On Fri, Aug 14, 2009 at 10:38:58AM +0100, Andrew Haley wrote: > Kevin Kofler wrote: > > We need to provide "architectural defaults" for plain i686, even crappy > > ones, they just need to work at all. > > I think there's a valid case for making an exception to this: when a > package is an accelerated version of a particular library. That is, > when the basic functionality of a library is available in a i686 > Fedora package, but a special SSEx version of the library makes use of > faster instructions. The run-time linker on i386 can already pick up an SSE2-specific build of a shared library if it's available -- the i386 gmp package makes use of this. Of course, both builds have to provide the same interfaces for that to work, and I'm not sufficiently familiar with this set of libraries to know if they do. If they do, and if the libc maintainers think it's worth doing, this support could be extended to cover SSE- and SSE3-specific versions. We could then package these optimized versions of the library with the fallback works-everywhere version and let ld.so sort out the details at runtime. That's quite a few "if"s, and I'm not one of the people who'd have to actually do the work, so I'll shut up now. HTH, Nalin -- fedora-devel-list mailing list fedora-devel-list@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/fedora-devel-list
