grokfield miscompiled with -O9

Richard Henderson rth@cygnus.com
Tue Jan 11 14:26:00 GMT 2000


On Tue, Jan 11, 2000 at 07:59:59AM -0700, Jeffrey A Law wrote:
> The whole point of that case is to deal with overlaps between the source
> and the dest.  In that case you need some kind of scratch register so that
> you do not stomp on your source operand before you're done reading it. 
> Right?

No.  The bit Alexandre is talking about is

      if (!cse_not_expected)
        tmp = gen_reg_rtx (QImode);
      else
        tmp = gen_lowpart (QImode, dest);

I'm moderately surprised that he got better results from disabling this.  

> I'd be more interested in fixing the code which pessimizes the code
> by setting the type to one instead of zero.  Unfortunately the comment
> doesn't provide information about the reload problem which led to that
> change.

I don't remember details, but I think it was an abort somewhere
under emit_reload_insns.  The comment doesn't give details because
I didn't really understand the problem at the time either.  It may
well have been fixed in the 10 months since I wrote that code.

As for type 0 vs type 1, type 0 is faster on P1, but they're
about the same on P2.



r~


More information about the Gcc-patches mailing list