Unrestricted values in primitive data types

[andrewthad]

Make the data types Double, Float, Char, Ptr, Int{N}, and Word{N} be unrestricted in the unboxed argument in their data constructor. This is related to the discussion in https://github.com/tweag/ghc/issues/179.

[andrewthad]

Weird. This works for the types defined in GHC.Int but not for things from GHC.Word or GHC.Types. In ghci, I can confirm this with:

>>> :info Word8
data Word8 = W8# ghc-prim-0.5.3:GHC.Prim.Word#
    -- Defined in ‘GHC.Word’
>>> :info Int8
data Int8 where
  I8# :: ghc-prim-0.5.3:GHC.Prim.Int# -> Int8
        -- Defined in ‘GHC.Int’
>>> :info Int
data Int = I# ghc-prim-0.5.3:GHC.Prim.Int#
        -- Defined in ‘ghc-prim-0.5.3:GHC.Types’

The data constructor for Int8 is unrestricted like it should be. The data constructors for Word8 and Int are not.

[aspiwack]

It may be an avatar of #169 , though (or another printing issue). I don't think I've modified any of this: do you mean that you have tried to make this modification and that's what you observe in your patch?

[andrewthad]

Yes. I've modified all of these data constructors in the PR. It's not just the printing. I get multiplicity errors when I try to consume Int linearly by pattern matching on it and using the internal Int# multiple times. However, if I do the same thing with Int64, it works fine. The weirdest part is the fixed-size integer types behave as I expect them to, but fixed-size word types don't. I tried make clean and totally rebuilding because I'm worried I just messed up the build somehow. I'm going to try on another computer this evening.

[aspiwack]

Ah… sorry, I had opened this from my email and didn't realise it was a PR rather than an issue. Obviously: sleep is neeed.

It could have happened if you had forgotten to add -XLinearTypes to some of these files. But it doesn't appear to be the case. So I'm equally puzzled.

Come to think of it, I don't think we can merge this yet, because the CPR worker-wrapper split (which is critical for numerical computation) is deactivated for unrestricted constructors. I'm, obviously, not allowed to merge such a performance regression.

[andrewthad]

Gotcha. I'll keep trying to figure out why this is messed up, and whenever worker-wrapper gets fixed, hopefully, this should be ready to merge.