Wired-in data-constructors with UNPACKed fields

[Herbert Valerio Riedel]

Hello Simon,

On 2014-08-19 at 00:01:17 +0200, Simon Peyton Jones wrote:

[...]

> But you can probably write the code in such a way as to be mostly
> independent (eg explicit UNPACK rather than rely on
> -funbox-strict-fields), or assume that some things won't happen
> (e.g. souce module will not be compiled with
> -fomit-interface-pragmas).  See MkId.mkDataConRep.

I was under the impression that even -O0 vs -O1+ makes a huge
difference:

As given the following program,

  {-# LANGUAGE MagicHash #-}
  module M where
  import GHC.Exts
  data T0 = C0 ByteArray#
  data T1 = C1 {-# UNPACK #-} !T0
          | C2 {-# UNPACK #-} !Int
          | C3 !Int
          | C4 Int

compilation with

  $ ../inplace/bin/ghc-stage2  -fforce-recomp  -ddump-types -O1 -c M.hs 
  TYPE SIGNATURES
  TYPE CONSTRUCTORS
    data T0 = C0 ByteArray#
    data T1
      = C1 {-# UNPACK #-}T0
      | C2 {-# UNPACK #-}Int
      | C3 {-# UNPACK #-}Int
      | C4 Int
  COERCION AXIOMS
  Dependent modules: []
  Dependent packages: [base, ghc-prim, integer-gmp2]

has everything but C4 unpacked as expected, but when using -O0, nothing
is UNPACKed at all:

  $ ../inplace/bin/ghc-stage2  -fforce-recomp  -ddump-types -O0 -c M.hs 
  TYPE SIGNATURES
  TYPE CONSTRUCTORS
    data T0 = C0 ByteArray#
    data T1 = C1 !T0 | C2 !Int | C3 !Int | C4 Int
  COERCION AXIOMS
  Dependent modules: []
  Dependent packages: [base, ghc-prim, integer-gmp2]

...am I interpreting the output `-ddump-types` incorrectly?


PS: adding a '!' in front of the 'ByteArray#' field in `T0` is not
    supposed to have any effect on primitive types, is it? If so, should
    GHC warn about the redundant '!'?

Cheers,
  hvr