2025-06-25 17:07:48
Smalltalk programs strictly alternate between data and behaviour. Messages (the only kind of data in Smalltalk) are implicit, constructed fresh at each point in the program method call syntax is used, and are shallow, meaning that the slots in each message object always contain references to objects, never other messages.1
This is in contrast to most other languages with true data, where (for example) lists may contain other data, and are not constrained to containing only object references / function values.
So, what would a Smalltalk be like without this strict alternation, where messages could appear as values? Suddenly the universe of Smalltalk values grows larger: previously, everything was an object, but now some things are data!
Of course, objects acting as reified messages may appear! ↩
2025-06-25 17:04:57
Usually, method lookup results in a single body to execute. This is analogous to the way pattern matching usually tries patterns in order, selecting just one continuation to execute.
What if, instead, we allowed all matching patterns from a bunch of alternatives to execute? (Or, in object-oriented terms: executed all methods potentially matching a given method call.)
Pattern alternatives do not become a kind of superposition, because there’s no notion of mutual exclusion; instead, they become a way of creating multiple concurrent branches of execution, somehow. (Not to say there’s any particular kind of interleaving or parallel execution of these branches that makes any sense! One could well limit consideration to sequential execution of each matching method, to start with.)
Can we recover true alternatives from this kind of every-match construct (it needs a name!)? One approach is to follow the idea from Alex’s, Mahdi’s and my PEG paper,1 treating alternation (/) as a kind of parallel match construct and using negative lookahead to cause a later branch to fail if some earlier branch succeeds.
T. Garnock-Jones, M. Eslamimehr, and A. Warth, “Recognising and Generating Terms using Derivatives of Parsing Expression Grammars,” Jan. 2018. https://arxiv.org/abs/1801.10490 ↩
2025-06-17 15:26:53
A fantasy abstract is a short piece of writing in the style of the abstract of an academic paper presenting the outline of a piece of research that has not (yet) been done. It acts as inspiration, as a means of communicating an idea, and as a place to nucleate further thinking on the topic, perhaps eventually kicking off the desired research.
It’s something I started doing during my PhD studies to help record and structure the relationships among ideas. I record a little bit of metadata about each abstract: not much more than the names of one or more broad research “themes” or threads that the idea might fit into.
For example, here’s one from January 2011, shortly after I started experimenting with the form. Fourteen years later it remains fantastic and unexplored (by me at least!):
Contracts for Protocols
Created: 2011-01-10
Thread
Network Languages
Abstract
Existing messaging middleware systems provide very low-level facilities to application developers, ranging from simple point-to-point datagram transfer up through simple stereotypical interaction patterns such as (optionally transacted) request-reply or publish-subscribe. These low-level facilities are then composed by the application developer into higher-level interactions, but without the benefit of any formal way of describing the higher-level interactions. This paper introduces contracts for messaging protocols implemented using messaging middleware, describes a prototype implementation, and discusses lessons learned.
Some things to note:
Citations are useful if you have them, but the main point is to capture the idea, not do an exhaustive background literature survey. In the example above there’s the ludicrous omission of any mention of session types, for example; what I had in mind was something akin to what, these days, are called “dynamic monitors” in the session types literature. Perhaps if I’d expanded this abstract into an actual paper at the time it’d have been a timely contribution :-) It’s a bit stale now…
It can be an absolute fantasy. Feel free to refer to nonexistent (but plausible?) research results. If you ever pick up the idea or gift it to someone else, it’ll be made rigorous and realistic then. Use the fantasy abstract to get the feeling of your idea.
2025-05-01 03:27:54
I always forget how to do this, so I’m writing it here in part as a reminder for myself next time I need to do this.
These instructions are for using daemontools and daemontools-run packages on Debian.
/etc/service/tonyg-services
Create a file /etc/service/tonyg-services/run containing
#!/bin/sh
exec setuidgid tonyg svscan /home/tonyg/services
Replace setuidgid with sudo -u if you want to preserve supplemental groups.
chmod a+x /etc/service/tonyg-services/run
/home/tonyg/services
2025-03-06 14:42:25
Small experiments in the use of libliftoff to try out the modern Linux graphics stack drove home quite how slow DRM “dumb buffers” can be, but also that it’s reading that’s slow, not writing.
Reading from a “dumb buffer” on my AMD GPU is orders of magnitude slower than reading from RAM. It can take seconds to read out a full 4k frame. It’s roughly a thousand times slower than reading RAM.12
Writing, by contrast, is quick.
While it is folklore that “dumb buffers are slow”, I found it challenging to find any
authoritative source on the matter. However, I did find something. In
/usr/include/drm/drm.h, we see the following comment, which sort of hints at the wider
situation:
/**
* DRM_CAP_DUMB_PREFER_SHADOW
*
* If set to 1, the driver prefers userspace to render to a shadow buffer
* instead of directly rendering to a dumb buffer. For best speed, userspace
* should do streaming ordered memory copies into the dumb buffer and never
* read from it.
*
* Note that this preference only applies to dumb buffers, it's irrelevant for
* other types of buffers.
*/
#define DRM_CAP_DUMB_PREFER_SHADOW 0x4Indeed, “for best speed […] never read from it.”
Update: Subsequent experimentation using gbm to allocate buffer objects shows that it
doesn’t help if you need to read or write pixel data to them (as opposed to, presumably, using
the GPU to render into them). Setting the GBM_BO_USE_WRITE flag when allocating a buffer
object, to allow subsequent writing of pixel data, causes the dri backend of
gbm
to simply allocate a “dumb buffer”!
Quick-and-dirty C experimentation shows speeds of ~2ms to read a full 3840×2160×32bit frame out of normal RAM. That’s about 16GB/s. Eyeballing the slow “dumb buffer” read times suggests then perhaps about 16MB/s for that! ↩
As a corollary to this realisation, I learned that attempting to use surfaces backed solely by “dumb buffers” to do fallback software composition is a losing proposition. Hence the whole idea of “shadow” buffers, presumably! ↩
2024-10-03 02:12:02
“Erlang supports change of code in a running system.”
However, the details are a bit fiddly. Here’s a cheat-sheet I used recently for a simple TCP service written using Erlang.
My program was a single module, running outside of any OTP application context. The
instructions here need minor emendation to either explicitly list modules to purge and reload
or to discover all modules within a single application; see the places in server-reload
below mentioning the atom my_server.
I did not use the -on_load() directive, because I wanted to be able to use multiple nodes
rather than controlling reloads from a single node’s shell repl, and I couldn’t figure out how
to make the two play nicely together.
I exported a code_change/0 from my module, to be called after loading a new version of the
module into a node. It sends a message code_change to each “global” actor in my program (in
this case, there was only one).
-export([code_change/0]).
code_change() ->
io:format("+ code_change~n"),
%% name registered previously with `global:register_name/2`:
global:send(name_of_my_global_actor, code_change),
ok.That actor distributes the notification on to any inferior actors it is managing, and then does an “MFA” self-call to upgrade its own codebase.
index(Connected) ->
receive
code_change ->
[P ! code_change || {_Peer, P} <- Connected],
?MODULE:index(Connected);
...
end.Similarly, all other notified actors perform “MFA” self-calls.
connection(Sock, Username, IndexPid) ->
receive
code_change ->
?MODULE:connection(Sock, Username, IndexPid);
...
end.Actors need to take care to manage upgrades of their state at the same time as they do the “MFA” self-calls.
I wanted it to be run by daemontools, so created the
following shell script called run, which daemontools will pick up to start a service:
#!/bin/sh
set -e
erlc -o ebin my_server.erl
exec erl \
-noshell \
-pa ebin \
-sname mainnode \
-setcookie f98b3a1e-80ec-11ef-b752-0b638e4de31c \
-s my_serverPick a fresh random cookie for the -setcookie argument. I used uuid(1).
Then, I created this script, server-reload:
#!/bin/sh
set -e
erlc -o ebin my_server.erl
exec erl \
-noshell \
-pa ebin \
-setcookie f98b3a1e-80ec-11ef-b752-0b638e4de31c \
-sname undefined \
-eval "
ServerNode = mainnode@$(hostname -s),
io:format(\"ServerNode: ~p~n\", [ServerNode]),
true = net_kernel:connect_node(ServerNode),
spawn(ServerNode, fun () ->
code:purge(my_server),
code:load_file(my_server),
ok = my_server:code_change()
end),
init:stop()"Running server-reload causes the source code to be compiled and hot-loaded into the running
server.
Then, I used a git post-receive hook to automatically recompile and reload the code on push to live:
#!/bin/sh
set -e
unset GIT_DIR
cd $HOME/location-of-checkout-of-server-repository
git pull --ff-only
./server-reloadThat’s all. The end result worked well: I used it to run a hotfix to my TCP service with many tens of live, active connections, and not one of them noticed a thing.
