On Tue, Sep 14, 2021 at 8:58 AM Andrew Pinski <pins...@gmail.com> wrote:
>
> On Wed, Sep 8, 2021 at 2:55 AM Hongtao Liu via Gcc-patches
> <gcc-patches@gcc.gnu.org> wrote:
> >
> > On Wed, Sep 8, 2021 at 5:33 PM Jakub Jelinek <ja...@redhat.com> wrote:
> > >
> > > On Wed, Sep 08, 2021 at 05:23:40PM +0800, Hongtao Liu wrote:
> > > > > Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
> > > > >
> > > > Patch LGTM.
> > >
> > > Thanks, committed.
> > >
> > > > PS:
> > > > I'm curious why we need the post_reload splitter @xorsign<mode>3_1
> > > > for scalar mode, can't we just expand them into and/xor operations in
> > > > the expander, just like vector modes did.
> > > > Let me do some experiments to see whether it is ok to remove the
> > > > splitter.
> > >
> > > I bet it is the question how should the code look like before RA.
> > > stv is somewhat related, but as that replaces whole chains, it can do:
> > > (insn 14 5 6 2 (set (subreg:V2DI (reg:DI 92) 0)
> > > (vec_concat:V2DI (mem/c:DI (symbol_ref:SI ("c") [flags 0x2]
> > > <var_decl 0x7f65a131fd80 c>) [1 c+0 S8 A64])
> > > (const_int 0 [0]))) "hohohou.c":6:9 -1
> > > (nil))
> > > on loads of memory.
> > > But it stv still does use paradoxical subregs:
> > > (insn 10 16 11 2 (set (subreg:V2DI (reg:DI 91) 0)
> > > (minus:V2DI (subreg:V2DI (reg:DI 87) 0)
> > > (subreg:V2DI (reg:DI 94) 0))) "hohohou.c":6:13 5003
> > > {*subv2di3}
> > > (expr_list:REG_DEAD (reg:DI 87)
> > > (expr_list:REG_UNUSED (reg:CC 17 flags)
> > > (nil))))
> > > (insn 11 10 0 2 (set (mem/c:DI (symbol_ref:SI ("a") [flags 0x2]
> > > <var_decl 0x7f65a131fc60 a>) [1 a+0 S8 A64])
> > > (reg:DI 91)) "hohohou.c":6:5 76 {*movdi_internal}
> > > (expr_list:REG_DEAD (reg:DI 91)
> > > (nil)))
> > > so perhaps just using paradoxical subregs everywhere would be ok?
> > Yes, I think so.
> > And I find paradoxical subreg like (subreg:V4SF (reg:SF)) are not
> > allowed by validate_subreg until r11-621.
> > That's why post_reload splitter is needed here.
>
> That is not exactly true. It has been around since before 2005. See
> https://gcc.gnu.org/PR24436 which is referencing the fixme comment in
> validate_subreg.
We also have things like (subreg:V4SF(reg:V2SF) 0), the problem of
defining post_reload splitter with V2SF is movv2sf is only defined
under TARGET_64BIT if there's no mmx(so should we also enable 64-bit
vector 32-bit mode?).
And for xorsign w/o post_reload splitter, the code is cleaner and even
more optimal.
>
> Thanks,
> Andrew Pinski
>
> Thanks,
> Andrew Pinski
>
> > > Jakub
> > >
> >
> >
> >
> > --
> > BR,
> > Hongtao
--
BR,
Hongtao
