Some lovely, old-Internet notes about interesting bit hackery you can use for all kinds of purposes. I really like these for implementing “scalar SIMD” (i.e. treating 64-bit integers as 8x8-bit integer vectors) optimizations.

An interesting variant of Solarized that feels more uniform in places. The idea of custom background colors, and tweaking things following OKLab, is quite interesting.

A fun read describing an important (and seemingly inherent) difficulty of compiler optimization. E-graphs get around this, but IIRC are significantly slower.

A really interesting discussion of array-oriented compilation architectures. Tries to answer the same sorts of questions I’ve been asking myself about making compilers faster – although I think I come to different conclusions right now.

For me, fills a hole between APL (high-performance CPU/GPU array manipulation) and Rust (strong type checking). Super interesting stuff.

Interesting concept, and they seem to have a good number of integrations already. It was a bit hard to find the actual documentation for how it works, but it otherwise looks cool.

Amazing dissertation by Aaron Wen-yao Hsu, demonstrating a novel memory layout for ASTs and advocating for bottom-up traversal patterns. The fact that the compiler is just 17 lines of APL (which I can’t read in the slightest) is even cooler.

I’ve been looking for a good replacement to SSA form for mid-level and low-level optimisations. This feels like the right direction forward, but I need to try it myself before I’m convinced.

Discussion of super-optimization based on equality graphs and finding equivalent expressions. Contrast with https://egraphs-good.github.io/.

Interesting study of vectorizing the tokenization of a complex language.

Lovely visual summaries of how prefetching works on modern Intel chips. I would guess that older chips follow similar patterns. Great for designing data structures and optimizing memory traversal.

I’ve thought about vectorized sorting for a while, but I didn’t know that bitonic sorts can be implemented efficiently on SIMD registers. This is an interesting approach.