On Wed, Jun 14, 2023 at 4:13 AM Peter Geoghegan <pg@bowt.ie> wrote:
On Tue, Jun 13, 2023 at 5:59 PM Thomas Munro <thomas.munro@gmail.com> wrote: > Trying to follow along here... you're doing the moral equivalent of > strxfrm(), so sort keys have the transitive property but direct string > comparisons don't? Or is this because LCIDs reach a different > algorithm somehow (or otherwise why do you need to use LCIDs for this, > when there is a non-LCID version of that function, with a warning not > to use the older LCID version[1]?)
I'm reminded of the fact that the abbreviated keys strxfrm() debacle (back when 9.5 was released) was caused by a bug in strcoll() -- not a bug in strxfrm() itself. From our point of view the problem was that strxfrm() failed to be bug compatible with strcoll() due to a buggy strcoll() optimization.
I believe that strxfrm() is generally less likely to have bugs than strcoll(). There are far fewer opportunities to dodge unnecessary work in the case of strxfrm()-like algorithms (offering something like ICU's pg_strnxfrm_prefix_icu() prefix optimization is the only one). On the other hand, collation library implementers are likely to heavily optimize strcoll() for typical use-cases such as sorting and binary search. Using strxfrm() for everything is discouraged [1].
Yes, I think the situation is quite similar to what you describe, with its WIN32 peculiarities. Take for example the attached program, it'll output:
s1 = s2 s2 = s3 s1 > s3 c1 > c2 c2 > c3 c1 > c3
As you can see the test for CompareStringEx() is broken, but we get a sane answer with LCMapStringEx().