Fix to regclass problem

[Jeffrey A Law]

  In message < 9911041056.AA22461@vlsi1.ultra.nyu.edu >you write:
  >       > I must say that I think the PA failures even when it was badly brok
  > en
  >       > are still bugs.  It should work if pseudos were *randomly* assigned
  >  to
  >       > register classes since reload should be able to fix it all up.
  > 
  >     Not for the shift register and not for insns created after reload.
  > 
  > Why?  Doesn't reload need to be able to handle reloading an operand that's
  > residing in *any* register valid for that mode?
The shift register is only 5 bits wide, any attempt to copy its contents to
another location is considered an error.

jeff