GHC and type-family rewriting?
benjamin.redelings at gmail.com
Sat Dec 17 15:17:01 UTC 2022
Thanks! This context of where the current GHC approach came from is
quite helpful. I think I see what you mean by reducing type families
only in "strict" positions... and maybe I see what you mean by trying to
make the whole approach less ugly. I still need to do quite a bit of
plumbing work before I get to anything that interesting.
But supposing I do get there, I'm curious if there are some papers on
term-rewriting that would be helpful to set the context? The OutsideIn
paper mentions Kapur (1997) "Shostak's congruence closure as completion"
in support of the flattening idea.
On 12/8/22 11:48 PM, Richard Eisenberg wrote:
>> On Nov 30, 2022, at 9:42 PM, Benjamin Redelings <benjamin.redelings at gmail.com> wrote:
>> (Q1) Did GHC evolve to this point starting from something fairly close to the OutsideIn paper?
>> (Q2) Is the new approach (i.e. eager type family rewriting) mostly to making rewriting faster?
> No. Simpler, not faster (and not slower). Or that was the intent.
>> (Q3) Does it sound reasonable to implement the approach from the OutsideIn paper, and than gradually transform it to look more like GHC?
> Sure, but I'm not sure what the advantage of doing so would be.
> This is all my doing: for years and years, GHC's treatment of type families was as described in OutsideIn. But I never could quite figure out why we needed to have flattening variables. And so I got rid of them -- this seemed like a simplification. I'm not sure it really panned out, though: without flattening variables, we need these cycle-breaker variables (which are pretty gross). On the flip side, I think the new approach might enable the possibility of reducing type families only in "strict" positions (e.g. the argument to another type family or perhaps a class during instance lookup). In the end, I don't think either the old way or the new way is the Right Answer. Maybe you can come up with something better than both!
Thanks! This is quite helpful.
More information about the ghc-devs