Lparallel overhaul

[Ambrevar]

Fixes #2.

This switches the whole thread management / concurrency logic from Calispel to Lparallel.

• It simplifies the code a lot (about 100 lines less!).
• It also makes it way less prone to race conditions, and much easier to debug.
• Also less risk of dangling threads.
• Finally, it may work on ECL. Still need to be tested.

To do:

• Test on ECL.
  Should work once the following issue "attempt to submit task to dead kernel" is fixed.
• Make sure there is no Attempt to submit task to dead kernel. error when running multiple times.
• Fix attribute-value API.
• Add wait-on-prompt-buffer to API.
  No need, it's now enough to force result-channel.
• Make sure errors can be lpara:task-handler-bind.
• Make sure no threads are left hanging, especially after running the test suite.

[Ambrevar]

@aartaka @aadcg This is ready for review.

The main point of discussion that's left is this attribute*-value* API. It's a bit clumsy, we need better names and possibly less convoluted functions.

[aartaka]

It's all clear and good overall. I mainly have some convenience suggestions:

• Move the kernel cleanup logic into a (defmethod (setf kernel) ((value null) prompter) ...).
• Add a wrapper kernel on source.

[aartaka]

Maybe a setf-method for initial-suggestions instead?

[Ambrevar]

Oops, this code was poorly ported :) I'll fix it.

That said, how could a setf-method help here?

[aartaka]

Adding an ensure-suggestions-list, for one.

[Ambrevar]

Actually the point of not having a writer is to hint that it should not be modified.
It's only ever set in this function.

[aartaka]

Note that there's no lpara:promise type, there's only lparallel.promise::%promise. So maybe pick a different type?

[Ambrevar]

result-channel renamed to result and blocking accessor added.

[Ambrevar]

Attribute-value API is somewhat fixed:

• We now export attribute-options and attributes-options methods so that the call site does not need to parse lists;
• attribute-default calls to attribute-value, so that the future is properly forced.
• attribute-value return "" when not ready, which is consistent with the previous behaviour.
• Future's computation starts when object-properties is called.

This is sufficient for the Lparallel overhaul.

There is more to fix in the attribute API, but I'll open a separate issue for that.

[Ambrevar]

Attribute API discussion: #21

[Ambrevar]

OK to merge? @aartaka @aadcg

[Ambrevar]

Wait before merging, we need to make sure atlas-engineer/nyxt#2998 (comment) is happy with this new API.

[aadcg]

It's OK to merge from my side. Splendid work @Ambrevar!

[aartaka]

Actually, why? Seems like mopu:slot-names definition is exactly what you're reproducing here:

(defun class-slot-names (thing)
  (let ((class (get-class thing)))
    (if class
      (mapcar 'mop:slot-definition-name
	      (mop:class-slots (finalize-class-if-necessary class)))
      (progn
	(warn "class for ~a not found)" thing)
	nil))))

EDIT: highlightling.

[aartaka]

Trusting André's review here :)

[Ambrevar]

Last commit (cdef02f) makes a radical change: it replaces update-notifier by update-hook (from Nhooks!).

The rationale is that update-notifier was clumsy to use (and indeed never used in Nyxt). With update-hook, it should become trivial to have Nyxt prompt-render-suggestions on each update, thus removing the need for a complex setup.

@jmercouris @aadcg @aartaka Thoughts?

[Ambrevar]

To do:

• Maybe define a new update-hook type where the handlers only accept a source argument, instead of using hook-any.

[aartaka]

• Maybe define a new update-hook type where the handlers only accept a source argument, instead of using hook-any.

The proper name would be hook-source, but yes, that won't hurt!

[Ambrevar]

Back to this: this has become a behemoth change and it's still not working...
Input buffering is not behaving as expected: either it's a mistake of mine, or there is something wrong with Lparallel.

In any case, I'll split this pull request in two parts:

• First part includes all the bug fixes, the switch to Lparallel, etc. EXCEPT input buffering. This we can probably merge quickly.
• Second part will be about input buffering, but it needs much more work.

[aadcg]

@Ambrevar sounds like an excellent plan!