gold speed for small programs

[Pádraig Brady]

On 14/12/10 07:04, Ralf Wildenhues wrote:
> Trying out gold (no LTO) on GCC's cc1, a log shows that the single
> biggest part (20%) of user time for such a medium sized program may be
> spent in SHA1 computation for the build id.  I think the SHA1 code in
> libiberty would benefit from an SSE implementation, such as the one
> found in git or described on <http://en.wikipedia.org/wiki/SHA-1>.
> 
> Adding bug-gnulib in Cc: because that's where the libiberty SHA1
> implementation comes from.

Linus' portable SHA1 implementation is much faster:
http://lists.gnu.org/archive/html/bug-coreutils/2009-08/threads.html#00118
There was talk of licensing changes, but nothing came of that.

There was also talk of an SSE version:
http://lists.gnu.org/archive/html/bug-coreutils/2010-04/msg00107.html
Although architecture specific, it's probably worth including given
the ubiquity of the architecture and the performance gain it provides.

cheers,
Pádraig.