Brendan Jurd <direvus(at)gmail(dot)com> writes:
> Here's version 7.
Applied with a couple of corrections: the numeric case wasn't dealing
with NaNs in the same way as the float cases, and the int8 case was
converting to float8 which would lose precision. I made it go through
numeric instead, which is pretty expensive but I doubt this is worth
expending extra code on.
regards, tom lane