Re: [RFD] optimized versions of openssl

["Marcelo E. Magallon" <mmagallo@debian.org>]

>> Michael Stone <mstone@debian.org> writes:

 > Now for the real overachiever, what would be really cool is if you
 > hacked openssl to do *runtime* detection of which optimizations to use. 

 That would be indeed much better.  I blindly assumed he was talking
 about compiler flags and I further assumed that openssl for some reason
 exhibits an actual speedup thanks to those flags -- as opposed to the
 usual 1% or so.  If OpenSSL actually has hand-crafted processor
 specific optimizations runtime detection would be indeed the way to go.
 Even further, abstracting all this CPU detection stuff inside a library
 would be the coolest thing.  Something like:

 struct cpu_optimized_function do_foobar_table[] = {
     { CPUDETECT_GENERIC, do_foobar_generic },
#if defined(INTEL...)
     { CPUDETECT_SSE, do_foobar_sse },
     { CPUDETECT_MMX, do_foobar_mmx },
#endif
#if defined(SPARC...)
     { CPUDETECT_SPARCV8, do_foobar_sv8 },
#endif
     { NULL, NULL }
 };

 /* ... */

 do_foobar = pick_optimized_function(do_foobar_table);

 Even better, hide the conditional inside a macro:

 struct cpu_optimized_function do_foobar_table[] = {
     CPUDETECT_GENERIC_ENTRY(do_foobar_generic)
     CPUDETECT_SSE_ENTRY(do_foobar_sse)
     CPUDETECT_MMX_ENTRY(do_foobar_mmx)
     CPUDETECT_SV8_ENTRY(do_foobar_sv8)
     CPUDETECT_LAST_ENTRY
 };

 something along those lines...

 My guess is that this would foster runtime CPU detection since it
 becomes a no-brainer.

-- 
Marcelo             | "I can bite your leg if you like."
mmagallo@debian.org |         -- Gaspode the wonder dog
                    |            (Terry Pratchett, Moving Pictures)