Error thunks not being exposed with "B" strictness

[simonpj]

At the moment GHC often floats error "urk" to the top level (which it should), after the strictness analyser. But since it's after strictness analysis, the exported thing doesn't have a strictness signature saying "I am a diverging term", which in turn loses useful optimisations in importing modules.

An example is test T3286. If you compile it with -O --ddump-simpl, you'll see stuff like

     case T3286b.$fFractionalLogFloat3
          `cast` (CoUnsafe T3286b.LogFloat GHC.Prim.Double#
                  :: T3286b.LogFloat ~ GHC.Prim.Double#)
     of ww2_aGk { __DEFAULT ->
     (GHC.Types.D# ww2_aGk)
     `cast` (sym T3286b.NTCo:LogFloat
             :: GHC.Types.Double ~ T3286b.LogFloat)

But if you look at T3286b.$fFractionalLogFloat3, it turns out to be bottom, so the case should be eliminated.

This is a long-standing infelicity; I'm making a ticket so I don't forget it.

[simonpj]

I believe this is fixed in the HEAD at least. See Note [Bottoming floats] in SetLevels.

Simon