On Wed, Sep 10, 2014 at 01:38:13PM -0500, Peter Bergner wrote:
> On Tue, 2014-09-09 at 19:28 -0400, David Edelsohn wrote:
> > On Tue, Sep 9, 2014 at 3:29 PM, Segher Boessenkool
> > > 2014-09-09 Segher Boessenkool <seg...@kernel.crashing.org>
> > >
> > > * config/rs6000/htm.md (tabort, tabortdc, tabortdci, tabortwc,
> > > tabortwci, tbegin, tcheck, tend, trechkpt, treclaim, tsr): Use xor
> > > instead of minus.
> > > * config/rs6000/vector.md (cr6_test_for_zero_reverse,
> > > cr6_test_for_lt_reverse): Ditto.
> >
> > This is okay with me, but let me give Peter a chance to comment if
> > there was a specific reason to use subfic instead of xori. This may
> > have been a carry-over from Z, which does not have the same CA clobber
> > issue.
>
> Actually, I just copied the usage in cr6_test_for_zero_reverse and
> cr6_test_for_lt_reverse, so I'm not against using xori...as long as
> compiling a "if __builtin_tbegin (0) {...}" still ends up with a
> .tbegin followed immediately by a branch (ie, no interleaving copy
> from CR and compare instruction).
Huh, interesting. I assumed 1-(0_or_1) and (0_or_1)^1 would look the
same to combine, but no.
With subfic, combine optimises it all to a branch on cr0. With xori,
for some reason combine has a much easier job, and it optimises the lot
to a copy of cr0 to some cc, and then branch on that. The RA of course
gets rid of the copy. The extra freedom will more likely help than hurt.
The simple testcase ends up as just "tbegin. 0; beqlr 0" in either case.
So, okay?
Segher
