GHC-8.4 related mega-issue

[phadej]

Because @RyanGlScott asked me to make one. I'll list my findings here

Semigroup-Monoid (needs releases)

• Could not deduce (Semigroup (Id a)) ad#68 (Hackage)
• Adapt to the Semigroup-Monoid Proposal constraints#65 and Reexport Div/Mod from GHC.TypeLits (introduced in GHC 8.4) constraints#70 (Hackage)
• No instance for (Semigroup (Predicate a)) contravariant#38, PR: Add Semigroup (Predicate a) instance contravariant#39 (Hackage)
• No instance for (Semigroup Clause) ersatz#37 (Hackage)
• make free forward-Wcompatible (again) free#155 (Hackage)
• Fix the build on GHC 8.4 graphs#10 (Hackage)
• Compatibility with Semigroup/Monoid proposal heaps#11 (Hackage)
• Adds Semigroup (Vector u v c) instance to make compatible with GHC >= 8.4 hybrid-vectors#9 (Hackage)
• No instance for (Semigroup Exceptions) hyphenation#13 (Hackage)
• Fix the build on GHC 8.4 intern#13 (Hackage)
• lca (Hackage)
• lens itself, Adapt the build to GHC 8.4.1 #779 (Hackage)
• Could not deduce (Apply m) lens-action#11 (Hackage)
• Adapt to the Semigroup–Monoid Proposal profunctors#58 (Hackage)
• Adapt to the Semigroup–Monoid Proposal reflection#37 (Hackage)
• Fix the build post-Semigroup–Monoid Proposal semigroupoids#61 (Hackage)
• No instance for (Semigroup (Jacket i a)) zippers#13 (Hackage)

Intra-dependencies (may need releases)

• ad (free-5) (Hackge)
• adjunctions (free-5) (Hackage)
• Allow ansi-terminal-0.8 ansi-wl-pprint#16 (ansi-terminal-0.8) (Hackage)
• kan-extensions (free-5) (Hackage)
• keys (free-5) (Hackage)
• recursion-schemes (free-5) (Hackage)
• Needs adaptation to wl-pprint-extras-3.6 wl-pprint-terminfo#10 (wl-pprint-terminfo-3.6, will need new release)

Lower-bounds (doesn't need releases)

• Bump base-orphans lower bound to >=0.5.2 distributive#43

build-tool-depends (doesn't need releases, but I'll do them anyway)

• Add build-tool-depends adjunctions#46 (Hackage)
• Add build-tool-depends bifunctors#66 (Hackage)
• Add build-tool-depends constraints#72 (Hackage)

new-build shenanigans (doesn't need releases)

• gl and cabal new-build don't mix gl#17

[phadej]

@RyanGlScott I get segmentation fault (locally, Ubuntu LTS-16.04) when trying to run doctests for approximate, bits, compensated built with

% ghc-8.4.1 --version                 
The Glorious Glasgow Haskell Compilation System, version 8.4.0.20171222
% ghc-8.4.1 --print-project-git-commit-id
c6cf13ca63f3a11a8da7c7e3bd69e673a8df5440

but e.g. lens doctests run fine. Could you check if you can reproduce?

[RyanGlScott]

I can take a look during my downtime this week.

Also, since I don't have to do this at the moment: which of the issues in the original comment will require releasing new versions of packages to Hackage? Some of them, like ekmett/distributive#43, don't seem to fall into this category, and it's not clear to me from a quick glance which ones do.

[RyanGlScott]

Hm... the bits doctests appear to be working for me on 8.4.1, as this Travis run demonstrates.

[RyanGlScott]

The approximate and compensated doctests also appear to be in working order.

[phadej]

@RyanGlScott

• I tagged sections for which will need releases (i.e. code changes required), else might be done with revisions only.
• Ok, so problem is probably in my machine, I'll try to run those tests on other box to verify test-pass.

EDIT it's nice that there aren't that much issues after all.

[RyanGlScott]

Some updates.

I've updated every. single. Kmett. package. to use head.hackage (well, at least those to which I have push access), so now we have some assurance that the Kmettiverse builds on GHC 8.4. The one package I wasn't able to convert was gl, as noted in ekmett/gl#17.

Now that everything builds, I've unchecked all of the entries in the Semigroup–Monoid category, and I'll check them off once they've had proper Hackage releases.

One challenge I'll need to overcome at some point: I don't have push access to keys or reflection.

Lastly, I was able to reproduce the segfault on the bits doctests using Ubuntu 17.04! (All my previous attempts were on Ubuntu 14.04.) This is both fascinating and terrifying. I'll have to investigate further.

[RyanGlScott]

I've managed to reduce the segfault down to:

module Bug where

{-# ANN module "I'm an annotation" #-}

$ ./doctest Bug.hs
Segmentation fault (core dumped)
$ ./doctest --version
doctest version 0.13.0
using version 8.4.0.20171222 of the GHC API
using /opt/ghc/8.4.1/bin/ghc-8.4.0.20171222
$ lsb_release -a
No LSB modules are available.
Distributor ID: Ubuntu
Description:    Ubuntu 17.04
Release:        17.04
Codename:       zesty

$ ./doctest Bug.hs
Examples: 0  Tried: 0  Errors: 0  Failures: 0
$ ./doctest --version
doctest version 0.13.0
using version 8.4.0.20171222 of the GHC API
using /opt/ghc/8.4.1/bin/ghc-8.4.0.20171222
$ lsb_release -a
No LSB modules are available.
Distributor ID: Ubuntu
Description:    Ubuntu 14.04.5 LTS
Release:        14.04
Codename:       trusty

[RyanGlScott]

And now I can reproduce the issue without doctest. See GHC Trac #14675.

[RyanGlScott]

New releases for all of the packages in the Semigroup-Monoid column have been uploaded to Hackage—except for free-5 and lens-4.16, since those will require lots of upper version bounds bumping in the ecosystem.

@ekmett @glguy: Can you think of any other changes we should put into lens-4.16 before a release?

[RyanGlScott]

Everything that needs a Hackage release to support GHC 8.4 has now been released. That leaves only two things from the checklist:

• Needs adaptation to wl-pprint-extras-3.6 wl-pprint-terminfo#10
• gl and cabal new-build don't mix gl#17

But neither of these are required to support GHC 8.4. I'll opt to close this issue and shift over to the respective remaining issues.

[phadej]

@RyanGlScott lens-4.16 has template-haskell <2.13. Is it intentional?

[RyanGlScott]

Nope, that certainly wasn't intentional. (I've just made a commit and a revision to compensate.)

But perhaps we should have a discussion about this. The reason why I didn't notice this in the first place is because our testing infrastructure for GHC 8.4.1 specifies allow-newer carte blanche for template-haskell (as well as base, Cabal, and ghc), so it's incredibly easy to let upper bounds issues for 8.4.1 slip by unnoticed. What would be ideal is if we could require that these libraries be up-to-date for lens itself, but relax them for downstream dependencies (as they haven't all caught up yet, so we can't simply drop allow-newer altogether).

(Relatedly, I've often found myself wishing that new-build would yell at you more if you are doing illicit things like this—see haskell/cabal#4627, for instance.)

[phadej]

@RyanGlScott We don't have a syntax for allow-newer to "relax base for everything except these packages". Either we have to explicitly list what needs to be relaxed, or than have somewhat blunt too. (it's still more precise than "--allow-newer" everything).

And relaxing for this package is exactly what we want for HEAD, because we don't want to have (in general) to relax upper bound on base or template-haskell so we don't accidentally release such version (when GHC release is far in the future).

So: bikeshed on the constraint syntax (in a new Cabal issue), maybe we can convince @hvr it has its uses. Maybe --allow-newer=~lens:template-haskell. The problem is that --allow-newer is disjunctive, so --allow-newer=~this:base,~that:base will do wrong thing...

[phadej]

% cabal --project-file=cabal-8.4.1.project new-build intern -j1
Build profile: -w ghc-8.4.0.20180204 -O1
In order, the following will be built (use -v for more details):
 - intern-0.9.1.4 (lib:intern) (first run)
Preprocessing library for intern-0.9.1.4..
Building library for intern-0.9.1.4..
[2 of 9] Compiling Data.Interned.IntSet ( Data/Interned/IntSet.hs, /home/ogre/Documents/acme-kmett/dist-newstyle/build/x86_64-linux/ghc-8.4.0.20180204/intern-0.9.1.4/build/Data/Interned/IntSet.o )

Data/Interned/IntSet.hs:221:10: error:
    • No instance for (Semigroup IntSet)
        arising from the superclasses of an instance declaration
    • In the instance declaration for ‘Monoid IntSet’
    |
221 | instance Monoid IntSet where

ping @RyanGlScott intern is broken too

[phadej]

src/Data/Graph/Algorithm.hs:68:10: error:
    • Could not deduce (Semigroup (GraphSearch g m))
        arising from the superclasses of an instance declaration
      from the context: (Graph g, Monoid m)
        bound by the instance declaration
        at src/Data/Graph/Algorithm.hs:68:10-56
    • In the instance declaration for ‘Monoid (GraphSearch g m)’
   |
68 | instance (Graph g, Monoid m) => Monoid (GraphSearch g m) where
   |          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
cabal: Failed to build graphs-0.7.

[phadej]

And other packages currently acme-kmett compile fine with GHC-8.4.1, I didn't build or run test-suites though.

[RyanGlScott]

I've uploaded intern-0.9.2 and graphs-0.7.1 to Hackage, so now I believe everything is 8.4-ready.