According to the spec, ideal floats do not have a -0. However, -0 is representable (and used) by the compiler's internal ideal float implementation. Fix the implementation to get rid of -0 (or hide it from the rest of the compiler).

See issue #12576

The compiler uses cmd/compile/internal/gc/mpgetflt{32} to do the conversion. Maybe a check inside this routine is enough?

We've had to sprinkle +0.0 around to get around this problem, like this:
s.constFloat32(n.Type, mpgetflt32(f)+0.0)