cliffle

Chaotic good artificer/ranger, sending you drive-by bugfix pull requests while sitting in a tree in the middle of nowhere.

My psych degree hasn't stopped me from regularly getting nerd-sniped into writing operating systems.

Making reliable things from unreliable parts, currently at Oxide Computer, previously Loon/GoogleX, NativeClient, etc

bad at (but occasionally posting in): español, (简体)中文, nederlands

he/him - Berkeley, CA

Public Key

npub19znrep7u2jrsfvqs5aa3s352l4ppfnxyun8pt28yulnkd7k9x08q9spj90

NIP-05 Address

cliffle@hachyderm-io.mostr.pub

Profile Code

nprofile1qqsz3f3uslw9fpcykqg2w7ccg6906ss5enzwfns44rjw0emxltzn8nspz3mhxue69uhhyetvv9ujuerpd46hxtnfduqs6amnwvaz7tmwdaejumr0dssys89m

Publishing to

relay.damus.io

nos.lol

Author Public Key

npub19znrep7u2jrsfvqs5aa3s352l4ppfnxyun8pt28yulnkd7k9x08q9spj90

Show more details

Last Notes

2026-10-23 18:05:27 UTC

Even though I'm not actively looking for a job, the "I don't join companies with no women in leadership positions" recruiter response is still scary to send.
Like, on the one hand, I want them to receive this feedback (I see what you're doing over there, with your page of faces that look suspiciously like mine). On the other hand, it's still scary to close off potential work, even though I'm employed.

2026-10-20 20:00:04 UTC

Today's question from my thought experiment in bizarro-world personal computing:
If storage is essentially free, the compiler is fast, and source code is not secret... what are the drawbacks from having a system boot from source code directly?
I've seen folks do this with tcc, if I remember correctly. In my case, the CFM bootstrap compiler fits in 8kiB without optimization, so it's easily ROMmable. Compiling the whole system takes about three seconds, which is an unfortunately slow boot but is potentially tolerable. So if the bootloader were the compiler, we could load it into high RAM, compile into low RAM, and then jump into the result.
This would dramatically improve the user's ability to inspect and modify the system...

2026-10-16 16:19:28 UTC

Y'know... I love the warm, undersaturated appearance of photos on my Pebble Time 2. It's such a cheerfully shit display, it reminds me of the display on my color Sony Clie circa 2000, when we were just impressed that it was color at all.
I feel like we could use more displays with an aesthetic other than "150% of sRGB" or whatever.
...this is totally what my parents sounded like when they said vinyl sounded better isn't it 💀

2026-10-15 18:06:40 UTC

This is why, when people say things like "omg the LLM can generate an embedded kernel for me!!!" I can't help but think
"Yeah, it was trained on _at least_ three of mine, plus immeasurable training load on my webserver serving the docs."
By claiming ownership of the result, or even just claiming to have managed the machine while it spit it out, you're basically just stripping all attribution for the actual work that went into its inputs. You might as well just clone my repo, remove the license, and remove my name. So no, I am not impressed.
Anyway, Anthropic just ripped off a bunch of Rain's work with no attribution, as they do.
#nevent1q…h42y

2026-10-14 18:06:13 UTC

So there's this thing I've been wanting, and can't find for purchase, so I guess I've gotta make it.
The Pi Pico has become a de-facto standard for "microcontroller you plug into something." But what if you want to make the CPU yourself?
There are plenty of boards that have an RP2xxx chip _and_ an FPGA, but they generally have a different pinout. I haven't been able to find one that's _just_ an FPGA and has the right pinout.
So, here's 5k LUTs, 128 kiB of SRAM, and several meg of flash. I'm hoping to put this in my PicoCalc so that it runs my CPUs natively.
https://media.hachyderm.io/media_attachments/files/117/440/578/311/928/823/original/eb00a3f1076a3028.png
https://media.hachyderm.io/media_attachments/files/117/440/580/825/863/541/original/7c83d5d74ab5c065.png

2026-10-09 15:36:31 UTC

I feel like Bryan's blog post buries the lede a little bit here with its choice of headline -- yeah Oxide took more money, but, the important thing is why:
We've been quietly profitable since the start of the year and have been receiving more orders than we can fill.
In this current economy of "give me more money so I can light it on fire" and companies filing to IPO proud to spend $3 to make $1, it's nice to be able to tell investors "if you give us some more money, we will make it a larger number in the near future."
https://oxide.computer/blog/our-445m-series-d

2026-10-04 22:05:27 UTC

So, uh... it turns out exFAT is actually a lot simpler than FAT32. A lot of legacy cruft was removed because it doesn't try to be DOS compatible.
It's still Microsoft, so everything's still UTF16, it's intended to be case-insensitive _so it includes a compressed Unicode case translation table in the filesystem,_ but, other than that.
Nothing actually prevents you from using exFAT on a 1 GiB disk, it appears. Seriously considering writing an implementation that omits FAT32/16 support.

2026-08-30 22:19:44 UTC

It's interesting that Google's compliance with non-law government directives regarding the names of bodies of water is so blazingly fast, because, they say, it's their responsibility under the law and they're just following orders...
...and yet their compliance with laws regarding labor rights, collective bargaining, sexual harrassment, and workforce diversity was and is glacial, despite being actual laws.
Every response is a choice.

2026-08-24 15:42:28 UTC

Physical therapist: "Do you have any resistance bands at home?"
Me: "Are you thinking more Rage Against The Machine, or, like, Woody Guthrie?"
Physical therapist: 😑

2026-08-21 19:46:49 UTC

- reply

@nprofile…mv7s We're running it at 100base rates in the management plane, so, just SGMII! Keeps the balance of system cheaper.

2026-08-21 19:43:50 UTC

- reply

@nprofile…mv7s SGMII is unfortunately really, really handy. We're using them in the Oxide management plane for this reason.

2026-08-21 19:38:05 UTC

- reply

@nprofile…mv7s oh man, a VSC-series chip. Shame they're kind of the only game in town for their particular blend of features, huh?

2026-08-09 08:16:40 UTC

In the early 90s, PKWare offered their Data Compression Library as a way to put basic compression into other applications. It uses a different algorithm from PKZIP. I encountered it because it's used by the proprietary recipe program that my mom used, and I've been recovering her database.
Turns out it's had an input-controlled buffer overrun since, well, forever! The library itself won't produce streams with the property required to trigger it, but it also doesn't defend against them.
But, because of the specifics of how its structs are laid out, it's probably not usefully exploitable. (The overrun stomps on a section of the input buffer that isn't likely to be used again.)

2026-08-07 01:33:14 UTC

- reply

@nprofile…mv7s I assume it's Aho-Corasick-as-a-Service?
You can't expect this handheld with *checks notes* gigabytes of RAM and eight cores to search emoji locally, can you?
😭

2026-08-07 01:30:25 UTC

I am bewildered and disappointed to report that Android's onboard emoji search stops working if you have a flaky Internet connection.
Meaning, it is not local.
Meaning, they're going out to the goddamn Internet to see if any of O(1000) records match the word "fish."
I can't even.

2026-08-05 18:51:24 UTC

Been studying some LZ77/LZSS-based compression algorithms recently, and I had a realization I thought was interesting:
If you apply entropy coding to literal (non-repeated) bytes in an LZSS stream, the probability estimate for each byte needs to be the likelihood of it being a _surprise,_ breaking any existing byte sequence in the history window.
The most common bytes in a certain ASCII corpus are what you'd expect: <space>ETAIONSHRDLU<newline>
But the most common _surprising_ bytes for LZSS with a 2kiB sliding window are different:
<newline>AOUIES<space>LCRPMD
The need for efficient encoding of newline makes sense, if you think about it. Given only a set of previous strings, it is very difficult to predict in English where a linebreak will fall -- because it ignores the properties that determine linebreak placement, namely, length of the current line so far and length of the next word.
I've found at least one text compressor in the wild that gets these frequency assumptions wrong, at a cost of several percent.

2026-07-19 04:58:13 UTC

This post captures a lot of my discomfort with generating code using LLMs: you might review it today, but so far everyone I know stops reviewing it eventually.
https://dustycloud.org/blog/faulty-towers-vibe-sickness-and-the-vibe-bobsled/
Seriously considering dusting off my degree, I could start being useful in about two years of school plus supervision.

2026-07-10 15:50:44 UTC

- reply

@nprofile…s8xx oh man, remember how much of a pain it is when you come out and now you have to change all of your passwords to start with "gay?"
Thank goodness the kids these days have password managers.

2026-07-08 20:53:56 UTC

- reply

@nprofile…r9vg yeah, this would be an intermediate compiler output -- it looks like version 0.32.1 of this tool has a corrupt package in some way, 0.32.0 installs.
The release notes for 0.32.1 are content-free and the commits are mostly updates to a markdown file being written for -- I guess? -- an agent of some sort? so it's kind of hard to tell what they broke, but they (or their agent) broke something.

2026-07-08 18:54:55 UTC

Well! Here's a rustc error message I've never seen before.
This is from `cargo install cookcli --no-default-features`
https://media.hachyderm.io/media_attachments/files/116/885/870/648/300/431/original/d1cb98381e5650e1.png

2026-07-07 01:30:14 UTC

- reply

@nprofile…r9vg 🦭

2026-07-06 20:49:46 UTC

Well, TIL about the STAR-MC1 core and the drama around ARM China kind of doing its own thing.
Apparently I'm wearing a STAR-MC1 right now (there's one in the Pebble Time 2). It's vaguely similar to a Cortex-M33.

2026-06-16 19:55:43 UTC

The ARM Cortex-M7 core can execute a basic delay loop of the form
ADDR: subs r0, #1
ADDR+2: bne ADDR
...in one cycle per iteration with the icache on.
...except when ADDR%8==6.
This means the cortex_m::asm::delay routine is still wrong on M7 (the delay is often too short by a factor of 2), but may become correct on some builds depending on the linker.
I found this using my M7 pipeline analysis harness I wrote last week, but it was actually motivated by firmware becoming flaky at work -- turns out, we had some code that was depending on being at ADDR%8==6 without realizing it.

2026-06-09 20:29:52 UTC

Got an inquiry from a recruiter who thought describing each of the company's founders' exploits would catch my attention. (Hint: it generally does not.) One founder was described as:
"famously built a nuclear fusion reactor as a kid"
I mean, by that logic, I built a spaceship as a kid -- in that neither of the things we built worked.
But it's a great way to communicate that your founders are into grandiose statements and are likely believers in the "Great Man" idea, which is a hard no from me.

2026-06-02 01:03:30 UTC

TIL that the cat-face emoji are not represented as [cat emoji] [zero width joiner] [non-cat emoji] and I find that surprising and disappointing

2026-05-21 10:38:00 UTC

If I had a dollar for every time that someone, during RustWeek, has said to me
"Hey, we used to work together, though I have a new name and gender now"
I would now have $5.50.
(One person left their name unchanged, so I'm only counting them for half.)
🦀 ❤️ :progresspride_flag:

2026-05-16 16:18:54 UTC

I feel like, right now, basically any project policy about "AI" (in this case generally code generation with LLMs) is going to provoke a lot of hate from two camps:
Those who will bring ad hominem bullshit if you are insufficiently pro-LLM
And those who will bring ad hominem bullshit if you are insufficiently anti-LLM
And yet the impact of LLM-driven genAI on projects right now is very real, making such policies necessary. But they've become amazing hot-take generators.
Every line of code in every project I maintain was written by a human (or generated by traditional codegen), and I plan on keeping it that way -- in part because I *like* writing code. Projects I co-maintain, like Hubris, may end up following a less clear-cut path (though obviously I'd prefer that we write and understand all the code). I'm really hesitant to post about this because I expect to be attacked.
So, consider that when you decide that replying to this message seems like a good idea. I am blocking freely.

2026-05-09 17:12:57 UTC

- reply

@nprofile…mv7s yeah, lilos is still on GitHub exclusively for the low-effort CI, but even that's starting to fail...

2026-05-09 17:00:52 UTC

Huh. That's odd. So I pushed a fix commit to lilos (on Github) on 4 May, and have records of the CI results.
As of this morning, the CI was failing again, and main had reverted to the prior commit.
Nobody else has admin access on the project. So it kind of seems like Github ... rewound my branch?
https://media.hachyderm.io/media_attachments/files/116/545/685/371/172/693/original/dc4c72a2cfd28bce.png

2026-04-27 17:31:24 UTC

hey
hey kid
yeah, you
wanna see Jean Paul Sartre kick Sam Altman in the nuts
https://existentialcomics.com/comic/652

2026-04-17 00:38:31 UTC

- reply

@nprofile…gr74 my X1G6 had it, I had to switch it over to the Windows one in the bios to improve sleep life. This would have been 2018 era.

2026-04-15 17:29:14 UTC

- reply

@nprofile…u76z I'm having complicated feelings about the prospect of a Klingon translation of Chuck Tingle (not his work, the human being)

2026-04-03 00:24:29 UTC

- reply

@nprofile…mv7s I think lately, the "dry ball" class of issues I've found on boards have usually been caused by, like, incomplete plugging of via-in-pad slurping the paste into the netherworld, or similar.
But yeah, Oxide's got some crazy stackups that really need to bake thoroughly, and we had some thermal inconsistencies in reflow early on. That one's at least easy to fix with heat.

2026-04-03 00:15:40 UTC

- reply

@nprofile…mv7s oh good. I don't have a CAT scanner at home. 😉

2026-04-03 00:12:07 UTC

- reply

@nprofile…mv7s missing a paste aperture on a BGA pad is such an evil defect.
I've had panels where the ball manages to bridge the gap on, like, 10% of inventory, or in one really nasty case, only along one edge of the panel (temperature gradient I guess?)

2026-03-02 00:30:51 UTC

- reply

@nprofile…mv7s well hey, at least you don't have to pare down your AICs

2026-02-26 07:15:14 UTC

- reply

@nprofile…mv7s if you can get it booted far enough to get through dimm training and report the results without those AICs, it would be interesting.
... to me, at least. 😁

2026-02-26 07:12:36 UTC

- reply

@nprofile…mv7s say, if you have any pcie aics in there, and/or integrated graphics, might be worth disabling/removing them and seeing if it affects the RAM detection... sometimes bioses misbehave if cards have weird opinions about BAR locations etc

2026-02-26 07:11:07 UTC

- reply

@nprofile…mv7s yeah, that should be new enough to avoid my concern. Guess we're back to just "bios bullshit" then.

2026-02-26 07:08:48 UTC

- reply

@nprofile…mv7s which Intel microarchitecture is the machine that only sees 384GiB? I recall some of them using channels in sets of three, and either capping out around there due to bios bullshit, or expecting you to jump to the next multiple of 3.

2026-02-18 02:52:13 UTC

- reply

@nprofile…s8xx incidentally this happens in my head every time I accidentally read the comments somewhere
https://media.hachyderm.io/media_attachments/files/116/089/359/476/894/235/original/c5260442bb290973.jpg

2026-02-18 02:38:52 UTC

- reply

@nprofile…s8xx yeah I need to curl up with a warm mug of something and these PoCs, thanks for putting this together

2026-02-18 02:37:42 UTC

- reply

@nprofile…s8xx that's, uh... not the hmac they're truncating by any chance, is it

2026-02-18 02:36:44 UTC

- reply

@nprofile…s8xx oh thank goodness

2026-02-18 02:34:05 UTC

Chewing on Soatok's last commentary on matrix crypto. I haven't implemented DH in a while, but:
If the system allows the other user to use the point at Infinity as their public key... Doesn't that mean if I'm in a position to tamper with both sides of the connection, I could tell each party separately that the other chose it, and they'd be none the wiser because they'd independently derive the same session key? But their traffic would be effectively unencrypted for any observer, not just me.
I sure hope there's some other feature of the protocol that prevents this.
#nevent1q…a70u

2026-02-16 17:29:48 UTC

At the last minute, I submitted a talk suggestion to RustWeek this year, and... they took it! 🎉
So, in May, I will be presenting on patterns in systems software that are not worth carrying into Rust because they go against the grain of the language. And I would like your help.
Have you tried to translate or recreate some systems-y software into Rust, only to be tripped up by a data structure, implementation pattern, or the like -- and found an alternative that unblocked you? I'm curious to hear about it. Surely I haven't seen them all.
Some examples from my journey below. 1/
#rustlang

2026-02-13 22:33:31 UTC

Hm, potential 45-base RNA quine discovered.
The conditions are _somewhat_ contrived but only somewhat.
https://arstechnica.com/science/2026/02/researchers-find-small-rnas-that-can-make-copies-of-themselves/

2026-02-11 00:02:45 UTC

- reply

@nprofile…mv7s in addition to diffusion, there's also the "cloud edge effect," where diffraction around clouds produces spikes in power when the sun passes behind their edges. I found this terribly counter-intuitive.

2026-01-30 02:40:20 UTC

- reply

@nprofile…mv7s another block option would be to include a longer 128 bit timestamp per block and use 64 bit relative within (or even narrower, 9200s is a long block). I did something like this for the timebase for Loon's Iridium telemetry.

2026-01-30 01:25:58 UTC

- reply

@nprofile…mv7s out of curiosity, what's your time precision here?

2026-01-29 23:58:55 UTC

Yup, this has happened to me, except in my case it was "echinoderms have ventral mouths and dorsal anuses, so they're literally poopyheads!"
https://www.smbc-comics.com/comic/nothing-3

2026-01-28 19:51:48 UTC

- reply

@nprofile…g0zc ran this by a Lutheran friend, she says it would have to contain at least one jello "salad" to qualify

2026-01-28 19:46:36 UTC

- reply

@nprofile…s8xx i love this as an escalation strategy when the recipient inevitably fails to respond to the initial private disclosure.
Like, alright NXP, we'll have these characters discuss your bootloader while enthusiastically ...

2026-01-28 18:34:07 UTC

Cleaning out my mom's house, I went through and collected several good cookbooks...
... and also whatever this is!
The 1970s in middle America were quite an era.
I'll post highlights below...
https://media.hachyderm.io/media_attachments/files/115/974/154/155/215/818/original/d765260c02085684.jpg

2026-01-22 18:52:33 UTC

Collected this weekend's updates to my Smalltalk-80 VM into a new wasm build. (You may need to shift-reload it to get it to work if you've visited before.)
https://cliffle.com/tmp-st80/
There's now a speed selector to experience what was considered "barely usable" back in 1983 (5000 bytecodes/sec, not remotely usable by modern standards). And the VM is complete enough to run "Benchmark testStandardTests."
If you crank up the speed knob (and have enough host CPU) the standard tests complete in 537 ms, which is pretty darn fast, and compares favorably with Rochus Keller's optimized C++ VM, despite my rather naive display routines.
#Smalltalk #rust

2026-01-18 17:47:29 UTC

- reply

@nprofile…3src tbh "this is kubernetes but for agents" is the kind of very clear warning label that I like to see on infohazards.

2026-01-17 18:44:13 UTC

Alright, slapped a basic egui wrapper on my Smalltalk-80 VM and now I can poke around in the browser and inspect its source code.
It's...very, very fast. Boots in less than one frame.
https://media.hachyderm.io/media_attachments/files/115/911/909/385/120/493/original/29ace8428248f64f.png

2026-01-16 17:08:40 UTC

- reply

@nprofile…mv7s I did this math recently and propane direct heat came out considerably ahead, assuming a reasonably recent propane furnace. Conversion losses in the genny and the inherent low efficiency of internal combustion could not be corrected by the COP of the heat pumps I was looking at.
I really wanted it to work.

2026-01-15 00:54:13 UTC

As of this morning, my toy Smalltalk-80 VM in Rust can make it through the first 68,139 bytecode instructions of the 1983 Xerox system image (before hitting a primitive operation I haven't implemented yet). Its behavior exactly matches dbanay's high quality C++ gloss of the book at an instruction trace level.
It's also 3.5x faster, and I'm not honestly sure why.
dbanay's code is deliberately not optimized for performance -- it's written to match the book. But for the most part, so is mine. I've taken a few small opportunities (like introducing that HeapOop type I mentioned the other day), but it's still doing essentially the same thing.
I've profiled both programs and see no smoking guns. gcc is making some questionable inlining/not decisions but I'm not even LTO'ing the Rust code (and I _have_ tried LTO'ing the C++ code).
There are a lot of reasons why mine _should_ be slower, naively: bounds checks, integer overflow checks, a lot more asserts added to the VM core during development... heck, I even use checked conversions to go from i16 to u16. And yet.
Normally I can explain why my program is faster than another, so this is bothering me.

2025-12-27 01:13:43 UTC

- reply

@nprofile…syqt ah, yeah, our parts are probably not integrated enough for your needs.

2025-12-27 01:03:11 UTC

- reply

@nprofile…syqt do you need the switching fets integrated on package?
We use an RAA228xxx multiphase part from 12V on Oxide sleds. It might be overkill, but I can get you the exact part number when I'm back home if it's potentially useful.

2025-12-21 03:13:08 UTC

There are a bunch of albums that I originally experienced in the early 2000s through piracy with bad metadata, with the side effect that my brain now _really_ expects their tracks to be in alphabetical order.
This has proven very difficult to unlearn.

2025-12-20 22:11:59 UTC

This use of AI is going to straight up kill someone. Yes please guess the ingredients of my food, said no one with food allergies.
Also, Mr LLM, that is clearly gailan, not bok choi. You had one job.
https://media.hachyderm.io/media_attachments/files/115/754/176/540/804/174/original/a07ff79489b5a01e.png

2025-12-18 06:41:44 UTC

- reply

@nprofile…stct nice. I'm thinking of doing this with my bucket of HP-35s.

2025-12-07 19:14:56 UTC

Oh shiiiiiiiit
It's glitchy but this one comes up, if I give it at least 80 mA at 3.75 V. (The batteries are long dead, and the factory charger wires got eaten by rats, so it's hooked up to my bench supply via test leads.)
I have such fond memories of these ridiculous tiny LED displays. This one is on an HP-35.
https://media.hachyderm.io/media_attachments/files/115/679/867/422/351/286/original/b51f41c10c1510ba.jpg

2025-12-07 04:22:05 UTC

- reply

@nprofile…syqt oh, yeah, I read your post about the scope limitations. I was just attempting to pun on "link training."

2025-12-07 04:14:36 UTC

- reply

@nprofile…syqt if you need to get that link speed up, that could also be "training weekend" 😉

2025-11-28 17:50:06 UTC

- reply

@nprofile…syqt @nprofile…dpjt I admit I was hoping that would be possible when the part was announced. 😞

2025-11-28 17:48:16 UTC

- reply

@nprofile…syqt @nprofile…dpjt so, the closest I've found is probably the SAMA5D2 series from Microchip, which has 128 kiB of L2 cache that can be switched to just be SRAM, and are also available with a nontrivial amount (~256 MiB) of eDRAM in package. I have some eval hardware but I haven't brought it up yet.

2025-11-28 17:26:17 UTC

- reply

@nprofile…syqt in case it's useful: one of the significant advantages of the dual AHB ports on the M33 is the ability to access vector table entries while stacking/unstacking interrupt frames. (Since the hardware does that on M.)
So you can shave several cycles off interrupt response by ensuring that your stack is in some sort of TCM accessible through the S port on the M33, and the vector table and ISRs are accessible through the C port. This will ensure that the hardware can do the operations in parallel. (Note that ST has a historical habit of putting fast SRAM in the C region that's tempting to use for stack. Recently they've gotten more explicitly with the ITCM/DTCM split like on the H7.)
Most applications aren't sensitive to latency to this degree, but mine often are, so I worry about these things.

2025-11-27 16:29:52 UTC

While it is not presented in the tutorial, it turns out Shift-X in Helix is the "select current line without surprises" command that normal-X _should be._
I want to learn key sequences for common operations and not worry about whether (in this case) x d will delete the _next_ line because the current line happens to be blank. So, shift-X for me.
(cc @nprofile…stct)

2025-11-24 16:50:36 UTC

- reply

@nprofile…cx5s @nprofile…syqt just wanna second the recommendation for a two-phase link, if you really want to do this with a linker. You can produce the initial ELF binary with the `-r` flag (if you're on GCC or Clang) and then feed the result back to ld after chewing on it.
But this is generally about the point where I wind up writing a program to munch on the ELF binaries, since we're getting a bit afield of what gnu ld is really good at. (The Hubris build system is arguably just a bigger linker wrapped around the linker, to do more or less this.)

2025-11-24 16:19:02 UTC

- reply

@nprofile…stct There was a scifi universe that has stuck with me, but I can't remember the name or author. In it, non-human aliens think humans are somewhat perverse for doing analytical calculus, because they've had (for a very long time) enough computing power to answer complex physics questions using discrete/stochastic simulations.
This seems increasingly prescient with the rise of LLMs.

2025-11-23 15:50:29 UTC

- reply

@nprofile…syqt @nprofile…dpjt ...incidentally, why end-of-RAM-minus-anything? The ARM stack is full-descending so it's fine for an empty stack pointer to point at the word just above RAM.

2025-11-23 15:39:27 UTC

- reply

@nprofile…syqt @nprofile…dpjt this was tweaked at some point during the development of M-profile, and older M parts actually have a bit for disabling the 8-byte alignment. I assume that ldrd/strd is the motivation, but I don't know that it's documented anywhere.
ISRs will automatically push an additional word if the SP is 4-byte-aligned on entry, and set a status bit to record that it happened so the hardware can skip it on unstacking. That part is done for you.
But particularly in your crt0 (or any asm that calls C code), if you don't maintain the alignment, things will eventually break in weird ways due to some code assuming stack alignment and not finding it. It generally won't crash.

2025-11-23 15:31:18 UTC

- reply

@nprofile…syqt @nprofile…dpjt yeah, it mostly works at 4-byte until it doesn't

2025-11-23 15:20:24 UTC

- reply

@nprofile…syqt fwiw, the M33 also wants 8 byte stack alignment.

2025-11-18 06:06:24 UTC

- reply

@nprofile…syqt are you doing this on the STM32MP257F, or is there a cheaper board?

2025-11-10 22:43:50 UTC

Frustrated with the size (and build quality) of my Framework 16 when traveling, I picked up this tiny (10 inch) Chuwi MinibookX for $350.
So far I'm pleased!
Thoughts in the replies.
https://media.hachyderm.io/media_attachments/files/115/527/813/351/238/337/original/796a933fcd65ff96.jpg

2025-11-10 07:22:32 UTC

- reply

@nprofile…syqt perfect for finding software that treats vector number as 8 bits!
(I've seen it in the wild.)

2025-11-06 16:36:38 UTC

Me: *writes an operating system shipping in Oxide's racks*
OSNews: *crickets*
Me: *writes an operating system used in a hyperscaler's datacenters*
OSNews: *crickets*
Me: *gets bored, ports pwd from V7 Unix*
OSNews: https://www.osnews.com/story/143690/v7-pwd-converted-to-modern-posix-systems/
I guess I need to give folks what they want! 😆

2025-11-01 19:06:20 UTC

I've come to the conclusion that "a description of a set of files" is actually the wrong way to think about tar.
It's an encoding of a sequence of verbs to be applied to a filesystem. And some of those verbs are not obvious, depending on the implementation.
For example, did you know you can use V7 tar as a generalized file deletion tool? It'll even delete directories if you're root.

2025-10-31 23:27:44 UTC

- reply

@nprofile…sxu8 Interesting, I'll reach out -- but I'm pretty sure tar.c started life before sccs had escaped the PWB Unix branch into Research. So I'd be surprised if it had revision history from before the V7 release.
I would be delighted to be wrong about this!

2025-10-31 22:18:02 UTC

Hey Unix history nerds:
Do we know who wrote tar(1)? Like, the original tar.c included in V7 Research Unix (which is the earliest one I can find). The list of potential culprits is pretty short, but there's no attribution I can find.
(If you have an answer, please clarify where you got it from, and reference any contemporaneous/primary sources if you've got them.)

2025-10-30 19:19:51 UTC

After a couple weeks of intermittent detective work, I appear to have hit gold:
Hal Jespersen, who edited POSIX.1, has found RCS files for the tar and cpio chapters, and emailed them to me.
This means I have local revision history (more granular than drafts!) for the POSIX standardization of tar, over five years starting April 1985. This has in turn allowed me to reconstruct the contents of the chapter at the time that GNU forked the format.

2025-10-29 21:59:27 UTC

Well! I've definitely seen my share of "idiosyncratic" C styles omitting braces and exploiting corner cases in 30+ years of having to read other peoples' C code, but this is a new one.
I had not thought about what would happen if you left the braces off a switch. Because, of course, doing so would be utterly unhinged.
I give you: the world's most obtuse version of "if (x == A || x == b)", indistinguishable from an editing mistake.
https://media.hachyderm.io/media_attachments/files/115/459/685/184/154/963/original/377fc4b51e36a8f3.png

2025-10-29 15:54:54 UTC

Subtle changes in ANSI C caused this line of code (from V7 Unix) to stop detecting errors.
read returns a signed int so it can indicate failure, as today. The comparison treats any too-short read _or a negative number_ as a failure. Good so far.
In V7 Unix (and before), sizeof returned an int, in part because it predates the addition of "unsigned' to the language.
During standardization, sizeof was changed to return an _unsigned_ number, since its result can never be negative. Makes sense.
The standard also firmed up the behavior for comparisons (like <) between signed and unsigned, in ways that seem like strict improvements...but with the result here that read's return value is promoted to unsigned before the comparison. Small negative numbers become large positive numbers, and all errors look like successes.
This idiom is not uncommon in the V7 Unix sources and must have posed a significant porting headache at the time!
Anyway, I thought that was kind of interesting. I had to fix this when porting v7pwd the other day.
https://media.hachyderm.io/media_attachments/files/115/458/237/431/244/577/original/d15a75159a65041e.png

2025-10-22 15:56:50 UTC

The POSIX standards, particularly older versions, are very hard to find online.
If you're not interested in sending money to the IEEE (I'm not!) it might be helpful to know that both POSIX 1988 and 1990 were incorporated into US federal standards, which means both are freely available as FIPS publication 151-1 and 151-2, respectively.
https://nvlpubs.nist.gov/nistpubs/Legacy/FIPS/fipspub151-1.pdf
https://nvlpubs.nist.gov/nistpubs/Legacy/FIPS/fipspub151-2.pdf

2025-10-18 15:48:41 UTC

I'm reverse-engineering the Android app for the solar charge controller in my trailer, to try and understand some corner cases of the protocol. I tossed the dex at jadx and started grepping for known hex addresses in its MODBUS register space... without luck?
In fact, I also couldn't find a CRC routine?
And then I noticed a lot of StringBuffer usage in the code and... well, this app is manipulating all the MODBUS packets as StringBuffers containing hex bytes, separated by spaces. It even does math on this form.
I know Java lacks unsigned bytes (a forever-crime) but this is a new one
https://media.hachyderm.io/media_attachments/files/115/395/951/559/165/252/original/088b9e2a2703e053.png

2025-10-18 05:25:38 UTC

In a sea of rapidly enshittifying e-commerce sites, McMaster-Carr remains a breath of fresh air.

2025-10-15 15:52:05 UTC

- reply

@nprofile…syqt @nprofile…2yfz I'm not super familiar with modern Intel implementations of x86, but it's possible you'll still need to trap into SMM for small things here and there, just because so much of the system assumes its existence. (Is it still involved in scaling/power?)
But you ought to be able to minimize it pretty aggressively!

2025-10-15 15:41:44 UTC

- reply

@nprofile…syqt @nprofile…2yfz this is what we wound up doing -- the x86 core is physically incapable of writing to the SPI flash. It can work. In our case, firmware updates have to go through the separate service processor, but for an end-user machine a physical switch would also be fine.
Assuming you're not trying to run an unmodified commodity operating system, of course. But we're not.

2025-10-07 14:52:03 UTC

- reply

@nprofile…8x2j well then! And here I thought it was for (M)icro(C)hip and trying to make a joke

2025-10-07 14:40:43 UTC

- reply

@nprofile…8x2j It's nice that they clearly designated both their Copyright and their...um...Mopyright

2025-10-03 16:30:18 UTC

- reply

@nprofile…59k6 next most challenging task: keeping it working across Neovim updates!

2025-06-19 02:03:35 UTC

Rolling a randomish character for a friend's D&D oneshot game, I used a dwarf name generator and got the best dwarf name.
I AM JSON PARSE SYNTAXERROR
https://media.hachyderm.io/media_attachments/files/114/707/567/299/660/245/original/6d9bfed7207c81c5.png

2025-05-30 18:54:04 UTC

- reply

@nprofile…pf0n you might already know this, but:
You can have a *GitHub specific* readme that GH will show in preference to a top level README.md or equivalent, by putting it in .github/.
I've used this to display a note on GitHub for a fork or whatever without having to crap up the main readme.

2025-04-30 15:33:20 UTC

- reply

@nprofile…59k6 I wonder if peoples' experience with the first board game was similar. They're also arguably non-obvious as a concept, not gated by any specific technological development (some of the older recognizable ones are just lines scratched in a rock), etc.
Though of course mass communications were slower back then.

2025-04-01 18:47:31 UTC

- reply

@nprofile…59k6 a stochastic parrot trained on people discussing math on the internet, huh... it seems like it would be hard to _prevent it_ from being kind of a condescending ass at times.
So, yes, I think it bolded that out of (emulated) condescension.

2025-03-30 19:30:36 UTC

A friend tipped me off to the existence of "heat set threaded inserts," which are brass threaded things with grippy teeth that you can press into plastic with an iron to have robust screw threads.
They're fantastic for making 3d printed parts take proper bolts and stuff.
However, they _also_ turn out to be really useful for fixing cheap plastic objects. A friend dropped the gate remote for his apartment complex, and tore out the screw (which was straight into the ABS plastic). I pushed a heat-set insert into the old screw hole and installed an M3 bolt. Maybe five minutes of work.
Highly recommended.