Use quick_stat with full_major instead of stat (which will be deprecated)

[dinosaure]

See this PR on ocaml/ocaml for more explanation:
ocaml/ocaml#12023

/cc @damiendoligez (to confirm if it's the right usage) and /cc @Julow if it does not disturb your metrics.

[Julow]

I think this doesn't give the right result.
Here's a repro:

I get 1844w/run (varying in each runs) with this PR, 3000w/run (always) with minor_allocated_precise (#54) and 1658w/run (varying too) with the released version. I think 3000 is the right number ?

open Bechamel
open Toolkit

let minor_allocated_precise =
  let module Ext = struct
    type witness = unit

    let load () = ()
    let unload () = ()
    let make () = ()
    let get () = Gc.minor_words ()
    let label () = "minor-allocated"
    let unit () = "word"
  end in
  let ext = Measure.register (module Ext) in
  Measure.instance (module Ext) ext

let input = List.init 1000 (fun i -> i)

let bench =
  Test.make ~name:"map"
    (Staged.stage @@ fun () -> ignore (List.map (fun x -> x + 1) input))

(* let instances = [ Instance.minor_allocated; minor_allocated_precise ] *)
let instances =
  if false then [ Instance.minor_allocated ] else [ minor_allocated_precise ]

let () =
  let cfg = Benchmark.cfg ~limit:2000 ~quota:(Time.second 0.1) () in
  let raw = Benchmark.all cfg instances bench in
  let ols =
    Analyze.ols ~bootstrap:0 ~r_square:true ~predictors:[| Measure.run |]
  in
  List.iter
    (fun inst ->
      Hashtbl.iter
        (fun name ols ->
          Format.printf "%s %s %a@." (Measure.label inst) name Analyze.OLS.pp
            ols)
        (Analyze.all ols inst raw))
    instances

[dinosaure]

Hmmhmm, so to be clear, this PR and what you propose on #54 is about two different questions:

• for this PR, the question is more how we can stabilize the major heap before to run a benchmark
• what you proposed on Minor_allocated: Use Gc.minor_words #54 is indeed interesting and the solution I would take to get metrics

So the question is much more: if this PR and #54 together produce a good result or not (but it's not like choose one or the other). From your code, I can reproduce this with the merge of #54 and this PR:

$ ./a.out
minor-allocated map { minor-allocated per run = 3000.000866; r² = Some 1 }

With bechamel.0.5.0, I can also have the same result, not sure how to reproduce the bad case 😕. I get also the same result with the upstream version.

$ ./a.out
minor-allocated map { minor-allocated per run = 3000.000946; r² = Some 1 }

With only #54, I have this result:

$ ./a.out
minor-allocated map { minor-allocated per run = 3000.001034; r² = Some 1 }

[dinosaure]

And, btw, it will be nice to also look into #55 🥲

[Julow]

I confirm the bad behavior with bechamel 0.5.0 and ocaml 5.3. I was expecting that the current PR fixes the same bug as #54 because both use Gc.quick_stat. Maybe that's explained by my ocaml version ?