


Interpreters, compilation, concurrency, and debugging tools.


The animation of the GC that I posted the other day revealed that I wasn't giving young objects time to become garbage. That version immediately promoted objects to old space, which causes it to grow even while no long-lived objects are being created. Fixed now.
Here's a visualization of the generational sliding GC of my modified SOM++ on a 16 bit computer. It just takes the whole of the free memory for the young gen. When the old space has grown too much, a full sliding GC collects the garbage and compacts down to the size of live data

Second installment in our interpreter optimization trilogy!
Chasing the perfect assembly: matching the speed of hand-written assembly with an interpreter written in Java: medium.com/graalvm/chasing-t… by @YudiZheng

Endive 1.2.0 is out 🚀, with a lot of community contributions: SIMD by default in the interpreter, major redline advancements with impressive numbers, faster memory for the Threads proposal, and a Component Model that's starting to be usable. Blog: endive.run/blog/endive-1.2.0

After the little setback, did some work on stack allocating instead of heap allocating frames <2M!. And some improvements to gcc to handle far branches better. Love being able to fix things in any layer: The compiler, the language runtime, gcc, the CPU design or even in the ISA.
Oh no I ran out of space on the ice40up5k. It's a fairly small FPGA. A tiny computer (even a virtual one) isn't much fun without USB and video output. Had to remove the BTB again. Still, the other progress has been pretty good.

Oh no I ran out of space on the ice40up5k. It's a fairly small FPGA. A tiny computer (even a virtual one) isn't much fun without USB and video output. Had to remove the BTB again. Still, the other progress has been pretty good.
Once I lengthened the pipeline I started wanting some branch prediction (of conditional branches) and some branch target prediction (for virtual C++ calls and interpreter loops). I actually managed to fit both in without breaking the LUT limit on this small FPGA.

Once I lengthened the pipeline I started wanting some branch prediction (of conditional branches) and some branch target prediction (for virtual C++ calls and interpreter loops). I actually managed to fit both in without breaking the LUT limit on this small FPGA.
Is it cheating to lengthen the pipeline of the CPU and make loads and stores faster, just so your Smalltalk VM goes faster? No, it isn't!

My Smalltalk (SOMpp fork) on Fructus is getting faster. < 4m cycles for my cut down DeltaBlue benchmark for the first time!
Fructus has a nice instruction set reference now!

Ahead-of-time Compilation of JavaScript to WebAssembly via Static Type Inference vimeo.com/1234257957 cfallin.org/pubs/vmil2026_ni… Chris Fallin (mastodon.world/@cfallin/1174…) @VMIL_Workshop 2026 (at @splashcon 2026)

I've been playing with making my own 16 bit CPU with its own instruction set. I call it Fructus because all computers were fruit-named when I was young. Right now I'm implementing integer division, and I'm doing it with an instruction I haven't seen as the primary insn.


I abused my position of power today to teach students that dynamically typed languages don't generally do type checks at run time.

My team is hiring! We develop the Verse programming language and related tech at Epic. See thread for more details about what we work on and how you can apply. I'll also be at REBASE this week in Oakland if you want to come have a chat in person.

There was a time where colleagues suggested that only code has value. I am glad, we are back at a point where explaining things is valuable again.
Ever heard of dynamic meta-compilation? Truffle uses partial evaluation to turn interpreters into optimized machine code without requiring you to write a compiler. We’ve added docs explaining how it works. Readable by humans, too: github.com/oracle/graal/blob…

Ever heard of dynamic meta-compilation? Truffle uses partial evaluation to turn interpreters into optimized machine code without requiring you to write a compiler. We’ve added docs explaining how it works. Readable by humans, too: github.com/oracle/graal/blob…

Brief weekend reading on another benchmarking pitfall: The Limitations of Running a Workload Generator In the Same JVM as the System-Under-Test norlinder.nu/posts/The-Limit… Spoiler: the GC may conspire against you and your data collection!