Re: question for gdb users...

[Richard Frith-Macdonald]

On 24 Sep 2009, at 08:50, Matt Rice wrote:

> On Thu, Sep 24, 2009 at 12:16 AM, Richard Frith-Macdonald
> <richard@tiptree.demon.co.uk> wrote:
> > On 24 Sep 2009, at 04:40, Matt Rice wrote:

<snip>

> The specific problem is that gdb resets breakpoints upon shared
> library load, in case symbols were overriden by c-linkage behaviour
>
> the 'canonical' version of 'main' is 'main', where the canonical
> version of '-[NSThread main]'
>
> when gdb resets the breakpoint on shared library load to check for a
> new address for 'main'. it will respond with more than one address,
> because main is ambiguous, so its unable to tell which of the symbols
> you actually want a breakpoint on.
>
> now imagine if the function was named init :)

This sounds like an issue with the way gdb stores its breakpoints ...  
shouldn't it be storing more information so that it can disambiguate?   
I guess probably understanding this requires a fairly detailed  
knowledge of the symbol representation used in the libraries and in  
gdb ... which I lack.  However, it's clear that at time point gdb is  
able to examine the binaries to get enough information to offer me a  
selection of places to break when I type 'break init' ... so there is  
sufficient information somewhere.

Now, for instance, if I type 'break init' I am presented with a  
numbered list of all the implementations of -init, and can type in a  
the number of an item to set a breakpoint in (for instance -[NSString  
init].   I would have expected gdb to record the fact that a  
breakpoint had been set in the NSString implementation of init rather  
than any other implementation, and be able to use that to determine  
which implementation in a newly loaded shared library should be used.
I  can also type in '1' to set a breakpoint in *all* implementations  
of init ... in which case, why would we care if there are multiple  
addresses in a newly loaded library .. we want all of them anyway.

Anyway ... perhaps it's not that straightforward, and perhaps even if  
it *is* that straightforward conceptually it's still a lot of work to  
implement changes., so I'm not trying to second-guess the basic  
objection.

<snip>

> > +[method]       a factory method of any class
> > -[method]   in instance method in any class
> > [method]  either a factory method or an instance method,
>
> these are good ideas, I was thinking
> +method/-method and was kind of perplexed on what to do for both,
> [method] might work, i was also thinking ?method, which has no real
> objc relation, but might be familiar to those using the shell?

The '?method' syntax would not have occurred to me.
I just thought that keeping on using square brackets would be familiar  
in objc.
Another possibility might be 'wildcarding' the class name rathe than  
simply omitting it
eg.
[* method] or [? method]
but I think I prefer the simple omission of the class name.

> > So 'break [method] would be like the current 'break method' syntax  
> > but
> > 'method' would be treated *only* as an ObjC method name, never as a  
> > function
> > or method in another language.
> >
> > If this would help resolve the issue, it would satisfy me.  Simply  
> > losing
> > the ability to set a breakpoint in any class would not.
>
> you could still do this with the rbreak command,
> by using 'rbreak .*methodname.*
>
> so even if we remove the function and we don't reintroduce an
> alternative syntax this is still possible.

I don't really consider rbreak a viable alternative as it's too clumsy  
for routine use ...

1. you have to remember that it's using a regular expression, and put  
in correct limitations to avoid getting far too many  matches (try  
'rbreak class' for instance).
2. it doesn't present you with a list of possible matches and ask you  
to select the one you want ... instead it tries to set breakpoints in  
all the matching methods and gives you a load of yes/no questions  
about methods in classes not yet loaded.

So yes, it's *possible* to use it, but it's not a good substitute for  
'break methodname'