Hi Simon,

Apologies for my slow reply. It's been nearly 15 years (!) since I wrote distributed-process and distributed-static (fun fact: it was my first task when I joined Well-Typed) and I felt I needed to set aside some time to properly page everything back in, as I haven't really used static pointers ever since (somewhat sadly..). I realize that the proposal how now been written and accepted, so this email comes too late in a way, but I finally had some time now to look at this properly. In hindsight, the proposal is simple enough, but I didn't realize that at the time.

Anyway. The proposal is fine with me :) I seems to me that it's no more than a minor inconvenience to the programmer. I have now prepared a Haskell Unfolder episode (which will likely be the next episode we air, episode #53) on static pointers, where I discuss the primitive building blocks provided by ghc, and show how we can implement some compositional infrastructure around that (including some thoughts on type classes, which I believe you and I discussed a very long time ago). I'll also mention the proposal and the upcoming changes in 9.14.2 (warning) and 9.16 (change implemented). I tried building it with your branch (wip/T26718) and that all worked fine. I also added one module in which your branch issues the warning expect, and that seemed good also.

By the way, I am not entirely sure why the proposal mentions the IsStatic class explicitly, but I just wanted to mention that it is really quite useful; I declare two instances in my Unfolder episode :)

-Edsko

On 11/19/25 13:02, Simon Peyton Jones wrote:

Thanks Laurent and Mathieu

Edsko, Duncan, Facundo: do you have any views?

I have now written a GHC proposal.  Could you add your thoughts to it?   I think it's a no-brainer myself, but we should go through the proper process.

Do any of you know people I should consult?   Eg authors of libraries that use StaticPtrs?

Thanks

Simon

On Fri, 7 Nov 2025 at 21:18, Simon Peyton Jones <simon.peytonjones@gmail.com> wrote:

Dear Laurent, Duncan, Mathieu, Facundo, Edsko

I have spent a little while digging into static pointers recently.  See my post below.   I wonder if you have any comments on my proposal?  

Do you know anyone else I should consult?

Thanks!

Simon

On Fri, 7 Nov 2025 at 18:13, Simon Peyton Jones (@simonpj) <gitlab@gitlab.haskell.org> wrote:

Simon Peyton Jones created an issue: #26556

Static pointers are not properly implemented. For example:

• #26545
• #24464
• #24773

among others. Moreover, the implementation is very messy, scattered about in FloatOut and elsewhere.

Let's fix it.

Discussion

I embarked on what I thought would be a simple refactor to

• Identify static bindings in the type checker
• Promote them to top level desugarer

thereby avoiding all the terribly painful static-form-floating stuff that been an ongoing source of breakage and irritation.

Sadly it was not as simple as I had thought. Merge request !14994 is my work in progress

• At first it seems simple: given static e

  • When typechecking e ensure that all its free variables are top-level defined
  • When desugaring, move e to top level

  Apparently simple!

• Complication 1. e might generate constraints. We don't want to solve those from locally-bound Givens, because they'll be out of scope when we promote to top level.

  Solution: wrap the constraints in an implication with SkolInfo of StaticFormSkol; and in the constraint solver zap all Givens when walking inside such an implication. That was done in

  commit 39d4a24beaa7874a69ffdc1528ca160818829169
  Author: Simon Peyton Jones <simon.peytonjones@gmail.com>
  Date:   Tue Sep 30 23:11:19 2025 +0100
  
    Build implication for constraints from (static e)
  
    This commit addresses #26466, by buiding an implication for the
    constraints arising from a (static e) form.  The implication has
    a special ic_info field of StaticFormSkol, which tells the constraint
    solver to use an empty set of Givens.

  So that complication wasn't at all bad.

• Complication 2. What if we have

  f x = let y = reverse "hello" in ...(static (y++y))...

  The free vars of the static are just {y}, and y is morally-top-level. It in turn has no free variables.

  Sadly (as it turns out) GHC tries to accept this case. When looking at the defn of y (with no static in sight yet) the typechecker marks it at a "static binding", meaning that it too can (and indeed must) be floated to top level.

  So if the desugarer moves the static to the top level, it must move y too. And that means it must mark the typechecked binding in some way, so the desugarer can identify it. Not so hard, but there is quite a bit of new plumbing.

• Complication 3. But what if y's RHS generates constraints, which use Givens (or solved dictionaries, which are very similar) from its context. E.g.

  f x = let p = x+1::Int; y = 2+3::Int in ...

  Now there may be a d :: Num Int lying around from dealing with p, and y may use it. Oh no! Now that'll be out of scope if we move y to top level.

  Plausible solution: use them same mechanism for static bindings as we did for static e expressions. That is, build an implication constraint whose SkolInfo says "zap Givens". This turned out to be considerably harder to implement than it was for Complication 1.

• Complication 4. What if y is not generalised, perhaps because of the Monomorphism Restriction? e.g.

  f :: Num a => a -> blah
  f x = let y = 3+3 in (x+y, static( ..y.. ))

  Now y is monomorphic and really does use the dictionary passed to f. So it really cannot appear in the static. Somehow y really isn't static after all. We must reject this program. Not only is it an implementation mess (Complications 1,2,3 are already imposing quite a signficant implemenation burden) but it becomes pretty hard to explain to the programmer just which uses of static are OK and which are not.

  What a swamp. At this point I threw up my hands and wrote this summary

Proposal

To me the solution is clear: the rule should be

• in static e, all the free vars of e should be bound at top level

That is a nice simple rule; it is easy to explain and easy to implement. It is also what the user manual says!

In retrospect, by addressing Complication 2 I was trying too hard! (And this extra feature is entirely undocumented.) I thought that I could deal with Complication 2 using the same mechanism as the one that deals with MonoLocalBinds. But I was wrong.

Making this change could perhaps break some programs. They would all be easy to fix, by moving bindings for any free variables to top level. But note that the status quo is not stable: it has bugs e.g #24464, #26545. What we have is at attempt to be clever that is simply wrong.

—
View it on GitLab.
You're receiving this email because of your activity on gitlab.haskell.org. Unsubscribe from this thread · Manage all notifications · Help