Implement Control.Concurrent.Counter.Unlifted in CMM

[Bodigrim]

Just an experiment to avoid any overhead for MutableByteArray inspired by @TerrorJack's https://gist.github.com/TerrorJack/4a48ed790155cc79619fffe9f5844521.

Caveats:

• Does not work in GHC JavaScript backend, which should probably retain the previous, pure Haskell implementation.

• Needs more care on 32-bit machines, current code is 64 bit only.

• ccall hs_atomic_add64 is suboptimal on x86_64, but at the moment hand-written CMM cannot access call MO_AtomicRMW W64 AMO_Add, see https://gitlab.haskell.org/ghc/ghc/-/issues/23206.

[Bodigrim]

Hmm, tests are failing on GHC 9.4.4, macOS x86_64:

 All
  Correctness:            OK (9.99s)
    +++ OK, passed 500 tests.
  Correctness, no delays: FAIL (0.27s)
    *** Failed! Falsified (after 3479 tests):
    Threads {unThreads = Thread {tDelay = Delay {unDelay = 6}, tIncrement = -950, tIterations = Iterations {unIterations = 7}} :| [Thread {tDelay = Delay {unDelay = 8}, tIncrement = -441, tIterations = Iterations {unIterations = 37}},Thread {tDelay = Delay {unDelay = 6}, tIncrement = -452, tIterations = Iterations {unIterations = 0}},Thread {tDelay = Delay {unDelay = 4}, tIncrement = -389, tIterations = Iterations {unIterations = 49}},Thread {tDelay = Delay {unDelay = 0}, tIncrement = -294, tIterations = Iterations {unIterations = 40}}]}
    -110710340 /= -53788
    Use --quickcheck-replay=512982 to reproduce.
    Use -p '/Correctness, no delays/' to rerun this test only.

Whatever the sequence of updates, the result must have been below five digits, so -110710340 suggests memory corruption.

@TerrorJack, any idea where I messed up?

[alt-romes]

Great! I was just thinking about this too. Thanks @Bodigrim

[TerrorJack]

Pong, will take a closer look next week.

[alt-romes]

I'm also interested in seeing this work, don't mind my second ping @TerrorJack

[TerrorJack]

Would you give https://gist.github.com/TerrorJack/e3f1231fa65b3d1f5ea63065dd33a87f a try and see if it fixes the issue? Honestly I'm no less puzzled than you why the original version has this memory issue, but this diff seems to help.

[Bodigrim]

Thanks @TerrorJack, this seems to work.

I think this is ready for review now.

[alt-romes]

FWIW LGTM

[Bodigrim]

Benchmarks:

Name	07f4b51	3dd86a3	Ratio
10 iterations and 1 threads	626901	650901	1.03828
100 iterations and 1 threads	1057955	1065232	1.00688
1000 iterations and 1 threads	4767930	4846413	1.01646
10000 iterations and 1 threads	40603875	40717926	1.00281
10 iterations and 2 threads	6199483	6215062	1.00251
100 iterations and 2 threads	6786286	6637318	0.978049
1000 iterations and 2 threads	36046876	37277081	1.03413
10000 iterations and 2 threads	466461621	459297509	0.984642
10 iterations and 4 threads	14304393	14116542	0.986868
100 iterations and 4 threads	22378799	21837321	0.975804
1000 iterations and 4 threads	241044897	221475244	0.918813
10000 iterations and 4 threads	2411656445	2126885546	0.881919
10 iterations and 6 threads	29380970	31528617	1.0731
100 iterations and 6 threads	51806632	49229728	0.950259
1000 iterations and 6 threads	421304931	389744873	0.92509
10000 iterations and 6 threads	10803454687	10544148437	0.975998
10 iterations and 8 threads	33264868	33220318	0.998661
100 iterations and 8 threads	70489263	68710925	0.974772
1000 iterations and 8 threads	751835473	614695391	0.817593
10000 iterations and 8 threads	18112116015	17322731250	0.956417
10 iterations and 12 threads	37138810	37928328	1.02126
100 iterations and 12 threads	111093969	99659307	0.897072
1000 iterations and 12 threads	2238697705	1766967041	0.789283
10000 iterations and 12 threads	24245685546	23640137500	0.975025
10 iterations and 16 threads	43222968	43582453	1.00832
100 iterations and 16 threads	145973196	137985633	0.945281
1000 iterations and 16 threads	3376071093	2974715039	0.881117
10000 iterations and 16 threads	35231724609	34480696093	0.978683
10 iterations and 20 threads	62185274	62320446	1.00217
100 iterations and 20 threads	183011389	173122430	0.945965
1000 iterations and 20 threads	4215415478	3965037304	0.940604
10000 iterations and 20 threads	42896585937	42560037500	0.992154
10 iterations and 32 threads	97156437	99794003	1.02715
100 iterations and 32 threads	260610671	260671594	1.00023
1000 iterations and 32 threads	7016384375	6485634765	0.924356
10000 iterations and 32 threads	73145804687	69175517187	0.945721
10 iterations and 64 threads	172624707	178914562	1.03644
100 iterations and 64 threads	610634515	591822558	0.969193
1000 iterations and 64 threads	14407456250	13955856250	0.968655
10000 iterations and 64 threads	151388475000	148245268750	0.979237
10 iterations and 128 threads	294979428	300389105	1.01834
100 iterations and 128 threads	1733554345	1293563940	0.746192
1000 iterations and 128 threads	29561675000	27645350000	0.935175
10000 iterations and 128 threads	303322475000	296239175000	0.976648
Geometric mean			0.96143

Interestingly the biggest speed up, up to 25%, is observed for 1000 iterations, while measurements for 10000 iterations are consistently better, but less impressive, up to 11%.

[sergv]

Impressive, thanks a lot!

[sergv]

NB for comparison of CMM against mutable prim array one can compare against Addr row in the benchmarks - it was on par with singleton array previously.

[sergv]

Released as 0.1.1.