For a long time, I’ve been annoyed that cabal tells you to add modules to your cabal file. It can detect missing modules, but won’t automatically add them. This is not a big deal for a single module, but if you like splitting code into many small modules, for example because you like fast compile times, then any refactor may cause many of these messages to appear, which becomes tedious to manage.

Currently, if you build a project with an extra module not listed in your cabal file, GHC emits a warning:

<no location info>: error: [-Wmissing-home-modules, -Werror=missing-home-modules]
    These modules are needed for compilation but not listed in your .cabal file's other-modules: 
        X

You’d say, why doesn’t cabal just add this module to the cabal file? Well, it can’t. Cabal is currently only able to parse Cabal files, and print them back out in a mangled¹ form. There are other programs providing module detection ², but nothing is integrated in cabal itself. Anyway, one day I was lamenting this problem on the internet and I was told by a forum user⁴ that this is in fact the cabal exact print issue. GHC is already able to do this for Haskell code, and defines it as follows:

Taking an abstract syntax tree (AST) and converting it into a string that looks like what the user originally wrote is called exact-printing. An exact-printed program includes the original spacing and all comments.

In any case, after some exploratory discussion in that thread, I got a maintainer to endorse an approach I made.

Previous attempts were abandoned. Or they revolved around creating a seperate AST³, which was against maintainer recommendation, and then abandoned. I’m doing a fully integrated design around the parseGenericPackageDescription function.

parseGenericPackageDescription :: ByteString 
                               -> ParseResult GenericPackageDescription 

This function takes a ByteString (cabal file), and returns a GenericPackageDescription. Which is the data type the rest of cabal works with. This is the cabal equivalent of an AST. To make sure the changes I make are helping, I setup several round trip test concerning the various properties we want see exact printed. For example, I’ll add a cabal file with comments:

cabal-version:           3.0
name:                    two-sections 
version:                 0
synopsis:                The -any none demo
build-type:              Simple

                         -- my awesome
-- library
library 
  default-language:  Haskell2010
  exposed-modules:   AnyNone
  build-depends:     base <5
                    , containers

We put that file through parseGenericPackageDescription, put the result of that function called GenericPackageDescription into the exactPrint function, and it should be the exact same. By doing this we should theoretically be allowed to modify GenericPackageDescription however we please, and get the result printed, which is a step towards solving our initial module not listed problem (and several others).

The implementation itself is simple. We add a field to GenericPackageDescription with all the exact print data, which we access trough Map‘s. The key is a new concept called a NameSpace. Values are ExactPostions for example. I modified the pretty printer to first try and get an exact position, before falling back to pretty printing. This instantly solved the tests for sections. Which was a good result!

However, there is still a long list of features we need to support, such as comments, common stanzas, better comma support, and conditional branches. Exact printing comments is hard because currently cabal filters them instead of parsing them. so the parser needs to be modified as well. Fortunately someone has already done some preliminary work on that, unfortunately it breaks my existing tests. Common stanzas allow users to declare some shared stanza, such as ghc-options or build depends fields, and re-use them across various targets such as the library or executable. This is also hard because the common stanzas are currently just absorbed into whatever they’re imported, they don’t exist in the GenericPackageDescription at all! Exact printing comma support seems a bit strange, but you can put commas in many different places and still have a valid cabal file. These locations need to be parsed and tracked somehow. I’ve not even spoken about conditional branches yet, which I’ve not thought about much, however the point is that there is a lot to support!

In general the issue why this exact printing is so hard is because cabal has a lot of features, all needing to be supported. And every new feature added will not have exact printing support until some solution gets merged. So over time we can expect exact printing to become even harder to implement.

Why I’m writing this instead of just implementing exact print? Back in November, I implemented some initial solution in anger, and with a draft PR created, my anger had all but dissipated. With clarity of mind, I realized spending all my free time implementing this feature isn’t sustainable. I’d burn out at that rate. Essentially I’m looking for a way to make this more sustainable.

A possibility is to create a Haskell tech proposal. This is a funding mechanism for long-standing issues in the Haskell community. With funding I could trade hours in my day job for hours working on this particular issue, without losing anything on mental health. Currently the issue is that these proposals look like a lot of work. And I’m not sure how much budget is even available for a project like this. So in essence this blog-post is a preproposal, to see what people think of it without going all the way.

Another possibility is recruiting volunteers. In general, however they’d run into the same issue I had! However I do think if I’d trade some money for some helping hands then this could work. The tests can be solved in parallel, a single person could work on comments, and another on common stanzas for example. But would this be allowed by the tech proposal system?

I believe this exact printing is solvable problem. However I lack the resources to solve this personally. I think a tech proposal may be able to provide said resources. Although the process involved is daunting enough for me to first ask the community, do people want this? Please reply if you’re bothered by this, or have any comments on my approach.