implement double-to-float narrowing in the x86 NCG (#4441)