On 26 April 2016 at 04:25, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> In short, these tests suggest that we need a coding pattern like
> this:
>
> volatile float8 asin_x = asin(x);
> return (asin_x / asin_0_5) * 30.0;
>
> We could drop the "volatile" by adopting -ffloat-store, but that
> doesn't seem like a better answer, because -ffloat-store would
> pessimize a lot of code elsewhere. (It causes a whole lot of
> changes in float.c, for sure.)
Agreed. That looks like the least hacky way of solving the problem. I
think it's more readable when the logic is kept local, and it's
preferable to avoid any compiler-specific options or global flags that
would affect other code.
6b1a213bbd6599228b2b67f7552ff7cc378797bf by itself looks like a
worthwhile improvement that ought to be more robust, even though it
didn't fix this problem.
Regards,
Dean
