This is mostly from code inspection, while investigating something else,
but I think that the current implementation of hw_watchpoint_used_count
is inaccurate. It counts the number of breakpoints which match the given
type.  But in fact, when I looked at the way the watchpoint is created,
watchpoints may have multiple bp_locations. So, a better approximation
would be to count the number of bp_locations, no?

Even so, this does not appear to be entirely accurate either. For
instance, if the user tries to watch an entity that does not fit
in one watchpoint, I think we can still end up with a situation
where we have one bp_location needing 2 h/w watchpoints.
In the end, it looks like the only way to really clean things up
would be to store for each watchpoint the associated number of
needed hardware watchpoints...

I thought about making the following change.  It's not clear that it is
a step in the right direction or not.  Pragmatically, I think it gives
more accurate results.  In terms of implementation, if we eventually
decide to store the mem_cnt in the breakpoint, then we'll have to
go back to iterating over all breakpoints (not bp_locations)...

When in doubt, do nothing?

        * breakpoint.c (hw_watchpoint_used_count): Compute the number
        of hardware watchpoints by iterating over all bp_locations
        instead of all breakpoints.

Tested on x86_64-linux. No regression...

-- 
Joel