It seems most mips64 hardware don't have FPU, yet one of
our tests test/bench/go1 is very heavy on floating point, so
even on a 2GHz mips64 core, that test needs >10mins alone,
whereas the whole all.bash completes in ~30mins. The problem
is that we're using kernel to trap and emulate every single FP
instruction, which has extremely large overhead.

I'm wondering if we should introduce softfloat for MIPS64 like
what we did for GOARM=5.

Implementation problems aside, the biggest question is how
to enable or disable softfloat on the toolchain level? Should we
introduce a GOMIPS environment variable? Or should we just
always use insert calls to _sfloat before any consecutive FP
instructions, and make _sfloat does nothing if the system has
a FPU? I don't like the way GOARM works and I think it confuses
more than it helped. (Esp. modern ARM chips usually contain
FPU whereas very few MIPS64 chips contain FPU.)

/cc @cherrymui