At Wed, 14 Oct 2020 23:06:28 -0400, Tom Lane <tgl@sss.pgh.pa.us> wrote in
> John Naylor <john.naylor@enterprisedb.com> writes:
> > With those points in mind and thinking more broadly, I'd like to try harder
> > on recomposition. Even several times faster, recomposition is still orders
> > of magnitude slower than ICU, as measured by Daniel Verite [1].
>
> Huh. Has anyone looked into how they do it?
I'm not sure it is that, but it would be that.. It uses separate
tables for decomposition and composition pointed from a trie?
That table is used after trying algorithmic decomposition/composition
for, for example, Hangul. I didn't look it any fruther but just for
information, icu4c/source/common/normalizer2impl.cpp seems doing that.
For example icu4c/srouce/common/norm2_nfc_data.h defines the static data.
icu4c/source/common/normalier2impl.h:244 points a design documentation
of normalization.
http://site.icu-project.org/design/normalization/custom
> Old and New Implementation Details
>
> The old normalization data format (unorm.icu, ca. 2001..2009) uses
> three data structures for normalization: A trie for looking up 32-bit
> values for every code point, a 16-bit-unit array with decompositions
> and some other data, and a composition table (16-bit-unit array,
> linear search list per starter). The data is combined for all 4
> standard normalization forms: NFC, NFD, NFKC and NFKD.
regards.
--
Kyotaro Horiguchi
NTT Open Source Software Center