Michael Paquier <michael.paquier@gmail.com> writes:
> ... Please note that if you added
> a volatile cast to "result" as well, then compilation was able to
> complete and regression tests passed...
Yeah, that squares with my analysis: the problem with the existing test
is that the compiler is throwing away the function call because its
output is unused.
If I haven't seen something from Andres shortly, I'll push a fix.
regards, tom lane