This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] combine: Convert subreg-of-lshiftrt to zero_extract properly (PR78390)
On Wed, Nov 30, 2016 at 03:40:32PM +0100, Michael Matz wrote:
> Hi,
>
> On Wed, 30 Nov 2016, Segher Boessenkool wrote:
>
> > > I don't think mode-changing _extracts are valid in this context. From the
> > > docu:
> > >
> > > `(sign_extract:M LOC SIZE POS)'
> > > ...
> > > The mode M is the same as the mode that would be used for LOC if
> > > it were a register.
> > >
> > > Probably it could be made to work just fine, but I'm not sure it'd be
> > > worth much, as then the targets would need to care for mode-changes
> > > occuring not just through subregs as usual, but also through extracts.
> >
> > The patch https://gcc.gnu.org/ml/gcc-patches/2016-11/msg02987.html I
> > submitted yesterday deals with this same issue, FWIW -- some ports
> > apparently already do mode-changing extracts.
>
> Yeah, saw that a bit later. So, hmmm. I'm not sure what to make of it,
> if the targets choose to use mode-changing extracts I guess that's fine,
> as they presumably will have written patterns that recognize them. But I
> don't think we should willy-nilly generate such patterns as we can't know
> if the target deals with them or not.
Just working on such a pattern for s390x, I had the impression
that combine uses the automatic mode change when it's there, and
otherwise it does things differently, that is, it tries different
combinations when it has the pattern than without. There seems to
be at least some kind of detection.
> We could of course always generate
> both variants: (subreg:M1 (extract:M2 (object:M2)) and
> (extract:M1 (object:M2)) and see if either matches, but that seems a bit
> too much work.
--
The insns that combine tried are:
(insn 7 4 8 2 (set (reg:DI 69)
(lshiftrt:DI (reg/v:DI 66 [ v_x ])
(const_int 48 [0x30])))
(insn 9 8 10 2 (parallel [
(set (reg:SI 68 [ v_or ])
(ior:SI (reg:SI 70 [ v_and1 ])
(subreg:SI (reg:DI 69) 4)))
(clobber (reg:CC 33 %cc))
])
A while ago combine handled the situation well, resulting in the
new "risbg" instruction, but for a while it's not been working.
It's a bit difficult to track that down to a specific commit
because of the broken "combine"-patch that took a while to fix.
Ciao
Dominik ^_^ ^_^
--
Dominik Vogt
IBM Germany