On Mon, Feb 24, 2003 at 11:26:39AM -0500, Camm Maguire wrote: > Greetings! > > Richard Zidlicky <[EMAIL PROTECTED]> writes: > > > On Tue, Feb 18, 2003 at 04:46:46PM -0500, Camm Maguire wrote: > > > Greetings, and thanks for the insight. GCL does a subbuild of part of > > > GMP. With the default configure script canonical host of > > > m68k-unknown-linux-gnu, gmp3 adds -m68000 to the command line. > > > > this is slightly broken - there has never been 68000 support > > in Linux. > > > > > Perhaps the autobuilders should use m68020-unknown-linux-gnu as the > > > lowest common denominator? > > > > this would work but not very well on 68060 models though.. here > > are the options: > > a) use m68k-*linux WITHOUT -m68000 flag. Will loose some > > performance on 68020-40 models, nearly optimal for 68060 > > b) m68020-linux will loose *lots* of performance on 68060 models. > > c) compile 2 versions of the library/program for 20-40 or 60 > > models. > > > > (a) is pretty acceptable, (b) would be probably the worst alternative. > > (c) would be nice and I think it is time to use multilibs for m68k-linux. > > > > OK, this won't get done before the next release, but what I'd like to > do is 1) benchmark the bignum code and see if these issues make a big > difference, and 2) assuming yes to 1), implement a dynamic library > alternative system as is done in atlas/lapack/blas on Debian > GNU/Linux, say by building a libgcl_gmp.so. I don't know if you are > familiar with how this works, but I'd like your feedback if you are > interested. Basically, one can install several binary-compatible > versions of the same library, and have the running cpu select the best > one for its capabilities at run time via ld.so.conf. I've made a > policy proposal, but thus far it has not received much interest: > > http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=120418
I agree that something like this is necessary, on m68k most if not all applications that use gmp, ssl and similar are in desperate need to come in 68020-40 and 68060 flavours. Incidentally glibc itself includes the gmp "muladd" code that would require a similar treatment. Richard
