Configuration layers development
Table of ContentsClose
1. Introduction
This document is intended as a tutorial for users who are interested in writing their first configuration layer, whether for private use or for contributing upstream. It should help clear up some confusion regarding how layers work and how Spacemacs (and Emacs) loads packages. For an overview of configuration layers with descriptions see Spacemacs layers list.
2. Nomenclature
Layers and packages. What gives?
- A set of Emacs Lisp files that, taken together, provide some feature. Packages may be available on a package repository, such as ELPA or MELPA or on a third-party service provider (such as github) or even locally on the disk.
- A collected unit of configuration that can be enabled (or disabled) in Spacemacs. A layer typically brings together one or more packages, as well as the glue configuration code required to make them play well with each other and Spacemacs in general.
Before writing a layer, it is helpful to consider what you are trying to achieve. Is there a package that provides the functionality you are after, and you want to integrate it in Spacemacs? If yes, you should write a layer. Are you trying to implement a new feature that would be useful for the Emacs community at large? In that case, consider whether it wouldn't be more appropriate to write a package first, and then a layer that uses your package.
3. The Emacs loading process
To understand how to best implement a layer, we have to investigate how Emacs loads code.
3.1. Emacs Lisp files
Emacs Lisp files contain code that can be evaluated. When evaluated, the functions, macros and modes defined in that file become available to the current Emacs session. Henceforth, this will be termed as loading a file.
One major problem is to ensure that all the correct files are loaded, and in the proper order. Another issue is to ensure that not too many files are loaded immediately. This causes startup to take too long. Instead, we want to make sure that files are loaded only as needed, and not all at once.
How is this done in Emacs, and how is it done in Spacemacs?
3.1.1. Loading a file
The simplest way to load a file is to call load-file
.
(load-file "~/elisp/foo.el")
This is as primitive as it comes. The path must be exact, and it does not have
to be in the Emacs load path (we'll get to that later). It will not look for a
byte-compiled .elc
file. It will simply load exactly what you tell it to.
3.2. Features
A better way to load what you need is to use features. A feature is a symbol
that typically has the same name as the file it resides in. Let us say you have
the following contents in a file called my-feature.el
.
;; Your code goes here ... (provide 'my-feature)
To have Emacs load this file, call require
, as such:
(require 'my-feature)
This checks whether the feature my-feature
has already been loaded. If not, it
looks for a file called my-feature.el
, my-feature.elc
or some such. If it
finds such a file, it will load it. When the call to provide
is evaluated, the
feature is added to the list of loaded features, so that subsequent calls to
require
will do nothing.
This will cause an error if no such file can be found.
The file my-feature.el
may very well contain other calls to require
, and in
fact this is quite a common way to ensure that dependencies are loaded before
your code runs.
Package authors should use this technique to make sure that dependencies are loaded before their code runs.
3.2.1. The load path
When loaded using require
, Emacs looks for files in its load path. This is
nothing more than a list of paths where elisp files can be found, and you can
inspect it through SPC h d v load-path
in Spacemacs. To add to the load path,
simply add to this list, e.g.
(add-to-list 'load-path "/some/path/")
3.3. Auto-loading
Calling require
is nothing more than a glorified way of calling load-file
.
It solves the problem of ensuring that files are loaded in the correct order,
and to some degree it solved the problem of where to find the files on disk but
a long list of calls to require
at startup would still cause Emacs to take
forever to load.
Emacs uses auto-loading to solve this problem. When a function is registered as auto-loading, an "empty" definition is provided. When that function is called, the file that provides the function is immediately loaded (along with all its required features). Finally, the "empty" function is substituted with the real one and called normally. The end user will see only a slight delay when first calling the function, while subsequent calls to that function (or any other function loaded as part of the same procedure) will be as quick as normal.
To register a function as auto-loadable, we call autoload
:
(autoload 'some-function "some-file")
This instructs Emacs that whenever some-function
is called, load
some-file.el
first, and then proceed.
After evaluating the above code, you can try to inspect some-function
by doing
SPC h d f some-function
. It will say it's an auto-loaded function, and that
nothing else is known about it until it is loaded. The call to autoload
can
optionally include more information, such as a doc-string, whether the function
can be called interactively, and so on. This provides more information to the
end-user without her having to actually load the file first.
Open your elpa
directory, go to helm
and look at the file
helm-autoloads.el
. This provides all the auto-loads for all the files in Helm.
However, this file is not written by hand. Instead, it is automatically
generated from "magic" comments in the source code of Helm. They look like this:
;;;###autoload (defun my-function () ;; Source code... )
The magic comment ;;;###autoload
instructs Emacs that the following definition
should be auto-loaded. This automatically generates an appropriate call to
autoload
.
Things that can be auto-loaded generally involve anything "definable", such as functions, macros, major or minor modes, groups, classes, and so on.
Magic comments also work on other things, such as variable definitions
(defvar
), but in that case, the definition is just copied verbatim into the
auto-loading file. For example, this code will load Helm on startup, long before
your file is actually evaluated, probably not what was intended:
;;;###autoload (require 'helm)
It is the responsibility of the package authors to ensure that their package can be appropriately auto-loaded, and most packages do this quite well.
Spacemacs makes thorough use of auto-loading. Almost everything in Spacemacs is loaded when needed instead of right away.
3.4. Eval after load
Often, we will want to configure packages after loading them. We may want to set
some variables or call some functions. This is trivial with require
, because
it loads immediately, but it can be tricky with autoloading, because the
configuration code must also be deferred.
Emacs offers with-eval-after-load
for this purpose. It can be used like this:
(with-eval-after-load 'helm ;; Code )
This arranges for the relevant code to be executed after Helm is loaded (using
either require
or an autoload), or if Helm is already loaded, the code is
executed immediately.
Since with-eval-after-load
is a macro and not a function, its argument does
not have to be quoted.
3.5. Use-package
For end users who are trying to put together an efficient Emacs configuration,
there is a very useful package called use-package
that provides a macro
which is also called use-package
which does a very good job of streamlining
the whole process of loading packages.
The aspiring layer author is recommended to have a look at the use-package
documentation. Some examples follow.
(use-package helm)
This simply loads Helm. It is essentially equivalent to (require 'helm)
.
(use-package helm :defer t)
This defers the loading of Helm using the auto-load facility and the auto-load commands provided by the Helm source code. It is, in fact, a no-op.
(use-package helm :defer t :init ;; Code to execute before Helm is loaded :config ;; Code to execute after Helm is loaded )
This form includes code to execute before and after Helm is loaded. The :init
section can be executed immediately, but since Helm is deferred, the :config
section is not executed until after loading, if ever. It is essentially
equivalent to simply running the :init
block, and then adding the :config
block in an with-eval-after-load
.
(use-package helm :commands (helm-find-files helm-M-x))
This creates auto-load references for additional commands, if you find that the package author has been slacking.
(use-package ruby-mode :mode "\\.rb\\'")
For packages that provide major modes, you can associate file extensions to that
mode by using the :mode
keyword. This adds an entry to auto-mode-alist
and
an auto-load for ruby-mode
. Typically this is not required, as ruby-mode
should already be auto-loadable, and the package should associate Ruby files
with itself already.
Use-package supports heaps of useful keywords. Look at the documentation for more.
4. Anatomy of a layer
A layer is simply a folder somewhere in Spacemacs's layer search path that usually contains these files (listed in loading order).
- declare additional layers
- the packages list and configuration
- all functions used in the layer should be declared here
- layer specific configuration
- general key bindings
Additionally, for each local package (see the next section), there should be a
folder <layer>/local/<package>/
containing the source code for that package.
Before initializing that package, Spacemacs will add this folder to the load
path for you.
4.1. layers.el
This file is the first file to be loaded and this is the place where additional layers can be declared as dependencies.
For instance, if layer A depends on some functionality of layer B, then in the
file layers.el
of layer A, we can add:
(configuration-layer/declare-layer-dependencies '(B))
The effect is that B is considered a used layer and will be loaded as if it
was added to dotspacemacs-configuration-layers
variables.
4.2. packages.el
It contains the list of packages of the layer and the actual configuration for the packages included in the layer.
This file is loaded after layers.el
.
It must define a variable called <layer>-packages
, which should be a list of
all the packages that this layer needs. Some valid package specifications are
as follows:
(defconst mylayer-packages '( ;; Get the package from MELPA, ELPA, etc. some-package (some-package :location elpa) ;; A local package (some-package :location local) ;; A local package to be built with Quelpa (some-package :location (recipe :fetcher local)) ;; A package recipe (some-package :location (recipe :fetcher github :repo "some/repo")) ;; An excluded package (some-package :excluded t) ))
The :location
attribute specifies where the package may be found. Spacemacs
currently supports packages on ELPA compliant repositories, local packages,
remote packages hosted on Git repositories (including specific helpers for
GitHub, GitLab, and Bitbucket) and MELPA recipes (through the Quelpa package).
Local packages should reside at <layer>/local/<package>/
.
For information about recipes see the MELPA documentation.
As you may have noticed from examples above, there are two ways to declare a
local package: using either :location local
or a Quelpa recipe with the
Spacemacs-specific pseudo-fetcher local
. The former is for the simplest
packages that declare no external dependencies, since it just adds the package
directory to the load-path
. The latter is for packages that do have external
dependencies declared and thus have to be built with Quelpa.
Packages may be excluded by setting the :excluded
property to true. This
will prevent the package from being installed even if it is used by another
layer.
For each included package, you may define one or more of the following functions, which are called in order by Spacemacs to initialize the package.
<layer>/pre-init-<package>
<layer>/init-<package>
<layer>/post-init-<package>
It is the responsibility of these functions to load and configure the package in question. Spacemacs will do nothing other than download the package and place it in the load path for you.
Note: A package will not be installed unless at least one layer defines an
init
function for it. That is to say, in a certain sense, the init
function
does mandatory setup while the pre-init
and post-init
functions do optional
setup. This can be used for managing cross-layer dependencies, which we will
discuss later.
4.3. funcs.el
It contains all the defined functions used in the layer.
This file is loaded after packages.el
and before config.el
.
It is good practice to guard the definition of functions to make sure a package is actually used. For instance:
(when (configuration-layer/package-used-p 'my-package) (defun spacemacs/my-package-enable () ...) (defun spacemacs/my-package-disable () ...))
By guarding these functions we avoid defining them when the package my-package
is not used.
4.4. config.el
This file configures the layer by declaring layer variables' default values and setting up some other variables related to the layer.
This file is loaded after funcs.el
.
4.5. keybindings.el
It contains general key bindings.
This is the last file loaded.
The word general here means independent of any package. Since the end user can exclude an arbitrary set of packages, you cannot be sure that, just because your layer includes a package, that package will necessarily be loaded. For this reason, code in these files must be generally safe, regardless of which packages are installed.
More on this in the next section.
5. The Spacemacs loading process
The Spacemacs loading process can be summarized as follows:
- Spacemacs goes through all the enabled layers and evaluates their files.
First
layers.el
is loaded to declare layer dependencies. Thenpackages.el
andfuncs.el
are loaded, but nothing happens from them since these files only define functions and variables, then the changes introduced byconfig.el
are applied. Spacemacs checks which packages should be downloaded and installed. To be installed, a package must be
- included by a layer that the user has enabled,
- not be excluded by any other layer that the user has enabled,
- not be excluded by the user herself, and
- there must be at least one
<layer>/init-<package>
function defined for it.
Alternatively, if a package is part of the end user's
dotspacemacs-additional-packages
, it will also be installed.- All packages which should be installed are installed in alphabetical order,
package.el
built-in Emacs library is in charge of implicit dependencies. Installed packages not following the rules of 2. are removed as well as their dependencies if possible. (This last behavior is optional but default.) - The
pre-init
,init
andpost-init
functions for each installed package are executed in turn.
It is step four that interests us. It is very important that a package is not
installed if no init
function is defined for it.
We say that a layer owns a package if it defines an init
function for it. A
layer does not own a package if it only defines pre-init
or post-init
functions.
Only one layer may own a package. Since layers are processed in order of specification in the user's dotfile, it is possible for layers to "seize" ownership of a package that was owned by a previously enabled layer.
6. Case study: auto-completion
Spacemacs provides a layer called auto-completion
which provides
auto-completion features in many modes. It does this using the package
company
. This layer owns the company
package, so it defines a function
called auto-completion/init-company
.
When a user enables the auto-completion
layer, Spacemacs locates it and finds
company
in the list of packages. Provided that company
is not excluded,
either by the user or another layer, Spacemacs then locates and runs the init
function for company
. This function includes a call to use-package
that sets
up the basic configuration.
However, auto-completion is a two-horse game. By its very nature, it is specific
to the major mode in question. It is pointless to expect the auto-completion
layer to include configuration for each conceivable major mode, and equally
futile to expect each programming language layer (python, ruby, etc.) to fully
configure company
on their own.
This is solved using the post-init
functions. The Python layer, for example,
includes the company
package and defines a function called
python/post-init-company
. This function is called after
auto-completion/init-company
, but it is not called if
- the
auto-completion
layer is not enabled, in which case noinit
function forcompany
will be found, or - the
company
package is excluded either by the user or another layer
As such, python/post-init-company
is the only safe place to put
configuration related to company
in Python mode.
If the Python layer had defined an init
function for company
, that package
would have been installed even if the auto-completion
layer had been disabled,
which is not what we want.
7. Layer tips and tricks
7.1. Cross-dependencies
Spacemacs provides a couple of additional useful functions you can use to check whether other layers or packages are included.
- check if a layer is enabled (
configuration-layer/layer-used-p
) - check if a package is or will be installed (
configuration-layer/package-used-p
)
These are useful in some cases, but usually you can get the desired result just
by using post-init
functions.
For layers that require another layers to be enabled, use the functions
configuration-layer/declare-layer
and configuration-layer/declare-layers
to
ensure that layers are enabled even if the user has not enabled them explicitly.
Calls to these functions must go in the layers.el
file.
7.2. Shadowing
Shadowing is the operation of replacing a used layer by another one. For instance if a used layer A can shadow a used layer B and the layer A is listed after the layer B in the dotfile then the layer A replaces the layer B and it is like only the layer A is being used.
Examples of this mechanism are helm/ivy layers or neotree/treemacs layers.
A layer can shadow other layers by calling in its layers.el
file the function
configuration-layer/declare-shadow-relation
. This function declares a
can-shadow
relation between all the layers.
can-shadow
is a commutative relation, if layer A can shadow layer B then layer
B can shadow layer A.
The shadow
operator is a binary operator accepting two layer names, it is not
commutative and the order of the operands is determined by the order of the
layers in the dotfile (like the ownership stealing mechanism).
If :can-shadow
property is set explicitly to nil
in the dotfile then the
layer won't shadow any layer.
For instance to install both ivy and helm layer:
(setq dotspacemacs-configuration-layers '( ivy (helm :can-shadow nil) )
note that due to the commutative relation can-shadow
the above example can
also be written like this (in this case, :can-shadow
should be read
:can-be-shawdowed
):
(setq dotspacemacs-configuration-layers '( (ivy :can-shadow nil) helm )
We will prefer the first form as it is more intuitive.
7.3. Use-package init and config
In the vast majority of cases, a package init
function should do nothing but
call to use-package
. Again, in the vast majority of cases, all the
configuration you need to do should be doable within the :init
or :config
blocks of such a call.
What goes where? Since :init
is executed before load and :config
after,
these rules of thumb apply.
In :config
should be
- Anything that requires the package to be already loaded.
- Anything that takes a long time to run, which would ruin startup performance.
The :init
block should contain setup for the entry points to the package. This
includes key bindings, if the package should be loaded manually by the user, or
hooks, if the package should be loaded upon some event. It is not unusual to
have both!
7.4. Use-package hooks
Spacemacs includes a macro for adding more code to the :init
or :config
blocks of a call to use-package
, after the fact. This is useful for pre-init
or post-init
functions to "inject" code into the use-package
call of the
init
function.
(spacemacs|use-package-add-hook helm :pre-init ;; Code :post-init ;; Code :pre-config ;; Code :post-config ;; Code )
Since a call to use-package
may evaluate the :init
block immediately, any
function that wants to inject code into this block must run before
the call to
use-package
. Further, since this call to use-package
typically takes place
in the init-<package>
function, calls to spacemacs|use-package-add-hook
always happen in the pre-init-<package>
functions, and not in
post-init-<package>
.
7.5. Best practices
If you break any of these rules, you should know what you are doing and have a good reason for doing it.
7.5.1. Package ownership
Each package should be owned by one layer only. The layer that owns the
package should define its init
function. Other layers should rely on
pre-init
or post-init
functions.
7.5.2. Localize your configuration
Each function can only assume the existence of one package. With some
exceptions, the pre-init
, init
and post-init
functions can only
configure exactly the package they are defined for. Since the user can exclude
an arbitrary set of packages, there is no a priori safe way to assume that
another package is included. Use configuration-layer/package-usedp
if you
must.
This can be very challenging, so please take this as a guideline and not something that is absolute. It is quite possible for the user to break her Spacemacs installation by excluding the wrong packages, and it is not our intention to prevent this at all costs.
7.5.3. Load ordering
In Spacemacs, layers are loaded in order of inclusion in the dotfile, and packages are loaded in alphabetical order. In the rare cases where you make use of this property, you should make sure to document it well. Many will assume that layers can be included in arbitrary order (which is true in most cases), and that packages can be renamed without problems (which is also in most cases).
Preferably, write your layer so that it is independent of load ordering. The
pre
- and post-init
functions are helpful, together with
configuration-layer/package-usedp
.
7.5.4. No require
Do not use require. If you find yourself using require
, you are almost
certainly doing something wrong. Packages in Spacemacs should be loaded through
auto-loading, and not explicitly by you. Calls to require
in package init
functions will cause a package to be loaded upon startup. Code in an :init
block of use-package
should not cause anything to be loaded, either. If you
need a require
in a :config
block, that is a sign that some other package is
missing appropriate auto-loads.
7.5.5. Auto-load everything
Defer everything. You should have a very good reason not to defer the loading of a package.