Overcoming Software

ghcid for the win!

Supercharge your Haskell development experience with ghcid!

ghcid is – at the current moment – the most important tool for Haskell development environments. It is fast, reliable, works on all kinds of projects, and is remarkably versatile. You can use it with any editor workflow, primarily by not integrating your editor! (though there are integrations available if you’re brave) For these reasons, whenever someone asks about a Haskell IDE, I tell them to ignore the siren song of ghc-mod, hdevtools, intero, haskell-ide-engine, etc1, and just stick with the old faithful GHCi and ghcid. Use whatever editor you want – make sure it has syntax highlighting, and open up GHCi and/or ghcid in a separate terminal.

Here are some things we’re going to do with it in this post:

As I think of additional “tricks” with ghcid, I will be updating this post and adding them here. If you have a suggestion or question, please open an issue on my blog’s GitHub :)

Basic warnings and errors

This is the bread and butter of what ghcid is good for. At this point, you’re probably used to running ghci and doing :reload to see whether or not your code compiles. ghci has some advantages over a cabal new-build or stack build or similar – it loads everything in interpreted byte code by default, which is much faster, and is capable of very intelligent module reloading to minimize work. This can cut the feedback time from compilation dramatically.

By default, ghcid will load with the flag -fno-code enabled. This turns off all code generation, and basically only gives you syntax and type checking. When you eventually customize your ghcid command, you will want to remember to either enable -fno-code or -fobject-code. You need -fobject-code in order to do stuff like run tests, check Template Haskell expressions, etc.

To customize your ghcid command, you do this:

$ ghcid --command "the command to start ghci"
# example, for a Template Haskell heavy project:
$ ghcid --command "stack ghci package:lib --ghci-options=-fobject-code"
# example, to pick a single executable target:
$ ghcid --command "stack ghci package:exe:main-node"
# example, to defer type errors:
$ ghcid --command "stack ghci --ghci-options=-fdefer-type-errors"

At IOHK, I wrote up a Makefile with the common ghcid commands I use when working on the new wallet. This command lets me say make ghcid in the wallet-new subdirectory and get lightning fast reloading of code, display of all warnings and errors, and lets me run through refactorings quite nice and quickly.

Fake type-of-expression support

Sometimes GHC feels like a reluctant wizard. It knows things. You know it knows things. It knows that you know that it knows things. But it doesn’t want to tell you!

A common question that IDE authors want to ask is “What’s the type of this expression?” ghc-mod, intero, all try to support this, to varying degrees of success and performance. But GHC is curious and easily distracted, and would much rather tell you that you’re wrong than answer a question. So let’s trick the wizard!

Just today, I was working on this snippet of code, pulled from my servant-persistent starter pack/example project:

main :: IO ()
main = do
    env  <- lookupSetting "ENV" Development
    port <- lookupSetting "PORT" 8081
    logEnv <- defaultLogEnv
    pool <- makePool env logEnv
    store <- serverMetricStore <$> forkServer "localhost" 8000
    waiMetrics <- registerWaiMetrics store
    metr <- M.initializeWith store
    let cfg = Config { configPool = pool
                     , configEnv = env
                     , configMetrics = metr
                     , configLogEnv = logEnv }
        logger = setLogger env
    runSqlPool doMigrations pool
    generateJavaScript
    run port $ logger $ metrics waiMetrics $ app cfg

I wanted to know what the type of the port variable was. With a more sophisticated toolchain, I might hover over port, and get a tooltip telling me what. But, we’re using the more primitive ghcid. Well, we know what it isn’t – It’s not (). So, in the olden tradition, let’s loudly be wrong and await correction:

    run (port :: ()) $ logger $ metrics waiMetrics $ app cfg

We fire up ghcid, making sure to include the executable package target:

$ ghcid --command "stack ghci servant-persistent:exe:perservant"

And we’re greeted with an error message:

/home/matt/Projects/servant-persistent/app/Main.hs:37:10: error:
    • Couldn't match type ‘()’ with ‘Int’
      Expected type: warp-3.2.22:Network.Wai.Handler.Warp.Types.Port
        Actual type: ()
    • In the first argument of ‘run’, namely ‘(port :: ())’
      In the expression: run (port :: ())
      In a stmt of a 'do' block:
        run (port :: ()) $ logger $ metrics waiMetrics $ app cfg
   |
37 |     run (port :: ()) $ logger $ metrics waiMetrics $ app cfg
   |          ^^^^^^^^^^

Ah, GHC expects it to be of type Int. There we go!

This works well with functions, too. Let’s say we want to know the type of run, instead:

    (run :: ()) port $ logger $ metrics waiMetrics $ app cfg

ghcid is happy to tell us how wrong we are:

/home/matt/Projects/servant-persistent/app/Main.hs:37:5: error:
    • Couldn't match expected type ‘Integer
                                    -> Network.Wai.Application -> IO ()’
                  with actual type ‘()’
    • The function ‘run :: ()’ is applied to one argument,
      but its type ‘()’ has none
      In the expression: (run :: ()) port
      In a stmt of a 'do' block:
        (run :: ()) port $ logger $ metrics waiMetrics $ app cfg
   |
37 |     (run :: ()) port $ logger $ metrics waiMetrics $ app cfg
   |     ^^^^^^^^^^^^^^^^
/home/matt/Projects/servant-persistent/app/Main.hs:37:6: error:
    • Couldn't match expected type ‘()’
                  with actual type ‘warp-3.2.22:Network.Wai.Handler.Warp.Types.Port
                                    -> Network.Wai.Application -> IO ()’
    • Probable cause: ‘run’ is applied to too few arguments
      In the expression: run :: ()
      In the expression: (run :: ()) port
      In a stmt of a 'do' block:
        (run :: ()) port $ logger $ metrics waiMetrics $ app cfg
   |
37 |     (run :: ()) port $ logger $ metrics waiMetrics $ app cfg
   |      ^^^

Note that we get a slightly inconsistent message. We’ve asserted that run :: (), and it has two expected types: one from definition, and one from inferred use. The inferred type is Integer -> Application -> IO (). The defined type is Port -> Application -> IO (). It infers Integer because, without run forcing port to be a Port, it has nothing else to tell it what to be, and therefore defaults to Integer.

Reload your web app on every edit

ghcid, in addition to the --command flag, also takes a --test flag. The flag name is somewhat too specific – upon a successful compile with no warnings or errors, it will issue that command to GHCi for you. It was initially intended for running tests, but we can do anything we like with it – and we are going to use it to get our web application reloading lightning fast.

This PR on the servant-persistent project includes the necessary changes to get this running. I have left a self-review on the PR, so I won’t explain too much here.

The ghcid command we use is:

ghcid \
    --command "stack ghci servant-persistent" \
    --test "DevelMain.update"

This calls the DevelMain.update function on every successful compile. The DevelMain module was mostly copied from the Yesod scaffold, with a few updates to make it work with this repo. In truth, all you need to provide is a development-oriented function IO Application that boots your application and gives you the WAI value. The DevelMain code uses foreign-store library to persist the state across GHCi sessions.

The ghcid README links to an article on threepenny-gui apps with a similar strategy.

This is really fast – because GHCi can reload only exactly what it needs, and doesn’t have to link anything, you get to see your changes almost immediately.

Run your test suite on every edit

What’s better than knowing your project compiles? Knowing that it passes the test suite!

We’ll use the --test command here, and we specify that we want to run the tests. In the cardano-sl repository, I put a Makefile command for test running:

ghcid-test: ## Have ghcid run the test suite for the wallet-new-specs on successful recompile
	ghcid \
	    --command "stack ghci cardano-sl-wallet-new:lib cardano-sl-wallet-new:test:wallet-new-specs --ghci-options=-fobject-code" \
	    --test "main"

There’s a tricky bit here: We have to tell stack ghci which package targets to load. I specify the library (cardano-sl-wallet-new:lib) so that it adds the library to the set of modules to watch for reloading. Then I specify the test-suite I want to run (cardano-sl-wallet-new:test:wallet-new-specs). Finally I use --ghci-options=-fobject-code, because this is fast, and I need to actually run the code (you get weird linker errors if you do -fno-code and try to run the nonexistent code).

Conclusion

ghcid is awesome and everyone owes Neil Mitchell a beverage.

If you have a suggested use case you want added, ping me on GitHub and I’ll credit you :)

  1. These are great projects. But they are flaky, partially because GHC’s API is difficult to interface with, and partially because GHCi’s interactive features have some performance issues with larger code bases. For small projects and libraries, they often work great. For larger projects, or more varied environments, they show their pain. You can spend a lot of time fussing with the editor integration and waiting on some command to finish, or you can just develop habits that don’t need them (like ghcid in a separate terminal). I say this as the author of the intero-neovim plugin.