[racket-dev] Things we could move out of the core

On Fri, Jun 28, 2013 at 5:17 PM, Matthew Flatt <mflatt at cs.utah.edu> wrote:
> At Fri, 28 Jun 2013 17:08:19 -0400, Sam Tobin-Hochstadt wrote:
>> On Fri, Jun 28, 2013 at 5:02 PM, Matthew Flatt <mflatt at cs.utah.edu> wrote:
>> > At Fri, 28 Jun 2013 14:38:03 -0400, Sam Tobin-Hochstadt wrote:
>> >> On Thu, Jun 27, 2013 at 8:45 PM, Robby Findler
>> >> <robby at eecs.northwestern.edu> wrote:
>> >> > Did you consider moving "#lang mzscheme" out as well?
>> >>
>> >> I've now created another pull request that does this, here:
>> >> https://github.com/plt/racket/pull/377
>> >>
>> >> There's one remaining question.  The `make-base-namespace` procedure
>> >> provided by `mzscheme` attaches the `mzscheme` module.  But this pull
>> >> request removes that module, so it can't be attached or required in
>> >> this code.  The alternatives are:
>> >>
>> >> 1. Just attach/require `scheme/mzscheme`.  Slightly incompatible in
>> >> some corner cases.
>> >> 2. Don't remove `mzscheme` from the core.
>> >> 3. Remove `make*-namespace` from `scheme/mzscheme` and implement them
>> >> in the `mzscheme` collection in the `mzscheme` package.
>> >>
>> >> I'm currently leaning toward 3 but I'd appreciate anyone else's thoughts.
>> >
>> > Is there some reason that `scheme/mzscheme' can't move to the
>> > "mzscheme" package (along with `racket/private/stxmz-body')?
>>
>> Because large portions of the core are written in the `mzscheme`
>> language (or `scheme/mzscheme`, after my patch), some of which feature
>> evaluating code in mzscheme-like namespaces.  If we can somehow get
>> around the latter problem, then the former is a Small Matter of
>> Programming, but it'll take a little while.
>
> Ah --- I had not actually looked at 9587a2f.
>
> I guess I'm confused on the goal, since I don't see changing `mzscheme'
> to `scheme/mzscheme' as a step forward. Can you say more about the
> intent of changing `mzscheme' to `scheme/mzscheme'?

The intent is to reduce the API surface area provided by the core.
Note that `scheme/mzscheme` is not a documented API.  Thus, moving the
`mzscheme` language and collection would require anyone who depended
on them to explicitly depend on another package. At some later point,
we finish implementing the core without using `scheme/mzscheme`, move
the actual implementation of `mzscheme` to the `mzscheme` package,
without any compatibility problems.

> Meanwhile, I worry that options 1 and 3 can create subtle and confusing
> bugs/incompatibilities. (I've spent a lot of time on problems that
> happened due to accidentally choosing similar options in the past.)

(1) definitely has the potential for subtle bugs. I believe that (3)
is semantics preserving *except* insofar as some part of the core was
using the `mzscheme` namespace creation functions, and would thus work
differently. However, there are only two places in the core that use
`make-namespace` from `mzscheme`.  They are (a) the `setup/unpack`
code using `eval` that you plan to remove and (b) a part of
`compiler/private/xform.rkt` for loading precompiled headers. I don't
know what the latter does, or if changing it to use a Racket namespace
would have the potential to introduce bugs.

But yes, there's definitely a worry about subtle bugs here.

Sam