[FFmpeg-devel] [PATCH] x86/hevc_sao: fix loading of RIP address

[Christophe Gisquet]

Hi,

2015-02-06 17:54 GMT+01:00 James Almer <jamrial at gmail.com>:
> pb_eo must be handled as a rip relative address for MSVC64, so an
> intermediate register is needed. Should fix link failures.

Seems ok on principle, passes fate on mingw64. I'm always wary of
those ABI, so if anyone could verify for msvc64 and unix64 (in
addition to yourself), that would be nice.

> +%if ARCH_X86_64
> +cglobal hevc_sao_edge_filter_%1_8, 4, 9, 8, dst, src, dststride, offset, eo, a_stride, b_stride, height, tmp
> +%define tmp2q heightq
>  %if WIN64
> -cglobal hevc_sao_edge_filter_%1_8, 4, 8, 8, dst, src, dststride, offset, a_stride, b_stride, height, tmp
> -%define  eoq heightq
>      movsxd           eoq, dword r4m
> -    movsx      a_strideq, byte [pb_eo+eoq*4+1]
> -    movsx      b_strideq, byte [pb_eo+eoq*4+3]
> -    imul       a_strideq, EDGE_SRCSTRIDE
> -    imul       b_strideq, EDGE_SRCSTRIDE
> -    movsx           tmpq, byte [pb_eo+eoq*4]
> -    add        a_strideq, tmpq
> -    movsx           tmpq, byte [pb_eo+eoq*4+2]
> -    add        b_strideq, tmpq
> -    mov          heightd, r6m
> -
> -%elif ARCH_X86_64
> -cglobal hevc_sao_edge_filter_%1_8, 5, 9, 8, dst, src, dststride, offset, eo, a_stride, b_stride, height, tmp
> -%define tmp2q heightq
> +%else
>      movsxd           eoq, eod
> +%endif
>      lea            tmp2q, [pb_eo]
>      movsx      a_strideq, byte [tmp2q+eoq*4+1]
>      movsx      b_strideq, byte [tmp2q+eoq*4+3]

The new common loading block could almost be abstracted in a single
macro (with the stride as parameter).

Something like:
%macro LOAD_EO_ARGS 1
%define tmp2q heightq
%if WIN64
    movsxd           eoq, dword r4m
%else
    movsxd           eoq, eod
%endif
    lea            tmp2q, [pb_eo]
    movsx      a_strideq, byte [tmp2q+eoq*4+1]
    movsx      b_strideq, byte [tmp2q+eoq*4+3]
    imul       a_strideq, %1
    imul       b_strideq, %1
    movsx           tmpq, byte [tmp2q+eoq*4]
    add        a_strideq, tmpq
    movsx           tmpq, byte [tmp2q+eoq*4+2]
    add        b_strideq, tmpq
    mov          heightd, r6m
%endmacro

(macro provided as example because it's probably clearer :D but not tested )

-- 
Christophe