Simon, I think we've been trying to be too clever...
The simple question is: for a given extension, what is the risk of
leaving it turned on by default?
Clearly we don't want extensions turned on that causes code to compile
but with a different meaning. We may not want extensions turned on that
cause most reasonable code not to compile.
But I would say neither risk is significant in the case of most
extensions. To use your examples:
* FFI doesn't not cause h98 code to compile to a different meaning. The
worst case is that code that uses 'foreign' as a function name doesn't
compile. That seems okay in that more code probably uses FFI than uses
foreign as a function name and the user can apply a language pragma to
turn it off if really desired.
* Existential types won't cause h98 code to compile with a different
meaning. The worst case is that code that uses 'forall' as a type
variable won't compile. That seems ok....
* TemplateHaskell also not compatible with h98. The worst case is the
loss "[d|" in list comprehensions.
* MagicHash: Does not appear in the ToC or the Index of the user's guide
so should probably be turned off. I have no idea what it does.
Note, in all cases where the extension is turned on by default, there
should be a language pragma to turn it off.
-Alex-
Simon Marlow wrote:
Alex Jacobson wrote:
Simon, from what I can tell, with GHC 6.8.1, use of foreign as a
function name or forall as a type variable or leaving out a space in a
list-comprehension doesn't "parse differently" when the relevant
extensions are enabled, it causes a parse error.
>
Extensions allow the same code to parse but with different meanings
need to be declared explicitly. But, extensions that are obvious from
syntax should be allowed to be declared simply from the use of that
syntax.
So for the first example I gave,
f x y = x 3# y
the "MagicHash" extension is one that you'd require to be explicitly
declared, because the expression parses both with and without the
extension.
Now, Let's take the Template Haskell example:
f x = [d|d<-xs]
So this is valid Haskell 98, but invalid H98+TH. You would therefore
like this example to parse unambiguously as H98, correct? But in order
to do that, our parser would need arbitrary lookahead: it can't tell
whether the expression is legal H98+TH until it gets to the '<-' in this
case. Certainly it's possible to implement this using a backtracking
parser, but Haskell is supposed to be parsable with a shift-reduce
parser such as the one GHC uses. Or we could try parsing the whole
module with various combinations of extensions turned on or off, but I'm
sure you can see the problems with that.
So basically the problem is that you need a parser that parses a strict
superset of Haskell98 - and that's hard to achieve.
Cheers,
Simon
I am not taking a position here on the merits of any extensions.
-Alex-
Simon Marlow wrote:
Alex Jacobson wrote:
Extensions that change syntax are effectively declared by the use of
that syntax. If you can parse the source, then you know which
extensions it uses.
I thought we'd already established that this isn't possible. Here
are some code fragments that parse differently depending on which
extensions are enabled:
f x y = x 3# y
f x = [d|d<-xs]
foreign x = x
f :: forall -> forall -> forall
You could argue that these syntax extensions are therefore badly
designed, but that's a separate discussion.
Cheers,
Simon
-Alex-
Duncan Coutts wrote:
On Fri, 2007-11-23 at 16:26 +0100, Wolfgang Jeltsch wrote:
Am Freitag, 23. November 2007 03:37 schrieben Sie:
On Fri, 2007-11-23 at 01:50 +0100, Wolfgang Jeltsch wrote:
Dont’t just think in terms of single modules. If I have a Cabal
package,
I can declare used extensions in the Cabal file. A user can
decide not
to start building at all if he/she sees that the package uses an
extension unsupported by the compiler.
Indeed. In theory Cabal checks all the extensions declared to be
used by
the package are supported by the selected compiler. In practise
I'm not
sure how well it does this or what kind of error message we get.
The problem is, of course, that you are not forced to specify all
used extensions in the Cabal file since you can still use language
pragmas. Sometimes it is even desirable to use LANGUAGE pragmas
instead of information in the Cabal file. For example, even if
some modules use undecidable instances, I might not want all
modules of the package to be compiled with -XUndecidableInstances
since this could hide problems with my class structure.
Our tentative plan there is to separate the extensions field into
those
used in some module, and those applied by cabal to every module. So
that
would allow you to specify a feature in one file but not all, while
still declaring to the outside world that the package uses the
feature.
As for enforcing that, that may come almost for free when we get
dependency chasing as we'll be looking for imports anyway. It
shouldn't
be much harder to look for language pragmas too.
Duncan
_______________________________________________
Glasgow-haskell-users mailing list
Glasgow-haskell-users@haskell.org
http://www.haskell.org/mailman/listinfo/glasgow-haskell-users
_______________________________________________
Glasgow-haskell-users mailing list
Glasgow-haskell-users@haskell.org
http://www.haskell.org/mailman/listinfo/glasgow-haskell-users
_______________________________________________
Glasgow-haskell-users mailing list
Glasgow-haskell-users@haskell.org
http://www.haskell.org/mailman/listinfo/glasgow-haskell-users
