Re: vast numbers of unimplemented MI commands.

[Jim Ingham <jingham at apple dot com>]

On Sep 18, 2003, at 8:30 AM, gdb-digest-help@sources.redhat.com wrote:

> > Hi,
> >
> > I guess this seems ok, I just hate to see stuff like this.
> >
> >    But if you can come up with a clean Interface(set of methods) of
> >    what you need we could work things and put it in the CDI and  
> > implement
> >    it in the GDB/MI plugin.  The annoying thing is that GDB/MI does  
> > not
> >    have any "mi" command for this, so we will have to send "cli"  
> > command
> >    and do weird parsing on the output.
> >
> > from the eclipse mailing lists. Seeing this makes me think that the
> > whole community is not understanding that we need to have clean
> > interfaces between GDB and the front ends.
>
> Amen to that!!
>
> > Developers that find a "missing hole" in GDB, don't seem to have the
> > desire to fix the hole when they can easily find a work around.
> > IMO, most developers don't have enough time to fix there own
> > front end and then go fix the debugger there integrating with.
>
> I hear you !!! 8-)
>
> > I think GDB should support a proper MI interface for front ends, and
> > thats the only way front ends should be able to communicate with it.
> > Adding the -interpreter console was probably the main cause in  
> > allowing
> > front ends to cheat there way past the MI interface.

I don't think that is right.  You could always issue console commands  
straight to the mi, THAT is the real way to cheat (and I agree that  
should be closed off).  The console interpreter (and -interpreter-exec  
console) fulfills a very useful purpose: satisfying folks who like a  
console interface to gdb as well as the GUI.  We would probably get  
assassinated right quick if we were to try to take this out of Xcode  
(used to be Project Builder - another Marketing person earns their  
salary!) .

Note that we have had to fix up a few things in the MI as we have gone  
along with Xcode, and most of those fixes are in the Apple sources not  
in the main FSF repository.  We are short-staffed for what we need to  
do here, which sort of explains why we haven't gotten to submitting our  
sources back to the FSF - experience has shown that to be very time  
consuming.  But you can get the Apple sources from the Apple Darwin  
site, or from the opendarwin site.  If you are serious about using the  
MI it might be worth your while to have a look here, since Xcode is the  
only fairly mature GUI the uses the MI...

> >
>
> Agreed, but not for the same reasons.  We do not use the  
> "-interpreter-exec console"
> because it turns out to be useless.  The first attempt was to satisfy  
> the CLI users
> within the IDE, so a prompt was given.  But the problem is to  
> synchronize the IDE,
> let see a scenario, the "show":
>
> - scenario 1 (show):
>
> # gdb -i mi hello
> ...
> (gdb)
> show auto-solib
> &"show auto-solib\n"
> ^done,value="on"
> (gdb)
>
> The problem here, and with many other commands, is that the  
> information is return in MI jargon
> only,  so the user will not see it.  To do this correctly we would  
> need a cli interpreter
> that will pretty print the output of any commands(no fun job, knowin  
> the annoying cli syntax of gdb).

I don't understand this comment.  This is exactly what  
-interpreter-exec console is for.  If the user issues this sort of  
console command, you just echo back whatever it sends to your console  
window.  You can also annotate if they do anything of other interest to  
the GUI (proceeding the inferior, setting breakpoints, etc...)  Note  
that to make parsing the output from gdb, I added a "console-quoted"  
interpreter that sends stuff back in the cli pretty-printing, but the  
output strings are packaged up in MI format.  So you get something  
like:

-interpreter-exec console-quoted "info func"
~"All defined functions:\n"
~"\nFile "
~"/SourceCache/Csu/Csu-46/crt.c:\n"
~"void _start(int, char **, char **);\n"
~"static void _call_mod_init_funcs(void);\n"
...
^done,time={wallclock="1.83584",user="1.02000",system="0.11000",start="1 
063903591.475599",end="1063903593.311440"}
(gdb)


This means that if the response from a command happens to start with  
^done or something like that, you won't get confused...

I still have some cleanup to do on this, because there are various  
places (like the show command) where the command ignores the uiout and  
prints straight to gdbstdout (grrr....)  But this makes it pretty easy  
to handle this sort of thing.

> - scenario 2 (show/next):
>
> The problem is reverse when doing -interpreter-exec, the information  
> not is not in MI,
> this is illustrate when doing a "next"
>
> (gdb)
> -interpreter-exec console next
> ~"41\t\tchar array[2][2] = { 'a', 'b', 'c'};\n"
> ^done
> (gdb)
>
> The front-end fall out of step with gdb and no longer has relevant  
> information.
> However doing CLI without -interpreter-exec is good.
>
> (gdb)
> next
> &"next\n"
> ^done,reason="end-stepping-range",thread- 
> id="0",frame={addr="0x0804846e",func="main",args=[{name="argc",value="1 
> "},{name="argv",value="0xbfffeb54"}],file="hello.c",line="40"}
> (gdb)
>
> The front-end can keep up.

This is just a hole in the interpreter-exec console implementation.  We  
tarted this up a bit on our end, so you get:

-interpreter-exec console-quoted next
^stepping
^running
(gdb)
*stopped,time={wallclock="0.01305",user="0.00000",system="0.02000",start 
="1063903739.568366",end="1063903739.581413"},reason="end-stepping- 
range",thread-id="1"

I forget why the GUI guys wanted the ^stepping as well as the ^running,  
for regularity it would probably be better to leave that out.  But with  
this modification, it is very easy to keep the GUI in sync with the  
CLI.  The Xcode console interpreter actually works pretty well, and  
very seldom gets out of sync with the GUI.

> > The problem is, every *real* world front end to GDB is doomed to end  
> > up
> > using a mix of MI commands and CLI commands. If GDB is ever released
> > in such a way that the CLI output is changed, all existing front ends
> > will break. Including the ones that use MI.

This has not been our experience.  Xcode doesn't use any CLI commands,  
provided you don't consider "-interpreter-exec" a console command.  It  
took some work on our part to get this all going, but it is very  
achievable.

> > What is the best solution? ...
>
> Good question.  Although I like to complain 8-)
> The move toward MI is a good one, it is much cleaner, kudos to gdb  
> folks.

We have had pretty good success with the MI so far...

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