Paradoxical subreg in compare

[Jeffrey A Law]

  In message < orsnv245h7.fsf@saci.lsd.ic.unicamp.br >you write:
 > This unusual paradoxical subreg seems wrong, according to the standard
  > semantics of subreg, but Jim Wilson says, in this case, it should be
  > equivalent to zero_extend.  Moreover, register_operand() accepts a
  > subreg of a mem before reload, because it is known for sure that
  > reload is going to fix it.  Unfortunately, a reload inheritance
  > problem prevents the job from being done correctly.
Then you need to be looking to fix reload inheritance, not combine.  Combine's
actions are OK given the current state of register_operand and friends.

  > Here's the patch.  Ok to install?
No, it's papering over the problem, not fixing it.

jeff