csjh (@csjh__) header
csjh (@csjh__) profile picture

university student who likes coding also check out https://teapaw.org/t.co/GAPzR23dHe former shopifella, current cloudflaridian

canada
Joined March 2024
Photo by csjh (@csjh__)Photo by csjh (@csjh__)Photo by csjh (@csjh__)Photo by csjh (@csjh__)Photo by csjh (@csjh__)Photo by csjh (@csjh__)Photo by csjh (@csjh__)Photo by csjh (@csjh__)Photo by csjh (@csjh__)Photo by csjh (@csjh__)
frontier models need to be trained on more BULLSHIT that gives you a stank face when you see it i'll know we have AGI when i see a model that spits out code that works because of a nasty central snippet and i'm like oooooo you nasty little thing
1
7
494
finally finished the blog post I've been working on for the past few months goes over some fun bits I've picked out of the webassembly runtimes I've made over the last couple years was a lot of fun to make, hope you enjoy csjh.blog/posts/a-tale-of-th…
1
2
9
799
Here's a detailed write-up of how I optimized a super simple AST-walking dynamic language interpreter (for a language I invented). It got 16x faster with ~20 changes. This puts it in the neighborhood of mature interpreters like CPython and QuickJS. zef-lang.dev/implementation
2
25
1
232
17,772
the masculine urge to solve all your problems with a bigass mmap
5
888
i couldn't think of an interesting idea in my post-project completion period so of course now i am just adding proper tiering up to said project the mitey wasm jit will soon be a multitier jit whether or not these tiers will be good is a different question
1
2
719
i'm trying out an idea i had where instead of doing method based jit u let the methods "grow" themselves kinda like a tracing jit i think but without explicit traces but i'm no expert in tracing jits not sure how well it'll work but it'll fill my time
1
431
made a silly little script for evaluating the branch predictor on my laptop it lets you check how predictable some buffers of 0s and 1s by generating a copy of a branchy function and running through the buffer doubt it's perfect but kinda cool gist.github.com/csjh/2628d8b…
6
701
i need an x64 laptop so bad restricting myself to armland seems a bit silly
6
1
17
1,358
pypy3 is so dope, really unfortunate that they're struggling to get dev support i've been using it a lot to iterate on stuff recently and it gets me giddy seeing 3x speed improvements from literally 0 effort
2
200
one cool thing about making a wasm jit is you can profile full codebases in any way you want (as long as they can compile to wasi) over the last couple days i've looked at how branches behave relative to what stack history precedes them (i.e. how functionA's branches compare when called from functionB vs. when called from functionC), i've looked at how often memory accesses are predictable, i've looked at how often indirect branch addresses are predictable, etc etc etc so cool
12
665
openai final interview tmr :)) feels like i should be more nervous than i am
40
3
466
42,294
i love how many projects i see nowadays with interesting claims / features that proceed to have absolutely no substance when you look at the codebase or test it out for yourself
6
4,279
didn't realize fuzzing was so easy to setup ran it overnight on the JIT and got surprisingly few crashes; mostly just had to add some limits (# of functions etc (which i should've already had because these are universally agreed-upon limits))
1
3
3,813
it does raise some annoyances about the size of the PROT_EXEC page i make though, i should really make some reusable concept of a vector that allocates a shitload of virtual memory and just errors out when you irresponsibly reach that shitload
1
1,355
finally did full libsodium eval on various wasm runtimes wasmer pretty good, but interestingly on my "real world" benchmark (compiling a big js file with esbuild) there's a significant perf difference (even after removing initialization time, <200ms) (some runtimes not listed because they crashed)
1
3
1,392
was writing my long-delayed blog post about my JIT and realized as i was trying to describe an implementation detail that it was a very bad implementation detail and proceeded to lose a day changing it lol
1
2
577
literally stopped mid sentence and went on a few hours long tangent
2
510
looked into this more... the go compiler wasm binary exclusively uses 1 param / 1 result function signatures, and never uses block result values! kinda wild
kinda interesting how different languages show themselves in wasm i'm testing my jit for the first time on golang binaries and they have an order of magnitude more br_tables than my other llvm-compiled test binaries like figma/spidermonkey
1
2
571
interesting from an optimizing (my) compiler performance perspective as well, things that are a hotpath in some binaries can be fully cold in others
1
285