------- Comment #2 from TabonyEE at austin dot rr dot com 2007-03-11 18:28 ------- Hhmm. I have made a few other changes and now when I don't define WORD_REGISTER_OPERATIONS, I get correct code. I get an HI load followed by a QI -> SI zero extend, not the preferred POST_MODIFY. I looked at all the changes I made since my last comment and I don't see how any could affect this. I think this is a latent bug waiting to be exposed that for some reason is difficult to trigger.
Now combine creates the subreg, but the subreg is spilled during the greg pass. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30798
