Performance testing

Werner Koch
Fri, 22 Nov 2002 19:17:49 +0100

Hi Ivo,

can you please give us a short overview on the architecture specific
packaging things?  I noticed that the runtime linker first tries a
subdirectories of /usr/lib to get a CPU optimized version.  Is that
scheme somewhere described and do you know whether there is some
support in libtool planned?

I have tought for quite some time on how to do CPU specific
optimizations for the assembler code.  It is tricky because we can't
simply use a switch or a function pointer due to the extra overhead.
The only solution within libgcrypt would be to duplicate larger parts
of the code - something I am not keen to do: It increases complexity,
introduces errors and we have to find a portable way to figure out the
best CPU to use.

Building extra libraries or splitting libgcrypt into a general and a
performance critical part seems to be a far better idea.