
> very weak.
>
> The MI was always going to have events and those events were going to 
> be separate from the results.  As a simple example (from the 5.2.1 
> doco):
> http://sources.redhat.com/gdb/download/onlinedocs/gdb_24.html#SEC223
>
> 	Command With Side Effects
> 	-> -symbol-file xyz.exe
> 	<- *breakpoint,nr="3",address="0x123",source="a.c:123"
> 	<- (gdb)
>
> [sic - its missing a ^done].  One reason for this separation of event 
> and result is that we felt we could see far enough ahead to at least 
> realise that the MI would have to handle things like:
>
> 	-interpreter cli "break foo"
> 	=breakpoint-created,...
> 	^done
>
> or even:
>
> 	-interpreter cli "source script"
> 	=breakpoint-created,...
> 	=breakpoint-deleted,...
> 	=running
> 	^done
> 	(gdb)
> 	=stopped,...
> 	=breakpoint-created,...
> 	=running
> 	(gdb)
>
> That is, in the general case, trying to tightly bind MI and events to 
> results wasn't going to be possible.
>
>

Seems better to say that an mi command will usually have a well defined 
result, but may well have unintended consequences as well.  We should 
be able to bind the result to the command token, but not the unintended 
consequences.  It's actually not impossible to do that either, the 
original implementation that I did had all the event-like things added 
as tagged bits to the command result, unless they came after the 
command returned.  But that's not the way Keith chose to implement it, 
which is fine.

Of course, -interpreter-exec (is Keith changing this to -interpreter, 
or did you just type it wrong?) has NO well defined results.  
Everything is an unintended consequence.  For convenience's sake, I 
think I report whether the command has started up the inferior or not, 
but even that could be gotten from events.

Jim
--
Jim Ingham                                   jingham@apple.com
Developer Tools
Apple Computer