sched/flow problem with memory corruption due to mark_set_resources

[Jeffrey A Law]

  > Do you have this tree handy?  If so, would you try out the following?
  > 
  > It seems we didn't _need_ to use the resource code, and in fact it's
  > a bit silly to do so here.
[ ... ]
  > 
  > 	* flow.c (new_insn_dead_notes): Use sets_reg_or_subreg not
  > 	mark_set_resources.
Hmmm, I wonder if this might account for the really weird haifa abort on the
PA64 port I've been putting off looking at for the last few days :-)

The symptom is flow creates a really wacky autoincrement address mode,
which later confuses the daylights out of the scheduler.

I'll give it a spin and let you know if it solves the problem.
jeff