block-sha1: improved SHA1 hashing
authorLinus Torvalds <torvalds@linux-foundation.org>
Sat, 8 Aug 2009 04:16:46 +0000 (21:16 -0700)
committerJunio C Hamano <gitster@pobox.com>
Sat, 8 Aug 2009 05:32:46 +0000 (22:32 -0700)
commit66c9c6c0fbba0894ebce3da572f62eb05162e547
tree44ca7ad4f2ae1ac18e8f03d2746d791fcec51c32
parent30d12d4c16abc052e8961c07651f97bea2c061bd
block-sha1: improved SHA1 hashing

I think I have found a way to avoid the gcc crazyness.

Lookie here:

#             TIME[s] SPEED[MB/s]
rfc3174         5.094       119.8
rfc3174         5.098       119.7
linus           1.462       417.5
linusas         2.008         304
linusas2        1.878         325
mozilla         5.566       109.6
mozillaas       5.866       104.1
openssl         1.609       379.3
spelvin         1.675       364.5
spelvina        1.601       381.3
nettle          1.591       383.6

notice? I outperform all the hand-tuned asm on 32-bit too. By quite a
margin, in fact.

Now, I didn't try a P4, and it's possible that it won't do that there, but
the 32-bit code generation sure looks impressive on my Nehalem box. The
magic? I force the stores to the 512-bit hash bucket to be done in order.
That seems to help a lot.

The diff is trivial (on top of the "rename registers with cpp" patch), as
appended. And it does seem to fix the P4 issues too, although I can
obviously (once again) only test Prescott, and only in 64-bit mode:

#             TIME[s] SPEED[MB/s]
rfc3174         1.662       36.73
rfc3174          1.64       37.22
linus          0.2523       241.9
linusas        0.4367       139.8
linusas2       0.4487         136
mozilla        0.9704        62.9
mozillaas      0.9399       64.94

that's some really impressive improvement. All from just saying "do the
stores in the order I told you to, dammit!" to the compiler.

Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
block-sha1/sha1.c