Bringing GNU Emacs to native code [video]
toobnix.org
toobnix.org
Great job all around!
I haven't had any stability issues with it, either, it just works.
last time i gave it a try opening a large file would completely kill performance, and iirc in particular really long lines (ex. 1000+ chars) would make the thing chug even if the actual file wasn't super big or anything.
If you don't have long lines but you are experiencing slowdowns on large buffers, your mode is doing something. As an example, large XML files bring my Emacs to a crawl, but if I switch to a text mode it's fine.
For long lines, I believe the latest Emacs has a fix: https://www.reddit.com/r/emacs/comments/ccoksw/solong_mitiga...
My uninformed guess is, this is bottlenecked by an inefficient algorithm that accesses the memory too much, i.e. it's nonlinear with respect to the number of characters. I doubt that tweaking the overall speed would fix it.
- inefficient font-lock regexes, which are rarely benchmarked / optimized since most files don't have long lines and because font-lock behavior is quite complicated to begin with;
- inefficient thing-at-point implementations / use, e.g. an O(n*n) thing-at-point called at various points by an O(n) function;
- modes using font-lock regexps where they really just need "fixed" styled text, but other Emacs architecture makes it difficult, e.g. compilation-mode and derivatives.
Emacs's inefficiency in handling files with long lines is due to two factors: (a) The primitive unit of work for the display engine (the code that determines how to combine text, font metrics, syntax highlighting, inline image display, etc.) is a line (a newline-delimited span of characters). (b) The redisplay routine is called very frequently—not just when the screen is repainted, but pretty much any time the screen location of a buffer element need to be calculated, and so, e.g., during navigation. So Emacs is constantly, under the hood, going back to the previous newline and re-calculating how the buffer contents from that point forward should be rendered.
(I have not tried so-long-mode and am mostly on Emacs 26 with some 25.)
Consider an ostensibly simple operation like determining the column in which a particular character appears. The answer to that isn't the number of characters since the last occurrence of `\n`, because whatever modes are active can inject arbitrary spacing, or display particular strings as something shorter or longer (a trivial example is the mode that causes `lambda` to be displayed as `λ`), or cause some characters to be displayed in a non-roman font where a single glyph is more than one column wide, and so on. To determine the character's column accurately you need to take into account everything that could affect how you'd paint the screen at that position, i.e., run the whole redisplay loop for some portion of a buffer. The richer the set of active modes, the more expensive it is to do that and the more often it is done.
Here is the corresponding paper: https://arxiv.org/pdf/2004.02504.pdf
Alternatively, it is easy to suppress the outline slide altogether (add "toc:nil" to "#+OPTIONS:").
It's just a bit of a giveaway as to how the file was made.
From an engineering perspective this is a excellent example of a direct path from interpreted to compiled code. The trade-offs are clear (heck, they are number 0-3), and while there is complexity, all the engineering time has been effectively concentrated inside a single project, rather than forced upon tens of thousands of maintainers and users. Bravo. I wonder what other bytecode interpreters could benefit from this toolchain. Compile times from qlop.
On a Intel(R) Core(TM) i7-4770K CPU @ 3.50GHz Without comp `2020-05-03T15:33:27 >>> app-editors/emacs: 6′40″` With compile `2020-05-03T18:35:43 >>> app-editors/emacs: 2:11:45`
0. https://github.com/tgbugs/tgbugs-overlay/blob/master/app-edi...
Pigs fly just fine with with enough thrust.
Example:
# Doing a simple magit refresh results in the following processes being spawned (tracked with dtrace)
# Notice how many calls are completely redundant
2020 May 4 12:46:22 11819 <67471> 64b /opt/local/bin/git --no-pager --literal-pathspecs -c core.preloadindex=true -c <...>
2020 May 4 12:46:22 11820 <67471> 64b /opt/local/bin/git --no-pager --literal-pathspecs -c core.preloadindex=true -c <...>
2020 May 4 12:46:22 11821 <67471> 64b /opt/local/bin/git --no-pager --literal-pathspecs -c core.preloadindex=true -c <...>
2020 May 4 12:46:22 11822 <67471> 64b /opt/local/bin/git --no-pager --literal-pathspecs -c core.preloadindex=true -c <...>
2020 May 4 12:46:22 11823 <67471> 64b /opt/local/bin/git --no-pager --literal-pathspecs -c core.preloadindex=true -c <...>
2020 May 4 12:46:22 11824 <67471> 64b /opt/local/bin/git --no-pager --literal-pathspecs -c core.preloadindex=true -c <...>
2020 May 4 12:46:22 11825 <67471> 64b /opt/local/bin/git --no-pager --literal-pathspecs -c core.preloadindex=true -c <...>
2020 May 4 12:46:22 11826 <67471> 64b /opt/local/bin/git --no-pager --literal-pathspecs -c core.preloadindex=true -c <...>
2020 May 4 12:46:22 11827 <67471> 64b /opt/local/bin/git --no-pager --literal-pathspecs -c core.preloadindex=true -c <...>
2020 May 4 12:46:22 11828 <67471> 64b /opt/local/bin/git --no-pager --literal-pathspecs -c core.preloadindex=true -c <...>
2020 May 4 12:46:22 11829 <67471> 64b /opt/local/bin/git --no-pager --literal-pathspecs -c core.preloadindex=true -c <...>
2020 May 4 12:46:22 11830 <67471> 64b /opt/local/bin/git --no-pager --literal-pathspecs -c core.preloadindex=true -c <...>
2020 May 4 12:46:22 11831 <67471> 64b /opt/local/bin/git --no-pager --literal-pathspecs -c core.preloadindex=true -c <...>
2020 May 4 12:46:22 11832 <67471> 64b /opt/local/bin/git --no-pager --literal-pathspecs -c core.preloadindex=true -c <...>
2020 May 4 12:46:22 11833 <67471> 64b /opt/local/bin/git --no-pager --literal-pathspecs -c core.preloadindex=true -c <...>
2020 May 4 12:46:22 11834 <67471> 64b /opt/local/bin/git --no-pager --literal-pathspecs -c core.preloadindex=true -c <...>
2020 May 4 12:46:22 11835 <67471> 64b /opt/local/bin/git --no-pager --literal-pathspecs -c core.preloadindex=true -c <...>
2020 May 4 12:46:22 11836 <67471> 64b /opt/local/bin/git --no-pager --literal-pathspecs -c core.preloadindex=true -c <...>
2020 May 4 12:46:22 11837 <67471> 64b /opt/local/bin/git --no-pager --literal-pathspecs -c core.preloadindex=true -c <...>
2020 May 4 12:46:22 11838 <67471> 64b /opt/local/bin/git --no-pager -c core.preloadindex=true -c log.showSignature=false <...>
2020 May 4 12:46:22 11839 <67471> 64b /opt/local/bin/git --no-pager --literal-pathspecs -c core.preloadindex=true -c <...>
2020 May 4 12:46:22 11840 <67471> 64b /opt/local/bin/git --no-pager --literal-pathspecs -c core.preloadindex=true -c <...>
2020 May 4 12:46:22 11841 <67471> 64b /opt/local/bin/git --no-pager -c core.preloadindex=true -c log.showSignature=false <...>
2020 May 4 12:46:22 11842 <67471> 64b /opt/local/bin/git --no-pager -c core.preloadindex=true -c log.showSignature=false <...>
2020 May 4 12:46:22 11843 <67471> 64b /opt/local/bin/git --no-pager -c core.preloadindex=true -c log.showSignature=false <...>
2020 May 4 12:46:22 11844 <67471> 64b /opt/local/bin/git --no-pager -c core.preloadindex=true -c log.showSignature=false <...>
2020 May 4 12:46:22 11845 <67471> 64b /opt/local/bin/git --no-pager -c core.preloadindex=true -c log.showSignature=false <...>
2020 May 4 12:46:22 11846 <67471> 64b /opt/local/bin/git --no-pager --literal-pathspecs -c core.preloadindex=true -c <...>Emacs 26.1 should perform much better at process creation as it uses vfork on macOS instead of fork.
I don’t think that caching is an easy solution. Git can do a lot of weird things so heuristics are bound to be unreliable, brittle, or both. Furthermore the wins from caching are small for the typical users who use systems that are capable of starting processes in good time.
Performance issues aside, a pure Lisp magit is the simplest and more universal setup which is why an attempt to fix the process spawning problems there should be made.
So yes, I'd much prefer C code that Daniel Colascione / John Wiegley / Eli Zaretskii / Stefan / Paul .. have reviewed over newly-written FFI code that hasn't been through that thresher. Everyone that jumped on the emacs-libvterm train early knows what I mean.
I keep Emacs running for months at a time and it's the foundation of pretty much everything I do on a computer. Other than the OS, it generally is the most stable, continuously running, continuously stressed piece of software that I have ever used.
Besides, why would a discerning engineer put up with magit creating all these wasteful processes knowing that most of them are redundant, regardless of performance feel? If I was a magit developer, I would surely try to fix this.
Also see https://stackoverflow.com/a/16902730/615245 (I think it's not needed with the latest versions of Git anymore, though).
> If I was a magit developer, I would surely try to fix this.
Since you're just a regular user, did you check out the issue tracker for related discussions?
It took me 124.98 mins to build the native-comp branch with a 4-cores/8-threads i7-4790k, 32GB RAM on a LXC instance while the master branch took me 244 seconds. Both branches are obtained from latest available git snapshot as 2020-5-4 20:20 CDT.
The following function is passed to (benchmark-run-compiled 10 ...) for each run.
;; -*- lexical-binding: t -*-
(require 'cl-lib)
(defun bf-1 nil
(/
(apply '+
(cl-loop repeat 300000
collect (cl-random 1.0)))
300000.0))
Before each run, (garbage-collect) function is called.
With Emacs's built-in core lisp functions, cl-lib functions and the above benchmark function are all native-compiled, it took 0.5823 second to complete.
With Emacs built-ins and cl-lib are native-compiled but the benchmark function is byte-compiled, it takes 0.6411 second to complete.
With byte-compiled Emacs built-in/cl-lib/benchmark functions, it takes 1.3574 second to complete.
With byte-compiled Emacs built-in/cl-lib functions and interpreted benchmark function, it takes 78.054 seconds with 1 GC taking 75.094 second, which implies the execution roughly takes 2.96 seconds.
I also ran same benchmark on a 4-core A10-6800k. and observed similar ratios on the builds from 2020-5-3.
No reaction at all. I am really curios to understand what I did wrong at the time.
That's an interesting question. My other guess would have been Gimp, and based on a quick glance at Debian's popcon, I'd say Gimp might be slightly in the lead.
Then again, like Open Firmware deployed Forth on millions of computers, right under our noses, it would not surprise me at all if there were a simple Lisp implementation hidden on every computer in the world.
Is there anything more recent than DSSSL?
Each successive step takes care of different optimizations, modifying the code as it goes down. At the last step, he converts LIMPLE to an IR (intermediate representation) that libgccjit understands, and hands it off to gcc for native compilation.
Could you just start with elisp and emit amd64 machine code in one step? Absolutely, but it would be hell to maintain, and then you lose out on all the pluggability of modern compilers. If you (consume and/or) emit standard(-ish) IRs, you get to participate in a pretty amazing ecosystem.
You can easily hack the readtable in Lisp, and rewrite Emacs Lisp sexps as Common Lisp sexps.
Some interesting things to consider are file-local variables, buffer-local variables, dynamic-scope/lexical-scope, how to handle floating point (-0.0e+NaN in Emacs Lisp, custom floating point rouding modes/traps in SBCL for example), but this looks like a reasonable approach overall.
For example (just a draft, this might be incorrect):
USER> (defun make-local-variable (symbol)
(eval `(define-symbol-macro ,symbol (bvar (quote ,symbol)))))
MAKE-LOCAL-VARIABLE
USER> (make-local-variable 'foo)
FOO
USER> (macroexpand-all '(list foo))
(LIST (BVAR 'FOO))
T
T
BVAR could expand into code that calls FFI functions, for compatibility. Using an FFI approach would allow to progressively rewrite some parts of the runtime into CL.
Just to clarify, I know this is a huge work to undertake, but this comes from the assumption that we want to keep all existing Elisp files running identically.
The funny thing about lisp is that all things are easy, but yet, other eco systems offer more finished libraries.
As for compilation to Common Lisp, I don't know how feasible it would be. Apparently they tried something similar with Guile and it ran into problems.
You maybe have some more inside info, but back in 2016 I remember a lot of complaints and that guile was a liability for the much bigger Emacs project. Maybe even people threatening to retire as maintainers
This implies you have either to convert back and forward everything or you just can't use the native datastructures of the new programming language (making the whole operation often quite pointless).
I must confess that most of the comments in this thread seems to start from the assumption that people working on this in the last two+ decades are probably dumb, and this is sad.
For example, there's Hemlock from CMUCL, and its descendants built into Lispworks and Clozure Common Lisp.
Those implementations aren't likely to work well as a substitute for GNU Emacs. For one thing, there's a substantial ecosystem of software that depends on specific APIs and other characteristics of GNU Emacs. Writing code to bridge GNU Emacs APIs with those available in the Hemlock descendants would be a lot of work.
There are some other obstacles as well. CCL's implementation of Hemlock uses the Cocoa text architecture, so it's not portable to platforms other than macOS.
The Lispworks implementation is portable across Windows and numerous UNIXEN, but the Lispworks license will not allow delivery of a proper substitute for GNU Emacs (it forbids building an application that can be construed as a Lisp development system, and when I pressed them about exactly what limits that policy implies, they explicitly used Emacs as an example of an application that would be forbidden).
The original CMUCL Hemlock and its portable version are designed to work with CLX. It could probably be made to work in a modern X environment, but making it fit well into modern GUI environments and porting it to all the platforms GNU Emacs works on would be a huge amount of work.
I'm probably overlooking some other Emacsen, but I don't know of any off the top of my head for which the situation is any easier.
So far, nobody -that I know of- has tried to move GNU Emacs to Common Lisp. GNU Emacs is a well-known, well-used platform and moving to Common Lisp could very well be worth the upfront costs.
Wishful thinking aside, the GNU Emacs development community has hunkered down behind Emacs Lisp, so improving it will bring immediate and tangible benefits to every GNU Emacs user. I love Common Lisp and do use SBCL for a lot of personal projects, but I also love GNU Emacs and I will take any improvements there that I can get. Watching Emacs Lisp become a more viable general purpose language (improved performance means an expanded set of problems that it can now address) is a great development. Stefan Monnier on emacs-devel:
The main benefit of such a compiler is not to run existing Elisp
code faster (99% of existing Elisp code runs fast enough that the
user won't notice if it runs faster) but to make it practical to
write other Elisp code which would otherwise be too slow. But for
that to work well, you want the new compiler to be
available "everywhere", rather than just on some platforms. So it
will only start being useful when it works on GNU/Linux, macOS,
Windows, ARM, RISC-V, x86, amd64, MIPS, younameit (AFAIK
lbgccjit's CPU coverage is already good enough for that, so the
main barrier here is the OS support), and when it's not just an
option at compile-time but when it's included in all builds.Hemlock is not a reimplementation of GNU Emacs; it's derived from earlier Emacsen, just as GNU Emacs is. Hemlock in particular is based on ZWEI, and was in version 0.99 around the time that work started on the GNU Emacs project. It's a slightly older sibling of GNU Emacs, rather than a descendant of it.
GNU Emacs has undoubtedly become the de facto standard Emacs, but I don't see why that disqualifies other branches of the Emacs family tree from using the name that all of them inherit from the venerable TECO implementation.
My point was that Emacsen have already been built in Common Lisp, but that they do not serve as substitutes for GNU Emacs--a point with which you appear to agree.
Maybe we even agree on the reason why: because no Emacs implementation can substitute for GNU Emacs without supporting its ecosystem. The amount of work needed to do that would be enormous, whether the proposed substitute is written in Common Lisp or in something else.
As you said, the value lies primarily in the ecosystem. Common Lisp is by far the better language, but GNU Emacs is by far the most empowering/practical environment. I don't think enormous work would be needed, there are plenty of powerful shortcuts one could take (some ideas are in this thread) and Emacs Lisp can be made to run on top of Common Lisp but it's safe to say that for whatever reasons the will/intent is simply not there.
Emacs Lisp is what we have to go on with and it's nice to see that it is not stagnating.
Any kind of rewrite of Emacs is going to be a lot of work, I am not denying that. But if we assume there is a layer where everything works as it currently does (the C primitives), the Emacs Lisp defined on top of that needs not be defined in C, as far as I know.
Also, the last time I checked, there were a ton of legacy stuff in the configure script for different platforms.
Refactor the existing emacs to work headless and become the legacy-core, while also establishing channels to attatch any other core and frontend.
Additionally, when you free up your architecture and loose up the internal bindings between different areas, you sometimes can open chances for significant improvments which prior where not possible, thus still be faster with a "worse" solution.
Considering how old GNU Emacs is, how much historical problems it has and how often you hear that this and that can't be optimized for this and that reason, this seems like such a situation where loosing up would open faster roads long term.
There are small syntactic differences between Emacs lisp and Common Lisp so a lexer that handled those would suffice.
There are some small semantic differences but they could be handled by porting code. And the Emacs c code could be rewritten in lisp or ported into the Common Lisp implementation, depending on your taste.
> There are some small semantic differences but they could be handled by porting code. I suspect there are more than just small differences. And someone would have to do the work to port the code. I suspect that if packages that people use don't work, they will just avoid using the new version of Emacs, as opposed to abandoning the package.
It's not an impossible project, but it's not clear to me it's easier/better than the proposed solution.