PR 8134: C++ crash

[Mark Mitchell <mark@codesourcery.com>]

PR 8134 is a crash in force_store_init_value on the branch; it is a
regression.

The root problem here is that we were not handling zero-initialization
of pointers to members correctly.  Alexandre tried to fix the problem,
but in the process introduced the crashes above.  This was resolved on
the mainline with a rather substantial reworking of class layout code;
that was my patch to create a separate base class variant of each type.

What should we do on the branch?

The obvious choices are:

a) Nothing

   In this case, stuff blows up badly.

b) Move my changes over.

   As far as we know, these are correct -- but they are substantial,
   and therefore risky.

c) Revert Alexandre's patch.

   In this case, we get back to GCC 3.0-like behavior; incorrect
   zero-initialization of some pointers-to-members.

I don't like any of these choices.  I think I lean towards (c),
merely as a "devil you know" kind of choice.  I can also do (b), if
people feel that's the right thing, but I'm nervous about somehow
making an inadvertant ABI change in the minor release.

Thoughts?

-- 
Mark Mitchell                mark@codesourcery.com
CodeSourcery, LLC            http://www.codesourcery.com