Re: [PATCH 2/2] mips: Optimize jit_ctor() / jit_ctzr()

[Paul Cercueil]

Hi Paulo,

Le mercredi 08 mars 2023 à 00:33 +0000, Paul Cercueil a écrit :
> The jit_ctzr() can be performed with just 5 instructions and 2
> temporary
> registers. The jit_ctor() can be performed by just inversing all bits
> then calculating the ctzr().
> 
> Signed-off-by: Paul Cercueil <paul@crapouillou.net>
> ---
>  lib/jit_mips-cpu.c | 61 +++++++++++++++++++-------------------------
> --
>  1 file changed, 25 insertions(+), 36 deletions(-)
> 
> diff --git a/lib/jit_mips-cpu.c b/lib/jit_mips-cpu.c
> index d71a5b5..7babc86 100644
> --- a/lib/jit_mips-cpu.c
> +++ b/lib/jit_mips-cpu.c
> @@ -1670,49 +1670,38 @@ _clzr(jit_state_t *_jit, jit_int32_t r0,
> jit_int32_t r1)
>  static void
>  _ctor(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
>  {
> -    if (jit_mips2_p()) {
> -       if (jit_mips6_p()) {
> -#if __WORDSIZE == 32
> -           BITSWAP(r0, r1);
> -           bswapr_ui(r0, r0);
> -           CLO_R6(r0, r0);
> -#else
> -           DBITSWAP(r0, r1);
> -           bswapr_ul(r0, r0);
> -           DCLO_R6(r0, r0);
> -#endif
> -       }
> -       else {
> -           fallback_bitswap(r0, r1);
> -           clor(r0, r0);
> -       }
> +    if (jit_mips6_p()) {
> +        rbitr(r0, r1);
> +        clor(r0, r0);
> +    }
> +    else {
> +        comr(r0, r1);
> +        ctzr(r0, r0);
>      }
> -    else
> -       fallback_cto(r0, r1);
>  }
>  
>  static void
>  _ctzr(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
>  {
> -  if (jit_mips2_p()) {
> -       if (jit_mips6_p()) {
> -#if __WORDSIZE == 32
> -           BITSWAP(r0, r1);
> -           bswapr_ui(r0, r0);
> -           CLZ_R6(r0, r0);
> -#else
> -           DBITSWAP(r0, r1);
> -           bswapr_ul(r0, r0);
> -           DCLZ_R6(r0, r0);
> -#endif
> -       }
> -       else {
> -           fallback_bitswap(r0, r1);
> -           clzr(r0, r0);
> -       }
> +    if (jit_mips6_p()) {
> +        rbitr(r0, r1);
> +        clzr(r0, r0);
> +    }
> +    else {
> +        jit_int32_t     t0, t1;
> +
> +        t0 = jit_get_reg(jit_class_gpr);
> +        t1 = jit_get_reg(jit_class_gpr);
> +
> +        negr(rn(t0), r1);
> +        andr(rn(t0), rn(t0), r1);
> +        clzr(r0, rn(t0));
> +        xori(rn(t1), r0, __WORDSIZE - 1);
> +        movnr(r0, rn(t1), rn(t0));

Thanks for merging.

I would just like to point out that this algorithm above is nothing
MIPS-specific, and other archs with a CLZ instruction (e.g. PPC) would
very much benefit from using this algorithm instead of the bit-reverse
one.

Cheers,
-Paul

> +
> +        jit_unget_reg(t0);
> +        jit_unget_reg(t1);
>      }
> -    else
> -       fallback_ctz(r0, r1);
>  }
>  
>  static void