Agreguesi i feed

I love Markdown with all my heart. It's a markup language so simple to understand that even people who are not software engineers can use it in a few minutes.

The flip side of that coin if that Markdown is limited. It can let you create various title levels, bold, italics, strikethrough, tables, links, and a bit more, but not so much.

When it comes to more complex documents, most people resort to full fledged office suite like Microsoft Office or LibreOffice. Both have their merits, but office file formats are brittle and heavy.

The alternative is to use another more complex markup language. Academics used to be into LaTeX but it's often tedious to use. Typst emerged more recently as a simpler yet useful markup language to create well formatted documents.

Tinymist is a language server for Typst. It provides the usual services a Language server provides, like semantic highlighting, code actions, formatting, etc.

But it really stands out by providing a live preview feature that keeps your cursor in sync in Helix when you are clicking around in the live preview!

I only had to install it with

$ brew install tinymist

I then configured Helix to use tinymist for Typst documents, enabling live preview along the way. This happens of course in ~/.config/helix/languages.toml

[language-server.tinymist] command = "tinymist" config = { preview.background.enabled = true, preview.background.args = ["--data-plane-host=127.0.0.1:23635", "--invert-colors=never", "--open"] } [[language]] name = "typst" language-servers = ["tinymist"]

A warm thank you to my lovely friend Felix for showing me the live preview mode of tinymist!

Bloomberg's Mark Gurman says Apple's foldable iPhone is still "on track" for a September unveiling alongside the iPhone 18 Pro lineup. 9to5Mac reports: The report notes that Apple's stock took a hit earlier today after Nikkei Asia indicated the iPhone Fold was having serious production issues. Clearly, sources within Apple were motivated to share positive news via Gurman. Not long ago, Gurman himself said that he was expecting an iPhone Fold release date that was a little bit later than iPhone 18 Pro. That's still very possible, but it sounds like Apple is internally feeling optimistic about its targeted September launch. The report continues: "While the complexity of the new display and materials may limit initial supply for several weeks, Apple is currently operating with a plan to put the device on sale around the same time -- or very soon after -- the new non-foldable models, the people said." Gurman adds an important qualifier: "Still, the release is six months away and production has yet to ramp up. That means the timing isn't final."

Read more of this story at Slashdot.

An anonymous reader quotes a report from The Drive: Farmers have been fighting John Deere for years over the right to repair their equipment, and this week, they finally reached a landmark settlement. While the agricultural manufacturing giant pointed out in a statement that this is no admission of wrongdoing, it agreed to pay $99 million into a fund for farms and individuals who participated in a class action lawsuit. Specifically, that money is available to those involved who paid John Deere's authorized dealers for large equipment repairs from January 2018. This means that plaintiffs will recover somewhere between 26% and 53% of overcharge damages, according to one of the court documents (PDF) -- far beyond the typical amount, which lands between 5% and 15%. The settlement also includes an agreement by Deere to provide "the digital tools required for the maintenance, diagnosis, and repair" of tractors, combines, and other machinery for 10 years. That part is crucial, as farmers previously resorted to hacking their own equipment's software just to get it up and running again. John Deere signed a memorandum of understanding in 2023 that partially addressed those concerns, providing third parties with the technology to diagnose and repair, as long as its intellectual property was safeguarded. Monday's settlement seems to represent a much stronger (and legally binding) step forward. The report notes that a judge's approval of the settlement is still required but likely to happen. John Deere also faces another lawsuit by the U.S. FTC, accusing the company of forcing farmers to use its authorized dealer network and driving up their costs for parts and repairs.

Read more of this story at Slashdot.

A $500,000 "Survivor"-style corporate retreat for 120 Plex employees in Honduras "turned into a week-long disaster involving illness, wild animals, armed guards, and employees stranded on a remote island," reports the Daily Beast. The CEO was bedridden by E. coli, staff were collapsing in brutal heat during Navy SEAL-led drills, there were fire ant attacks, uncooked food, and failing utilities. At one point, a porcupine even crashed through the ceiling of a guest's room. Here's an excerpt from the report: Tech media company Plex flew its 120 employees to a Honduran resort in 2017 for what was billed as a Survivor-style getaway. They called it "Plexcon." The first harbinger of trouble was an email that arrived before the group departed, informing them that the hotel manager and chef had both quit within days of each other. Things went sharply downhill from there. CEO Keith Valory, 54, had flown out a day early, intending to channel his inner Jeff Probst and welcome his staff off the buses like a game show host. Instead, he spent the arrival morning flat on his back. "I got E. coli, which is maybe the worst thing you could get, possibly, ever," Valory told the Wall Street Journal this week. "Just as people were arriving on the buses, I was like, 'Uh oh.' I lost 8 or 10 pounds. They had a doctor come to me, which apparently is pretty standard. They nailed an IV bag to the bedpost." With the CEO incapacitated, chief product officer and co-founder Scott Olechowski, 52, stepped in to run proceedings -- beginning with a forced eating challenge in which one employee had to consume a dead tarantula. [...] Sean Hoff, 42, founder of Moniker Partners, the independent retreat agency that planned the trip, was running himself ragged attempting damage control -- the showers, water, and electricity kept cutting out. [...] Meanwhile, senior software engineer Rick Phillips, 53, was trying to sleep when he heard a crash in his room. He ignored it until morning. "I got up and went over to get in the shower, and there was a porcupine," he said. "It must have climbed a tree and fallen through the ceiling."

Read more of this story at Slashdot.

The FBI says (PDF) Iran-linked hackers disrupted internet-connected systems used by U.S. oil, gas, and water companies. Even with the recent two-week ceasefire between Iran and the United States and Israel, hackers backing Tehran say they won't end their retaliatory cyberattacks. The Hill reports: The report warned that similar companies across the country should be aware of an increased push by hackers to take over programmable logic controller (PLC) systems, which can be used to digitally control physical machinery from remote locations. Secure internet access for PLCs from one company, Rockwell Automation, were removed by Iran-linked coders who then "maliciously interacted with project files and altered data," according to the report. Hackers first gained access to some of the platforms in January of last year. All access to compromised platforms ended in March, the report said. The FBI said the move resulted in "operational disruption" and "financial loss." [...] Rockwell Automation wasn't the only company to recently face cyberattacks from Iran-linked hackers. Stryker, a major U.S. medical device maker, was targeted by Iran-affiliated coders in mid-March. It was unclear if physical operations were affected by the security breach. FBI Director Kash Patel was personally impacted by hackers who leaked his emails and records related to his personal travels and business from more than 10 years ago. [...] The FBI urged companies to adopt network defenders and multifactor authentication to prevent future attacks. Tuesday's report was published alongside the National Security Agency, the Department of Energy, and the Cybersecurity and Infrastructure Security Agency. "Government and experts have been warning about internet connected systems for years, and how vulnerable they are," one source familiar with the federal investigation into the hacks told CNN. Many companies have "ealready removed those systems and followed the guidance," the person added.

Read more of this story at Slashdot.

• To work early, writing and planning, team call. Dug up an old mail of mine on inflated claims around tenders which seem to become a topic.
• Encouraging partner call, sync with Chris and then Philippe.
• Published the next strip: A Community Organizer - "aligning the community"
• Had the first Collabora Office Technical Committee call - missing Miklos' careful preparation, next week will be better no doubt.
• Interesting chat with an old FLOSS hand.
• Band practice in the evening.

A New York Times investigation by John Carreyrou claims a British cryptographer named Adam Back is the strongest circumstantial candidate yet for being Satoshi Nakamoto. The report citing overlaps in writing style, ideology, technical background, and old posts that outlined key parts of Bitcoin years before its launch. Carreyrou is a renowned investigative journalist and author, best known for exposing the massive fraud at Theranos while at the Wall Street Journal. Here's an excerpt from the report: ... As anyone steeped in Bitcoin lore will tell you, Satoshi was a master at the art of maintaining anonymity on the internet, leaving few, if any, digital footprints behind. But Satoshi did leave behind a corpus of texts, including a nine-page white paper (PDF) outlining his invention and his many posts on the Bitcointalk forum, an online message board where users gathered to discuss the digital currency's software, economics and philosophy. And that corpus, it turned out, had expanded significantly during the impostor's civil trial when Martti Malmi, a Finnish programmer who collaborated with Satoshi in Bitcoin's early days, released a trove of hundreds of emails he had exchanged with him. Emails Satoshi sent to other early Bitcoin adopters had surfaced before, but none came close in volume to the Malmi dump. If Satoshi was ever going to be found, I was convinced the key lay somewhere in these texts. Then again, others must have gone down this road before me. Journalists, academics and internet sleuths had been trying to identify Satoshi for 16 years. During that span, more than 100 names had been put forward, including those of an Irish cryptography student, an unemployed Japanese American engineer, a South African criminal mastermind and the mathematician portrayed in the movie "A Beautiful Mind." The most alluring theories had focused on coincidences that aligned with what little was known about Satoshi: a particular code-writing style, a mysterious work history, an expertise in Bitcoin's key technical concepts, an anti-government worldview. But they had run aground under the weight of an alibi or some other piece of inconsistent or contrary evidence. Each failure had been met with glee by many members of the Bitcoin community. As they liked to point out, only Satoshi could definitively prove his identity by moving some of his coins. Any evidence short of that would be circumstantial. It seemed foolish to think that I could somehow crack a case that had confounded so many others. But I craved the thrill of a big, challenging story. So I decided to try once more to unmask Bitcoin's mysterious creator. Back, for his part, denies being Satoshi, writing in a post on X: "i'm not satoshi, but I was early in laser focus on the positive societal implications of cryptography, online privacy and electronic cash, hence my ~1992 onwards active interest in applied research on ecash, privacy tech on cypherpunks list which led to hashcash and other ideas."

Read more of this story at Slashdot.

Starting May 20th, Amazon will stop Kindle Store access for Kindle and Kindle Fire devices released in 2012 and earlier. After that date, those devices will "no longer be able to purchase, borrow, or download new content." Owners can still read content already on the device, but if an affected device is reset or deregistered after the cutoff, it can't be re-registered. The Verge reports: The complete list of affected devices goes all the way back to the original Kindle that launched in 2007 with a full keyboard and scroll wheel. [...] Amazon will be notifying affected users over email ahead of May 20th with an explanation of what their older devices can and cannot do. Pre-2012 Kindle Fire devices will be subjected to the same limitations as Kindle e-readers when it comes to books, but other apps and Amazon services on those devices won't be impacted. For longtime users wanting to take the opportunity to upgrade to newer Kindle hardware, Amazon will offer a 20 percent discount on new Kindle devices and a $20 ebook credit that will be added to their accounts after upgrading, valid until June 20th, 2026, at 11:59PM PT. Their older purchases will be available on new devices as long as they log in to the same account they've been using for the past 14 years or more.

Read more of this story at Slashdot.

An anonymous reader quotes a report from the Financial Times: Iran will demand that shipping companies pay tolls in cryptocurrency for laden oil tankers passing through the Strait of Hormuz (source paywalled; alternative source), as it seeks to retain control over passage through the key waterway during the two-week ceasefire. Hamid Hosseini, a spokesperson for Iran's Oil, Gas and Petrochemical Products Exporters' Union, told the FT on Wednesday that Iran wanted to collect tolling fees from any tanker passing and to assess each ship. "Iran needs to monitor what goes in and out of the strait to ensure these two weeks aren't used for transferring weapons," said Hosseini, whose industry association works closely with the state. "Everything can pass through, but the procedure will take time for each vessel, and Iran is not in a rush," he added. [...] Hosseini said that each tanker must email authorities about its cargo, after which Iran will inform them of the toll to be paid in digital currencies. He said that the tariff is $1 per barrel of oil, adding that empty tankers can pass freely. "Once the email arrives and Iran completes its assessment, vessels are given a few seconds to pay in Bitcoin, ensuring they can't be traced or confiscated due to sanctions," Hosseini added.

Read more of this story at Slashdot.

Meta has launched Muse Spark, its first major AI model under Alexandr Wang's leadership. The model was built over the past nine months and is being positioned as a significant step up from Llama 4. Axios reports: Muse Spark will power queries in the Meta AI app and Meta.ai website immediately, with plans to expand across Facebook, Instagram and WhatsApp. The model accepts voice, text and image inputs, but produces text-only output. [...] Meta plans to release a version of Muse Spark under an open-source license. The model uses a fast mode for casual queries and several reasoning modes. A "shopping mode" highlights how Meta hopes to differentiate itself. It combines large language models with data on user interests and behavior. Over time, the model will also power "features that cite recommendations and content people share across Instagram, Facebook, and Threads," Meta said in a blog post. Wang, the 29-year-old entrepreneur who co-founded Scale AI, joined Meta's "superintelligence" unit last year to help Meta catch up to rival models from OpenAI and Anthropic.

Read more of this story at Slashdot.

Microsoft has apparently terminated the account VeraCrypt uses to sign its Windows drivers and bootloader, leaving the encryption project unable to publish Windows updates and throwing future releases into doubt. VeraCrypt's developer says Microsoft gave no clear explanation or warning for the move. "I didn't receive any emails from Microsoft nor any prior warnings," Mounir Idrassi, VeraCrypt's developer, told 404 Media. From the report: VeraCrypt is an open-source tool for encrypting data at rest. Users can create encrypted partitions on their drives, or make individual encrypted volumes to store their files in. Like its predecessor TrueCrypt, which VeraCrypt is based on, it also lets users create a second, innocuous looking volume if they are compelled to hand over their credentials. Last week, Idrassi took to the SourceForge forums to explain why he had been absent for a few months. The most serious challenge, he wrote, "is that Microsoft terminated the account I have used for years to sign Windows drivers and the bootloader." "Regarding VeraCrypt, I cannot publish Windows updates. Linux and macOS updates can still be done but Windows is the platform used by the majority of users and so the inability to deliver Windows releases is a major blow to the project," he continued. "Currently I'm out of options." Idrassi told 404 Media the termination happened in mid-January. "I was surprised to discover that I could no longer use my account," he said. On the forum and in the email to 404 Media, Idrassi shared what he said was the only message he received connected to the account shutdown. "Based on the information you have provided to date, we have determined that your organization does not currently meet the requirements to pass verification. There are no appeals available, we have closed your application," it reads. Idrassi told 404 Media the message is concerning his company IDRIX. "As you can read in their message, they say that the organization (IDRIX) doesn't meet their requirements, but I don't see which requirement IDRIX suddenly stopped meeting," he said. Idrassi said he has tried contacting Microsoft support, but he received automated responses that he believes contained AI-generated text.

Read more of this story at Slashdot.

People often think of containers as locked boxes that keep software separate from the rest of the computer. In reality, that safety depends on a chain of digital gatekeepers. If one gatekeeper misses a signal, the whole box opens up.

alternative_right quotes a report from the New York Post: The CIA used a futuristic new tool called "Ghost Murmur" to find and rescue the second American airman who was shot down in southern Iran, The Post has learned. The secret technology uses long-range quantum magnetometry to find the electromagnetic fingerprint of a human heartbeat and pairs the data with artificial intelligence software to isolate the signature from background noise, two sources close to the breakthrough said. It was the tool's first use in the field by the spy agency -- and was alluded to Monday afternoon by President Trump and CIA Director John Ratcliffe at a White House briefing. "It's like hearing a voice in a stadium, except the stadium is a thousand square miles of desert," a source briefed on the program told The Post. "In the right conditions, if your heart is beating, we will find you." The relatively barren landscape made for "an ideal first operational use" of Ghost Murmur, the first source noted. "Normally this signal is so weak that it can only be measured in a hospital setting with sensors pressed nearly against the chest," the source said. "But advances in a field known as quantum magnetometry -- specifically sensors built around microscopic defects in synthetic diamonds -- have apparently made it possible to detect these signals at dramatically greater distances." "The capability is not omniscient. It works best in remote, low-clutter environments and requires significant processing time," this person added.

Read more of this story at Slashdot.

BrianFagioli writes: Artificial intelligence has now run directly on a satellite in orbit. A spacecraft about 500km above Earth captured an image of an airport and then immediately ran an onboard AI model to detect airplanes in the photo. Instead of acting like a simple camera in space that sends raw data back to Earth for later analysis, the satellite performed the computation itself while still in orbit. The system used an NVIDIA Jetson Orin module to run the object detection model moments after the image was taken. Traditionally, Earth observation satellites capture images and transmit large datasets to ground stations where computers process them hours later. Running AI directly on the satellite could reduce that delay dramatically, allowing spacecraft to analyze events like disasters, infrastructure changes, or aircraft activity almost immediately. "This success is a glimpse into the future of what we call Planetary Intelligence at scale," said Kiruthika Devaraj, VP of Avionics & Spacecraft Technology. "By running AI at the edge on the NVIDIA Jetson platform, we can help reduce the time between 'seeing' a change on Earth and a customer 'acting' on it, while simultaneously minimizing downlink latency and cost. This shift toward integrated AI at the edge is a technological leap that can help differentiate solutions like Planet's Global Monitoring Service (GMS), providing valuable insights for our customers and enabling rapid response times when it matters most."

Read more of this story at Slashdot.

An anonymous reader quotes a report from TechCrunch: A group of Russian government hackers have hijacked thousands of home and small business routers around the world as part of an ongoing campaign aimed at redirecting victim's internet traffic to steal their passwords and access tokens, security researchers and government authorities warned on Tuesday. [...] The hacking group targeted unpatched routers made by MikroTik and TP-Link using previously disclosed vulnerabilities according to the U.K. government's cybersecurity unit NCSC and Lumen's research arm Black Lotus Labs, which released new details of the campaign Tuesday. According to the researchers, the hackers were able to spy on large numbers of people over the course of several years by compromising their routers, many of which run outdated software, leaving them vulnerable to remote attacks without their owners' knowledge. The NCSC said that these operations are "likely opportunistic in nature, with the actor casting a wide net to reach many potential victims, before narrowing in on targets of intelligence interest as the attack develops." Per the researchers and government advisories, the Russian hackers hacked routers to modify the device's settings so that the victim's internet requests are surreptitiously passed to infrastructure run by the hackers. This allows the hackers to redirect victims to spoof websites under their control, then steal passwords and tokens that let the hackers log in to that victim's online accounts without needing their two-factor authentication codes. Black Lotus Labs said that Fancy Bear compromised at least 18,000 victims in around 120 countries, including government departments, law enforcement agencies, and email providers across North Africa, Central America, and Southeast Asia. Microsoft, which also released details of the campaign on Tuesday, said in a blog post that its researchers identified over 200 organizations and 5,000 consumer devices affected by these hacking operations, including at least three government organizations in Africa. The Justice Department said Tuesday it neutralized compromised routers in the U.S. under court authorization. As the DOJ put it, the FBI "developed a series of commands to send to compromised routers" to collect evidence, reset settings, and prevent hackers from breaking back in.

Read more of this story at Slashdot.

Apple may have a supply problem on its hands with the MacBook Neo... The laptop reportedly relies on "binned" A18 Pro chips with one GPU core disabled, and demand is so strong that the supply of those cheaper leftover chips could run out before the next model is ready. That leaves Apple choosing between lower margins, shifting production plans, or changing the lineup to keep its $599 hit product in stock. MacRumors reports: The all-new MacBook Neo has been such a hit that Apple is facing a "massive dilemma," according to Taiwan-based tech columnist and former Bloomberg reporter Tim Culpan. [...] In the latest edition of his Culpium newsletter today, Culpan said the MacBook Neo is selling so well that Apple's supply of the binned A18 Pro chips with a 5-core GPU will "run out" before the company is able to fully satisfy demand for the laptop. Apple's initial plan was to have suppliers build around five to six million MacBook Neo units before ceasing production of the model with the A18 Pro chip, he said, but it sounds like demand is so strong that Apple might run out of A18 Pro chips to put in the MacBook Neo before the second-generation MacBook Neo with an A19 Pro chip is ready next year. Apple is unlikely to mark the MacBook Neo as temporarily sold out, so it may be forced to take action, but profit margins might be affected. A18 Pro chips are manufactured with TSMC's second-generation 3nm process, known as N3E, and Culpan said TSMC's N3E production lines are currently operating at maximum capacity. As a result, he said that Apple may have to pay a premium to restart A18 Pro chip production for the MacBook Neo, which would lower its profit margins. Apple would have to disable a GPU core on these chips to ensure that they have only a 5-core GPU, like all other MacBook Neo units sold to date. Alternatively, Culpan said that Apple could reallocate some of its chip production that was originally planned for other devices, but he said the cost would still be higher than what it paid for its initial batch of A18 Pro chips. Culpan speculated that Apple could also opt to discontinue the $599 model with 256GB of storage, leaving the $699 model with 512GB of storage and a Touch ID button as the only configuration available. This is unlikely to happen any time soon, in our view, given how heavily Apple has been promoting the MacBook Neo's affordability. Apple might also be able to move up the release of a MacBook Neo with the iPhone 17 Pro's A19 Pro chip, but that too would be a costlier option, at least until the company achieves a sufficient stockpile of binned A19 Pro chips with a 5-core GPU. In any case, Apple could opt to keep the starting price of current and future MacBook Neo models at $599 and simply accept lower profit margins on the laptop, especially given that it attracts customers to the macOS and broader Apple ecosystem.

Read more of this story at Slashdot.

"Anthropic has unveiled Claude Mythos, a new AI model capable of discovering critical vulnerabilities at scale," writes Slashdot reader wiredmikey. "It's already powering Project Glasswing, a joint effort with major tech firms to secure critical software. But the same capabilities could also accelerate offensive cyber operations." SecurityWeek reports: Mythos is not an incremental improvement but a step change in performance over Anthropic's current range of frontier models: Haiku (smallest), Sonnet (middle ground), and Opus (most powerful). Mythos sits in a fourth tier named Copybara, and Anthropic describes it as superior to any other existing AI frontier model. It incorporates the current trend in the use of AI: the modern use of agentic AI. "The powerful cyber capabilities of Claude Mythos Preview are a result of its strong agentic coding and reasoning skills... the model has the highest scores of any model yet developed on a variety of software coding tasks," notes Anthropic in a blog titled Project Glasswing -- Securing critical software for the AI era. In the last few weeks, Mythos Preview has identified thousands of zero-day vulnerabilities with many classified as critical. Several are ten or 20 years old -- the oldest found so far is a 27-years old bug in OpenBSD. Elsewhere, a 16-years old vulnerability found in video software has survived five million hits from other automated testing tools without ever being discovered. And it autonomously found and chained together several in the Linux kernel allowing an attacker to escalate from ordinary user access to complete control of the machine. [...] Anthropic is concerned that Mythos' capabilities could unleash cyberattacks too fast and too sophisticated for defenders to block. It hopes that Mythos can be used to improve cybersecurity generally before malicious actors can get access to it. To this end, the firm has announced the next stage of this preparation as Project Glasswing, powered by Mythos Preview. Given the rate of AI progress, it will not be long before such capabilities proliferate, potentially beyond actors who are committed to deploying them safely. "Project Glasswing is a starting point. No one organization can solve these cybersecurity problems alone: frontier AI developers, other software companies, security researchers, open-source maintainers, and governments across the world all have essential roles to play." Claude Mythos Preview is described as a general-purpose, unreleased frontier model from Anthropic that has nevertheless completed its training phase. The firm does not plan to make Mythos Preview generally available. The implication is that 'Preview' is a term used solely to describe the current state of Mythos and the market's readiness to receive it, and will be dropped when the firm gets closer to general release.

Read more of this story at Slashdot.

Over on his excellent blog, Matt Keeter posts some results from having ported a bytecode virtual machine to tail-calling style. He finds that his tail-calling interpreter written in Rust beats his switch-based interpreter, and even beats hand-coded assembly on some platforms.

He also compares tail-calling versus switch-based interpreters on WebAssembly, and concludes that performance of tail-calling interpreters in Wasm is terrible:

1.2× slower on Firefox, 3.7× slower on Chrome, and 4.6× slower in wasmtime. I guess patterns which generate good assembly don't map well to the WASM stack machine, and the JITs aren't smart enough to lower it to optimal machine code.

In this article, I would like to argue the opposite: patterns that generate good assembly map just fine to the Wasm stack machine, and the underperformance of V8, SpiderMonkey, and Wasmtime is an accident.

some numbers

I re-ran Matt’s experiment locally on my x86-64 machine (AMD Ryzen Threadripper PRO 5955WX). I tested three toolchains:

• Compiled natively via cargo / rustc

• Compiled to WebAssembly, then run with Wasmtime

• Compiled to WebAssembly, then run with Wastrel

For each of these toolchains, I tested Raven as implemented in Rust in both “switch-based” and “tail-calling” modes. Additionally, Matt has a Raven implementation written directly in assembly; I test this as well, for the native toolchain. All results use nightly/git toolchains from 7 April 2026.

My results confirm Matt’s for the native and wasmtime toolchains, but wastrel puts them in context:

We can read this chart from left to right: a switch-based interpreter written in Rust is 1.5× slower than a tail-calling interpreter, and the tail-calling interpreter just about reaches the speed of hand-written assembler. (Testing on AArch64, Matt even sees the tail-calling interpreter beating his hand-written assembler.)

Then moving to WebAssembly run using Wasmtime, we see that Wasmtime takes 4.3× as much time to run the switch-based interpreter, compared to the fastest run from the hand-written assembler, and worse, actually shows 6.5× overhead for the tail-calling interpreter. Hence Matt’s conclusions: there must be something wrong with WebAssembly.

But if we compare to Wastrel, we see a different story: Wastrel runs the basic interpreter with 2.4× overhead, and the tail-calling interpreter improves on this marginally with a 2.3x overhead. Now, granted, two-point-whatever-x is not one; Matt’s Raven VM still runs slower in Wasm than when compiled natively. Still, a tail-calling interpreter is inherently a pretty good idea.

where does the time go

When I think about it, there’s no reason that the switch-based interpreter should be slower when compiled via Wastrel than when compiled via rustc. Memory accesses via Wasm should actually be cheaper due to 32-bit pointers, and all the rest of it should be pretty much the same. I looked at the assembly that Wastrel produces and I see most of the patterns that I would expect.

I do see, however, that Wastrel repeatedly reloads a struct memory value, containing the address (and size) of main memory. I need to figure out a way to keep this value in registers. I don’t know what’s up with the other Wasm implementations here; for Wastrel, I get 98% of time spent in the single interpreter function, and surely this is bread-and-butter for an optimizing compiler such as Cranelift. I tried pre-compilation in Wasmtime but it didn’t help. It could be that there is a different Wasmtime configuration that allows for higher performance.

Things are more nuanced for the tail-calling VM. When compiling natively, Matt is careful to use a preserve_none calling convention for the opcode-implementing functions, which allows LLVM to allocate more registers to function parameters; this is just as well, as it seems that his opcodes have around 9 parameters. Wastrel currently uses GCC’s default calling convention, which only has 6 registers for non-floating-point arguments on x86-64, leaving three values to be passed via global variables (described here); this obviously will be slower than the native build. Perhaps Wastrel should add the equivalent annotation to tail-calling functions.

On the one hand, Cranelift (and V8) are a bit more constrained than Wastrel by their function-at-a-time compilation model that privileges latency over throughput; and as they allow Wasm modules to be instantiated at run-time, functions are effectively closures, in which the “instance” is an additional hidden dynamic parameter. On the other hand, these compilers get to choose an ABI; last I looked into it, SpiderMonkey used the equivalent of preserve_none, which would allow it to allocate more registers to function parameters. But it doesn’t: you only get 6 register arguments on x86-64, and only 8 on AArch64. Something to fix, perhaps, in the Wasm engines, but also something to keep in mind when making tail-calling virtual machines: there are only so many registers available for VM state.

the value of time

Well friends, you know us compiler types: we walk a line between collegial and catty. In that regard, I won’t deny that I was delighted when I saw the Wastrel numbers coming in better than Wasmtime! Of course, most of the credit goes to GCC; Wastrel is a relatively small wrapper on top.

But my message is not about the relative worth of different Wasm implementations. Rather, it is that performance oracles are a public good: a fast implementation of a particular algorithm is of use to everyone who uses that algorithm, whether they use that implementation or not.

This happens in two ways. Firstly, faster implementations advance the state of the art, and through competition-driven convergence will in time result in better performance for all implementations. Someone in Google will see these benchmarks, turn them into an OKR, and golf their way to a faster web and also hopefully a bonus.

Secondly, there is a dialectic between the state of the art and our collective imagination of what is possible, and advancing one will eventually ratchet the other forward. We can forgive the conclusion that “patterns which generate good assembly don’t map well to the WASM stack machine” as long as Wasm implementations fall short; but having shown that good performance is possible, our toolkit of applicable patterns in source languages also expands to new horizons.

Well, that is all for today. Until next time, happy hacking!

In an earlier blog post we found out that Pystd's simple sorting algorithm implementations were 5-10% slower than their stdlibc++ counterparts. The obvious follow up nerd snipe is to ask "can we make the Pystd implementation faster than stdlibc++?"

For all tests below the data set used was 10 million consecutive 64 bit integers shuffled in a random order. The order was the same for all algorithms.

Stable sort

It turns out that the answer for stable sorting is "yes, surprisingly easily". I made a few obvious tweaks (whose details I don't even remember any more) and got the runtime down to 0.86 seconds. This is approximately 5% faster than std::stable_sort. Done. Onwards to unstable sort.

Unstable sort

This one was not, as they say, a picnic. I suspect that stdlib developers have spent more time optimizing std::sort than std::stable_sort simply because it is used a lot more.

After all the improvements I could think of were done, Pystd's implementation was consistently 5-10% percent slower. At this point I started cheating and examined how stdlibc++'s implementation worked to see if there are any optimization ideas to steal. Indeed there were, but they did not help.

Pystd's insertion sort moves elements by pairwise swaps. Stdlibc++ does it by moving the last item to a temporary, shifting the array elements onwards and then moving the stored item to its final location. I implemented that. It made things slower.

Stdlibc++'s moves use memmove instead of copying (at least according to code comments).  I implemented that. It made things slower.

Then I implemented shell sort to see if it made things faster. It didn't. It made them a lot slower. So did radix sort.

Then I reworked the way pivot selection is done and realized that if you do it in a specific way, some elements move to their correct partitions as a side effect of median selection. I implemented that and it did not make things faster. It did not make them slower, either, but the end result should be more resistant against bad pivot selection so I left it in.

At some point the implementation grew a bug which only appeared with very large data sets. For debugging purposes I reduce the limit where introsort switches from qsort to insertion sort from 16 to 8. I got the bug fixed but the change made sorting a lot slower. As it should.

But this raises a question, namely would increasing the limit from 16 to 32 make things faster? It turns out that it did. A lot. Out of all perf improvements I implemented, this was the one that yielded the biggest improvement by a fairly wide margin. Going to 64 elements made it even faster, but that made other algorithms using insertion sort slower, so 32 it is. For now at least.

After a few final tweaks I managed to finally beat stdlibc++. By how much you ask? Pystd's best observed time was 0.754 seconds while stdlibc++'s was 0.755 seconds. And it happened only once. But that's enough for me.