bug#71655: Eshell external commands do not work under GNU Emacs for Windows

[Eli Zaretskii <eliz@gnu.org>]

> Date: Wed, 19 Jun 2024 22:34:02 -0700
> Cc: 71655@debbugs.gnu.org, james@literate-devops.io
> From: Jim Porter <jporterbugs@gmail.com>
> 
> On 6/19/2024 9:53 PM, Eli Zaretskii wrote:
> >> Date: Wed, 19 Jun 2024 12:40:12 -0700
> >> Cc: 71655@debbugs.gnu.org, james@literate-devops.io
> >> From: Jim Porter <jporterbugs@gmail.com>
> >>
> >> It's trying to find a shebang, which I guess(?) is so that Eshell can
> >> support shebangs on MS-Windows. What's strange is that 'file-readable-p'
> >> is non-nil, but 'insert-file-contents-literally' fails.
> > 
> > It fails because winget.exe is not a regular file, and
> > insert-file-contents barely supports non-regular files (and even that
> > almost exclusively on Posix systems).
> 
> 'file-regular-p' for that file is also non-nil. Should we change that?

Only if that is useful.  For now, I'm not sure I see a reason to do
that, since the code to support that will not be trivial, and will
have to include full support for these files in file-attributes and
similar APIs as well.

> > Not here.  The native cat.exe says "Invalid argument", just like
> > Emacs, and the one from MSYS says "Permission denied".  I get similar
> > errors from other utilities, for example wc.  And MSYS ls shows it as
> > a regular file of size zero.
> 
> My version of ls reports it as a symlink, interestingly enough.

It isn't a symlink.  It is a reparse point of type APPEXECLINK, which
has different attributes and different data structure describing the
target.  We could represent it as a symlink (since Posix has no direct
equivalent), but the implementation under the hood will need to be
different.

> > So my opinion on this is that Eshell should really skip reading files
> > whose size is zero when it looks for an interpreter, since we will
> > never find anything useful that way.
> 
> Well, I don't think size=0 is the only way that we could skip over files 
> like this. If 'file-regular-p' or 'file-readable-p' returned nil, we 
> could use that to skip it. We could also skip files ending in ".exe". We 
> could skip files that signal from 'insert-file-contents-literally'.

I agree that all those other conditions (including the .exe test) seem
to be reasonable, in addition to zero-size.

> I don't mind checking for size=0 if that's what we decide, but my 
> reading of the existing 'eshell-script-interpreter' suggests that it 
> should have already worked. If there's a bug in 'file-regular-p' (or 
> some other function Eshell uses here), I think it's work fixing it at 
> the source so that we squash the bug once and for all.

Fixing file-regular-p (and all the related APIs) for this purpose
sounds like a lot of work for little or no gain.  But if someone wants
to work on that and provide a clean patch, I don't mind.

> Otherwise, some other Lisp code might try to do something similar
> one day (maybe a Lisp version of file(1)?) and get bitten by this
> bug.

When they do, we'll have another situation to consider.

In general, you must understand that the depth and breadth of
emulating Posix assumptions and concepts on MS-Windows are driven by
practical needs, not by theoretical possibilities and potential
breakage that _might_ happen in some hypothetical Lisp.  Especially as
I don't quite see people with such patches lining up...

So if a simpler change in the (so far) single application which bumped
into this could fix the problem, I'm all for it.