Re: [PATCH 01/12] target/ppc: Moved VMLADDUHM to decodetree and use gvec

[Richard Henderson]

On 9/23/22 21:47, Lucas Mateus Castro(alqotel) wrote:
> From: "Lucas Mateus Castro (alqotel)"<lucas.araujo@eldorado.org.br>
>
> This patch moves VMLADDUHM to decodetree a creates a gvec implementation
> using mul_vec and add_vec.
>
> rept    loop    master             patch
> 8       12500   0,01810500         0,00903100 (-50.1%)
> 25      4000    0,01739400         0,00747700 (-57.0%)
> 100     1000    0,01843600         0,00901400 (-51.1%)
> 500     200     0,02574600         0,01971000 (-23.4%)
> 2500    40      0,05921600         0,07121800 (+20.3%)
> 8000    12      0,15326700         0,21725200 (+41.7%)
>
> The significant difference in performance when REPT is low and LOOP is
> high I think is due to the fact that the new implementation has a higher
> translation time, as when using a helper only 5 TCGop are used but with
> the patch a total of 10 TCGop are needed (Power lacks a direct mul_vec
> equivalent so this instruction is implemented with the help of 5 others,
> vmuleu, vmulou, vmrgh, vmrgl and vpkum).
>
> Signed-off-by: Lucas Mateus Castro (alqotel)<lucas.araujo@eldorado.org.br>
> ---
>   target/ppc/helper.h                 |  2 +-
>   target/ppc/insn32.decode            |  2 ++
>   target/ppc/int_helper.c             |  3 +-
>   target/ppc/translate.c              |  1 -
>   target/ppc/translate/vmx-impl.c.inc | 48 ++++++++++++++++++-----------
>   5 files changed, 35 insertions(+), 21 deletions(-)

Reviewed-by: Richard Henderson <richard.henderson@linaro.org>

r~