Today I'm asking the #Gentoo arch testers to stop testing stuff using 387 arithmetic. Yep, the one that causes random differences in rounding by using 80-bit registers (vs 64 bits for a regular double), and therefore spams us with useless test errors. Sure, the test suites are broken in the first place by expecting exact results but many upstreams just don't care — and we'd rather focus on real issues. I mean, too often they don't even care about 32-bit arches at all, and bothering them about ancient FPU won't help.

That said, we've already switched the 32-bit multilib builds on amd64 to use `-mfpmath=sse`. The next step would to do the same in new #x86 profiles. While at it, we're also going to need to raise the baseline to SSE2 (e.g. `-march=pentium-m`, `-march=pentium4` or just `-msse2`).

mail-archive.com/gentoo-dev@li