Re: [PATCH] Always create a new value object in valpy_do_cast

[Tom Tromey <tom@tromey.com>]

>>>>> "Simon" == Simon Marchi <simark@simark.ca> writes:

>> I think what we probably need is to change how the value_mark API works,
>> so that it isn't relying on a value pointer being in all_values, but
>> rather return some other type/object that isn't invalidated in the same
>> way.

Simon> Is there an advantage of that value chain thing vs using value_ref_ptr
Simon> throughout?  Not replace all `value *` with value_ref_ptr, but have
Simon> functions return and accept value_ref_ptr where that makes sense.

The chain is used to implement the watchpoint semantics.  It's needed, I
think, but it would be better if it were opt-in -- that is, some caller
has to request it explicitly.  However that requires switching to
value_ref_ptr everywhere...

.. but please don't do that, I'm nearly done with a series to move value
to value.h and use methods, and it's reasonably large and would be
painful to rebase.

Meanwhile I think we have a fix for this bug, the problem is finding a
test case.  Though... the one we have does fail for Hannes so perhaps
that's enough.

Tom