[2/2,i386] STV changes, reg-copy as vector

The following splits out the change that makes the live-range split /
mode change copy pseudos vector mode pseudos.  I've tried reproducing
the issue I faced with SImode chains but failed likely because we're
never using simple moves in the end while for SImode we can end up
with

(set (subreg:V4SI (reg:SI 42) 0) (subreg:V4SI (reg:SI 41) 0))

which we happily propagate out.  So pursuing this patch independently
doesn't make much sense.  Still the main change is in
dimode_scalar_chain::make_vector_copies where 'vreg' is now V2DImode
instead of DImode and the change to emit_move_insn (vreg, tmp)
hints at the change done to the above problematic insn.

The rest of the changes deal with some regs already being in
the appropriate vector mode.

I realize I might have avoided my original issue by emitting not
the above reg-reg move but

 (set (subreg:V4SI (reg:SI 42) 0)
   (vec_merge:V4SI
      (vec_duplicate:V4SI (reg:SI 41))
      (const_vec [0 0 0 0])
      (const_1)))

but I didn't try (the full patch contained a testcase for the
issue).  Still it seems odd to use DImode pseudos when all uses
are wrapped in (subreg:V2DI ...).

As (unrelated) change I propose to error with fatal_insn_not_found
in case recog doesn't recognize the altered insns which make it
easier to spot errors early.

Bootstrap / regtest running on x86_64-unknown-linux-gnu ontop of [1/2], OK 
for trunk if that succeeds?

Thanks,
Richard.

2019-08-13  Richard Biener  <rguenther@suse.de>

	* config/i386/i386-features.c
	(dimode_scalar_chain::replace_with_subreg): Elide the subreg if the
	reg is already vector.
	(dimode_scalar_chain::convert_op): Likewise.
	(dimode_scalar_chain::make_vector_copies): Use a vector-mode pseudo as
	destination to avoid later passes copy-propagating it out.
	(dimode_chain::convert_insn): Convert the source of a
	memory op in case it is not of the appropriate scalar mode.
	Raise fatal_insn_not_found if the converted insn is not recognized.

On Tue, Aug 13, 2019 at 2:33 PM Richard Biener <rguenther@suse.de> wrote:
>
>
> The following splits out the change that makes the live-range split /
> mode change copy pseudos vector mode pseudos.  I've tried reproducing
> the issue I faced with SImode chains but failed likely because we're
> never using simple moves in the end while for SImode we can end up
> with
>
> (set (subreg:V4SI (reg:SI 42) 0) (subreg:V4SI (reg:SI 41) 0))
>
> which we happily propagate out.  So pursuing this patch independently
> doesn't make much sense.  Still the main change is in
> dimode_scalar_chain::make_vector_copies where 'vreg' is now V2DImode
> instead of DImode and the change to emit_move_insn (vreg, tmp)
> hints at the change done to the above problematic insn.
>
> The rest of the changes deal with some regs already being in
> the appropriate vector mode.
>
> I realize I might have avoided my original issue by emitting not
> the above reg-reg move but
>
>  (set (subreg:V4SI (reg:SI 42) 0)
>    (vec_merge:V4SI
>       (vec_duplicate:V4SI (reg:SI 41))
>       (const_vec [0 0 0 0])
>       (const_1)))
>
> but I didn't try (the full patch contained a testcase for the
> issue).  Still it seems odd to use DImode pseudos when all uses
> are wrapped in (subreg:V2DI ...).

It looks to me that the above is the way to go. make_vector_copies
creates a scalar pseudo, and all other support functions expect a
scalar that will be processed with "replace_with_subreg". I think that
the safest way is to emit the code above for SImode for the
problematic move insn, and

> +           emit_move_insn (gen_rtx_SUBREG (V2DImode, vreg 0),
> +                           gen_rtx_VEC_MERGE (V2DImode,
> +                                              gen_rtx_VEC_DUPLICATE (V2DImode,
> +                                                                     tmp),
> +                                              CONST0_RTX (V2DImode),
> +                                              GEN_INT (HOST_WIDE_INT_1U)));

for DImode. This way, you won't need other changes (conditional
generation of SUBREGs), the above should be enough.

Oh, and don't use emit_move_insn, emit insn with gen_rtx_SET
(gen_rtx_SUBREG (...)) should do the trick.

Uros.

> As (unrelated) change I propose to error with fatal_insn_not_found
> in case recog doesn't recognize the altered insns which make it
> easier to spot errors early.
>
> Bootstrap / regtest running on x86_64-unknown-linux-gnu ontop of [1/2], OK
> for trunk if that succeeds?
>
> Thanks,
> Richard.
>
> 2019-08-13  Richard Biener  <rguenther@suse.de>
>
>         * config/i386/i386-features.c
>         (dimode_scalar_chain::replace_with_subreg): Elide the subreg if the
>         reg is already vector.
>         (dimode_scalar_chain::convert_op): Likewise.
>         (dimode_scalar_chain::make_vector_copies): Use a vector-mode pseudo as
>         destination to avoid later passes copy-propagating it out.
>         (dimode_chain::convert_insn): Convert the source of a
>         memory op in case it is not of the appropriate scalar mode.
>         Raise fatal_insn_not_found if the converted insn is not recognized.
>
> Index: gcc/config/i386/i386-features.c
> ===================================================================
> --- gcc/config/i386/i386-features.c     (revision 274328)
> +++ gcc/config/i386/i386-features.c     (working copy)
> @@ -573,7 +573,8 @@ rtx
>  dimode_scalar_chain::replace_with_subreg (rtx x, rtx reg, rtx new_reg)
>  {
>    if (x == reg)
> -    return gen_rtx_SUBREG (V2DImode, new_reg, 0);
> +    return (GET_MODE (new_reg) == V2DImode
> +           ? new_reg : gen_rtx_SUBREG (V2DImode, new_reg, 0));
>
>    const char *fmt = GET_RTX_FORMAT (GET_CODE (x));
>    int i, j;
> @@ -627,7 +628,7 @@ void
>  dimode_scalar_chain::make_vector_copies (unsigned regno)
>  {
>    rtx reg = regno_reg_rtx[regno];
> -  rtx vreg = gen_reg_rtx (DImode);
> +  rtx vreg = gen_reg_rtx (V2DImode);
>    df_ref ref;
>
>    for (ref = DF_REG_DEF_CHAIN (regno); ref; ref = DF_REF_NEXT_REG (ref))
> @@ -641,7 +642,12 @@ dimode_scalar_chain::make_vector_copies
>                             gen_rtx_SUBREG (SImode, reg, 0));
>             emit_move_insn (adjust_address (tmp, SImode, 4),
>                             gen_rtx_SUBREG (SImode, reg, 4));
> -           emit_move_insn (vreg, tmp);
> +           emit_move_insn (vreg,
> +                           gen_rtx_VEC_MERGE (V2DImode,
> +                                              gen_rtx_VEC_DUPLICATE (V2DImode,
> +                                                                     tmp),
> +                                              CONST0_RTX (V2DImode),
> +                                              GEN_INT (HOST_WIDE_INT_1U)));
>           }
>         else if (TARGET_SSE4_1)
>           {
> @@ -841,7 +847,8 @@ dimode_scalar_chain::convert_op (rtx *op
>             gcc_assert (!DF_REF_CHAIN (ref));
>             break;
>           }
> -      *op = gen_rtx_SUBREG (V2DImode, *op, 0);
> +      if (GET_MODE (*op) != V2DImode)
> +       *op = gen_rtx_SUBREG (V2DImode, *op, 0);
>      }
>    else if (CONST_INT_P (*op))
>      {
> @@ -931,6 +938,8 @@ dimode_scalar_chain::convert_insn (rtx_i
>      case MEM:
>        if (!REG_P (dst))
>         convert_op (&src, insn);
> +      else if (GET_MODE (src) != DImode)
> +       src = gen_rtx_SUBREG (DImode, src, 0);
>        break;
>
>      case REG:
> @@ -977,7 +986,9 @@ dimode_scalar_chain::convert_insn (rtx_i
>    PATTERN (insn) = def_set;
>
>    INSN_CODE (insn) = -1;
> -  recog_memoized (insn);
> +  int patt = recog_memoized (insn);
> +  if (patt == -1)
> +    fatal_insn_not_found (insn);
>    df_insn_rescan (insn);
>  }
>

On Tue, 13 Aug 2019, Uros Bizjak wrote:

> On Tue, Aug 13, 2019 at 2:33 PM Richard Biener <rguenther@suse.de> wrote:
> >
> >
> > The following splits out the change that makes the live-range split /
> > mode change copy pseudos vector mode pseudos.  I've tried reproducing
> > the issue I faced with SImode chains but failed likely because we're
> > never using simple moves in the end while for SImode we can end up
> > with
> >
> > (set (subreg:V4SI (reg:SI 42) 0) (subreg:V4SI (reg:SI 41) 0))
> >
> > which we happily propagate out.  So pursuing this patch independently
> > doesn't make much sense.  Still the main change is in
> > dimode_scalar_chain::make_vector_copies where 'vreg' is now V2DImode
> > instead of DImode and the change to emit_move_insn (vreg, tmp)
> > hints at the change done to the above problematic insn.
> >
> > The rest of the changes deal with some regs already being in
> > the appropriate vector mode.
> >
> > I realize I might have avoided my original issue by emitting not
> > the above reg-reg move but
> >
> >  (set (subreg:V4SI (reg:SI 42) 0)
> >    (vec_merge:V4SI
> >       (vec_duplicate:V4SI (reg:SI 41))
> >       (const_vec [0 0 0 0])
> >       (const_1)))
> >
> > but I didn't try (the full patch contained a testcase for the
> > issue).  Still it seems odd to use DImode pseudos when all uses
> > are wrapped in (subreg:V2DI ...).
> 
> It looks to me that the above is the way to go. make_vector_copies
> creates a scalar pseudo, and all other support functions expect a
> scalar that will be processed with "replace_with_subreg". I think that
> the safest way is to emit the code above for SImode for the
> problematic move insn, and
> 
> > +           emit_move_insn (gen_rtx_SUBREG (V2DImode, vreg 0),
> > +                           gen_rtx_VEC_MERGE (V2DImode,
> > +                                              gen_rtx_VEC_DUPLICATE (V2DImode,
> > +                                                                     tmp),
> > +                                              CONST0_RTX (V2DImode),
> > +                                              GEN_INT (HOST_WIDE_INT_1U)));
> 
> for DImode. This way, you won't need other changes (conditional
> generation of SUBREGs), the above should be enough.
> 
> Oh, and don't use emit_move_insn, emit insn with gen_rtx_SET
> (gen_rtx_SUBREG (...)) should do the trick.

OK, that seems to work on the testcases, which makes splitting out this
patch pointless.  Integrated the change with the full patch.

Thanks,
Richard.