Allocation in VS up 500%

Somewhere in between the commits fee253fc and 99c61e22 nofib’s VS benchmark has allocations increased by 500%. perf.haskell.org is currently defuct (linker errors due to recent changes to the ffi stuff, I believe Moritz is trying to find that out), so I cannot just look there to see which commit broke it. But maybe someone feels like bisecting this.

If not maybe perf.haskell.org will be able to tell us once its fixed (although keeping up with GHC commits there is a bit of a pain these days.)

Trac metadata

Trac field	Value
Version	8.3
Type	Bug
TypeOfFailure	OtherFailure
Priority	low
Resolution	Unresolved
Component	Compiler
Test case
Differential revisions
BlockedBy
Related
Blocking
CC
Operating system
Architecture

changed weight to 3

added Tbug Trac import labels

I can restrict the problem to the range 33452dfc..99c61e22 (only 340 commits left)

Where is nofib's VS benchmark? What's the latest released compiler with good perf? 8.2? 8.0?

It’s in real/eff/VS, just added this August, so no released compiler has it in the suite.

But I can of course test an old compiler with make NoFibRuns=1 mode=slow HC=ghc-8.0. I get

8.0       483117568
8.2       483117520
7ea5d41d 3123117584

added runtime perf label

added Plow label

I haven't been able to reproduce the increased allocations with any recent compiler release:

$ make  NoFibRuns=1 mode=slow HC=(which ghc-8.0.2)
...
<<ghc: 483117568 bytes, 767 GCs (759 + 8), 0/45018104 avg/max bytes residency (8 samples), 54096 bytes GC work, 83M in use, 0.000 INIT (0.000 elapsed), 0.168 MUT (0.168 elapsed), 0.060 GC (0.060 elapsed), 0.030 GC(0) (0.030 elapsed), 0.030 GC(1) (0.030 elapsed), 1 balance :ghc>>
$ make  NoFibRuns=1 mode=slow HC=(which ghc-8.2.2)
...
<<ghc: 483117520 bytes, 383 GCs (375 + 8), 0/55340048 avg/max bytes residency (8 samples), 41160 bytes GC work, 83M in use, 0.000 INIT (0.000 elapsed), 0.175 MUT (0.176 elapsed), 0.050 GC (0.051 elapsed), 0.050 GC(0) (0.050 elapsed), 0.001 GC(1) (0.001 elapsed), 1 balance :ghc>>
$ make  NoFibRuns=1 mode=slow HC=(which ghc-8.4.4)
...
<<ghc: 483117520 bytes, 383 GCs (375 + 8), 0/55340072 avg/max bytes residency (8 samples), 41448 bytes GC work, 83M in use, 0.000 INIT (0.000 elapsed), 0.175 MUT (0.175 elapsed), 0.055 GC (0.055 elapsed), 0.028 GC(0) (0.028 elapsed), 0.026 GC(1) (0.026 elapsed), 1 balance :ghc>>
$ make  NoFibRuns=1 mode=slow HC=(which ghc-9.0.1)
...
<<ghc: 483117328 bytes, 383 GCs (375 + 8), 0/55339848 avg/max bytes residency (8 samples), 39712 bytes GC work, 83M in use, 0.000 INIT (0.000 elapsed), 0.161 MUT (0.161 elapsed), 0.073 GC (0.073 elapsed), 0.040 GC(0) (0.040 elapsed), 0.033 GC(1) (0.033 elapsed), 1 balance :ghc>>

So I suspect this is fixed and can be closed.

Simon Jakobi says the regression has disappeared again, thus closing.

closed