Re: [HACKERS] another locale problem

>>>Tom Lane said:
> Daniel Kalchev <daniel(at)digsys(dot)bg> writes:
> > To summarize the problem. If key contains (equivalent cyrillic
> > letters) 'ABC', 'ABCD', 'DAB' and 'ABX' and the query is:
>
> > SELECT key FROM t WHERE key ~* '^AB';
>
> > index scan will be used and the correct tuples ('ABC', 'ABCD' and
> > 'ABX') will be returned. If the query is
>
> > SELECT key FROM t WHERE key ~* '^ab';
>
> > index scan will be used and no tuples will be returned.
>
> Hm. Is it possible that isalpha() is doing the wrong thing on your
> machine? makeIndexable() currently assumes that isalpha() returns true
> for any character that is subject to case conversion, but I wonder
> whether that's a good enough test.

In fact, after giving it some though... the expression in gram.y

(strcmp(opname,"~*")
== 0 && isalpha(n->val.val.str[pos])))

is wrong. The statement in my view decides that a regular expression is not
indexable if it contains special characters or if it contains non-alpha
characters. Therefore, the statement should be written as:

(strcmp(opname,"~*")
== 0 && !isalpha((unsigned char)n->val.val.str[pos])))

(two fixes :) This makes indexes work for '^abc' (lowercase ASCII). But does
not find anything, which means regex does not work. It does not work for both
ASCII and non-ASCII text/patterns. :-(

> The other possibility is that regexp's internal handling of
> case-insensitive matching is not right.

I believe it to be terribly wrong, and some releases ago it worked with 8-bit
characters by just compiling it with -funsigned-char. Now this breaks things...

Daniel