Re: [Monotone-devel] Re: zlib vs gzip

[Jack Lloyd]

On Wed, Oct 31, 2007 at 08:00:11AM -0700, Zack Weinberg wrote:

> I'd like to point out that the #1 easiest thing anyone could do right
> now to speed up monotone is get hand-optimized assembly SHA1
> implementations into our copy of botan.  That one function shows up as
> at least 7% of runtime every single time I profile it, doesn't matter
> what the task is.

To confirm: that one function == SHA_160::hash() ? Or calculate_ident?

Do you see a measurable change here if you switch to the (current) asm
version and/or using the OpenSSL engine? Do you see a change if you
manually unroll the two loops in SHA_160::hash?

What CPU(s) are you profiling on? What compiler and version?

There is no obvious (to me: suggestions are welcome) way of making the
C++ code faster, except taking advantage of x86(-64)'s bswap
instruction and ability to do unaligned loads (which has been written
and lightly tested but not checked in).

-Jack