Doom saves a lot of startup time by avoiding package-initialize, because
it loads every packages' autoloads file, which is expensive.
Unfortunately, these autoloads files are necessary for some plugins with
abnormal file structures (such as ESS). Previously, I was content with
loading them myself, but these occurrences have become more frequent, so
it would be safest if I relied on package-initialize more.
So doom-initialize will now do so. However, it will cache the load-path
(and Info-directory-list, fixing the lost info buffers) on first run.
This makes byte-compiling Doom almost useless, but it makes startup just
as fast as Doom would be if you had.
This needs more testing.
Autoload paths used to be relative to arbitrary roots (usually
doom-modules-dir). This commit changes doom//reload-autoloads to iterate
over autoloads in doom-autoloads-file and replace their file paths with
an absolute one.
Theoretically this should make Emacs feel marginally faster, but we
*probably* get more of a benefit from a shorter load-path (because we no
longer need doom-modules-dirs, doom-core-dir or doom-psuedo-module-paths
in the load-path).
This is highly experimental however!
Any module can now use a doom! call to declare a module tree root.
This means that if you have a doom! block in
~/.emacs.d/modules/lang/org/init.el, then you can have submodules in
~/.emacs.d/modules/lang/org/modules/MODULE/SUBMODULE if you wanted to
for some reason.
This is only really truly useful for private modules. A doom! block in
~/.doom.d/init.el will recognize and enable modules in
~/.doom.d/modules/.
~/.doom.d/modules is now a full module tree, like ~/.emacs.d/modules.
Symlinks are no longer involved.
Private modules can now shadow Doom modules. e.g.
~/.doom.d/modules/lang/org will take precendence over
~/.emacs.d/modules/lang/org.
Also, made doom--*-load-path variables public (e.g. doom--site-load-path
=> doom-site-load-path), and rearranged the load-path for a 10-15%
startup boost.
Originally, I built the load-path with site-lisp paths first, then
packages. There was a modest ~10% startup boost doing this, because
there were considerably more site packages loaded at startup than
plugins.
However, this meant built-in packages would get precedence over plugins,
which is undesirable. In org's case, I simply modified the load-path
in lang/org/init.el. However, this issue has cropped up again in #340.
Evidently, that 10% boost may not be worth the risk it imposes, so I've
rearranged the load-path with packages first.
+ map-delete is shorter and faster than assq-delete-all
+ map-put is simpler than the delete-then-set workflow
+ map-merge is great for merging default and user settings
`get-file-buffer` returns nil if no buffer for that file already exists.
In an edge case on Windows, `update-file-autoloads` doesn't leave
a hanging buffer, possibly ignoring its SAVE-AFTER argument. Using
`find-file-noselect` fixes this, which will open a buffer if one doesn't
already exist.
Now that the org ELPA archive has https support, we can add it to
package-archives. This fixes some 'org is unavailable' errors when
installing org packages that have declared earlier versions of org as
a dependency.
This also makes installing a newer version of org-mode much simpler.
Woo!
This causes errors because of arbitrary module load order. This happens
in Emacs 25.x, because hash-table-keys there uses maphash, while in
Emacs 26+ it uses cl-loop.
Local cache and data files were stored in a parent folder named after
your host, because I'd symlink the same config across computers via
dropbox. This would cause byte-compile issues in packages, so I stopped
doing this.
Now that each computer gets its own clone, host-namespaced local
directories are unnecessary. There's no other benefit to having them.
