@vouillon, I looked at updating the runtime but I'm not familiar enough with this part.

I've finally managed to find my way out int the runtime.
I've tried to change the current design as little as possible but It might not be the best approach.

@vouillon, I've reworked the implementation of effects in the js runtime so that it looks more like what's done upstream in the c runtime. Can you take a look ?

The deep_state benchmark (included in this PR in compiler/tests-jsoo/lib-effects/deep_state.ml) runs significantly faster with this PR

$ hyperfine -L v before,after --warmup 3 'node {v}.js 64 500_000'
Benchmark 1: node before.js 64 500_000
  Time (mean ± σ):      2.745 s ±  0.046 s    [User: 2.779 s, System: 0.036 s]
  Range (min … max):    2.671 s …  2.822 s    10 runs
 
Benchmark 2: node after.js 64 500_000
  Time (mean ± σ):      1.724 s ±  0.044 s    [User: 1.761 s, System: 0.033 s]
  Range (min … max):    1.639 s …  1.771 s    10 runs
 
Summary
  node after.js 64 500_000 ran
    1.59 ± 0.05 times faster than node before.js 64 500_000
$ hyperfine -L v before,after --warmup 3 'node {v}.js 128 500_000'
Benchmark 1: node before.js 128 500_000
  Time (mean ± σ):      5.348 s ±  0.096 s    [User: 5.386 s, System: 0.041 s]
  Range (min … max):    5.225 s …  5.473 s    10 runs
 
Benchmark 2: node after.js 128 500_000
  Time (mean ± σ):      3.636 s ±  0.247 s    [User: 3.691 s, System: 0.042 s]
  Range (min … max):    2.986 s …  3.845 s    10 runs
 
Summary
  node after.js 128 500_000 ran
    1.47 ± 0.10 times faster than node before.js 128 500_000

The latest commit (that implement perform and reperform separately) is even faster.

$ hyperfine -L v before,next --warmup 3 'node {v}.js 128 500_000'
Benchmark 1: node before.js 128 500_000
  Time (mean ± σ):      5.249 s ±  0.128 s    [User: 5.290 s, System: 0.045 s]
  Range (min … max):    5.117 s …  5.468 s    10 runs
 
Benchmark 2: node next.js 128 500_000
  Time (mean ± σ):      2.383 s ±  0.043 s    [User: 2.448 s, System: 0.040 s]
  Range (min … max):    2.286 s …  2.437 s    10 runs
 
Summary
  node next.js 128 500_000 ran
    2.20 ± 0.07 times faster than node before.js 128 500_000

@vouillon, thanks for the review. I've applied you suggestion and removed the caml_fiber_stack to only keep caml_current_stack.

@vouillon, should we merge in the current state or should we wait for the wasm runtime?

I would really prefer that #1461 get merged first, rather than asking @OlivierNicole to rebase it once more.
But I don't think we need to wait for the wasm runtime.