Fix another miscompilation exposed by lencode, where we lowered i64->f32

conversions to __floatdidf instead of __floatdisf on targets that support
f32 but not i64 (e.g. sparc).

llvm-svn: 26254