On 03/01/11 06:36, Vince Weaver wrote: > > On Mon, 28 Feb 2011, Gabe Black wrote: >> Second, this is actually an interesting coincidence because I was just about >> to start looking at getting spec2006 to work better. A guy named Vince >> Weaver did some work testing our x86 implementation a while ago, and he has >> some patches that haven't yet made it into the code base. You can find those >> here: >> >> http://www.csl.cornell.edu/~vince/projects/m5/ > I'm still lurking here. I've been meaning to clean up those pages and > make sure the patches still compile, just haven't had time. > > One issue is that I think it was decided that the syscall code needed a > major re-write, so there was some resistance to merging syscall emulation > patches that didn't fix the underlying problem. > > The other main issue with spec2k/spec2k6 is x87 support. By compiling > with -msse2 to force SSE code generation you can avoid most of the current > unimplemented instruction errors, but there are a few x87 instructions > (mostly transcendental functions) that will still be generated. My > patches address enough of those to get spec2k going, but I don't think an > official set of x87 uops were ever specified, which makes the minimal x87 > supoort a bit of a hack. > > Vince > _______________________________________________ > m5-users mailing list > [email protected] > http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
I'm pretty sure gcc uses sse by default when it's building a 64 bit binary because it assumes if a CPU can run 64 bit code it also has SSE. If you're running 32 bit binaries (which I don't necessarily recommend but should work) then using -msse2 is a good idea. Spec2k has been working fine for a while now, at least for the binaries I could get to build in a modern environment, and are part of our nightly regressions. It could be, though, that recompiling them with a more recent gcc and/or glibc would pull in some of these system calls or instructions which aren't supported. In any case, it would be good to get those patches on review board. There are a few issues with some of them, but once those are fixed I'd like to get them into the tree. Gabe Gabe _______________________________________________ m5-users mailing list [email protected] http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
