143356876
Denys Vlasenko
2022-01-08
Subject
libbb/sha1: add a comment
Binary
No binary (build failed)
Architecture & Compiler Flags
—
Build Log
(log yok)
Commit Diff GitHub →
libbb/hash_md5_sha_x86-64.S.sh+22 -0
| @@ -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 ### | |