Skip to content
Proposed compiler and language changes for GHC and GHC/Haskell
Python Makefile Shell Nix
Branch: master
Clone or download

README.rst

GHC Proposals

This repository contains specifications for proposed changes to the Glasgow Haskell Compiler. The purpose of the GHC proposal process and of the GHC Steering Committee, is to broaden the discussion of the evolution of GHC.

What is a proposal?

A GHC Proposal is a document describing a proposed change to the compiler, the GHC/Haskell language, or the libraries in the GHC.* module namespace. These include,

  • A syntactic change to GHC/Haskell (e.g. the various ShortImports proposals, do expressions without $)
  • A major change to the user-visible behaviour of the compiler (e.g. the recent change in super-class solving, and -Wall behavior)
  • The addition of major features to the compiler (e.g. -XTypeInType, GHCi commands, type-indexed Typeable representations)

Changes to the GHC API or the plugin API are not automatically within the scope of the committee, and can be contributed following the usual GHC workflow. Should the GHC maintainers deem a change significant or controversial enough to warrant that, they may, at their discretion, involve the committee and ask the contributor to write a formal proposal.

The life cycle of a proposal

This section outlines what stages a proposal may go through. The stage is identified by a GitHub label, which is identified in the following list.

  1. (No label.) The author drafts a proposal.

    What is a proposal?What should a proposal look like?

  2. (No label.) The author submits the proposal to the wider Haskell community for discussion, as a pull request against this repository.

    How to submit a proposal

  3. (No label.) The wider community discusses the proposal in the commit section of the pull request, while the author refines the proposal. This phase lasts as long as necessary.

    Discussion goalsHow to comment on a proposal≡ List of proposals under discussion

  4. Label: Pending shepherd recommendation. Eventually the proposal author brings the proposal before the committee for review.

    How to bring a proposal before the committeeWho is the committee?≡ List of proposals waiting for shepherd

  5. Label: Pending committee review. One committee member steps up as a shepherd, and generates consensus within the committee within four or five weeks.

    Committee processReview criteria≡ List of proposals under review

  6. Eventually, the committee rejects a proposal (label: Rejected), or passes it back to the author for review (label: Needs revision), or accepts it (label: Accepted).

    Acceptance of the proposal implies that the implementation will be accepted into GHC provided it is well-engineered, well-documented, and does not complicate the code-base too much.

    ≡ List of accepted proposals≡ List of proposals being revised≡ List of rejected proposals

  7. Label: Dormant. If a proposal sees no activity for along time, it is marked as “dormant”, and eventually closed.

    What is a dormant proposal?≡ List of dormant proposals

  8. Label: Implemented. Once a proposal is accepted, it still has to be implemented. The author may do that, or someone else. We mark the proposal as “implemented” once it hits GHC’s master branch (and we are happy to be nudged to do so by email or GitHub issue).

    ≡ List of implemented proposals

Do not hesitate to contact us if you have questions.

How to start a new proposal

Proposals are written in either ReStructuredText or Markdown. While the proposal process itself has no preference, keep in mind that the GHC Users Guide uses ReStructuredText exclusively. Accepted proposals written in ReStructuredText thus have the slight benefit that they can be more easily included in the official GHC documentation.

Proposals should follow the structure given in the ReStructuredText template, or the Markdown template. (The two are identical except for format.)

See the section Review criteria below for more information about what makes a strong proposal, and how it will be reviewed.

To start a proposal, create a pull request that adds your proposal as proposals/0000-proposal-name.rst or proposals/0000-proposal-name.md. Use the corresponding proposals/0000-template file as a template.

If you are unfamiliar with git and github, you can use the GitHub web interface to perform these steps:

  1. Load the proposal template using this link (ReStructuredText) or this link (Markdown).
  2. Change the filename and edit the proposal.
  3. Press “Commit new file”

The pull request summary should include a brief description of your proposal, along with a link to the rendered view of proposal document in your branch. For instance,

This is a proposal augmenting our existing `Typeable` mechanism with a
variant, `Type.Reflection`, which provides a more strongly typed variant as
originally described in [A Reflection on
Types](http://research.microsoft.com/en-us/um/people/simonpj/papers/haskell-dynamic/index.htm)
(Peyton Jones, _et al._ 2016).

[Rendered](https://github.com/bgamari/ghc-proposals/blob/typeable/proposals/0000-type-indexed-typeable.rst)

Discussion goals

Members of the Haskell community are warmly invited to offer feedback on proposals. Feedback ensures that a variety of perspectives are heard, that alternative designs are considered, and that all of the pros and cons of a design are uncovered. We particularly encourage the following types of feedback,

  • Completeness: Is the proposal missing a case?
  • Soundness: Is the specification sound or does it include mistakes?
  • Alternatives: Are all reasonable alternatives listed and discussed. Are the pros and cons argued convincingly?
  • Costs: Are the costs for implementation believable? How much would this hinder learning the language?
  • Other questions: Ask critical questions that need to be resolved.
  • Motivation: Is the motivation reasonable?

How to comment on a proposal

To comment on a proposal you need to be viewing the proposal's diff in "source diff" view. To switch to this view use the buttons on the top-right corner of the Files Changed tab.

The view selector buttons.

Use the view selector buttons on the top right corner of the "Files Changed" tab to change between "source diff" and "rich diff" views.

Feedback on a open pull requests can be offered using both GitHub's in-line and pull request commenting features. Inline comments can be added by hovering over a line of the diff.

The ``+`` button appears while hovering over line in the source diff view.

Hover over a line in the source diff view of a pull request and click on the + to leave an inline comment

For the maintenance of general sanity, try to avoid leaving "me too" comments. If you would like to register your approval or disapproval of a particular comment or proposal, feel free to use GitHub's "Reactions" feature.

How to bring a proposal before the committee

When the discussion has ebbed down and the author thinks the proposal is ready, they

  1. Review the discussion thread and ensure that the proposal text accounts for all salient points. Remember, the proposal must stand by itself, and be understandable without reading the discussion thread.
  2. Add a comment to the a pull request, briefly summarizing the major points raised during the discussion period and stating your belief that the proposal is ready for review. In this comment, tag the committee secretary (currently @nomeata).

The secretary will then label the pull request with Pending shepherd recommendation and start the committee process. (If this does not happen within a day or two, please ping the secretary or the committee.)

What is a dormant proposal?

In order to keep better track of actively discussed proposals, proposals that see no activity for an extended period of time (a month or two) might be marked as “dormant”. At any time the proposer, or someone else can revive the proposal by picking up the discussion (and possibly asking the secretary to remove the dormant tag).

You can see the list of dormant proposals.

Who is the committee?

You can reach the committee by email at ghc-steering-committee@haskell.org:

The current members, including their GitHub handle, when they joined and their role, are listed at:

Christopher Allen @bitemyapp 2017/02  
Vitaly Bragilevsky @bravit 2018/09  
Joachim Breitner @nomeata 2017/02 secretary
Iavor Diatchki @yav 2017/02  
Richard Eisenberg @goldfirere 2017/02  
Sandy Maguire @isovector 2019/07  
Simon Marlow @simonmar 2017/02 co-chair
Simon Peyton-Jones @simonpj 2017/02 co-chair
Eric Seidel @gridaphobe 2018/09  
Arnaud Spiwack @aspiwack 2019/07  

The committee members have committed to adhere to the Haskell committee guidelines for respectful communication.

We would also like to thank our former members:

Ryan Newton @rrnewton 2017/02 - 2018/09
Roman Leshchinskiy @rleshchinskiy 2017/02 - 2018/11
Ben Gamari @bgamari 2017/02 - 2019/07
Manuel M T Chakravarty @mchakravarty 2017/02 - 2019/07

Committee process

The committee process starts once the secretary has been notified that a proposal is ready for decision.

The steps below have timescales attached, so that everyone shares the same expectations. But they are only reasonable expectations. The committee consists of volunteers with day jobs, who are reviewing proposals in their spare time. If they do not meet the timescales indicated below (e.g they might be on holiday), a reasonable response is a polite ping/enquiry.

  • The secretary nominates a member of the committee, the shepherd, to oversee the discussion. The secretary

    • labels the proposal as Pending shepherd recommendation,
    • assigns the proposal to the shepherd,
    • drops a short mail on the mailing list, informing the committee about the status change.
  • Based on the proposal text (but not the GitHub commentary), the shepherd decides whether the proposal ought to be accepted or rejected or returned for revision. The shepherd should do this within two weeks.

  • If the shepherd thinks the proposal ought to be rejected, they post their justifications on the GitHub thread, and invite the authors to respond with a rebuttal and/or refine the proposal. This continues until either

    • the shepherd changes their mind and supports the proposal now,
    • the authors withdraw their proposal,
    • the authors indicate that they will revise the proposal to address the shepherds point. The shepherd will label the pull request as Needs Revision.
    • the authors and the shepherd fully understand each other’s differing positions, even if they disagree on the conclusion.
  • Now the shepherd proposes to accept or reject the proposal. To do so, they

    • post their recommendation, with a rationale, on the GitHub discussion thread,
    • label the pull request as Pending committee review,
    • re-title the proposal pull request, appending (under review) at the end. (This enables easy email filtering.)
    • drop a short mail to the mailing list informing the committee that discussion has started.
  • Discussion among the committee ensues, in two places

    • Technical discussion takes place on the discussion thread, where others may continue to contribute.
    • Evaluative discussion, about whether to accept, reject, or return the proposal for revision, takes place on the committee's email list, which others can read but not post to.

    It is expected that every committee member express an opinion about every proposal under review. The most minimal way to do this is to "thumbs-up" the shepherd's recommendation on GitHub.

    Ideally, the committee reaches consensus, as determined by the secretary or the shepherd. If consensus is elusive, then we vote, with the Simons retaining veto power.

    This phase should conclude within a month.

  • For acceptance, a proposal must have at least some enthusiastic support from member(s) of the committee. The committee, fallible though its members may be, is the guardian of the language. If all of them are luke-warm about a change, there is a presumption that it should be rejected, or at least "parked". (See "evidence of utility" above, under "What a proposal should look like".)

  • A typical situation is that the committee, now that they have been asked to review the proposal in detail, unearths some substantive technical issues. This is absolutely fine -- it is what the review process is for!

    If the technical debate is not rapidly resolved, the shepherd should return the proposal for revision. Further technical discussion can then take place, the author can incorporate that conclusions in the proposal itself, and re-submit it. Returning a proposal for revision is not a negative judgement; on the contrary it might connote "we absolutely love this proposal but we want it to be clear on these points".

    In fact, this should happen if any substantive technical debate takes place. The goal of the commitee review is to say yes/no to a proposal as it stands. If new issues come up, they should be resolved, incorporated in the proposal, and the revised proposal should then be re-submitted for timely yes/no decision. In this way, no proposal should languish in the committee review stage for long, and every proposal can be accepted as-is, rather than subject to a raft of ill-specified further modifications.

    The author of the proposal may invite committee collaboration on clarifying technical points; conversely members of the committee may offer such help.

    When a proposal is returned for revision, GitHub labels are updated accordingly and the (under review) suffix is removed from the title of the PR.

  • The decision is announced, by the shepherd or the secretary, on the Github thread and the mailing list.

    Notwithstanding the return/resubmit cycle described above, it may be that the shepherd accepts a proposal subject to some specified minor changes to the proposal text. In that case the author should carry them out.

    The secretary then tags the pull request accordingly, and either merges or closes it. In particular

    • If we say no: The pull request will be closed and labeled Rejected.

      If the proposer wants to revise and try again, the new proposal should explicitly address the rejection comments.

      In the case that the proposed change has already been implemented in GHC, it will be reverted.

    • If we say yes: The pull request will be merged and labeled Accepted. Its meta-data will be updated to include the acceptance date. A link to the accepted proposal is added to the top of the PR discussion, together with the sentence “The proposal has been accepted; the following discussion is mostly of historic interest.”.

      At this point, the proposal process is technically complete. It is outside the purview of the committee to implement, oversee implementation, attract implementors, etc.

      The proposal authors or other implementors are encouraged to update the proposal with the implementation status (i.e. ticket URL and the first version of GHC implementing it.)

Review criteria

Here are some characteristics that a good proposal should have.

  • It should be self-standing. Some proposals accumulate a long and interesting discussion thread, but in ten years' time all that will be gone (except for the most assiduous readers). Before acceptance, therefore, the proposal should be edited to reflect the fruits of that dicussion, so that it can stand alone.

  • It should be precise, especially the "Proposed change specification" section. Language design is complicated, with lots of interactions. It is not enough to offer a few suggestive examples and hope that the reader can infer the rest. Vague proposals waste everyone's time; precision is highly valued.

    We do not insist on a fully formal specification, with a machine-checked proof. There is no such baseline to work from, and it would set the bar far too high. On the other hand, for proposals involving syntactic changes, it is very reasonable to ask for a BNF for the changes.

    Ultimately, the necessary degree of precision is a judgement that the committee must make; but authors should try hard to offer precision.

  • It should offer evidence of utility. Even the strongest proposals carry costs:

    • For programmers: most proposals make the language just a bit more complicated;
    • For GHC maintainers: most proposals make the implementation a bit more complicated;
    • For future proposers: most proposals consume syntactic design space add/or add new back-compat burdens, both of which make new proposals harder to fit in.
    • It is much, much harder subsequently to remove an extension than it is to add it.

    All these costs constitute a permanent tax on every future programmer, langauge designer, and GHC maintainer. The tax may well be worth it (a language without polymorphism would be simpler but we don't want it), but the case should be made.

    The case is stronger if lots of people express support by giving a "thumbs-up" in GitHub. Even better is the community contributes new examples that illustrate how the proposal will be broadly useful. The committee is often faced with proposals that are reasonable, but where there is a suspicion that no one other than the author cares. Defusing this suspicion, by describing use-cases and inviting support from others, is helpful.

  • It should be copiously illustrated with examples, to aid understanding. However, these examples should not be the specification.

Below are some criteria that the committee and the supporting GHC community will generally use to evaluate a proposal. These criteria are guidelines and questions that the committee will consider. None of these criteria is an absolute bar: it is the committee's job to weigh them, and any other relevant considerations, appropriately.

  • Utility and user demand. What exactly is the problem that the feature solves? Is it an important problem, felt by many users, or is it very specialised? The whole point of a new feature is to be useful to people, so a good proposal will explain why this is so, and ideally offer evidence of some form. The "Endorsements" section of the proposal provides an opportunity for third parties to express their support for the proposal, and the reasons they would like to see it adopted.

  • Elegant and principled. Haskell is a beautiful and principled language. It is tempting to pile feature upon feature (and GHC Haskell has quite a bit of that), but we should constantly and consciously strive for simplicity and elegance.

    This is not always easy. Sometimes an important problem has lots of solutions, none of which have that "aha" feeling of "this is the Right Way to solve this"; in that case we might delay rather than forge ahead regardless.

  • Does not create a language fork. By a "fork" we mean

  • It fails the test "Is this extension something that most people would be happy to enable, even if they don't want to use it?";
  • And it also fails the test "Do we think there's a reasonable chance this extension will make it into a future language standard?"; that is, the proposal reflects the stylistic preferences of a subset of the Haskell community, rather than a consensus about the direction that (in the committee's judgement) we want to push the whole language.
The idea is that unless we can see a path to a point where everyone has the extension turned on, we're left with different groups of people using incompatible dialects of the language. A similar problem arises with extensions that are mutually incompatible.
  • Fit with the language. If we just throw things into GHC willy-nilly, it will become a large ball of incoherent and inconsistent mud. We strive to add features that are consistent with the rest of the language.

  • Specification cost. Does the benefit of the feature justify the extra complexity in the language specification? Does the new feature interact awkwardly with existing features, or does it enhance them? How easy is it for users to understand the new feature?

  • Implementation cost. How hard is it to implement?

  • Maintainability. Writing code is cheap; maintaining it is expensive. GHC is a very large piece of software, with a lifetime stretching over decades. It is tempting to think that if you propose a feature and offer a patch that implements it, then the implementation cost to GHC is zero and the patch should be accepted.

    But in fact every new feature imposes a tax on future implementors, (a) to keep it working, and (b) to understand and manage its interactions with other new features. In the common case the original implementor of a feature moves on to other things after a few years, and this maintenance burden falls on others.

How to build the proposals?

The proposals can be rendered by running:

nix-shell shell.nix --run "make html"

this will then create a directory _build which will contain an index.html file and the other rendered proposals. This is useful when developing a proposal to ensure that your file is syntax correct.

Questions?

Feel free to contact any of the members of the GHC Steering Committee with questions. Email and IRC (#ghc on irc.freenode.net) are both good ways of accomplishing this.

You can’t perform that action at this time.