cdepillabout/pretty-simple

Cross-compilation broken by custom Setup.hs used for doctests

vmchale opened this issue ยท 16 comments

Currently, the custom Setup.hs used for doctests causes cross-compilation to fail due to a loingstanding cabal bug.

I set write-ghc-environment-files to always in my cabal.project and then just run doctest from the command-line as a workaround.

Hmm, yeah, I would like to get away from using a custom setup if it's at all possible. There are a lot of issues around them, and this isn't the kind of library that justifies it. The pretty-simple tests quite often fail to run for me on GHC 8.10 anyway.

I'm sure there was talk at some point of a proper cabal doctest command? I'll have a look for that discussion.

Running doctest from the project route is giving me cannot satisfy -package-id pretty-simple-3.3.0.0-inplace, even when I have an environment file which contains precisely that (generated by cabal repl --write-ghc-environment-files=always.

@vmchale How are you getting this to work?

I put them in my cabal.project file as a line like

write-ghc-environment-files: always

(perhaps you've run into another cabal bug ๐Ÿ˜ฌ )

It doesn't seem to make a difference how I specify the flag (file vs. command line), which to be honest is a good thing...

I'm not very familiar with doctest though. To be clear, you're just running doctest, with no args, from the pretty-simple directory? If so then I guess I am hitting a bug.

FWIW, I'm using doctest 0.17 on pretty-simple's HEAD (763d769), with GHC 8.10.2, on Linux.

My workflow is

cabal build --write-ghc-environment-files always
doctest src

For whatever reason doctest by itself doesn't work. And I set build-type: Simple in the .cabal file

(I think cabal repl doesn't quite do what one wants; it doesn't build pretty-simple itself)

(I think cabal repl doesn't quite do what one wants; it doesn't build pretty-simple itself)

Aha, you're right! I'd missed that subtlety.

The problem with doctest is that it doesn't know how to find the required package database when running with the various Haskell build tools (stack, cabal-v1, cabal-v2, haskell environment from nixpkgs, haskell.nix).

cabal-doctest is the only reliable solution right now (although it still doesn't work with haskell.nix).

I don't want to move away from cabal-doctest until there is some other way to reliably run doctests.

However, as you've seen, the custom Setup.hs is only for the doctests, so as a workaround you should be able to delete the Setup.hs, change the build-type to Simple, and disable the doctests.

If you wanted to make this process a little easier, please feel free to send a PR putting the doctests under a flag that you can easily disable.

dpwiz commented

BTW, lens did a similar update recently.

(An older version of lens and this package breaks with global -split-sections ghc flag.)

Thanks @dpwiz, that link is very handy.

Looking at the cabal-docspec manual, I think that could be the perfect solution.

If you wanted to make this process a little easier, please feel free to send a PR putting the doctests under a flag that you can easily disable.

Having just tried this, simply changing the one line to build-type: Simple is enough to get cross-compilation working. But it isn't something that can be put behind a flag AFAICT.

There's a long thread on the Cabal issue tracker about doctests: haskell/cabal#4500

Recently sol commented about using cabal repl directly with doctest to run doctests: haskell/cabal#4500 (comment)

I'd be interested to see a comparison between cabal repl --with-ghc=doctest, cabal-docspec and cabal-doctests (including whether they easily work with stack, Nixpkgs haskell stuff, and haskell.nix).

Looks like there is one more doctest contender in the mix, doctest-parallel:

One feature of doctest-parallel is that it apparently parses cabal files to discover modules, so you don't need a custom Setup.hs.

It would be great add a comparison to doctest-parallel to the list above

There has apparently recently been a PR to cabal that adds a new doctest-related feature:

I haven't looked at this thoroughly, but this might be the best option so far, since it will officially be supported by cabal.

edit: This has been released in Cabal-3.8.1.0: https://discourse.haskell.org/t/just-released-cabal-3-8-1-0/4920