Weak test revisions

[jmid]

This PR revises our Weak tests...

The Weak module depends on the state of the GC.
As such, we have to be extra careful to start from "as clean a starting point as possible".
Just having run sequential tests pollutes the heap state - so better run a Gc.full_major () initially - and a cheaper Gc.minor () in between attempts.

Secondly, the PR switches away from strings to more easy-to-shrink int64s.

Finally I think running Lin tests of Weak is misguided, as I don't think it makes sense to perform a fragile GC-dependent parallel run and compare its outputs against more fragile GC-dependent sequential runs.
Parallel Weak STM tests however work well as stress tests - we previously caught run-time bugs that way (e.g., #181)

Hopefully this will solve #299.

I can see a 5.2 s390x run even completed within the timelimit with this: 🎉
https://ocaml-multicoretests.ci.dev:8100/job/2023-06-09/150708-ci-ocluster-build-44e4ee

[shym]

This looks very good! 😄

The only thing I would have done differently is the support for 5.0. I’ve used this in the past: shym/ortac@cecf0c5 which has the nice side-effect that it doesn’t override the standard module when not needed. But I’m not sure this makes such a difference.

[jmid]

Good point about overriding 👍
Actually, now that you mention it, I realize by reordering the include I would use the 5.1 and 5.2 bindings when they are available. I've previously used that trick in QCheck but forgotten about it: 😅
https://github.com/c-cube/qcheck/blob/9424d94c8fa682ef90ad8bf2d487c852d8c4f38d/src/core/QCheck.ml#L12-L19

[shym]

Nice trick 👍

[jmid]

The shadowing is handled in 9b85211.
I've also gone for removing the Lin tests entirely as I don't see us needing them in the future.
We may have to disable the parallel src/weak/stm_tests_hashset test on Windows bytecode where it doesn't seem to consistently trigger.

[jmid]

CI summary for 720028c

• a failure to trigger a parallel weak/stm_tests_hashset on Windows bytecode 5.1.0~alpha2
• a timeout due to excessive Lin Ephemeron shrinking on Cygwin part 1 5.1.0~alpha2
• a timeout due to excessive Lin Out_channel shrinking on Windows bytecode trunk 5.2
• a timeout due to threadomain livelock on Windows 5.1.0~alpha2 and on Windows trunk 5.2 [ocaml5-issue] Windows failures on threadomain #203
• a timeout due to a on domain_spawntree with Atomic on Windows trunk 5.2 [ocaml5-issue] Windows trunk bytecode domain_spawntree crash or deadlock #354
• a timeout due to long-running int ref STM and Thread tests (several 1500+ secs) on Cygwin part 2 trunk 5.2
• a failure to trigger a Lin Out_channel test on Linux arm64 5.1 Reliability of parallel Lin Out_channel tests #360

Modulo the first one, none of this is related to this PR.

[shym]

Small suggestion for the last commit: rather than disabling the warning globally, it can be disabled only where it might trigger, it’s easier to understand why it is required.

@@ -41,6 +41,7 @@ struct
 
   module Int64 =
   struct
+    [@@@warning "-unused-value-declaration"]
     (* support Int64.hash added in 5.1 *)
     external seeded_hash_param :
       int -> int -> int -> 'a -> int = "caml_hash" [@@noalloc]

[jmid]

CI summary of latest run

• 2 Windows bytecode runs on 5.1.0~alpha2 and 5.2 trunk timed out due to threadomain [ocaml5-issue] Windows failures on threadomain #203
• 2 Cygwin part 1 runs (1 on 5.1.0~alpha2 and 1 trunk) were cancelled before running 🤔 potentially getting confused by the midway push...

[jmid]

Small suggestion for the last commit: [...]

Fair enough. No need to silence more than necessary. Should be addressed in f3d0160 👍

[jmid]

CI summary for the latest run:

• 1 Windows bytecode 5.1.0~alpha2 threadomain timeout and 1 crash - with a new variant [ocaml5-issue] Windows failures on threadomain #203
• 1 Windows bytecode trunk 5.2 unexpected Thread Queue counterexample Lin Thread Queue test reliability #363
• 2 timeouts on Cygwin part 1 5.1 and 1 Linux s390x 5.1 due to several (primary Lin) tests taking 1000+ seconds
• 1 macOS arm64 5.2 failure to trigger STM int ref test parallel asymmetric 'STM _ ref test parallel asymmetric' failure to trigger #364

I'll go ahead and merge.