[02/17,APX,NDD] Restrict TImode register usage when NDD enabled

  Under APX NDD, previous TImode allocation will have issue that it was
originally allocated using continuous pair, like rax:rdi, rdi:rdx.

This will cause issue for all TImode NDD patterns. For NDD we will not
assume the arithmetic operations like add have dependency between dest
and src1, then write to 1st highpart rdi will be overrided by the 2nd
lowpart rdi if 2nd lowpart rdi have different src as input, then the write
to 1st highpart rdi will missed and cause miscompliation.

To resolve this, under TARGET_APX_NDD we'd only allow register with even
regno to be allocated with TImode, then TImode registers will be allocated
with non-overlapping pairs.

There could be some error for inline assembly if it forcely allocate __int128
with odd number general register.

gcc/ChangeLog:

	* config/i386/i386.cc (ix86_hard_regno_mode_ok): Restrict even regno
	for TImode if APX NDD enabled.
---
 gcc/config/i386/i386.cc | 10 ++++++++++
 1 file changed, 10 insertions(+)
  

On Tue, Dec 5, 2023 at 3:29 AM Hongyu Wang <hongyu.wang@intel.com> wrote:
>
> Under APX NDD, previous TImode allocation will have issue that it was
> originally allocated using continuous pair, like rax:rdi, rdi:rdx.
>
> This will cause issue for all TImode NDD patterns. For NDD we will not
> assume the arithmetic operations like add have dependency between dest
> and src1, then write to 1st highpart rdi will be overrided by the 2nd
> lowpart rdi if 2nd lowpart rdi have different src as input, then the write
> to 1st highpart rdi will missed and cause miscompliation.
>
> To resolve this, under TARGET_APX_NDD we'd only allow register with even
> regno to be allocated with TImode, then TImode registers will be allocated
> with non-overlapping pairs.

Perhaps you could use earlyclobber with __doubleword instructions:

(define_insn_and_split "*add<dwi>3_doubleword"
  [(set (match_operand:<DWI> 0 "nonimmediate_operand" "=ro,r")
    (plus:<DWI>
      (match_operand:<DWI> 1 "nonimmediate_operand" "%0,0")
      (match_operand:<DWI> 2 "x86_64_hilo_general_operand" "r<di>,o")))
   (clobber (reg:CC FLAGS_REG))]

For the above pattern, you can add earlyclobbered &r output
alternative that guarantees that output won't be allocated to any of
the input registers.

Uros.

> There could be some error for inline assembly if it forcely allocate __int128
> with odd number general register.
>
> gcc/ChangeLog:
>
>         * config/i386/i386.cc (ix86_hard_regno_mode_ok): Restrict even regno
>         for TImode if APX NDD enabled.
> ---
>  gcc/config/i386/i386.cc | 10 ++++++++++
>  1 file changed, 10 insertions(+)
>
> diff --git a/gcc/config/i386/i386.cc b/gcc/config/i386/i386.cc
> index 93a9cb556a5..3efeed396c4 100644
> --- a/gcc/config/i386/i386.cc
> +++ b/gcc/config/i386/i386.cc
> @@ -20873,6 +20873,16 @@ ix86_hard_regno_mode_ok (unsigned int regno, machine_mode mode)
>         return true;
>        return !can_create_pseudo_p ();
>      }
> +  /* With TImode we previously have assumption that src1/dest will use same
> +     register, so the allocation of highpart/lowpart can be consecutive, and
> +     2 TImode insn would held their low/highpart in continuous sequence like
> +     rax:rdx, rdx:rcx. This will not work for APX_NDD since NDD allows
> +     different registers as dest/src1, when writes to 2nd lowpart will impact
> +     the writes to 1st highpart, then the insn will be optimized out. So for
> +     TImode pattern if we support NDD form, the allowed register number should
> +     be even to avoid such mixed high/low part override. */
> +  else if (TARGET_APX_NDD && mode == TImode)
> +    return regno % 2 == 0;
>    /* We handle both integer and floats in the general purpose registers.  */
>    else if (VALID_INT_MODE_P (mode)
>            || VALID_FP_MODE_P (mode))
> --
> 2.31.1
>
  

Uros Bizjak <ubizjak@gmail.com> 于2023年12月5日周二 18:46写道：

>
> On Tue, Dec 5, 2023 at 3:29 AM Hongyu Wang <hongyu.wang@intel.com> wrote:
> >
> > Under APX NDD, previous TImode allocation will have issue that it was
> > originally allocated using continuous pair, like rax:rdi, rdi:rdx.
> >
> > This will cause issue for all TImode NDD patterns. For NDD we will not
> > assume the arithmetic operations like add have dependency between dest
> > and src1, then write to 1st highpart rdi will be overrided by the 2nd
> > lowpart rdi if 2nd lowpart rdi have different src as input, then the write
> > to 1st highpart rdi will missed and cause miscompliation.
> >
> > To resolve this, under TARGET_APX_NDD we'd only allow register with even
> > regno to be allocated with TImode, then TImode registers will be allocated
> > with non-overlapping pairs.
>
> Perhaps you could use earlyclobber with __doubleword instructions:
>
> (define_insn_and_split "*add<dwi>3_doubleword"
>   [(set (match_operand:<DWI> 0 "nonimmediate_operand" "=ro,r")
>     (plus:<DWI>
>       (match_operand:<DWI> 1 "nonimmediate_operand" "%0,0")
>       (match_operand:<DWI> 2 "x86_64_hilo_general_operand" "r<di>,o")))
>    (clobber (reg:CC FLAGS_REG))]
>
> For the above pattern, you can add earlyclobbered &r output
> alternative that guarantees that output won't be allocated to any of
> the input registers.
>

Yes, it does resolve the dest/src overlapping issue we met, thanks!
I tried it and no fails in gcc-testsuite and spec. Suppose for
different src1/src2 RA can handle them correctly.

Will update in V3 patches with the changes of get_attr_isa (insn) == ISA_APX_NDD
  

On Wed, Dec 6, 2023 at 2:31 AM Hongyu Wang <wwwhhhyyy333@gmail.com> wrote:
>
> Uros Bizjak <ubizjak@gmail.com> 于2023年12月5日周二 18:46写道：
>
> >
> > On Tue, Dec 5, 2023 at 3:29 AM Hongyu Wang <hongyu.wang@intel.com> wrote:
> > >
> > > Under APX NDD, previous TImode allocation will have issue that it was
> > > originally allocated using continuous pair, like rax:rdi, rdi:rdx.
> > >
> > > This will cause issue for all TImode NDD patterns. For NDD we will not
> > > assume the arithmetic operations like add have dependency between dest
> > > and src1, then write to 1st highpart rdi will be overrided by the 2nd
> > > lowpart rdi if 2nd lowpart rdi have different src as input, then the write
> > > to 1st highpart rdi will missed and cause miscompliation.
> > >
> > > To resolve this, under TARGET_APX_NDD we'd only allow register with even
> > > regno to be allocated with TImode, then TImode registers will be allocated
> > > with non-overlapping pairs.
> >
> > Perhaps you could use earlyclobber with __doubleword instructions:
> >
> > (define_insn_and_split "*add<dwi>3_doubleword"
> >   [(set (match_operand:<DWI> 0 "nonimmediate_operand" "=ro,r")
> >     (plus:<DWI>
> >       (match_operand:<DWI> 1 "nonimmediate_operand" "%0,0")
> >       (match_operand:<DWI> 2 "x86_64_hilo_general_operand" "r<di>,o")))
> >    (clobber (reg:CC FLAGS_REG))]
> >
> > For the above pattern, you can add earlyclobbered &r output
> > alternative that guarantees that output won't be allocated to any of
> > the input registers.
> >
>
> Yes, it does resolve the dest/src overlapping issue we met, thanks!
> I tried it and no fails in gcc-testsuite and spec. Suppose for
> different src1/src2 RA can handle them correctly.

Yes, and when memory input operand is used in doubleword patterns, you
need earlyclobber anyway, otherwise nothing prevents the compiler from
clobbering address registers. When addr registers are dead, the
compiler can (and will) allocate output register to the same regno as
address register.

Uros,

> Will update in V3 patches with the changes of get_attr_isa (insn) == ISA_APX_NDD