CSV parsing and other issues (Re: LC_NUMERIC)

[Maxim Nikulin]

On 03/06/2021 01:54, Boruch Baum wrote:
> Please consider having the elisp 'format' function adopt the
> single-quote and 'I' flags. Each is already implemented in both the GNU
> C printf command and the linux printf command. The single-quote flag is
> part of the 'Single UNIX Specification' and the 'I' flag has been part
> of glibc since version 2.2 [ref: man(3) printf].
>
> If function 'format' uses 'printf' as its backend, this would seem to be
> a matter of exposing an existing feature.

I do not know the story why Emacs does not support locale-aware number 
formats, but I suspect that relying on libc is opening a can of worms. 
Once setlocale(LC_NUMERIC, "") is invoked, one is never sure if printf- 
and scanf-like functions deal with default "C" representation or with 
formatted accordingly to current locale numbers. Some numbers related to 
communication protocols must be always formatted using "C" locale. I do 
not remember if it happened with XFree86 or with Xorg, but at certain 
moment users experienced problems. X11 could not start at all due to 
invalid configs. The source of problem was "," as decimal separator in 
some locales and wrong expectations concerning numbers in config files.

Recently I found the following fixup_locale function:
http://git.savannah.gnu.org/cgit/emacs.git/tree/src/emacs.c#n2861

    setlocale (LC_NUMERIC, "C");

I was surprised that impossible to determine current decimal separator 
from elisp. At the same time e.g. `string-collate-lessp' has LOCALE 
argument.

A month ago some patches were submitted to Org mode with intention to 
improve import of tables, see https://debbugs.gnu.org/47885 A part of 
discussion is missed in the bug tracker: 
https://lists.gnu.org/archive/html/emacs-orgmode/2021-05/msg00693.html 
Org mode has a piece of code that tries to guess if the file has commas 
or tabs as field separator (CSV or TSV format). The suggested change 
adds e.g. semicolon. (Sidenote: probably csv-mode is a better place than 
org-mode for such code.)

The problem is that office software uses semicolon for locales where 
comma serves as decimal separator for floating-point numbers (e.g. 
de_DE, es_ES, fr_FR, ru_RU, etc.):

    A;1,2;3,4

So semicolon should be tried with higher priority than comma if in 
current locale numbers are represented as e.g 1,2. Unfortunately the 
only way to get such information from Emacs is to call some external 
application. Maintaining own mapping of locale to separator is 
unnecessary burden.

Besides office software, there are some equipment that always use "C" 
number formatting, so a user can have a mix of files with various 
dialects of CSV. Thus locale info is not enough, some heuristics is 
required anyway.

More subtle questions rise on the next step. Org allows to perform 
calculations on table cells (and there is calc). Should numbers be 
converted to "C" locale representation during import? Should conversion 
happen when passing cell content as argument and the result converted 
back to current locale? I anticipate that buffer-local setting will be 
requested. There was even discussion of mixed-language documents in 
emacs-orgmode mail list, however numbers were not mentioned.

So locale-aware number formatting would be a great improvement for 
Emacs. On the other hand, it should be implemented with great care to 
avoid localized numbers in some cases. Maybe locale argument should be 
passed to functions that deal with numbers. Formatting of integer 
numbers is not enough, floating point numbers should be handled as well. 
Parsing numbers formatted accordingly to locale rules should be 
addressed too. A function similar to `locale-info' is highly desired to 
get properties of locale (e.g. decimal_point from result of localeconv). 
Some decision is required whether calc & Co should operate with
localized numbers.