Plugin generates unhygienic code

[parsonsmatt]

Our custom prelude exports error as an alias for errorRequireCallStack, which causes the large-record plugin to fail with this compiler error message:

src/Handler/Perks/Types.hs:16:14-25: error: [GHC-64725]
    • Add RequireCallStack to your function context or use provideCallStack
    • In the expression:
        error
          "Pattern match failure in vectorToPerksSecrets: vector with invalid number o
f elements."
      In a case alternative:
          _ -> error
                 "Pattern match failure in vectorToPerksSecrets: vector with invalid n
umber of elements."
      In the expression:
        case Data.Record.Plugin.Runtime.anyArrayToList _x of

We can work-around this by hiding custom error and using Prelude.error but I think it'd be better if the plugin were hygienic about the terms it uses

[edsko]

Using Exact names in source plugins in a pain in the ass: you need to provide the location of where the symbol is defined; you cannot provide the location of a module that merely re-exports it. For this reason large-records comes with a Data.Record.Plugin.Runtime module that defines type aliases and term aliases for nearly everything we need. Unfortunately, we cannot define class aliases; if you do

type MyEq = Eq

then ghc will not accept

instance MyEq T where ..

This means that for classes we must define the precise location of the class; for the Prelude names I wasn't doing this, and was just using whatever was in scope, to sidestep this issue. I have now pushed a fix to #164 that imports these names from GHC.Classes (Eq and Ord) and GHC.Show (Show) instead. Unfortunately, the former comes from ghc-prim, so the (small) price to pay is that that is now yet another dependency to add (for a total of three: ghc-prim, large-records, and record-hasfield). Once that PR passes CI I will make a release.

[edsko]

Released as https://hackage.haskell.org/package/large-records-0.4.2 .