Welcome to El Compilador, a compiler for Emacs Lisp.
The compiler can now generate C code that can be compiled as part of Emacs. Using the bubble sort benchmark from http://www.emacswiki.org/emacs/EmacsLispBenchmark (with the list bumped to 1000 elements), with 100 runs, I got some timings:
Approach | Seconds |
---|---|
interpreted | 54.874574673000005 |
byte-compiled | 13.390510359999999 |
el-compilador | 4.312016277000001 |
I've long wanted to write a compiler for Emacs Lisp. Here it is. Well, the start of it. In the long term I have a few goals for Emacs and Emacs Lisp that are served by this project:
I think Emacs should move more strongly toward self-hosting. Too much of Emacs is written in C, and in the long term this should be migrated to lisp. Beyond just being more fun to hack, having Emacs written in Emacs Lisp would make it simpler to upgrade the language implementation.
There are plenty of functions currently written in C which were either
translated for performance (widget-apply
) or just because some other
part of the core needed to call it. These would stop being acceptable
reasons to write in C.
The C core is also badly behaved about making direct calls. This is
ok for primitives like cons
, but not ok for functions that one might
reasonably want to advise or rewrite, like read
. Normally this lack
of indirection is just because it is a pain to write out in C -- but
automatic translation could eliminate this problem.
I'm also interested in using the compiler to either write a JIT or a new register-based bytecode interpreter. These could be done without modifying Emacs once the new FFI code lands.
Finally, it is bad and wrong that Emacs has three bytecode interpreters (the Emacs Lisp one, the regexp engine, and CCL). There should be only one, and we can use this work to push Emacs toward that goal.
You can use the function in loadup.el
to load the compiler and then
use the two handy entry points:
elcomp--do
. The debugging entry point. This takes a form, compiles it, and then dumps the resulting IR into a buffer. For example, you can try this on a reasonably direct translation ofnthcdr
fromfns.c
:
(elcomp--do '(defun nthcdr (num list)
(cl-check-type num integer)
(let ((i 0))
(while (and (< i num) list)
(setq list (cdr list))
(setq i (1+ i)))
list)))
- You can pass
elcomp--c-translate
as the third argument toelcomp--do
to use the "C" back end. At least some forms of the output will compile. It targets the API used by the Emacs source tree (not the Emacs dynamic module API). Some constructs don't have the needed back end support yet, so not everything will work.
El Compilador is an
SSA-based
compiler. The objects in the IR are described in elcomp.el
. EIEIO
or cl-defstruct
are used for most things.
The compiler provides a number of optimization passes:
-
Jump threading,
elcomp/jump-thread.el
. This also does some simple optimizations on predicates, likenot
removal. This can sometimes turn athrow
into agoto
when it is caught in the samedefun
. -
Exception handling cleanup,
elcomp/eh-cleanup.el
. This removes useless exception edges. -
Block coalescing,
elcomp/coalesce.el
. This merges basic blocks when possible. -
Constant and copy propagation,
elcomp/cprop.el
. This also evaluates pure functions. -
Dead code elimination,
elcomp/dce.el
. -
Type inference,
elcomp/typeinf.el
. This is a flow-sensitive type inferencer.
There are any number of bugs. There are some notes about them in various files. Some are filed in the github issues.
The into-SSA pass is written in the stupidest possible way. Making this smarter would be nice.