Optimize SPSC queue

[polytypic]

This PR optimizes the SPSC queue.

See the individual commits to better understand the motivation behind the optimizations.

Here is a run of the benchmarks on my M3 Max before the optimizations:

➜  saturn git:(rewrite-bench) ✗ dune exec --release -- ./bench/main.exe -budget 1 Single | jq '[.results.[].metrics.[] | select(.name | test("over")) | {name, value}]'
[                                    
  {
    "name": "messages over time/2 workers, capacity 1",
    "value": 6.500299359098922
  },
  {
    "name": "messages over time/2 workers, capacity 8",
    "value": 12.632201914378621
  },
  {
    "name": "messages over time/2 workers, capacity 64",
    "value": 14.90261050891111
  },
  {
    "name": "messages over time/2 workers, capacity 512",
    "value": 14.960086955505428
  },
  {
    "name": "messages over time/2 workers, capacity 4096",
    "value": 20.06888141824778
  },
  {
    "name": "messages over time/2 workers, capacity 32768",
    "value": 22.94386410808166
  }
]

Here is a run of the benchmarks after the optimizations:

➜  saturn git:(optimize-spsc-queue) ✗ dune exec --release -- ./bench/main.exe -budget 1 Single | jq '[.results.[].metrics.[] | select(.name | test("over")) | {name, value}]'
[                                    
  {
    "name": "messages over time/2 workers, capacity 1",
    "value": 6.868795322207784
  },
  {
    "name": "messages over time/2 workers, capacity 8",
    "value": 18.420713908583064
  },
  {
    "name": "messages over time/2 workers, capacity 64",
    "value": 30.26195314050038
  },
  {
    "name": "messages over time/2 workers, capacity 512",
    "value": 38.7165046104334
  },
  {
    "name": "messages over time/2 workers, capacity 4096",
    "value": 123.22902197665648
  },
  {
    "name": "messages over time/2 workers, capacity 32768",
    "value": 270.82312457121435
  }
]

TODO:

• Add comments on the use/lack of fences in the code.

[lyrm]

I am unsure the commit Avoid float array pessimization and a bit of the previous one (removing the indirection of the array and thus doing initialization with Obj.magic) are a good idea. They introduce a lot of Obj.magic which I feel we should be careful with, especially because it could cause issues with flambda.

Also, what happen if you are actually putting float in the queue ? I am guessing the array float special case will not be triggered ?

I have no immediate solution for float array pessimization, but for initialization, we could have a make : size_exponent:int -> 'a -> 'a function instead of the create function that requires the user to provide an inhabitant for type 'a.

It may just move the issue to the user lever that may want to either use the same trick with Obj.magic or use and option type though...

[lyrm]

Thanks for the PR ! All the optimizations (except maybe the ones mentioned in my previous message :p) are great and thanks again for making the git history so readable. And obviously, the gain in performance is amazing !

[lyrm]

I should have added that the optimization with the cashes is very cool !!

[lyrm]

The changes proposed in this PR have already been merged in two different PRs (#133 and #135) that separate "safe" and "unsafe" optimizations. Thanks @polytypic for all this work!