On Thu, Jun 20, 2019 at 7:36 AM Hongtao Liu <crazy...@gmail.com> wrote:
>
> On Sat, Jun 8, 2019 at 4:12 AM Uros Bizjak <ubiz...@gmail.com> wrote:
> >
> > On 6/7/19, H.J. Lu <hjl.to...@gmail.com> wrote:
> >
> > >> > > +/* Register pair. */
> > >> > > +VECTOR_MODES_WITH_PREFIX (P, INT, 2); /* P2QI */
> > >> > > +VECTOR_MODES_WITH_PREFIX (P, INT, 4); /* P2HI P4QI */
> > >> > >
> > >> > > I think
> > >> > >
> > >> > > INT_MODE (P2QI, 16);
> > >> > > INT_MODE (P2HI, 32);
> > >> > >
> > >> > > with the above subreg approach should work.
> > >> > >
> > >> >
> > >> > I don't think subreg works on pseudo registers with non-zero
> > >> > offset. validate_subreg has
> > >> >
> > >> > if (maybe_lt (osize, regsize)
> > >> > && ! (lra_in_progress && (FLOAT_MODE_P (imode) || FLOAT_MODE_P
> > >> > (omode))))
> > >> > {
> > >> > /* It is invalid for the target to pick a register size for a
> > >> > mode
> > >> > that isn't ordered wrt to the size of that mode. */
> > >> > poly_uint64 block_size = ordered_min (isize, regsize);
> > >> > unsigned int start_reg;
> > >> > poly_uint64 offset_within_reg;
> > >> > if (!can_div_trunc_p (offset, block_size, &start_reg,
> > >> > &offset_within_reg)
> > >> > || (BYTES_BIG_ENDIAN
> > >> > ? maybe_ne (offset_within_reg, block_size - osize)
> > >> > : maybe_ne (offset_within_reg, 0U)))
> > >> > return false;
> > >>
> > >> It works with SImode subregs of DImode values on 32bit targets. Please
> > >> look for calls to gen_highpart, one concrete example is in
> > >> atomic_compare_and_swap<mode>.
> > >>
> > >
> > > It works because of
> > >
> > > #define REGMODE_NATURAL_SIZE(MODE) UNITS_PER_WORD
> > >
> > > and only works for the high part of SImode of DImode.
> > >
> > > P2QI and P2HI are 2 special modes of mask register pair for
> > > 2 instructions. Do we want to make them more generic?
> >
> > If enhancing the referred define means that we don't need two
> > artificial instructions and leave all heavy lifting to the existing
> Do you mean that we take P2HI and P2QI as normal vector modes,
> and reuse ix86_expand_vector_* things?
> But still two artificial instructions can't be avoided.
> > generic functionality, then this is the way to go.
No, declare them as integer modes and use subregs to access high and
low register. This should work in the same way as SImode hard
registers are accessed in DImode pair for 32bit targets.
Uros.
