Did you use the -O3 and --check-bounds=no flags, as well as rebuild the system image for your architecture?
Actually, almost half the time is being spent in log, which should get a lot better soon by switching from the openlibm implementation to a pure Julia implementation: https://github.com/JuliaLang/julia/pull/24750. You can actually already use the new algorithm in 0.6.2 by replacing log with Base.Math.JuliaLibm.log.