Clarification of HsBang and isBanged

[Simon Peyton Jones]

OK, so one thing I failed to explain in the comment is this: for imported DataCons, the dcSrcBangs field is precisely the [HsImplBang] decisions computed when compiling the defining module.

So if the defining module was compiled with –O –funbox-strict-fields, GHC will make one set of choices.  Those [HsImplBang] choices are recorded in the IfaceBangs in the interface file. Those IfaceBangs in turn get put back into the dcSrcBangs field of the DataCon constructed by TcIface when reading the interface file.

So when the dcSrcBangs are in fact HsImpBangs, they should be followed slavishly, because the decisions have already been taken.  Even if the –O or –funbox-strict-fields flags differ in the importing module from the defining module.

I’ve rewritten the comment below.

Maybe the field should not be called dcSrcBangs but dcOrigBangs?

Simon

{- Note [Bangs on data constructor arguments]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider
  data T = MkT !Int {-# UNPACK #-} !Int Bool

When compiling the module, GHC will decide how to represent
MkT, depending on the optimisation level, and settings of
flags like -funbox-small-strict-fields.

Terminology:
  * HsSrcBang:  What the user wrote
                Constructors: HsNoBang, HsUserBang

  * HsImplBang: What GHC decided
                Constructors: HsNoBang, HsStrict, HsUnpack

* If T was defined in this module, MkT's dcSrcBangs field
  records the [HsSrcBang] of what the user wrote; in the example
    [ HsSrcBang Nothing True
    , HsSrcBang (Just True) True
    , HsNoBang]

* However, if T was defined in an imported module, MkT's dcSrcBangs
  field gives the [HsImplBang] recording the decisions of the
  defining module.  The importing module must follow those decisions,
  regardless of the flag settings in the importing module.

* The dcr_bangs field of the dcRep field records the [HsImplBang]
  If T was defined in this module, Without -O the dcr_bangs might be
    [HsStrict, HsStrict, HsNoBang]
  With -O it might be
    [HsStrict, HsUnpack, HsNoBang]
  With -funbox-small-strict-fields it might be
    [HsUnpack, HsUnpack, HsNoBang]


From: Johan Tibell [mailto:johan.tibell at gmail.com]
Sent: 11 January 2015 17:28
To: Simon Peyton Jones
Cc: ghc-devs at haskell.org
Subject: Re: Clarification of HsBang and isBanged

Those comments and the renaming really help. Here are a couple of more questions I got after exploring some more:

DsMeta.repBangTy look wrong to me:

    repBangTy :: LBangType Name -> DsM (Core (TH.StrictTypeQ))
    repBangTy ty= do
      MkC s <- rep2 str []
      MkC t <- repLTy ty'
      rep2 strictTypeName [s, t]
      where
        (str, ty') = case ty of
                       L _ (HsBangTy (HsSrcBang (Just True) True) ty) -> (unpackedName,  ty)
                       L _ (HsBangTy (HsSrcBang _     True) ty)       -> (isStrictName,  ty)
                       _                                              -> (notStrictName, ty)

Shouldn't the second case look at whether -funbox-strict-fields or -funbox-small-strict-fields is set and use unpackedName instead of isStrictName if so? What is repBangTy for?

A related question, in MkId.dataConArgRep we have:

    dataConArgRep _ _ arg_ty HsStrict
      = strict_but_not_unpacked arg_ty

Here we're not looking at -funbox-strict-fields and -funbox-small-strict-fields. Is it the case that we only need to look at these flags in the case of HsSrcBang, because HsStrict can only be generated by us (and we presumably looked at the flags when we converted a HsSrcBang to a HsStrict)?

On Thu, Jan 8, 2015 at 4:09 PM, Simon Peyton Jones <simonpj at microsoft.com<mailto:simonpj at microsoft.com>> wrote:
I’m glad you are getting back to strictness.

Good questions.

I’ve pushed (or will as soon as I have validated) a patch that adds type synonyms, updates comments (some of which were indeed misleading), and changes a few names for clarity and consistency.  I hope that answers all your questions.

Except these:


•         Why is there a coercion in `HsUnpack` but not in `HsUserBang (Just True) True`?  Because the former is implementation generated but the latter is source code specified.

•         Why isn't this information split over two data types.  Because there’s a bit of overlap. See comments with HsSrcBang

Simon

From: Johan Tibell [mailto:johan.tibell at gmail.com<mailto:johan.tibell at gmail.com>]
Sent: 08 January 2015 07:36
To: ghc-devs at haskell.org<mailto:ghc-devs at haskell.org>
Cc: Simon Peyton Jones
Subject: Clarification of HsBang and isBanged

HsBang is defined as:
    -- HsBang describes what the *programmer* wrote
    -- This info is retained in the DataCon.dcStrictMarks field
    data HsBang
      = HsUserBang   -- The user's source-code request
           (Maybe Bool)       -- Just True    {-# UNPACK #-}
                              -- Just False   {-# NOUNPACK #-}
                              -- Nothing      no pragma
           Bool               -- True <=> '!' specified

      | HsNoBang              -- Lazy field
                              -- HsUserBang Nothing False means the same as HsNoBang

      | HsUnpack              -- Definite commitment: this field is strict and unboxed
           (Maybe Coercion)   --    co :: arg-ty ~ product-ty

      | HsStrict              -- Definite commitment: this field is strict but not unboxed

This data type is a bit unclear to me:

 * What are the reasons for the following constructor overlaps?
   * `HsNoBang` and `HsUserBang Nothing False`
   * `HsStrict` and `HsUserBang Nothing True`
   * `HsUnpack mb_co` and `HsUserBang (Just True) True`

* Why is there a coercion in `HsUnpack` but not in `HsUserBang (Just True) True`?

* Is there a difference in what the user wrote in the case of HsUserBang and HsNoBang/HsUnpack/HsStrict e.g are the latter three generated by the compiler as opposed to being written by the user (the function documentation notwithstanding)?

A very related function is isBanged:
    isBanged :: HsBang -> Bool
    isBanged HsNoBang                  = False
    isBanged (HsUserBang Nothing bang) = bang
    isBanged _                         = True

What's the meaning of this function? Is it intended to communicate what the user wrote or whether result of what the user wrote results in a strict function?

Context: I'm adding a new StrictData language pragma [1] that makes fields strict by default and a '~' annotation of fields to reverse the default behavior. My intention is to change HsBang like so:

    -       Bool               -- True <=> '!' specified
    +       (Maybe Bool)       -- True <=> '!' specified, False <=> '~'
    +                          -- specified, Nothing <=> unspecified

1. https://ghc.haskell.org/trac/ghc/wiki/StrictPragma

-- Johan

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.haskell.org/pipermail/ghc-devs/attachments/20150111/4c9e355c/attachment-0001.html>