[PATCH 0/4] Smart pointer wrapper for frame_info

[Tom Tromey <tom@tromey.com>]

GDB occasionally gets bugs where a frame_info is kept alive across a
call to reinit_frame_cache.  This causes a use-after-free and, if
you're lucky, a crash.

This series aims to make this setup more "reliable", in the sense that
you'll always get a crash if you break the rules.  This is done by
wrapping frame_info in a smart pointer class, and having
reinit_frame_cache invalidate all the pointers.

I looked into a more ambitious approach: having the frame_info
wrappers automatically "reinflate" on demand after invalidation.
However, I couldn't readily make this work.  I am not completely sure,
but I think there are a few issues.

First, sometimes reinit_frame_cache is called after other global state
has been changed.  This means that one can't call get_frame_id from
there.

Second, I think computing the frame_id requires some unwinding.  This
means that it doesn't make sense to eagerly compute the frame_id when
constructing a frame_info.

This brings up another point -- most frame_info_ptr objects will not
really be needed after a reinit.  So, saving them all is an
unnecessary expense.  But accepting this means, essentially, accepting
the status quo.

It's possible to add a "stash" method to the new smart pointer class,
to make selected instances auto-inflate.  However I haven't done so.
That may be more expensive as well, as these objects are often passed
by value in the current series; and again it doesn't seem hugely
better than the status quo.

Regression tested on x86-64 Fedora 34.

Let me know what you think.  I tend to think that at least patch #2
could go in regardless of whether the rest of this series does.

Tom