[PATCH] MI: Free values when updating

[Daniel Jacobowitz]

On Wed, Jan 24, 2007 at 12:02:12AM +1300, Nick Roberts wrote:
>  > > it's just that this patch stops calling it at other times 
>  > > when it's needed.  Without any change, do enable timings (if you have that
>  > > patch), create a variable object of a large array and all its children then
>  > > repeatedly do "-var-update *".  It should take longer and longer to execute.
>  > 
>  > Why? Is it because the memory consumption of gdb grows, or because the list
>  > of released values grows without ever being cleared, or for some other
>  > reason?
> 
> The latter, I think.

Except that there isn't a list of released values.  So what is GDB
doing that is taking longer and longer?

The call to release_value does a linear walk over all non-released
values.  So if we have a lot of things which aren't being released,
then your patch which calls free_all_values is probably the right thing
to do - that should clean it up.

>  > -      if (gdb_evaluate_expression (var->root->exp, &new_val))
>  > -	{
>  > -	  release_value (new_val);
>  > -	}
>  > -
>  > +      gdb_evaluate_expression (var->root->exp, &new_val);
>  >        return new_val;
>  >      }
> 
> I think if you also remove the (3) calls to release_value in c_value_of_child
> and cplus_value_of_child this is equivalent to my change (and more tidy).

No, those are different.  They come from things like the call to
gdb_value_ind in c_describe_child.  That creates a new value, which is
returned to the caller (the MI front end, to be printed and later
released).  It's the ones in c_value_of_root which matter, because we
save them in the varobj.

You're probably right about the increasing time though - releasing
something already released will be slow.  I wonder if we should make
that an internal error somehow.

-- 
Daniel Jacobowitz
CodeSourcery