I wrote:
> Let me finish investigating the 2.x series and get back to you on that.
2.1 indeed seems to be the only version that emits the busted semicolon.
I found that 2.2 and 2.3 both fail one of their "make check" tests on my
machine --- if that's widespread it might explain a slow uptake rate for
these versions :-(. But the failure is in one of the GLR tests and
shouldn't affect us. All four 2.x versions produce gram.c files that
compile and pass regression for me.
So I'd vote for "allow 1.875 and up, except 2.1 specifically".
regards, tom lane
