bug#37633: Column part interpreted wrong in compilation mode

[Eli Zaretskii]

> From: Bernd Paysan <bernd@net2o.de>
> Cc: 37633@debbugs.gnu.org, anton@mips.complang.tuwien.ac.at
> Date: Sat, 05 Oct 2019 19:05:26 +0200
> 
> We can likely assume that the auto-detected encoding is the correct one, i.e. 
> buffer-file-coding-system can be used (the default for the optional encoding 
> system parameter for bufferpos-to-filepos and filepos-to-bufferpos).

Encoding of subprocess output is generally not auto-detected, it uses
the defaults derived from the locale.  I don't recommend
auto-detecting, because that's quite fragile (and is not needed here
anyway, IMO).

> Problem with precision: "exact" requires encoding the entire file, so it's 
> slow for large files.  Particularly with automatically generated files, this 
> is likely not acceptable, so "approximate" could be good enough.

We cannot use 'exact' here because there's no file per se: we only
have the compiler output.  We must use 'approximate'.