On 09/22/2012 07:05 PM, Tom Lane wrote:
> I wrote:
>> ... It strikes me though that if
>> we're going to invent an opt_if_not_exists production in the grammar,
>> there are a lot of other places where it should be used too, for
>> consistency if nothing else.
> BTW, I tried to do this and realized that it doesn't work, because IF
> is not a reserved word. The only way that opt_if_not_exists isn't
> ambiguous is if it must appear before something that's not an
> identifier, which is to say it works in ALTER TYPE ADD VALUE ... Sconst
> and nowhere else. Otherwise you have to spell it out with duplicate
> productions so that bison doesn't have to make a shift/reduce decision
> till it's seen the whole phrase.
>
> If we're ever forced to make IF reserved for other reasons, we could
> clean up a lot of both IF EXISTS and IF NOT EXISTS productions.
>
> There are other ways we could refactor the productions involved to
> reduce duplication; for instance I think that we could make it work for
> CREATE TABLE IF NOT EXISTS by defining a nonterminal that expands to
> either "qualified_name" or "IF NOT EXISTS qualified_name". But that
> seems ugly enough to not be much of an improvement, not least because
> the nonterminal would need to return two separate pieces of info,
> and that's not terribly easy in bison.
>
>
:-(
I remember running into this when I did the DINE stuff. I was actually
pleasantly surprised that it worked with the enum command.
cheers
andrew
