bug#63288: 30.0.50; Emacs 30 packages fail to build with native comp on some machines

bug#63288: 30.0.50; Emacs 30 packages fail to build with native comp on some machines

[Brian Leung]

In https://github.com/nix-community/emacs-overlay/issues/318, we
received a report about failed external-package builds when using
Nix, which has some additional plumbing including
https://github.com/nixos/nixpkgs/blob/abcc3146aeb06b8d2ded75bddac3a63544a6be7e/pkgs/build-support/emacs/melpa2nix.el#L26-L32.

It was observed that reverting
https://git.savannah.gnu.org/cgit/emacs.git/commit/?id=ea9831bb3cb4878273f6f848051c9b8c3c76d5f1
would resolve our issue.

On my own machine, I don't always encounter the build failure (happened
to me once a few weeks ago, and then again today).

bug#63288: 30.0.50; Emacs 30 packages fail to build with native comp on some machines

[Eli Zaretskii]

> From: Brian Leung <leungbk@posteo.net>
> Date: Fri, 05 May 2023 03:59:22 +0000
> 
> In https://github.com/nix-community/emacs-overlay/issues/318, we
> received a report about failed external-package builds when using
> Nix, which has some additional plumbing including
> https://github.com/nixos/nixpkgs/blob/abcc3146aeb06b8d2ded75bddac3a63544a6be7e/pkgs/build-support/emacs/melpa2nix.el#L26-L32.
> 
> It was observed that reverting
> https://git.savannah.gnu.org/cgit/emacs.git/commit/?id=ea9831bb3cb4878273f6f848051c9b8c3c76d5f1
> would resolve our issue.
> 
> On my own machine, I don't always encounter the build failure (happened
> to me once a few weeks ago, and then again today).

Adding Mattias.

bug#63288: 30.0.50; Emacs 30 packages fail to build with native comp on some machines

[Mattias Engdegård]

>> In https://github.com/nix-community/emacs-overlay/issues/318, we
>> received a report about failed external-package builds when using
>> Nix, which has some additional plumbing including
>> https://github.com/nixos/nixpkgs/blob/abcc3146aeb06b8d2ded75bddac3a63544a6be7e/pkgs/build-support/emacs/melpa2nix.el#L26-L32.
>> 
>> It was observed that reverting
>> https://git.savannah.gnu.org/cgit/emacs.git/commit/?id=ea9831bb3cb4878273f6f848051c9b8c3c76d5f1
>> would resolve our issue.
>> 
>> On my own machine, I don't always encounter the build failure (happened
>> to me once a few weeks ago, and then again today).

That's remarkable -- not only is it difficult to see anything wrong with that change (ea9831bb3c), it actually reverts byte-code generation for `ignore` forms in general so that the bytecode generated for `package-read-from-string` on master is again identical to that on emacs-29.

Andrea, does the native compiler somehow miscompile package-read-from-string?

Since the failure appears to be nondeterministic, perhaps there is a native-comp build race involved?

bug#63288: 30.0.50; Emacs 30 packages fail to build with native comp on some machines

[Andrea Corallo]

Mattias Engdegård <mattiase@acm.org> writes:

>>> In https://github.com/nix-community/emacs-overlay/issues/318, we
>>> received a report about failed external-package builds when using
>>> Nix, which has some additional plumbing including
>>> https://github.com/nixos/nixpkgs/blob/abcc3146aeb06b8d2ded75bddac3a63544a6be7e/pkgs/build-support/emacs/melpa2nix.el#L26-L32.
>>> 
>>> It was observed that reverting
>>> https://git.savannah.gnu.org/cgit/emacs.git/commit/?id=ea9831bb3cb4878273f6f848051c9b8c3c76d5f1
>>> would resolve our issue.
>>> 
>>> On my own machine, I don't always encounter the build failure (happened
>>> to me once a few weeks ago, and then again today).
>
> That's remarkable -- not only is it difficult to see anything wrong
> with that change (ea9831bb3c), it actually reverts byte-code
> generation for `ignore` forms in general so that the bytecode
> generated for `package-read-from-string` on master is again identical
> to that on emacs-29.
>
> Andrea, does the native compiler somehow miscompile package-read-from-string?

Hi, (sorry for being late) not that I'm aware!  But in this case the
failer should be deterministic no?

> Since the failure appears to be nondeterministic, perhaps there is a native-comp build race involved?

Mmmh, or maybe related to the presence or not of some eln?

Best Regards

  Andrea

bug#63288: 30.0.50; Emacs 30 packages fail to build with native comp on some machines

[Mattias Engdegård]

>> Andrea, does the native compiler somehow miscompile package-read-from-string?
> 
> Hi, (sorry for being late) not that I'm aware!  But in this case the
> failer should be deterministic no?

Well I haven't studied the native compiler enough to know how many dice it rolls...

>> Since the failure appears to be nondeterministic, perhaps there is a native-comp build race involved?
> 
> Mmmh, or maybe related to the presence or not of some eln?

Could be. I'm going out on a limb here and declare the byte-code compiler completely innocent in this case.

Would you work with the original reporters so that they don't have to maintain some terrible hack that may be hiding a deeper bug? If anything turns out to be pointing in my direction I'll be happy to help out.