This is the mail archive of the
mailing list for the GCC project.
Re: PR57052, rs6000.md subregs
- From: David Edelsohn <dje dot gcc at gmail dot com>
- To: GCC Patches <gcc-patches at gcc dot gnu dot org>, Alan Modra <amodra at gmail dot com>
- Date: Wed, 24 Apr 2013 10:11:10 -0400
- Subject: Re: PR57052, rs6000.md subregs
- References: <20130424090135 dot GN22536 at bubble dot grove dot modra dot org>
On Wed, Apr 24, 2013 at 5:01 AM, Alan Modra <firstname.lastname@example.org> wrote:
> Practically all of the patterns in rs6000.md having a low-part subreg
> use zero as the byte number. That's correct when little-endian, but
> not when big-endian, so combine loses optimization opportunities.
> This patch duplicates the insns for big-endian. I know Segher has
> been working on removing these subreg patterns, but I think he'll need
> to stop combine creating low-part subregs for that to work. So until
> his patch materializes, is this OK for mainline? Bootstrapped and
> regression tested powerpc-linux.
> PR target/57052
> * config/rs6000/rs6000.md (rotlsi3_internal7): Rename to
> rotlsi3_internal7le and condition on !BYTES_BIG_ENDIAN.
> (rotlsi3_internal8be): New BYTES_BIG_ENDIAN insn.
> Repeat for many other rotate/shift and mask patterns using subregs.
> Name lshiftrt insns.
> (ashrdisi3_noppc64): Rename to ashrdisi3_noppc64be and condition
> on WORDS_BIG_ENDIAN.
This is okay, but I eagerly await Segher's patch that removes all of this.