On Thu, Oct 4, 2012 at 11:19 PM, Jacques Carette <carette@mcmaster.ca> wrote:

On 04/10/2012 7:15 PM, bob zhang wrote:

Has anyone have the experience using polymorphic variant for a big Ast?
The benefit I can think of is open recursion, global namespace(not in a module). Did anyone give a try?

I have -- mostly because I needed subtying, since the language I was modelling had a kind of subtyping that polymorphic variants could track 'for free'.

It works. But it can be a real pain too: depending on your use cases, you may need a fair amount of annotations (casts). And if you make a mistake, the error messages are truly frightening, especially when you have an AST with 4 mutually recursive parts, totaling about 25 cases, and the mistake is 3 or 4 levels deep -- the error messages can go on for pages and pages. Buried in there will be the information you need to fix the mistake, but finding it can be extremely challenging.

Thanks for your warning, that's really helpful

I would say: use it only if you really really need what polymorphic variants 'buy' you, else stay away. In my original code, I have rewritten most of it to use normal variants (except for one case) and use explicit open recursion (i.e. extra type variable + tying the knot) to get the job done. The error messages are sane now.

I am considering change Camlp4 Ast node into Polymorphic Variants(I know it's hard, suppose I could make it) The problem with normal Algebraic Data Type to me is that I have to use functors to abstract over types, and functors are really intrusive, you need to write a large amount of signatures everywhere.

To make things much worse, the ADT depends on the qualified module names which makes the bootstrapping system fragile. So I think polymorphic variant may fit this area nicely.

As you said, the error message is something I worried about. How about performance issues?

Note that I don't think the error messages were incorrect in any way (I am sure they were not). It might have been possible to have made them friendlier / more precise, but I am not even sure of that.

Jacques