Michael Albinus [2022-04-03 17:44:55] wrote:
> Stefan Monnier <monnier@iro.umontreal.ca> writes:
>> Note: this is not a compile error.  It's an error that happened during
>> expansion of a compiler-macro.  These are optional, so those errors only
>> affect the performance of the code but shouldn't affect its behavior.
>>
>> IOW this sound more serious than it is.
>>
>> `tramp-file-name-method--cmacro` is supposed to be a function which
>> inlines the code of calls to `tramp-file-name-method`.
>>
>> More specifically `tramp-file-name-method--cmacro` is auto-generated
>> (and put on the `compiler-macro` property of the
>> `tramp-file-name-method` symbol) by the `cl-defsubst` used to define
>> `tramp-file-name-method`.
>>
>> So I guess the source of the error is that unloading `tramp` undefines
>> the functions (both `tramp-file-name-method--cmacro` and
>> `tramp-file-name-method`) but leaves the `compiler-macro` property of the
>> `tramp-file-name-method` symbol.
>>
>> We really should make it so package unloading knows how to undo
>> top-level `put`s.
>
> Honestly, I have no knowledge about the byte-compiler. I've reopened the
> bug, so you can work on it.

The main point is that putting the `cl-defstruct` into the
`tramp-loaddefs.el` is really ugly and I think it's asking for trouble.
In contrast the OP's "errors" aren't actual errors, they're just
messages indicating that some optimization could not be applied.
We should really change the wording to clarify that they're warnings.

(see first hunk in patch below).

But in the mean time I'd much rather live with a harmless
"compiler-macro error" message than with a `cl-defstruct` in
a loaddefs file.

Lars Ingebrigtsen [2022-09-02 13:43:48] wrote:
> Stefan Monnier <monnier@iro.umontreal.ca> writes:
>> We really should make it so package unloading knows how to undo
>> top-level `put`s.
> Do you have any ideas for how to do that?

There are a few ways:
- We could let `unload-feature` remove compiler macros when present, as
  in the second hunk of the patch below.
- We could make `cl-defstruct` use `define-symbol-prop` rather than
  `put`, as in the third hunk of the patch below.


        Stefan