[bug#53878] [PATCH v3 09/15] gnu: Add racket-vm-cgc.

[Philip McGrath <philip@philipmcgrath.com>]

[-- Attachment #1: Type: text/plain, Size: 5202 bytes --]

Hi,

On Sunday, February 20, 2022 11:48:01 AM EST Liliana Marie Prikler wrote:
> Am Sonntag, dem 20.02.2022 um 09:09 -0500 schrieb Philip McGrath:
> > Hi,
> > 
> > On Sunday, February 20, 2022 4:03:26 AM EST Liliana Marie Prikler
> > 
> > wrote:
> > > So here's my suggestion:
> > > 
> > > Inside chez-and-racket-bootstrap, define (make-<package>) functions
> > > for
> > > the following:
> > > - chez-bootstrap-bootfiles, chez-for-racket-bootstrap-bootfiles:
> > >   Taking version and origin.
> > > - racket-vm-cgc: Taking version and origin.
> > > - racket-vm-bc: Taking racket-vm-cgc.
> > > - racket-vm-cs: Taking racket-vm-bc.
> > > 
> > > Inside chez, define chez-scheme, as well as non-bootstrapped
> > > versions of stex et al.  Also define make-chez-scheme-for-racket,
> > > taking version and origin as parameter.  Finally, define chez-
> > > scheme-for-system, which uses (resolve-interface '(gnu packages
> > > racket)) to get racket's version and
> > > origin.
> > > 
> > > Inside racket, define %racket-version, %racket-origin, racket-
> > > minimal
> > > and racket.  It'd also be good if you made local definitions
> > > (define racket-vm-cgc (make-racket-vm-cgc %racket-version %racket-
> > > origin))
> > > (define racket-vm-bc (make-racket-vm-bc racket-vm-cgc))
> > > ...
> > > in this file.
> > 
> > My understanding—which is not very good!—is that this would have the
> > same problem we do currently. It would be analogous to my example
> 
> > from <https://issues.guix.gnu.org/53878#93>:
> Well, for one this claim is both verifiable and falsifiable by way of
> implementation.  For another, I don't really see the issue however.
> Since those bindings would be local to racket.scm and not used anywhere
> else, I don't think there is a cycle to be found anywhere.
> 

While working on v4, I did try using `racket-vm-for-system` at what Ludo’ 
called "the top level of a module", and I encountered the same error.

> > > But  Ludo’'s examples show that's wrong: those uses of `chez
> > > scheme` are in what the "expansion contexts" model would call
> > > "expression contexts".
> > > 
> > > Instead, I think rule № 2 prohibits any reference to a variable
> > > imported from another (gnu packages ...) module that will be
> > > evaluated when the (gnu packages ...) modules are—visited?
> > > instantiated? [2][3]—IDK when exactly, but, for practical purposes,
> > > any variable reference that is not underneath a lambda abstraction.
> > > 
> > > If that's right, IIUC, it would mean that:
> > > 
> > >     (define chez-scheme-for-racket
> > >       (make-chez-scheme-for-racket ...))
> > > 
> > > would also be prohibited.
> > > 
> > > On the other hand, uses of `(racket-vm-for-system)` and `(chez-
> > > scheme-for-system)` in an `imports` field should still be fine,
> > > thanks to the implicit thunks.
> > 
> > The reference to `make-chez-scheme-for-racket` or `make-racket-vm-cs`
> > or any such procedure defined in "chez-and-racket-bootstrap.scm"
> > would be evaluated when "racket.scm" is instantiated—or whenever
> > precisely it is that causes the  problem.
> 
> I don't think that'd be a problem since make-chez-scheme-for-racket is
> itself a function.  If it still is, one pair of brackets makes it not
> so.  This is a well-explored technique of resolving chains, used for
> example in our build system code.  chez-and-racket-bootstrap.scm should
> imo not be imported anywhere but chez.scm and racket.scm, so it by
> itself can not form a cycle.  Only chez.scm and racket.scm can, but
> there are ways of making those well-formed.

AFAICT the error is not from applying the function bound to the variable 
`racket-vm-for-system` (or `make-racket-vm-bc`, or whatever), but from 
evaluating the reference to a variable imported from a sibling
`(gnu packages ...)` module, regardless of the value to which the variable is 
bound.

My current best guess at Guile's semantics for mutually-recursive modules is 
that it's something like the `letrec` restriction, e.g. the difference between 
these two examples (given an implementation that fully enforces `letrec` 
semantics, as opposed to `letrec*`):

--8<---------------cut here---------------start------------->8---
$ scheme 
Chez Scheme Version 9.5.6
Copyright 1984-2021 Cisco Systems, Inc.

> (letrec ((a (lambda () 1))
           (b a))
    2)
Exception: attempt to reference undefined variable a
Type (debug) to enter the debugger.
> (letrec ((a (lambda () 1))
           (b (lambda () (a))))
    (b))
1
--8<---------------cut here---------------end--------------->8---

> 
> See (standard-packages) in guix/build-system/gnu.scm pointing to (gnu
> packages commencement) or (default-python) in guix/build-
> system/python.scm for pointers.

Given Ludo’'s explanation, I think the difference (or at least an important 
difference) is that those functions are defined in `(guix ...)` modules, as 
opposed to `(gnu packages ...)` modules.

But I wish I knew with any degree of certainty *why* this would be true, if 
indeed it is.

Maybe Maxime knows?

-Philip




[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 833 bytes --]