It seems to me that fixity information behaves more like semantics than like syntax. For instance, fixities may be imported, and obey namespacing rules. Knowing and correctly handling these rules seems beyond the scope of a mere parser: I would hope that a single Haskell file could be parsed without reference to any files, and fixity declarations seem to be just about the only thing which prevent this -- hence my suggestion to change the AST in order to regain this property. The use I envision of it is as I described: writing a quasiquoter using HSE to parse the user's Haskell expressions. The problem is that, for such a case, HSE (or any other parser) is forced to parse infix expressions for which it cannot possibly know the correct fixities. Any result with more information than the list form I gave would be a lie. I realise that I don't know how fixities are implemented in Haskell compilers, so perhaps I'm misunderstanding how they are treated. Cheers, Reiner On Sat, Nov 22, 2008 at 11:54 PM, Niklas Broberg wrote:

...

Of course, this would require a change to Template Haskell, so a second-best solution would be to forbid unparenthesised expressions in my quasiquoter. Then, parsing can proceed correctly without knowing the fixities. This would be easiest to do if haskell-src-exts changed its AST in a similar way to described above for Template Haskell.

I'm not sure I follow you here. In what way would it be simpler if HSE changes its AST to a less-information constructor? I won't do that, for the same reason you point out with TH and disadvantages when using it as a library, though I'm still curious what uses you envision and how it would be made easier. I'm still in the process of designing the fixity support for HSE, and all input is valuable. :-)

Cheers,

/Niklas

It turns out that there is at least a (partial) solution to my quasiquote problem. Template Haskell's "reify" function can be used to find an operator's fixity, although it seems not for all cases. However, for the purposes of this discussion, suppose I can write a function userFixity :: String -> Q Fixity which takes a operator used in user code and returns its fixity, in the Q monad (Template Haskell's monad). I want this information to be used somehow when creating the Template Haskell AST, so that the operators used have the correct fixities. If I use HSE for parsing Haskell expressions, then I want it to tell me where it unsure of the fixities, so that I can insert the correct ones there. For this use case, I would want HSE to use an AST like I suggested, because it allows the parser to say, "I'm not sure what the correct fixity is". Cheers, Reiner On Sun, Nov 23, 2008 at 8:23 AM, John A. De Goes wrote:

Though many see it as "losing" information, I agree wholeheartedly with your proposal to change the AST.

It's better to have an AST that conveys less information, but truthfully, than to have an AST that purports to convey more information, when in fact that information is false.

In most languages, some things just can't be known at parse time. They need to be resolved later.

In this case, the most important thing is following the principle of least surprise: A Haskell expression inside a quasiquote should work the same as a Haskell expression outside a quasiquote.

Violating the "principle of least surprise" is one of the most grievous mistakes language (and interface) designers make.

Regards,

John A. De Goes N-BRAIN, Inc. http://www.n-brain.net [n minds are better than n-1]

On Nov 22, 2008, at 9:02 AM, Reiner Pope wrote:

...

It seems to me that fixity information behaves more like semantics than like syntax. For instance, fixities may be imported, and obey namespacing rules. Knowing and correctly handling these rules seems beyond the scope of a mere parser: I would hope that a single Haskell file could be parsed without reference to any files, and fixity declarations seem to be just about the only thing which prevent this -- hence my suggestion to change the AST in order to regain this property.

The use I envision of it is as I described: writing a quasiquoter using HSE to parse the user's Haskell expressions. The problem is that, for such a case, HSE (or any other parser) is forced to parse infix expressions for which it cannot possibly know the correct fixities. Any result with more information than the list form I gave would be a lie.

I realise that I don't know how fixities are implemented in Haskell compilers, so perhaps I'm misunderstanding how they are treated.

Cheers,

Reiner

On Sat, Nov 22, 2008 at 11:54 PM, Niklas Broberg wrote:

...

...

Of course, this would require a change to Template Haskell, so a second-best solution would be to forbid unparenthesised expressions in my quasiquoter. Then, parsing can proceed correctly without knowing the fixities. This would be easiest to do if haskell-src-exts changed its AST in a similar way to described above for Template Haskell.

I'm not sure I follow you here. In what way would it be simpler if HSE changes its AST to a less-information constructor? I won't do that, for the same reason you point out with TH and disadvantages when using it as a library, though I'm still curious what uses you envision and how it would be made easier. I'm still in the process of designing the fixity support for HSE, and all input is valuable. :-)

Cheers,

/Niklas

_______________________________________________ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe

_______________________________________________ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe

On Mon, Nov 24, 2008 at 12:39 AM, Niklas Broberg wrote:

...

I want this information to be used somehow when creating the Template Haskell AST, so that the operators used have the correct fixities. If I use HSE for parsing Haskell expressions, then I want it to tell me where it unsure of the fixities, so that I can insert the correct ones there. For this use case, I would want HSE to use an AST like I suggested, because it allows the parser to say, "I'm not sure what the correct fixity is".

As noted above, I really don't like that change. If you use HSE for parsing expressions, it would *never* know the correct fixities, and you would always get a completely left-biased tree. Why would that be harder to work with? I understand the argument about least surprise, and that this feature must be strongly documented (which it currently isn't), but for practical purposes I don't see why the current state would be problematic. It should even be trivial to convert the left-biased tree into an intermediate structure exactly like the one you suggest, no?

No, I believe it wouldn't. The left-biased tree cannot distinguish where parentheses have been used from where HSE inserted its own left fixities. For instance, if we have the expressions xs ++ ys ++ zs (xs ++ ys) ++ zs Then HSE will return something like (I'm using strings for the subexpression parses, for simplicity) InfixE (InfixE "xs" "++" "ys") "++" "zs" for both the first and second parses. However, if I then use the knowledge that ++ is right infix, I will want to convert the first, but not the second parses into right infix. I can't do this, because they are both parsed the same way. I would also like to point out that a list representation as I suggested can in fact encode the correct fixities if they are known to HSE. This is true simply because the list constructor is isomorphic to the current constructor in the special case where the list of operators has length 1. For instance, in the first example above, if HSE somehow knew that ++ is right infix, it should return a parse result of InfixE ["xs", InfixE ["ys", "zs"] ["++"]] ["++"] rather than just InfixE ["xs", "ys", "zs"] ["++", "++"] Cheers, Reiner

Cheers,

/Niklas