Avoid BImode in reload_cse_move2add

[Richard Kenner]

    So I would appreaciate it if you did go back and figure out
    what in the world this patch is supposed to do.

I found the problem with my access to gcc.gnu.org (my PC here was somehow
caching its old IP address) and tried to search the archives to find the old
discussion on the topic.  I couldn't find it, but did find another patch to
the same area which was due to REG being BImode.  The author speculated that
the code "is not prepared to really see an BImode here", with which I
obviously agree.  This was PR 21144.

The fact that I tested the patch on x86_64 does not, of course, mean that
I originally detected the problem on that target and it's likely I saw it
on IA64.  I know it was an ICE and I think, but am not sure, that it was
an ACATS test.  I had all this written on a piece of paper, which I've been
unable to find.