bug#41646: Startup in Windows is very slow when load-path contains many

[Lin Sun <sunlin7.mail@gmail.com>]

On Mon, Oct 21, 2024 at 2:34 PM Stefan Monnier <monnier@iro.umontreal.ca> wrote:
>
> > The `load-hints` can reduce searching attempts by putting the matched
> > paths on the top of `load-path`, it won't break the original
> > load-path; and the patch for the `package.el` will put the installed
> > files into the `load-hints`.
>
> The downside is that it can break existing setups for users who use
> `package.el` but also modify their `load-path` "by hand" in the
> init file, and it doesn't help users who don't use `package.el`.

The `load-hints' do nothing with its default value nil.

The package manager can use the `load-hints' or ignore it. The
"package.el" is emacs builtin package manager, can support or ignore
the `load-hints' by setting `package-enable-load-hints` to t or nil.
Other package managers can continue without any change, or do some
work to get performance benefits by supporting the `load-hints'.

> Note also that your `load-hints` could grow large, so scanning it could
> take a significant amount of time.  Maybe it would make sense to turn it
> into a hash table for those entries that don't use the "*" special
> thingy (and maybe use a radix-tree for those entries using the "*"
> special thingy)?

The `load-hints` could grow as the `load-path', or larger than the
`load-path', but it should not be too much.

For the hash table, it can not support the '*', or the hash-table
requires every file to have an entry explicitly, which costs a lot on
building the table.

I had checked the radix-tree at the beginning, it's not user-friendly,
or it's not easy to dump the radix tree for an end user to understand
which is obviously matching the entry or not.

The `load-hints' in the list are easy to understand / maintain by the end user.

> But your prefix idea makes me think maybe we can aim for a significantly
> smaller table, where we basically record only one entry per
> package/directory, like for "~/.emacs.d/elpa/helm-core-VERSION/" we just
> record "helm" because all the `.el` files share the "helm" prefix.
> I.e. keep for each dir the corresponding longest-common-prefix.
> If we're careful to consider only those files with a `.el` suffix, then
> I think we can reduce the hint to such a longest-common-prefix.  I.e. an
> info which doesn't say just "you can find FOO* files here" but "you can
> find *only* FOO* files here".

I had searched all 200+ packages in my test env, most of the packages
use their feature name as the prefix, only 11 packages have
exceptions. But I didn't understand how it works toward the
`load-hints'.

> Then we should be able to create quickly (so it can be recomputed on the
> fly whenever `load-path` changes) a radix-tree that maps a relative
> file name to the list of directories from `load-path` where it is
> worthwhile to look (by filtering out those dirs whose
> longest-common-prefix doesn't match).  We'd only do it for MUST_SUFFIX
> is specified, of course.
>
> > 1. On my local linux test env, disable load-hints on the package.el,
> > the test cli spends 6.327s; and enable load-hints then it spends
> > 5.392s.
> >
> > 2. On my local Windows test env, disable load-hints on the package.el,
> > the test cli spends 11.769s, and enable load-hints then it spends
> > 7.279s.
>
> Is that with or without using `package-quickstart`?

The `package-quickstart' does not help in this scenario, the key point
is the `load-path` count will times the read attempt.
A simple "(require 'X)" will lead the emacs walks through the
`load-path' to attempt opening the "X.so, X.so.gz, X.elc, X.elc.gz,
X.el, X.el.gz" one by one, if the `load-path' has 200 entries, emacs
will try search 200x6=1200 times for the worst case.

The `load-hints` will help put the matched path to the top of
`load-path` then emacs can find the X on the top entries of
`load-path' then returns shortly.

> BTW, in your patch, you change `locate-file-internal` which seems wrong,
> since that function is not specific to loading ELisp files, it's also
> used for $MANPATH, $PATH, and things like that.
>
> Similarly, I wasn't able to convince myself that your patch does the
> right thing when `require` or `load` is used such that MUST_SUFFIX is
> not specified.

I'm going to search the cases carefully. Thank you for all the
comments, appreciate it.