[anematode]

Nice post :)

Last year I was working on a tail-call interpreter (https://github.com/anematode/b-jvm/blob/main/vm/interpreter2...) and found a similar regression on WASM when transforming it from a switch-dispatch loop to tail calls. SpiderMonkey did the best with almost no regression, while V8 and JSC totally crapped out – same finding as the blog post. Because I was targeting both native and WASM I wrote a convoluted macro system that would do a switch-dispatch on WASM and tail calls on native.

Ultimately, because V8's register allocation couldn't handle the switch-loop and was spilling everything, I basically manually outlined all the bytecodes whose implementations were too bloated. But V8 would still inline those implementations and shoot itself in the foot, so I wrote a wasm-opt pass to indirect them through a __funcref table, which prevented inlining.

One trick, to get a little more perf out of the WASM tail-call version, is to use a typed __funcref table. This was really horrible to set up and I actually had to write a wasm-opt pass for this, but basically, if you just naively do a tail call of a "function pointer" (which in WASM is usually an index into some global table), the VM has to check for the validity of the pointer as well as a matching signature. With a __funcref table you can guarantee that the function is valid, avoiding all these annoying checks.

[functional_dev]

The article shows WASM being 1.2-3.7x slower, and your experience confirms it.

Do you have any idea which operations regress the most?

[anematode]

Based on looking at V8's JITed code, there seemed to be a lot of overhead with stack overflow checking, actually. The function prologues and epilogues were just as bloated in the tail-call case. I'll upload some screenshots if I can find them.