sqrt' can convert meters to seconds with ghc8

[expipiplus1]

Among other things

ghc-7.10.3 seems to be fine.

This code compiles with ghc-8.0.1:

{-# LANGUAGE QuasiQuotes #-}
{-# LANGUAGE DataKinds #-}
{-# OPTIONS_GHC -fplugin Data.UnitsOfMeasure.Plugin #-}

module Bug where

import Data.UnitsOfMeasure.Defs     ()
import Data.UnitsOfMeasure

bug :: Floating a => Quantity a [u|s|] -> Quantity a [u|m|]
bug = sqrt' 

{- Neither of these compile
bug' :: Floating a => Quantity a u -> Quantity a [u|m|]
bug' = sqrt' 

bug'' :: Floating a => Quantity a [u|s|] -> Quantity a u
bug'' = sqrt' 
-}

[expipiplus1]

compiling with ghc instead of ghci panics:

[1 of 1] Compiling Bug              ( Bug.hs, Bug.o )
ghc: panic! (the 'impossible' happened)
  (GHC version 8.0.1 for x86_64-unknown-linux):
        StgCmmEnv: variable not found
  irred_a8Lt
  local binds for:
  $trModule
  $trModule1_r8L9
  $trModule2_r8LJ
  $dFloating_s8LP

Please report this as a GHC bug:  http://www.haskell.org/ghc/reportabug

[adamgundry]

Hmm, this is surprising, thanks for the report. It looks like the constraint Base "s" ~ Base "m" ^: 2 is somehow getting forgotten. I'll investigate.

[expipiplus1]

Thanks!

[expipiplus1]

It's also possible to leave off the Floating a constraint too

[adamgundry]

Okay, I think I see what is going on here. Here's an explanation for posterity...

The plugin is correctly reporting the constraint Base "s" ~~ Base "m" ^: 2 as insoluble (note the use of ~~, which is a type family provided by uom-plugin that is a pseudo-synonym of ~). However, it turns out that the GHC error-reporting machinery requires a subtle invariant, namely that only certain kinds of constraint can be regarded as insoluble. Equalities are included, but not irreducible predicates such as ~~. Failure to respect this invariant means that errors are silently ignored.

Here's my plan to work around this:

• for an insoluble wanted, neglect to report the constraint as insoluble, so it will be reported as unsolved instead (we can't change the actual wanted, otherwise it will be reported twice);
• for an insoluble given, replace ~~ with ~ when reporting the constraint as insoluble, so it will be picked up by GHC.

This isn't completely satisfactory. It might be possible to do better using the new custom type errors machinery, but I haven't had time to look at that yet.

I suspect it would be a good idea to change GHC to always report insoluble constraints, even unusual ones arising from plugins. Otherwise this seems like a good source of subtle bugs.

[expipiplus1]

Thanks!