bug#59882: Multiple versions of Org in load-path problem

[Max Nikulin <manikulin@gmail.com>]

On 18/12/2022 22:40, Stefan Monnier wrote:
> Could Someone™ investigate why that code
> doesn't do its job.
> 
> The code in question is in `package--reload-previously-loaded` and
> my understanding is that it should cause `org-macs` to be reloaded

I do not see obvious flaws in `package--reload-previously-loaded' code. 
My guess is that we get reports from users who have installed released 
versions of emacs.

Notice that my complain concerning failure `byte-recompile-directory', 
hidden from its caller, is still valid.

On 25/12/2022 16:48, Ihor Radchenko wrote:
> Max Nikulin writes:
> 
>> I have tried the following in Emacs-27. For Emacs >= 29 add -L to Org <=
>> 9.5.

Stefan convinced me that I was too optimistic concerning Emacs-29.

>> - "emacs -l org" to simulate a regular user session with some opened files.
>> - Update to Org-9.6 from ELPA.
>> - Close emacs and start it again "emacs -l org"
>>
>> Result:
>> byte-code: Invalid function: org-assert-version
> 
> I tried to follow these steps, but unfortunately I am unable to
> reproduce. Everything works fine using Emacs 27 on my side. Strange.

I tried it once more in a minimal LXC container with Ubuntu-22.04 LTS 
jammy. This time I even removed the elpa-org-9.5.2 system package, so 
built-in version of org is really used. So, emacs-27

- emacs -l org
- M-x list-packages RET
- / n org RET
- move cursor to org, install it
- i x y
- compilation buffers pops up with warning and errors.

Due to "-l org" some libraries like org, org-macs, ob are loaded. Files 
that `require' already loaded libraries generate warnings

Compiling file /home/emcs/.emacs.d/elpa/org-9.6/ob-R.el at Sun Dec 18 
04:30:49 2022

In end of data:
ob-R.el:562:1:Warning: the function ‘org-assert-version’ is not known to be
     defined.

because `org-assert-version' is defined in org-macs.el, but the old 
version without it is already loaded. Call to `org-assert-version' is 
just added to the .elc file.

E.g. org-keys.el produce the same warning. Another case is

Compiling file /home/emcs/.emacs.d/elpa/org-9.6/ob-matlab.el at Sun Dec 
18 04:30:50 2022
ob-matlab.el:43:1:Error: Symbol’s function definition is void: 
org-assert-version

This file contains "(require 'ob-octave)". Attempt of load new version 
of ob-octave fails because during execution of it contents causes an 
error due to the `org-assert-version' call.

In particular, compilation of org.el fails because the new version has 
new "(require 'org-fold)" and the latter also tries to call undefined 
`org-assert-version'.

So at this point we have partially compiled broken install. Some files 
are not compiled, others have call to `org-assert-version' instead of 
its expansion.

Let's try to start emacs again and load org. Since org.el is not 
compiled, "(org-assert-version)" inside it is unable to catch any 
problem due to mixed-version compilation. Now `org-release' returns new 
version. The failure happens during "(require 'org-keys)" that was 
compiled with the warning.

Debugger entered--Lisp error: (invalid-function org-assert-version)
   org-assert-version()
   byte-code("\300\301!\210\302 \210\300\303!\210\300\304!\207" [require 
org-macs org-assert-version cl-lib oc] 2)
   require(org-keys)

So we have an obscure load time error.  I would say that compile *error* 
messages are obscure as well because it is unclear that errors happen in 
`require'. E.g. when compiling org-matlab.el the reported line number is 
for "(require 'ob-maxima)" while compiling of ob.maxima.el finished with 
a warning, not an error and it is really confusing. "ob-octave:34:" in 
addition to "ob-matlab:43:" during compiling of ob-matlab.el would be 
more informative.