Re: [PATCH 1/3] SH-2A FDPIC: New pattern: use_initial_val

[Richard Henderson <rth at redhat dot com>]

On 08/19/2010 11:55 AM, Bernd Schmidt wrote:
> AFAIR no code is output for it.  If that's not true, then that at least
> needs to be revisited.  Does the splitting code handle the case where we
> return no insn at all?

Yes.  Lots of other ports do that regularly.

I think the correct fix is

(define_insn_and_split "use_initial_val"
  [(unspec [(match_operand 0 "pseudo_register_operand" "r")] UNSPEC_INITVAL)]
  ""
  "#"
  "reload_completed"
  [(const_int 0)])

+  { DONE; })

Otherwise you'll split to (const_int 0), which is defined elsewhere
as the "nop" pattern.  You can't elide the [] block from the split
pattern though; you have to have something there.

>> What advantage does this unspec have over a (USE reg) insn?
> 
> I wasn't certain a USE wouldn't just get deleted, now or in the future,
> if it was the only remaining use of a pseudo.

I'm pretty sure they get deleted during/after reload, but they
hang around until then.

> I don't have the build trees anymore, and I don't recall all the details
> - but I think it was generating a reference to a symbol, which required
> the PIC register.  I believe it was __fpscr_values - see emit_fpu_switch
> in sh.c.

That does seem like a likely candidate.  Which means that the new use
can occur from essentially any random FP insn.  I agree that you'd not
want to annotate each and every one of those.

I wonder if a better solution might be to enhance df_get_exit_block_use_set
to keep your pseudo alive for the entire function?  We currently have 
EPILOGUE_USES, but that's limited to hard registers.  I wonder if the
macro could be replaced by a hook that, instead of returning true for
any specific regno, sets bits in a bitset.

You'd want to have the hook disable the forced use sometime just reload,
presumably, in order to let the pseudo die properly.  I don't know enough
about the current state of the transition to df-scan to know if you'd
need to rebuild REG_DEAD notes before reload, or what.


r~