reload_indi with scratch == target?

[Ulrich Weigand]

Hello,

I just debugged a bootstrap failure on s390x which turned out to be caused
by a reload_indi insn getting passed the same register for both the scratch
and the target.  This is apparently because the primary reload has type
RELOAD_FOR_INPUT while the secondary has type RELOAD_FOR_INPUT_ADDRESS,
and the reload_reg_free_p routine assumes that two reloads of these
types for the same operand can share a register (without consideration
for the secondary reload case).

Now, from the description in the info pages I had assumed that I'd get a
scratch that is distinct from everything else in the insn, input and
output.  On the other hand, the sparc64 code contains comments that
this situation can happen, and implements a workaround.

Did I misread the manual or is this a bug in reload?  Should I implement
a similar workaround to sparc64 for our platform?


Mit freundlichen Gruessen / Best Regards

Ulrich Weigand

--
  Dr. Ulrich Weigand
  Linux for S/390 Design & Development
  IBM Deutschland Entwicklung GmbH, Schoenaicher Str. 220, 71032 Boeblingen
  Phone: +49-7031/16-3727   ---   Email: Ulrich.Weigand@de.ibm.com