[FFmpeg-devel] [PATCH 06/10] lavu/aes: add x86 AESNI optimizations

[Henrik Gramner]

I changed the asm a bit and made it about 1 cycle faster on Haswell
and slightly smaller (-48 bytes overall incl. alignment on 64-bit
Linux).

%macro AES_CRYPT 1
cglobal aes_%1rypt, 6,6,2
    shl      r3d, 4
    add      r5d, r5d
    add       r0, 0x60
    add       r2, r3
    add       r1, r3
    neg       r3
    pxor      m1, m1
    test      r4, r4
    je .block
    movu      m1, [r4] ; iv
.block:
    movu      m0, [r2+r3] ; state
%ifidn %1, enc
    pxor      m0, m1
%endif
    pxor      m0, [r0+8*r5-0x60]
    cmp      r5d, 24
    je .rounds12
    jl .rounds10
    aes%1     m0, [r0+0x70]
    aes%1     m0, [r0+0x60]
.rounds12:
    aes%1     m0, [r0+0x50]
    aes%1     m0, [r0+0x40]
.rounds10:
    aes%1     m0, [r0+0x30]
    aes%1     m0, [r0+0x20]
    aes%1     m0, [r0+0x10]
    aes%1     m0, [r0+0x00]
    aes%1     m0, [r0-0x10]
    aes%1     m0, [r0-0x20]
    aes%1     m0, [r0-0x30]
    aes%1     m0, [r0-0x40]
    aes%1     m0, [r0-0x50]
    aes%1last m0, [r0-0x60]
    test      r4, r4
    je .noiv
%ifidn %1, enc
    mova      m1, m0
%else
    pxor      m0, m1
    movu      m1, [r2+r3]
%endif
.noiv
    movu [r1+r3], m0
    add       r3, 16
    jl .block
%ifidn %1, enc
    test      r4, r4
    je .ret
    movu    [r4], m0
.ret:
%endif
    REP_RET
%endmacro

%if HAVE_AESNI_EXTERNAL
INIT_XMM aesni
AES_CRYPT enc
AES_CRYPT dec
%endif