/use-package

A use-package declaration for simplifying your .emacs

Primary LanguageEmacs Lisp

use-package

The use-package macro allows you to isolate package configuration in your .emacs file in a way that is both performance-oriented and, well, tidy. I created it because I have over 80 packages that I use in Emacs, and things were getting difficult to manage. Yet with this utility my total load time is around 2 seconds, with no loss of functionality!

Notes for users upgrading to 2.x are located at the bottom.

The basics

Here is the simplest use-package declaration:

(use-package foo)

This loads in the package foo, but only if foo is available on your system. If not, a warning is logged to the *Messages* buffer. If it succeeds, a message about "Loading foo" is logged, along with the time it took to load, if it took over 0.1s.

Use the :init keyword to execute code before a package is loaded. It accepts one or more forms, up until the next keyword:

(use-package foo
  :init
  (setq foo-variable t))

Similarly, :config can be used to execute code after a package is loaded. In cases where loading is done lazily (see more about autoloading below), this execution is deferred until after the autoload occurs:

(use-package foo
  :init
  (setq foo-variable t)
  :config
  (foo-mode 1))

As you might expect, you can use :init and :config together:

(use-package color-moccur
  :commands (isearch-moccur isearch-all)
  :bind ("M-s O" . moccur)
  :init
  (bind-key "M-o" 'isearch-moccur isearch-mode-map)
  (bind-key "M-O" 'isearch-moccur-all isearch-mode-map)
  :config
  (use-package moccur-edit))

In this case, I want to autoload the commands isearch-moccur and isearch-all from color-moccur.el, and bind keys both at the global level and within the isearch-mode-map (see next section). When the package is actually loaded (by using one of these commands), moccur-edit is also loaded, to allow editing of the moccur buffer.

Key-binding

Another common thing to do when loading a module is to bind a key to primary commands within that module:

(use-package ace-jump-mode
  :bind ("C-." . ace-jump-mode))

This does two things: first, it creates an autoload for the ace-jump-mode command and defers loading of ace-jump-mode until you actually use it. Second, it binds the key C-. to that command. After loading, you can use M-x describe-personal-keybindings to see all such keybindings you've set throughout your .emacs file.

A more literal way to do the exact same thing is:

(use-package ace-jump-mode
  :commands ace-jump-mode
  :init
  (bind-key "C-." 'ace-jump-mode))

When you use the :commands keyword, it creates autoloads for those commands and defers loading of the module until they are used. Since the :init form is always run -- even if ace-jump-mode might not be on your system -- remember to restrict :init code to only what would succeed either way.

The :bind keyword takes either a cons or a list of conses:

(use-package hi-lock
  :bind (("M-o l" . highlight-lines-matching-regexp)
         ("M-o r" . highlight-regexp)
         ("M-o w" . highlight-phrase)))

The :commands keyword likewise takes either a symbol or a list of symbols.

Modes and interpreters

Similar to :bind, you can use :mode and :interpreter to establish a deferred binding within the auto-mode-alist and interpreter-mode-alist variables. The specifier to either keyword can be a cons cell, a list, or just a string:

(use-package ruby-mode
  :mode "\\.rb\\'"
  :interpreter "ruby")

;; The package is "python" but the mode is "python-mode":
(use-package python
  :mode ("\\.py\\'" . python-mode)
  :interpreter ("python" . python-mode))

If you aren't using :commands, :bind, :bind*, :bind-keymap, :bind-keymap*, :mode, or :interpreter (all of which imply :defer; see the docstring for use-package for a brief description of each), you can still defer loading with the :defer keyword:

(use-package ace-jump-mode
  :defer t
  :init
  (autoload 'ace-jump-mode "ace-jump-mode" nil t)
  (bind-key "C-." 'ace-jump-mode))

This does exactly the same thing as the other two commands above.

Notes about lazy loading

In almost all cases you don't need to manually specify :defer t. This is implied whenever :bind or :mode or :interpreter is used. Typically, you only need to specify :defer if you know for a fact that some other package will do something to cause your package to load at the appropriate time, and thus you would like to defer loading even though use-package isn't creating any autoloads for you.

You can override package deferral with the :demand keyword. Thus, even if you use :bind, using :demand will force loading to occur immediately and not establish an autoload for the bound key.

Information about package loads

When a package is loaded, and if you have use-package-verbose set to t, or if the package takes longer than 0.1s to load, you will see a message to indicate this loading activity in the *Messages* buffer. The same will happen for configuration, or :config blocks that take longer than 0.1s to execute. In general, you should keep :init forms as simple and quick as possible, and put as much as you can get away with into the :config block. This way, deferred loading can help your Emacs to start as quickly as possible.

Additionally, if an error occurs while initializing or configuring a package, this will not stop your Emacs from loading. Rather, the error will be captured by use-package, and reported to a special *Warnings* popup buffer, so that you can debug the situation in an otherwise functional Emacs.

Conditional loading

You can use the :if keyword to predicate the loading and initialization of modules. For example, I only want edit-server running for my main, graphical Emacs, not for other Emacsen I may start at the command line:

(use-package edit-server
  :if window-system
  :init
  (add-hook 'after-init-hook 'server-start t)
  (add-hook 'after-init-hook 'edit-server-start t))

The :disabled keyword can turn off a module you're having difficulties with, or stop loading something you're not using at the present time:

(use-package ess-site
  :disabled t
  :commands R)

When byte-compiling your .emacs file, disabled declarations are ommitted from the output entirely, to accelerate startup times.

Byte-compiling your .emacs

Another feature of use-package is that it always loads every file that it can when .emacs is being byte-compiled. This helps to silence spurious warnings about unknown variables and functions.

However, there are times when this is just not enough. For those times, use the :defines and :functions keywords to introduce dummy variable and function declarations solely for the sake of the byte-compiler:

(use-package texinfo
  :defines texinfo-section-list
  :commands texinfo-mode
  :init
  (add-to-list 'auto-mode-alist '("\\.texi$" . texinfo-mode)))

If you need to silence a missing function warning, you can use :functions:

(use-package ruby-mode
  :mode "\\.rb\\'"
  :interpreter "ruby"
  :functions inf-ruby-keys
  :config
  (defun my-ruby-mode-hook ()
    (require 'inf-ruby)
    (inf-ruby-keys))

  (add-hook 'ruby-mode-hook 'my-ruby-mode-hook))

Prevent a package from loading at compile-time

Normally, use-package will load each package at compile time before compiling the configuration, to ensure that any necessary symbols are in scope to satisfy the byte-compiler. At times this can cause problems, since a package may have special loading requirements, and all that you want to use use-package for is to add a configuration to the eval-after-load hook. In such cases, use the :no-require keyword, which implies :defer:

(use-package foo
  :no-require t
  :config
  (message "This is evaluated when `foo' is loaded"))

Extending the load-path

If your package needs a directory added to the load-path in order to load, use :load-path. This takes a symbol, a function, a string or a list of strings. If the path is relative, it is expanded within user-emacs-directory:

(use-package ess-site
  :load-path "site-lisp/ess/lisp/"
  :commands R)

Note that when using a symbol or a function to provide a dynamically generated list of paths, you must inform the byte-compiler of this definition so the value is available at byte-compilation time. This is done by using the special form eval-and-compile (as opposed to eval-when-compile). Further, this value is fixed at whatever was determined during compilation, to avoid looking up the same information again on each startup:

(eval-and-compile
  (defun ess-site-load-path ()
    (shell-command "find ~ -path ess/lisp")))

(use-package ess-site
  :load-path (lambda () (list (ess-site-load-path)))
  :commands R)

Diminishing minor modes

use-package also provides built-in support for the diminish utility -- if you have that installed. Its purpose is to remove strings from your mode-line that provide no useful information. It is invoked with the :diminish keyword, which is passed either a minor mode symbol, a cons of the symbol and its replacement string, or just a replacement string, in which case the minor mode symbol is guessed to be the package name with "-mode" appended at the end:

(use-package abbrev
  :diminish abbrev-mode
  :config
  (if (file-exists-p abbrev-file-name)
      (quietly-read-abbrev-file)))

For package.el users

You can use use-package to load packages from ELPA with package.el. This is particularly useful if you share your .emacs among several machines; the relevant packages are downloaded automatically once declared in your .emacs. The :ensure keyword causes the package(s) to be installed automatically if not already present on your system (set (setq use-package-always-ensure t) if you wish this behavior to be global for all packages):

(use-package magit
  :ensure t)

If you need to install a different package from the one named by use-package, you can specify it like this:

(use-package tex-site
  :ensure auctex)

Lastly, when running on Emacs 24.4 or later, use-package can pin a package to a specific archive, allowing you to mix and match packages from different archives. The primary use-case for this is preferring packages from the melpa-stable and gnu archives, but using specific packages from melpa when you need to track newer versions than what is available in the stable archives is also a valid use-case.

By default package.el prefers melpa over melpa-stable due to the versioning (> evil-20141208.623 evil-1.0.9), so even if you are tracking only a single package from melpa, you will need to tag all the non-melpa packages with the appropriate archive. If this really annoys you, then you can set use-package-always-pin to set a default.

If you want to manually keep a package updated and ignore upstream updates, you can pin it to manual, which as long as there is no repository by that name, will Just Work(tm).

use-package throws an error if you try to pin a package to an archive that has not been configured using package-archives (apart from the magic manual archive mentioned above):

Archive 'foo' requested for package 'bar' is not available.

Example:

(use-package company
  :ensure t
  :pin melpa-stable)

(use-package evil
  :ensure t)
  ;; no :pin needed, as package.el will choose the version in melpa

(use-package adaptive-wrap
  :ensure t
  ;; as this package is available only in the gnu archive, this is
  ;; technically not needed, but it helps to highlight where it
  ;; comes from
  :pin gnu)

(use-package org
  :ensure t
  ;; ignore org-mode from upstream and use a manually installed version
  :pin manual)

NOTE: the :pin argument has no effect on emacs versions < 24.4.

Extending use-package with new or modified keywords

Starting with version 2.0, use-package is based on an extensible framework that makes it easy for package authors to add new keywords, or modify the behavior of existing keywords.

First step: Add the keyword

The first step is to add your keyword at the right place in use-package-keywords. This list determines the order in which things will happen in the expanded code. You should never change this order, but it gives you a framework within which to decide when your keyword should fire.

Second step: Create a normalizer

Define a normalizer for your keyword by defining a function named after the keyword, for example:

(defun use-package-normalize/:pin (name-symbol keyword args)
  (use-package-only-one (symbol-name keyword) args
    (lambda (label arg)
      (cond
       ((stringp arg) arg)
       ((symbolp arg) (symbol-name arg))
       (t
        (use-package-error
         ":pin wants an archive name (a string)"))))))

The job of the normalizer is take a list of arguments (possibly nil), and turn it into the single argument (which could still be a list) that should appear in the final property list used by use-package.

Third step: Create a handler

Once you have a normalizer, you must create a handler for the keyword:

(defun use-package-handler/:pin (name-symbol keyword archive-name rest state)
  (let ((body (use-package-process-keywords name-symbol rest state)))
    ;; This happens at macro expansion time, not when the expanded code is
    ;; compiled or evaluated.
    (if (null archive-name)
        body
      (use-package-pin-package name-symbol archive-name)
      (use-package-concat
       body
       `((push '(,name-symbol . ,archive-name)
               package-pinned-packages))))))

Handlers can affect the handling of keywords in two ways. First, it can modify the state plist before recursively processing the remaining keywords, to influence keywords that pay attention to the state (one example is the state keyword :deferred, not to be confused with the use-package keyword :defer). Then, once the remaining keywords have been handled and their resulting forms returned, the handler may manipulate, extend, or just ignore those forms.

The task of each handler is to return a list of forms representing code to be inserted. It does not need to be a progn list, as this is handled automatically in other places. Thus it is very common to see the idiom of using use-package-concat to add new functionality before or after a code body, so that only the minimum code necessary is emitted as the result of a use-package expansion.

Fourth step: Test it out

After the keyword has been inserted into use-package-keywords, and a normalizer and a handler defined, you can now test it by seeing how usages of the keyword will expand. For this, temporarily set use-package-debug to t, and just evaluate the use-package declaration. The expansion will be shown in a special buffer called *use-package*.

Some timing results

On my Retina iMac, the "Mac port" variant of Emacs 24.4 loads in 0.57s, with around 218 packages configured (nearly all of them lazy-loaded). However, I experience no loss of functionality, just a bit of latency when I'm first starting to use Emacs (due to the autoloading). Since I also use idle-loading for many packages, perceived latency is typically reduced overall.

On Linux, the same configuration loads in 0.32s.

If I don't use Emacs graphically, I can test the absolute minimum times. This is done by running:

time emacs -l init.elc -batch --eval '(message "Hello, world!")'

On the Mac I see an average of 0.36s for the same configuration, and on Linux 0.26s.

Upgrading to 2.x

Semantics of :init is now consistent

The meaning of :init has been changed: It now always happens before package load, whether :config has been deferred or not. This means that some uses of :init in your configuration may need to be changed to :config (in the non-deferred case). For the deferred case, the behavior is unchanged from before.

Also, because :init and :config now mean "before" and "after", the :pre- and :post- keywords are gone, as they should no longer be necessary.

Lastly, an effort has been made to make your Emacs start even in the presence of use-package configuration failures. So after this change, be sure to check your *Messages* buffer. Most likely, you will have several instances where you are using :init, but should be using :config (this was the case for me in a number of places).

:idle has been removed

I am removing this feature for now because it can result in a nasty inconsistency. Consider the following definition:

(use-package vkill
  :commands vkill
  :idle (some-important-configuration-here)
  :bind ("C-x L" . vkill-and-helm-occur)
  :init
  (defun vkill-and-helm-occur ()
    (interactive)
    (vkill)
    (call-interactively #'helm-occur))

  :config
  (setq vkill-show-all-processes t))

If I load my Emacs and wait until the idle timer fires, then this is the sequence of events:

:init :idle <load> :config

But if I load Emacs and immediately type C-x L without waiting for the idle timer to fire, this is the sequence of events:

:init <load> :config :idle

It's possible that the user could use featurep in their idle to test for this case, but that's a subtlety I'd rather avoid.

:defer now accepts an optional integer argument

:defer [N] causes the package to be loaded -- if it has not already been -- after N seconds of idle time.

(use-package back-button
  :commands (back-button-mode)
  :defer 2
  :init
  (setq back-button-show-toolbar-buttons nil)
  :config
  (back-button-mode 1))

Add :preface, occurring before everything except :disabled

:preface can be used to establish function and variable definitions that will 1) make the byte-compiler happy (it won't complain about functions whose definitions are unknown because you have them within a guard block), and 2) allow you to define code that can be used in an :if test.

Note that whatever is specified within :preface is evaluated both at load time and at byte-compilation time, in order to ensure that definitions are seen by both the Lisp evaluator and the byte-compiler, so you should avoid having any side-effects in your preface, and restrict it merely to symbol declarations and definitions.

Add :functions, for declaring functions to the byte-compiler

What :defines does for variables, :functions does for functions.

use-package.el is no longer needed at runtime

This means you should put the following at the top of your Emacs, to further reduce load time:

(eval-when-compile
  (require 'use-package))
(require 'diminish)                ;; if you use :diminish
(require 'bind-key)                ;; if you use any :bind variant