bug#33653: 27.0.50; Change Gnus obarrays-as-hash-tables into real hash tables

[Eric Abrahamsen]

On 03/27/19 13:54 PM, Katsumi Yamaoka wrote:
> On Tue, 26 Mar 2019 14:44:32 -0700, Eric Abrahamsen wrote:
>> Thanks for this report. When you have a moment will you please try the
>> attached patch and see if it fixes the problem?
>
> Works great.  Now I can enter and read the nnml:テスト group.
> Thank you.  But there seem to be some more fix needed.
>
> The group level of nnml:テスト is 1 and there are some unread
> articles in it, however the group is not listed in the Group
> buffer when I launch Gnus by `C-u 1 M-x gnus RET' (it also
> happened when I worked on non-ASCII group names years ago, but
> I don't recall how I fixed it, sorry).  Typing L helps it.

I think I've found the source of the problem. The error you and others
are seeing comes when the group has an entry in gnus-newsrc-hashtb (ie,
you're subscribed to the group) but not in gnus-active-hashtb (Gnus
can't tell if the group has articles).

Gnus writes "nnml:テスト" to .newsrc.eld as
"nnml:\343\203\206\343\202\271\343\203\210", and that's used as the key
in gnus-newsrc-hashtb. But when it reads the active info from nnml, the
(unibyte) buffer contains:

nnml:\343\203\206\343\202\271\343\203\210

As a symbol, not a string. My current code then uses `symbol-name' to
turn it into a string, but it's crucially *not the same string* as the
one saved to.newsrc.eld, even though it looks the same in edebug, the
*scratch* buffer, etc.

Hence a mismatch between the two keys in the two hash-tables, and things
going haywire.

Obviously this arises from a misunderstanding on my part about how
bytes and strings interact. The code in question is at line 2145 in
gnus-start.el -- it uses `read' to pick pieces out of the buffer.

Right now my best idea is to swap out `read' for a some sort of regexp
crawl, but I am open to suggestions.

Thanks,
Eric