On Thu, Feb 09, 2017 at 11:27:49PM +0100, Johannes Schindelin wrote: > 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. > > OpenSSL 1.0.2 has made considerable performance improvements and > support the Intel hardware acceleration features. See: > https://software.intel.com/en-us/articles/improving-openssl-performance > https://software.intel.com/en-us/articles/intel-sha-extensions > > To test this I added/staged a single file in a gigantic > repository having a 450MB index file. The code in read-cache.c > verifies the header SHA as it reads the index and computes a new > header SHA as it writes out the new index. Therefore, in this test > the SHA code must process 900MB of data. Testing was done on an > Intel I7-4770 CPU @ 3.40GHz (Intel64, Family 6, Model 60) CPU. I think this is only half the story. A heavy-sha1 workload is faster, which is good. But one of the original reasons to prefer blk-sha1 (at least on Linux) is that resolving libcrypto.so symbols takes a non-trivial amount of time. I just timed it again, and it seems to be consistently 1ms slower to run "git rev-parse --git-dir" on my machine (from the top-level of a repo). 1ms is mostly irrelevant, but it adds up on scripted workloads that start a lot of git processes. Whether it's a net win or not depends on how much sha1 computation you do in your workload versus how many processes you start. I don't know what that means for Windows, though. My impression is that process startup is so painfully slow there that the link time may just be lost in the noise. It may just always be a win there. So not really an objection to your patch, but something you may want to consider. (Of course, it would in theory be possible to have the best of both worlds either by static-linking openssl, or by teaching block-sha1 the same optimizations, but both of those are obviously much more complex). -Peff