Am 13.02.2017 um 18:16 schrieb Johannes Schindelin:
On Sat, 11 Feb 2017, Johannes Sixt wrote:
Am 10.02.2017 um 00:41 schrieb Junio C Hamano:
Johannes Schindelin <johannes.schindelin@xxxxxx> writes:
From: Jeff Hostetler <jeffhost@xxxxxxxxxxxxx>
Use OpenSSL's SHA-1 routines rather than builtin block-sha1
routines. This improves performance on SHA1 operations on Intel
processors.
...
Signed-off-by: Jeff Hostetler <jeffhost@xxxxxxxxxxxxx>
Signed-off-by: Johannes Schindelin <johannes.schindelin@xxxxxx>
---
Nice. Will queue as jh/mingw-openssl-sha1 topic; it is a bit too
late for today's integration cycle to be merged to 'next', but let's
have this by the end of the week in 'master'.
Please don't rush this through. I didn't have a chance to cross-check the
patch; it will have to wait for Monday. I would like to address Peff's
concerns about additional runtime dependencies.
I never meant this to be fast-tracked into git.git. We have all the time
in our lives to get this in, as Git for Windows already carries this patch
for a while, and shall continue to do so.
I've been working with this patch for the past few days, and I did not
notice any disadvantage during interactive work even though there is a
new dependency on libcrypto.dll.
Here are some unscientific numbers collected during test suite runs:
bash -c "time make -j4 -k test"
with this patch:
real 34m47.242s
user 9m55.827s
sys 25m20.483s
without this patch:
real 34m2.330s
user 9m56.556s
sys 25m5.520s
It looks like BLK_SHA1 has some advantage, but I would not count on
these figures too much. (I certainly did not sit idly in front of the
workstation during these tests, for example. That may have skewed the
numbers somewhat.) (And, no, I'm not going to measure best-of-five
timings, not even best-of-two. ;)
In summary: Interactive response times do not decline noticably. I do
not object the patch.
-- Hannes