bug#40407: [PATCH] slow ENCODE_FILE and DECODE_FILE

[Eli Zaretskii]

> From: Mattias Engdegård <mattiase@acm.org>
> Date: Sun, 5 Apr 2020 17:35:55 +0200
> Cc: 40407@debbugs.gnu.org
> 
> 5 apr. 2020 kl. 17.03 skrev Mattias Engdegård <mattiase@acm.org>:
> 
> > Actually they can, as far as I can tell. Have a look yourself.
> 
> To clarify, calls to {EN,DE}CODE_FILE are probably safe by design since they 
> already return their argument in several cases.

In theory, yes.  In practice, not really.  The cases where they return
their argument are those when we didn't yet set up any encoding for
file names.  This happens only very early into startup, and frankly,
we have nothing else to do at that point.

Once we do set up default-file-name-coding-system, these macros will
never return their argument (unless someone forcefully sets the
encoding to nil, in which case they deserve what they get).  Do you
agree?

> Some calls to code_convert_string_norecord are not safe because they are used 
> with auto lisp strings; I'm going through them to find out just which ones. 
> This can be done piecemeal.

I'm okay with making the safe cases faster, but we'd need to clearly
comment each one, because later changes might make them unsafe.  Any
code that uses the un-encoded string after encoding it, or the
un-decoded string after decoding it, could become broken if these two
are the same string.

And let's please keep in mind that on most modern platforms in most
use cases default-file-name-coding-system is utf-8, so encoding is
fast, and thus we don't need to go overboard here.  IOW, if there's a
doubt, there's no doubt.

Thanks.