Peter Eisentraut <pe...@eisentraut.org> writes:
> but after the recent improvements to pg_locale_t handling, we don't need 
> all three anymore.  All the information we have is contained in 
> pg_locale_t, so we just need to keep that one.  This allows us to 
> structure the locale-using regular expression code more similar to other 
> locale-using code, mainly by provider, avoiding another layer that is 
> specific only to the regular expression code.  The first patch 
> implements that.

I didn't read that patch in detail; somebody who's more familiar than
I with the recent locale-code changes ought to read it and confirm
that no subtle behavioral changes are sneaking in.  But +1 for
concept.

> The second patch removes a call to pg_set_regex_collation() that I think 
> is unnecessary.

I think this is actively wrong.  pg_regprefix is engaged in
determining whether there's a fixed prefix of the regex, which
at least involves a sort of symbolic execution.  As an example,
whether '^x' has a fixed prefix surely depends on whether the locale
is case-insensitive.  (It may be that we get such cases wrong today,
since pg_regprefix was written before we had ICU locales and I don't
know if anyone has revisited it with this in mind.  But removing this
pg_set_regex_collation call is surely not going to make that better.
In any case, the gain of removing it must be microscopic.)

> (I don't have any plans to get rid of the remaining global variable. 
> That would certainly be nice from an intellectual point of view, but 
> fiddling this into the regular expression code looks quite messy.  In 
> any case, it's probably easier with one variable instead of three, if 
> someone wants to try.)

Yeah.  Those global variables are my fault.  I did try hard to avoid
having them, but came to the same conclusion that it was not worth
contorting the regex code to pass a locale pointer through it.
Maybe if we ever completely give up on maintaining code similarity
with the Tcl version, we should just bull ahead and do that; but for
now I don't want to.

                        regards, tom lane


Reply via email to