#1172 broke company integration

[arybczak]

#1172 makes company (and presumably also auto-complete) not work at all, because it overwrites completion-at-point-functions here.

I don't know about auto-complete, but company-mode with company-dabbrev-code and company-etags backends provides great auto-completion facilities (and in case of company-etags ability to jump to source with C-w directly from the list of completions) and currently it's got silently suppressed by a piece of code that offers completion of keywords/pragmas and simple declarations alone.
#1172 should've been implemented as either auto-complete or company backend (as was already mentioned by @cocreature in #215) instead of hard-coding it into haskell-mode as it effectively hides company/auto-complete functionality from users.

At the very least, if the completion code is kept in its current state, at least make overwriting completion-at-point-functions optional and disable it by default. But really it should be made into company backend so it composes with other backends.

[geraldus]

@arybczak can you create and maintain backends you talking about both for company and AC?

[arybczak]

I can't, because my elisp writing skills are close to none.

But having these as company/ac backends is just the end goal, the main problem is actually that current behaviour prevents company from working (and I don't know why exactly that happens either, all I know is that commenting out the statement I pointed to solves the problem).

[gracjan]

Is there somewhere in the vast internet a guide how to properly implement completions in Emacs? It sounds more involved than it should be.

I agreed with cocreature that we should leave dabbrev alone and I thought we already do skip that. Aren't we?

If it goes about me we can just support conditional dependency on company and autocomplete. The issue is of course work required to build those integrations and then maintain them.

Any volunteers?

[arybczak]

Is there somewhere in the vast internet a guide how to properly implement completions in Emacs? It sounds more involved than it should be.

It looks like writing company backends is quite straightforward: https://github.com/company-mode/company-mode/wiki/Writing-backends

I agreed with cocreature that we should leave dabbrev alone and I thought we already do skip that. Aren't we?

I'm not sure, but for the record, company offers company-dabbrev backend, which works well.

BTW, I later found out that company supposedly integrates with completion-at-point-functions (by company-capf backend), but for some reason it doesn't work in this particular case, i.e. when haskell-mode sets completion-at-point-functions, other backends stop working. I'm not sure whether it's a bug in company or the implementation of haskell-mode's completion function.

[gracjan]

I tried to understand company mode integration and I do not see how this is different than completion-at-point-functions. company mode invokes backends in order, the first one that returns non-nil is the one selected as source for completions. Same mechanism as completion-at-point-functions. I do not understand the need for existence of company mode other than a GUI.

[arybczak]

You can combine backends together, e.g. (company-dabbrev-code company-etags) will give you completions from both backends.

[gracjan]

@arybczak: what is the value of your company-backends variable?

[arybczak]

(company-c-headers company-bbdb company-nxml company-css company-eclim company-semantic company-xcode company-cmake company-capf company-files
                   (company-dabbrev-code company-gtags company-etags company-keywords)
                   company-oddmuse company-dabbrev)

[arybczak]

@gracjan

I also have

(eval-after-load 'company-etags
      '(progn
         (add-to-list 'company-etags-modes 'haskell-mode)))

To enable etags completion in haskell-mode.

and for haskell-mode I just use defaults:

;; enable haskell-mode
(defconst haskell-mode-directory "~/.emacs.d/haskell-mode/")
(add-to-list 'load-path haskell-mode-directory)
(require 'haskell-mode-autoloads)
(add-to-list 'Info-default-directory-list haskell-mode-directory)

You can see the whole configuration here if that helps.

[gracjan]

Looks like you have company-capf in the list on the first level. This means company-mode will call backends in turn company-c-headers ... up to company-capf. Then company-capf will return a non-empty list because completions-at-point-functions returns a non-empty list. At this point the process stops.

If your company-backend looked like this:

(company-c-headers ... company-files
                   (company-capf company-dabbrev-code ... company-dabbrev)

Then all backends from the inner list would be called and a set-union of results would be made. I guess this is what you want, right?

[arybczak]

Hmm, yeah, putting company-capf inside the group does the job, but it's baffling as if completion-at-point-functions doesn't return any completion candidates, company-capf should presumably let subsequent backends do the job, but that is not the case. Maybe it's actually a bug in company-capf :/

[gracjan]

Maybe. I'm closing this one.