Where can BangPatterns appear

Question

From GHC user guide it seems like most Pat can be PBangPat, but there are a few exceptions. e.g. top-level bangs in a module (like !main) aren't allowed and x : !xs fails to parse x : (!xs) parses thanks @chi. What is the formal specification about where the bangs can be added? I've looked into some chapters of the user guide and the Report but found nothing.

Platinocyanide

Good question, though in doubt it's probably best not to use BangPatterns at all but good old seq. – Zola

Answer 1

There is no accepted formal specification for BangPatterns since they are not a part of any Haskell Report. The closest thing we have to a specification is the User's Guide along with the haskell-prime proposal it links to.

Both of those sources explicitly mention that a bang pattern is not allowed at the top level of a module.

As for x : !xs, the User's Guide has this to say about the syntax of bang patterns:

We add a single new production to the syntax of patterns:

pat  ::= !pat

It should be read in conjunction with the Haskell 2010 Report:

pat  ::= lpat qconop pat
       | lpat

lpat ::= apat
       | - (integer | float)
       | gcon apat_1 ... apat_k

apat ::= var [ @ apat]
       | ...
       | ( pat )
       | ...

According to these rules x : !xs actually should parse (since !xs is a pat, the whole thing is lpat qconop pat). So either the User's Guide (and the haskell-prime proposal) is wrong or GHC is wrong on this point.

I believe that in practice the syntax accepted by GHC is "anything that looks like a valid expression" including interpreting (!x) as a section of the operator !. For example (! Just x) is accepted as a pattern but (! ! x) is not.

Academy

Actually the relevant section of the User's Guide is so old that it links to the Haskell 98 report, and the BNF grammar at that time treated precedence differently. For patterns in particular, adding pat ::= !pat means to add it at the lowest precedence. – Define

Although disturbingly, by this reading (! ! x) should work but doesn't. – Define

@ØrjanJohansen, oh good point. I didn't try following the links to the Report since I already had the 2010 Report open in another tab. That does make the description closer to reality, though as you say still not quite the same. – Academy

See also ghc.haskell.org/trac/ghc/ticket/1087 and ghc.haskell.org/trac/ghc/ticket/10732. – Wheaten