Re: [feature/dll-only-windows] A new windows build, comments wanted

[Phillip Lord <phillip.lord@russet.org.uk>]

Eli Zaretskii <eliz@gnu.org> writes:

>> From: Phillip Lord <phillip.lord@russet.org.uk>
>> Date: Sat, 09 Jan 2021 19:57:00 +0000
>> 
>> So I have tried a new technique which is now on
>> feature/dll-only-windows. Essentially, I just name all of the DLLs that
>> Emacs uses directly and then figure out any dependencies of these using
>> `ntldd`. I've tried `ntldd` directly on emacs.exe which does not
>> work.
>
> Of course, it won't: we load all the DLLs dynamically at run time, we
> don't link against their import libraries.  The only exceptions I know
> of are (1) GMP, and (2) libgccjit (in the native-comp build).

Indeed.

>> The practical upshot of this is that the dependencies file is just
>> a pile of DLLs.
>
> Did you verify that the DLLs include all of _their_ dependencies?  If
> yes, how did you do that?

It's totally dependent on ntldd being correct, of course. It comes with
a --recursive option which I used rather than reimplementing it
myself. As ever, the only tests I run on the windows package for Emacs
are to unpack, start and then run "w32-feature.el" from etc. These all
pass.



>> 2) harfbuzz is currently not working on i686.
>
> Is that a bug in the MinGW64 HarfBuzz port, or is that a bug in Emacs?
> I'm using a 32-bit Emacs build with HarfBuzz all the time, but it's
> HarfBuzz I built myself (it's available from the ezwinports site).

I don't know, I haven't debugged it yet:-)

>> Before I fix this, I think it is worth asking whether I still need
>> to produce a i686 binary.
>
> I cannot tell you what to do, but it would be nice to have binaries
> that can be run on older Windows versions.  So if the build supports
> XP and older Windows, keeping the 32-bit build would be an advantage.

I know that you use a very old windows, but I doubt you use my packages!
I have no other data of how many people this would impact, but my guess
would be not many, so I am asking for opinions.

One solution could be to stop making the 32-bit build and see if anyone
complains. It could be restored, either before or after the Emacs-28
release.


>> 3) Currently the "no-deps" version actually includes libXpm. Emacs
>> starts without it, but looks ugly. I would like to no longer special
>> case libXpm and just make the "no-deps" download really include no
>> deps. I think this is reasonable, because this download is now, really
>> special purpose and "with-deps" is the default.
>
> If we believe no one will want the no-deps download, why have it at
> all?  If we think someone will want it, I don't think they should be
> punished by having BW icons on the tool bar.

Indeed. My guess is that currently if people use -no-deps they do so
because they want the smaller download. This change, of course, we
reduce the difference and regardless the installer version is smaller
still (by a half).

Given all that, the only real audience for the -no-deps version would be
those who have a mingw64 installation already. They won't get BW icons
(assuming that libXpm is installed). Of course, that audience is likely
to be technically skilled and they could just delete the DLLs in the
"with-deps" download.

On balance, therefore, I would say we don't need the -no-deps version
and it would clean up the download site and reduce confusion.


>> 4) Currently, native-comp is an "--with-nativecomp" option even on the
>> native-comp branch. It is likely to be merged this way to master?
>
> Yes, I think so.

Are you worried about stability or newness? I guess it will become the
default at some point. It seems to have few disadvantages, other than
hammering the CPU at bit at initial start up.


>
>> Currently, I build the Windows distribution of Emacs with all the
>> default options.
>
> Which non-default options of practical importance does that leave out?

Currently, none. If --with-nativecomp is not default, then, I think it
would become the example.

>> I can make an alternative release "--with-nativecomp" but needs some
>> effort; it also does not address the question of how I should build
>> Emacs for Windows when a full release of Emacs-28 happens. If, the
>> full release of Emacs-28 will be native-comp, I'd rather start
>> building snapshots with it as soon as it is merged.
>
> I think you should build --with-nativecomp.  People can always
> uninstall libgccjit or rename it if they don't want to use native
> compilation.

Okay. I guess "make NATIVE_FULL_AOT=1" would be the thing as well,
depending on how long it takes.


>> 5) Why is it "--with-nativecomp", shouldn't it be "--with-native-comp".
>
> Probably, but we didn't yet get to splitting such thin hair in that
> branch ;-)

Yeah, can't help it. Splitting hairs is my job.

Phil