diff options
| -rwxr-xr-x | libbb/hash_md5_sha_x86-64.S.sh | 22 |
1 files changed, 22 insertions, 0 deletions
diff --git a/libbb/hash_md5_sha_x86-64.S.sh b/libbb/hash_md5_sha_x86-64.S.sh index eef0095..901896e 100755 --- a/libbb/hash_md5_sha_x86-64.S.sh +++ b/libbb/hash_md5_sha_x86-64.S.sh @@ -6,6 +6,28 @@ # also contains the diff of the generated file. exec >hash_md5_sha_x86-64.S +# There is a way to use XMM registers (which always exist for x86-64!) for W[] +# For example, if we load W as follows: +# %xmm0: w[0x0] w[0x1] w[0x2] w[0x3] +# %xmm4: w[0x4] w[0x5] w[0x6] w[0x7] +# %xmm8: w[0x8] w[0x9] w[0xa] w[0xb] +# %xmm12: w[0xc] w[0xd] w[0xe] w[0xf] +# then the xor'ing operation to generate next W[0..3] is: +# movaps %xmm0, %xmmT2 +# palignr $0x8, %xmm4, %xmmT2 # form (w[0x2],w[0x3],w[0x4],w[0x5]) +# # Right-shifts xmm4:xmmT2 by 8 bytes. Writes shifted result to xmmT2. SSSE3 insn. +# movaps %xmm0, %xmmT13 +# palignr $0x4,%xmm0,%xmmT13 # form (w[0xd],w[0xe],w[0xf],w[0x0]) +# xmm0 = xmm0 ^ t2 ^ xmm8 ^ t13 +# xmm0 = rol32(xmm0,1) # no such insn, have to use pslld+psrld+or +# and then results can be extracted for use: +# movd %xmm0, %esi # new W[0] +# pextrd $1, %xmm0, %esi # new W[1] +# # SSE4.1 insn. Can use EXTRACTPS (also SSE4.1) +# pextrd $2, %xmm0, %esi # new W[2] +# pextrd $3, %xmm0, %esi # new W[3] +# ... but this requires SSE4.1 and SSSE3, which are not universally available on x86-64. + echo \ '### Generated by hash_md5_sha_x86-64.S.sh ### |