Re: [bug fix] Savepoint-related statements terminates connection

I wrote:
> ... PFA a patch
> that invents a notion of an "implicit" transaction block.

On further consideration, I think the control logic I added in
exec_simple_query() is a shade bogus. I set it up to only force
an implicit transaction block when there are at least two statements
remaining to execute. However, that has the result of allowing, eg,

begin\; select 1\; commit\; vacuum;

Now in principle it's perfectly OK to allow that, since the vacuum
is alone in its transaction. But it feels more like an implementation
artifact than a good design. The existing code doesn't allow it,
and we might have a hard time duplicating this behavior if we ever
significantly rewrote the transaction infrastructure. Plus I'd hate
to have to explain it to users. I think we'd be better off enforcing
transaction block restrictions on every statement in a multi-command
string, regardless of the location of any COMMIT/ROLLBACK within the
string.

Hence, attached a v2 that does it like that. I also fully reverted
4f896dac1 by undoing its changes to PreventTransactionChain; other
than that, the changes in xact.c are the same as before.

regards, tom lane