NoRecursion plugin

A GHC plugin to remove support for recursion

General recursion can be the cause of a lot of problems. This removes recursion from GHC, allowing you to guarantee you’re using other mechanisms, like recursion schemes.

usage

Add no-recursion to your build dependencies.

Add -fplugin NoRecursion to your GHC options. This can be done per-module with

{-# options_ghc -fplugin NoRecursion #-}

Now, any recursion in that module will result in a compilation failure.

NB: This won’t prevent you from using recursive functions imported from other modules, but inlined definitions from other modules will be checked.

allowing some recursion

The recommended way to re-enable recursion at the module level is to add

{-# options_ghc -fplugin-opt=NoRecursion:allow-recursion:true #-}

at the beginning of the file.

If you want to re-enable it for specific definitions, the best way is to use

{-# options_ghc -fplugin-opt=NoRecursion:ignore-decls:recDef #-}

recDef :: a -> b
recDef = recDef

You can also do it with a source annotation

recDef :: a -> b
recDef = recDef
{-# ann recDef "Recursion" #-}

Unfortunately, the ann pragma isn’t allowed by Safe Haskell, so any module that uses it will be inferred as Unsafe. So you have to decide between an Unsafe (or Trustworthy) module and not enabling recursion for the entire module. This can be mitigated by putting recursive definitions in a module by themselves.

NoRecursion supports two source annotations: "Recursion" and "NoRecursion".

You can re-enable recursion for an entire module with

{-# ann module "Recursion" #-}

And then you can re-disable recursion for individual names with

nonRecDef :: a -> a
nonRecDef = id
{-# ann nonRecDef "NoRecursion" #-}

If both '"Recursion"' and "NoRecursion" annotations exist on the same name (or module), it’s treated as NoRecursion.

NB: If multiple names are mutually recursive, then they must all have recursion enabled to avoid being flagged by the plugin.

ANN has some caveats:

• If you enable the OverloadedStrings language extension, you will have to specify the type in the annotation, like

  {-# ann module "Recursion" :: String #-}
  

plugin options

The plugin currently supports four options

• allow-recursion: (true|false) whether to allow recursion by default. As mentioned above, this is the best way to re-enable recursion for a single module, but you can do the reverse and specify allow-recursion:true globally, then use allow-recursion:false per-module.

• ignore-method-cycles: (true|false) whether to ignore cycles between method definitions.the method level. This crops up a lot with errors about things like $csconcat.

• ignore-methods: (list of method names) ignores the named methods. Very useful for silencing errors about default method definitions.

• ignore-decls: (list of decl names) ignores the named decls. This is good to put at the top of a module where you have intentionally written a recursive definition.

suggestions

in $csconcat, the following bindings were recursive: go1

This particular message occurs when you define a Semigroup instance that didn’t have an explicit sconcat implementation. The default definition is recursive, and NoRecursion catches that. Similar messages occur with default definitions for other classes as well.

You can’t apply ann to methods, so here are some ways to get around this issue:

1. write an explicit non-recursive definition, or
2. add {-# options_ghc -fplugin-opt=NoRecursion:ignore-methods:sconcat #-} to the top of the module, which will ignore this method module-wide.

Unfortunately, because sconcat (and mconcat) require lazy lists ([]), it’s not possible to write a total definition for these.

versioning

This project largely follows the Haskell Package Versioning Policy (PVP), but is more strict in some ways.

The version always has four components, A.B.C.D. The first three correspond to those required by PVP, while the fourth matches the “patch” component from Semantic Versioning.

Here is a breakdown of some of the constraints:

sensitivity to additions to the API

PVP recommends that clients follow these import guidelines in order that they may be considered insensitive to additions to the API. However, this isn’t sufficient. We expect clients to follow these additional recommendations for API insensitivity

If you don’t follow these recommendations (in addition to the ones made by PVP), you should ensure your dependencies don’t allow a range of C values. That is, your dependencies should look like

yaya >=1.2.3 && <1.2.4

rather than

yaya >=1.2.3 && <1.3

use package-qualified imports everywhere

If your imports are package-qualified, then a dependency adding new modules can’t cause a conflict with modules you already import.

avoid orphans

Because of the transitivity of instances, orphans make you sensitive to your dependencies’ instances. If you have an orphan instance, you are sensitive to the APIs of the packages that define the class and the types of the instance.

One way to minimize this sensitivity is to have a separate package (or packages) dedicated to any orphans you have. Those packages can be sensitive to their dependencies’ APIs, while the primary package remains insensitive, relying on the tighter ranges of the orphan packages to constrain the solver.

transitively breaking changes (increments A)

removing a type class instance

Type class instances are imported transitively, and thus changing them can impact packages that only have your package as a transitive dependency.

widening a dependency range with new major versions

This is a consequence of instances being transitively imported. A new major version of a dependency can remove instances, and that can break downstream clients that unwittingly depended on those instances.

A library may declare that it always bumps the A component when it removes an instance (as this policy dictates). In that case, only A widenings need to induce A bumps. B widenings can be D bumps like other widenings, Alternatively, one may compare the APIs when widening a dependency range, and if no instances have been removed, make it a D bump.

breaking changes (increments B)

restricting an existing dependency’s version range in any way

Consumers have to contend not only with our version bounds, but also with those of other libraries. It’s possible that some dependency overlapped in a very narrow way, and even just restricting a particular patch version of a dependency could make it impossible to find a dependency solution.

restricting the license in any way

Making a license more restrictive may prevent clients from being able to continue using the package.

adding a dependency

A new dependency may make it impossible to find a solution in the face of other packages dependency ranges.

non-breaking changes (increments C)

adding a module

This is also what PVP recommends. However, unlike in PVP, this is because we recommend that package-qualified imports be used on all imports.

other changes (increments D)

widening a dependency range for non-major versions

This is fairly uncommon, in the face of ^>=-style ranges, but it can happen in a few situations.

deprecation

NB: This case is weaker than PVP, which indicates that packages should bump their major version when adding deprecation pragmas.

We disagree with this because packages shouldn’t be publishing with -Werror. The intent of deprecation is to indicate that some API will change. To make that signal a major change itself defeats the purpose. You want people to start seeing that warning as soon as possible. The major change occurs when you actually remove the old API.

Yes, in development, -Werror is often (and should be) used. However, that just helps developers be aware of deprecations more immediately. They can always add -Wwarn=deprecation in some scope if they need to avoid updating it for the time being.

licensing

This package is licensed under The GNU AGPL 3.0 or later. If you need a license for usage that isn’t covered under the AGPL, please contact Greg Pfeil.

You should review the license report for details about dependency licenses.

comparisons

Other projects similar to this one, and how they differ.

WartRemover

WartRemover is a Scala linting tool. A Recursion wart was added in 2017, and I’ve been meaning to write this plugin ever since. It only took seven years to find a few hours to make it happen …