CI and tests

[Ambrevar]

There is a lot to backport from Nyxt in this area:

• Check for compilation warnings
• Check for unbound exported symbols.
• Validate make-instance symbols

Nyxt currently only compiles against SBCL and CCL.
But libraries could (and should?) be tested against ECL, ABCL, CLISP and Clasp.

In my experience, there are a few blockers though:

• ABCL 1.9 (the latest as of this writing) fails to compile FSet. It's fixed on master: Error compiling FSET armedbear/abcl#246
• CLISP does not support package-local-nicknames. There is a fix though: https://gitlab.com/gnu-clisp/clisp/-/merge_requests/3

[aartaka]

There's a piece of CI abstracted from Nyxt CI, on ntemplate. There also is a sensible template for testing. What I'd like to add is a benchmarking template akin to prompter benchmark from Nyxt.

And then, most implementations I've tried adding to CI—they all fail. ABCL and CLISP in particular. And ECL on Ubuntu is breaking with almost 100% probability.

I guess that this can be closed.

[Ambrevar]

Why do ABCL and CLISP fail? Do you have examples to share?

Something else that could be added: merge atlas-engineer/nyxt#2305, then update ntemplate to normalize the testing process of compilation warning checks, unexported symbol checks, etc.

[aartaka]

Why do ABCL and CLISP fail? Do you have examples to share?

CLISP: See this njson commit CI atlas-engineer/njson@4449701

ABCL: I haven't actually tested much of it, but there were problems with it at the installation stage IIRC. It refused to install basically.

Something else that could be added: merge atlas-engineer/nyxt#2305, then update ntemplate to normalize the testing process of compilation warning checks, unexported symbol checks, etc.

Yes!

[Ambrevar]

Note that we could use Guix to install CLISP and ABCL.

CLISP will fail loading local package nicknames. Unless we compile this fork instead https://gitlab.com/gnu-clisp/clisp/-/merge_requests/3

[aartaka]

Note that we could use Guix to install CLISP and ABCL.

Yes, but consistency wouldn't hurt.

CLISP will fail loading local package nicknames. Unless we compile this fork instead https://gitlab.com/gnu-clisp/clisp/-/merge_requests/3

It's good that we don't use them much in the extracted libraries :)

[Ambrevar]

Yes, but consistency wouldn't hurt.

Could use Guix all over the place then, it's very mature for CL support these days.

The only compiler that's lacking is Clasp.

[aartaka]

Yes, sounds fun :D

[Ambrevar]

Most of this is fixed by #7. What's left to do:

• Enable tests for ECL, Clisp (with local nicknames?), maybe Allegro, Clasp and ABCL.
• "Validate make-instance symbols" is not done because we need to figure out a universal, undisruptive way of doing it.