bug#39137: 26.3; ispell-really-hunspell sometimes fails in init

[Eli Zaretskii <eliz@gnu.org>]

> From: Bob Rogers <rogers@modulargenetics.com>
> Date: Tue, 14 Jan 2020 16:25:07 -0500
> 
>    Spell checking can die with "(wrong-type-argument stringp nil)" when
> ispell-really-hunspell is true.  To reproduce (if it reproduces at all
> on your system), just do "emacs -Q" with Hunspell installed and M-$ on
> any of the words in the startup screen.

Granted, this doesn't reproduce on my system.  And I have difficulty
understanding how could it reproduce on _any_ system, especially in
"emacs -Q".

>    On some of the systems I use, it fails reliably when initializing; on
> others, it works reliably.  All systems are running the openSUSE Leap
> 15.1 release of GNU/Linux, fully updated, with the identical locally-
> built RPM version of Emacs 26.3 (with no local patches) and the distro
> Hunspell 1.6.2.  The key seems to be whether ispell-dictionary-alist is
> initialized before ispell-start-process is called

Can you step through the ispell.el's initialization code, and explain
how come ispell-dictionary-alist ends up being uninitialized by the
time ispell-start-process is called for spell-checking?  This
shouldn't happen.

>   ;; Parse hunspell affix file if using hunspell and entry is uninitialized.
>   (if ispell-really-hunspell
>       (or (cadr (assoc ispell-current-dictionary ispell-dictionary-alist))
> 	  (ispell-hunspell-fill-dictionary-entry ispell-current-dictionary)))
> 
> It appears that "nil" is valid for ispell-current-dictionary and means
> "default", but ispell-hunspell-fill-dictionary-entry expects only a
> string.  On that basis, I would claim that there is a bug that has
> perhaps been hidden by being on a rare execution path.  But if
> ispell-dictionary-alist is already initialized with a "nil" entry, then
> ispell-hunspell-fill-dictionary-entry never gets called, leading to the
> "reproducible irreproducibility" of this bug.  And if
> ispell-dictionary-alist should *always* be initialized and contain the
> current entry, then this is dead code.

It isn't dead code, it just assumes that either
ispell-current-dictionary is nil and there's a nil entry in
ispell-dictionary-alist, or it is a string that specifies a new
dictionary, which is then used to update ispell-dictionary-alist.

We need to establish how come the first assumption is false on those
affected systems.

Thanks.