Re: insensitive collations

	Peter Eisentraut wrote:

[v7-0001-Collations-with-nondeterministic-comparison.patch]

+GenericMatchText(const char *s, int slen, const char *p, int plen, Oid
collation)
 {
+  if (collation && !lc_ctype_is_c(collation) && collation !=
DEFAULT_COLLATION_OID)
+  {
+    pg_locale_t    locale = pg_newlocale_from_collation(collation);
+
+    if (locale && !locale->deterministic)
+      ereport(ERROR,
+	   (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+	    errmsg("nondeterministic collations are not supported for
LIKE")));
+  }

This test gets side-stepped when pattern_fixed_prefix() in selfuncs.c
returns Pattern_Prefix_Exact, and the code optimizes the operation by
converting it to a bytewise equality test, or a bytewise range check
in the index with Pattern_Type_Prefix.

Here's a reproducer:

===
create collation ciai (locale='und-u-ks-level1', deterministic=false,
provider='icu');

create table w(t text collate "C");

insert into w select md5(i::text) from generate_series(1,10000) as i;
insert into w values('abc');

create index indexname on w(t );

select t from w where t like 'ABC' collate ciai;
 t 
---
(0 rows)

select t from w where t like 'ABC%' collate ciai;
 t 
---
(0 rows)

===

For the LIKE operator, I think the fix should be that like_fixed_prefix() 
should always return Pattern_Prefix_None for non-deterministic collations.

For regular expressions, pg_set_regex_collation() is called at some
point when checking for a potential prefix, and since it errors out with
non-deterministic collations, this issue is taken care of already.


Best regards,
-- 
Daniel Vérité
PostgreSQL-powered mailer: http://www.manitou-mail.org
Twitter: @DanielVerite