Re: Safety of elisp-flymake-byte-compile (Was Re: [Emacs-diffs] scratch/allow-custom-load-paths)

[Stefan Monnier <monnier@iro.umontreal.ca>]

>>> 1. Create a elisp-flymake-maybe-enable function that checks the buffer
>>>    for top-level forms that _could_ make it unsafe for byte-compiling on
>>>    the fly.  This would include, but not limited to, eval-and-compile,
>>>    eval-when-compile, defmacro, cl-defmacro, any "unknown" top-level
>>>    form.  This will generate a lot of false positives (positive meaning
>>>    "unsafe") but perhaps it could be made to generate 0 false negatives
>>>    and still successfully vet a good number of elisp files.
>>
>> I was thinking that we can probably do it without a separate check:
> You mean do it directly in elisp-flymake--batch-compile-for-flymake?
> Yeah, that's definitely a good idea.

I'm saying "without another traversal of the code".
I.e. elisp-flymake--batch-compile-for-flymake would just call the
byte-compiler in a different way (either via a new entry-point, or by
let-binding some new variable) to cause it to be more careful (and not
worry about generating correct&efficient code).

>> In bytecomp.el, when working in "flymake" mode (a mode in which the
>> output bytecode is not actually needed) we'd treat eval-when/and-compile
>> just like `progn`, and we'd mark some other macros as "unsafe"
>> (in which case we'd treat the corresponding calls as if they expanded
>> to nil) and when we see a defmacro, we use `unsafep` to decide whether
>> that macro should be treated as unsafe
>
> Trying to parse this... You mean when we see a call to a macro, not when
> we see a `defmacro' top-level form, right?

Right.

>> (so we could still macro expand
>> locally defined macros as long as they're simple enough).
>
>> Also, in that mode, we'd likely skip byte-opt altogether as well as
>> compiler macros.
>
> OK. What's byte-opt BTW? Optimization?  

See lisp/emacs-lisp/byte-opt.el
It's two byte-optimizers: one applied on the Elisp code after
macro-expansion, and another applied on the byte-code.

>> The most obvious remaining holes there would be macros defined by
>> installed packages and whose expansion includes execution of some of its
>> arguments (cl-eval-when being the most obvious one, but there are many
>> more subtle ones).
>
> So basically, in your proposal, package authors would use sth. like
> (declare (flymake-safe t)) in their defmacros?

The macros defined in the flymake'd file wouldn't need any declaration:
they'd all be treated as suspicious and checked via `unsafep`.

For the macros defined in already-loaded packages, I'm not sure
what would be the better option between a whitelist (such as the `flymake-safe`
declaration above) or a blacklist or something in-between.

> But then we would have to prompt the user to accept or reject these
> marks right?

Since they're in the files we already trust I don't think that's needed.

> I don't think I completely understood your idea: what about, for
> example, eglot.el's macrology that checks LSP interface destructuring at
> compile-time?  There are some eval-when-compile's there right now:
>
>   (eval-and-compile
>     (defvar eglot--lsp-interface-alist
>       `(
>         (CodeAction (:title) (:kind :diagnostics :edit :command))
>         (Command (:title :command) (:arguments))
>         ...)
>         "Alist of (INTERFACE REQUIRED-KEYS OPTIONAL-KEYS)"))
>
> And then
>
>    (eglot--dbind ((Command) title not-really-a-key) some-lsp-object
>      (do-something-with title not-really-a-key))
>
> gives me a nice warning that not-really-a-key isn't in the "Command"
> interface.  How would that work in your new elisp-flymake-byte-compile?

It would only work once `eglot--dbind` is defined by a pakage in
`load-path` (and after that package defined eglot--lsp-interface-alist).

But not if you just open eglot.el.

> Also, what do you think of my option 2 to disable most of the system
> interface when flymaking?

Providing ways to run Elisp in a confined environment would be useful in
various circumstances, but it's non-trivial.


        Stefan