Re: [PATCH] RTL_SSA: Relax PHI_MODE in phi_setup

[Richard Sandiford <richard.sandiford@arm.com>]

Juzhe-Zhong <juzhe.zhong@rivai.ai> writes:
> Hi, Richard.
>
> RISC-V port needs to add a bunch VLS modes (V16QI,V32QI,V64QI,...etc)
> There are sharing same REG_CLASS with VLA modes (VNx16QI,VNx32QI,...etc)
>
> When I am adding those VLS modes, the RTL_SSA initialization in VSETVL PASS (inserted after RA) ICE:
> rvv.c:13:1: internal compiler error: in partial_subreg_p, at rtl.h:3186
>    13 | }
>       | ^
> 0xf7a5b1 partial_subreg_p(machine_mode, machine_mode)
>         ../../../riscv-gcc/gcc/rtl.h:3186
> 0x1407616 wider_subreg_mode(machine_mode, machine_mode)
>         ../../../riscv-gcc/gcc/rtl.h:3252
> 0x2a2c6ff rtl_ssa::combine_modes(machine_mode, machine_mode)
>         ../../../riscv-gcc/gcc/rtl-ssa/internals.inl:677
> 0x2a2b9a4 rtl_ssa::function_info::simplify_phi_setup(rtl_ssa::phi_info*, rtl_ssa::set_info**, bitmap_head*)
>         ../../../riscv-gcc/gcc/rtl-ssa/functions.cc:146
> 0x2a2c142 rtl_ssa::function_info::simplify_phis()
>         ../../../riscv-gcc/gcc/rtl-ssa/functions.cc:258
> 0x2a2b3f0 rtl_ssa::function_info::function_info(function*)
>         ../../../riscv-gcc/gcc/rtl-ssa/functions.cc:51
> 0x1cebab9 pass_vsetvl::init()
>         ../../../riscv-gcc/gcc/config/riscv/riscv-vsetvl.cc:4578
> 0x1cec150 pass_vsetvl::execute(function*)
>         ../../../riscv-gcc/gcc/config/riscv/riscv-vsetvl.cc:4716
>
> The reason is that we have V32QImode (size = [32,0]) which is the mode set as regno_reg_rtx[97]
> When the PHI input def comes from ENTRY BLOCK (index =0), the def->mode () = V32QImode.
> But the phi_mode = VNx2QI for example (I use VLA modes intrinsic write the codes).
> Then combine_modes report ICE.
>
> In this situation, I relax it and let it use phi_mode directly.

The idea is that phi_mode must be:

(a) big enough to store all possible inputs without losing significant bits
(b) something that occupies the right number of registers

I think the patch loses property (a).

I suppose it would be difficult to find a "real" mode that is known to
contain both V32QI and VNx2QI without losing property (b).

There is some support for using BLKmode as a wildcard mode for registers.
Does it work if you add:

  if (!ordered_p (GET_MODE_SIZE (mode1), GET_MODE_SIZE (mode2)))
    return BLKmode;

before the call to wider_subreg_mode in combine_modes?

Thanks,
Richard

>
> Is it correct ?
>
> Thanks.
>
> gcc/ChangeLog:
>
>         * rtl-ssa/functions.cc (function_info::simplify_phi_setup): Relax combine in PHI setup.
>
> ---
>  gcc/rtl-ssa/functions.cc | 14 +++++++++++++-
>  1 file changed, 13 insertions(+), 1 deletion(-)
>
> diff --git a/gcc/rtl-ssa/functions.cc b/gcc/rtl-ssa/functions.cc
> index c35d25dbf8f..0793598ab1d 100644
> --- a/gcc/rtl-ssa/functions.cc
> +++ b/gcc/rtl-ssa/functions.cc
> @@ -143,7 +143,19 @@ function_info::simplify_phi_setup (phi_info *phi, set_info **assumed_values,
>        // If the input has a known mode (i.e. not BLKmode), make sure
>        // that the phi's mode is at least as large.
>        if (def)
> -	phi_mode = combine_modes (phi_mode, def->mode ());
> +	{
> +	  /* For target like RISC-V, it applies both variable-length
> +	     and fixed-length to the same REG_CLASS.
> +
> +	     It will cause ICE for these 2 following cases:
> +	       1. phi_mode: variable-length.
> +		  def->mode (): fixed-length.
> +	       2. phi_mode: fixed-length.
> +		  def->mode (): variable-length.  */
> +	  if (!(GET_MODE_SIZE (phi_mode).is_constant ()
> +		^ GET_MODE_SIZE (def->mode ()).is_constant ()))
> +	    phi_mode = combine_modes (phi_mode, def->mode ());
> +	}
>      }
>    if (phi->mode () != phi_mode)
>      phi->set_mode (phi_mode);