bug#19390: 25.0.50; `package-activate' is too slow

[Artur Malabarba]

2014-12-16 12:31 GMT-02:00 Dmitry Gutov > On 12/16/2014, Artur Malabarba wrote:
>
>> I didn't notice anything as aggravating as the report, but I agree my
>> latest patch is a performance concern.
>
> Do you maybe have fewer packages installed?

Perhaps.
But the reason is beyond the point, :-) I agree this patch needs
performance improvements.

> Any particular reason you're using file-truename', instead of
>expand-file-name'? Replacing the former with the latter already gives
> ~twofold performance improvement.

Yes, I originally used `expand-file-name', but
`package--list-loaded-files' would miss some files if there was a
symlink in the path. (I specifically observed this, it's not just
theory).
The problem is that a file returned by `find-libary' could be the same
as another file contained in  `load-history' but have a different name
(because of symlinks).


>> 2. Creating the `history' variable as below is a problem.
>>
>> (mapcar (lambda (x) (file-name-sans-extension
>>                                    (file-truename (car x))))
>>                      load-history)
>
>
> With the above change, this piece of code is relatively fast.
>
> For instance, (package--list-loaded-files default-directory) in the Helm
> directory takes 140 ms here, whereas
>
>             (mapcar (lambda (x) (file-name-sans-extension
>                                  (expand-file-name (car x))))
>              load-history)
>
> only takes ~10 ms. Still not fast enough not to need the other patch, of
> course.

Yes, I agree `file-truename' is slow. The foolproof way would be to
have a variable that stores the `file-truename's, so we wouldn't have
to calculate them.
Using `expand-file-name' is a valid way to improve performance here,
but it won't *always* reload the files for all users.