Re: In PG12, query with float calculations is slower than PG11

... and pushed.  One other change I made beyond those suggested
was to push the zero-divide ereport's out-of-line as well.

I did not do anything about adding unlikely() calls around the
unrelated isinf tests in float.c.  That seemed to me to be a separate
matter, and I'm not quite convinced it'd be a win anyway.

            regards, tom lane