Make resuming a continuation more efficient

[hhugo]

Fix #1658
See corresponding OCaml PR ocaml/ocaml#12735

• some comments needs to be updated.
• The wasm part has not been optimized

[hhugo]

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

[hhugo]

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.

[hhugo]

@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 ?

[hhugo]

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

[hhugo]

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

[hhugo]

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

[hhugo]

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

[vouillon]

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.

Copilot reviewed 4 out of 13 changed files in this pull request and generated no comments.

Files not reviewed (9)

• compiler/lib/effects.ml: Language not supported
• compiler/lib/parse_bytecode.ml: Language not supported
• compiler/tests-check-prim/main.4.14.output: Language not supported
• compiler/tests-check-prim/unix-Unix.4.14.output: Language not supported
• compiler/tests-check-prim/unix-Win32.4.14.output: Language not supported
• compiler/tests-jsoo/lib-effects/deep_state.ml: Language not supported
• compiler/tests-jsoo/lib-effects/dune: Language not supported
• runtime/wasm/effect.wat: Language not supported
• compiler/tests-full/stdlib.cma.expected.js: Evaluated as low risk

Comments suppressed due to low confidence (4)

runtime/js/effect.js:87

• The condition 'if (last === 0)' should be checked to ensure that 'last' is correctly initialized.

if (last === 0) {

runtime/js/effect.js:139

• [nitpick] The variable name 'last_fiber' is ambiguous. It should be renamed to something more descriptive, such as 'current_fiber'.

var last_fiber = caml_current_stack;

runtime/js/effect.js:132

• The new behavior introduced in the 'caml_perform_effect' function should be covered by tests.

function caml_perform_effect(eff, k0) {

runtime/js/effect.js:156

• The new behavior introduced in the 'caml_reperform_effect' function should be covered by tests.

function caml_reperform_effect(eff, cont, last, k0) {