On 15/01/2019 02:52, John Naylor wrote:
The majority of cases are measurably faster, and the best case is at least 20x faster. On the whole I'd say this patch is a performance win even without further optimization. I'm marking it ready for committer.
I read through the patch one more time, tweaked the comments a little bit, and committed. Thanks for the review!
I did a little profiling of the worst case, where this is slower than the old approach. There's a lot of function call overhead coming from walking the string with pg_mblen(). That could be improved. If we inlined pg_mblen() into loop, it becomes much faster, and I think this code would be faster even in the worst case. (Except for the very worst cases, where hash table with the new code happens to have a collision at a different point than before, but that doesn't seem like a fair comparison.)
I think this is good enough as it is, but if I have the time, I'm going to try optimizing the pg_mblen() loop, as well as similar loops e.g. in pg_mbstrlen(). Or if someone else wants to give that a go, feel free.
- Heikki
