Runs much slower on GHC-9.6

[leonschoorl]

With GHC-9.6.2 doctest-parallel to runs much slower then before :(

$ time cabal run -w ghc-9.6.2 -- clash-prelude:doctests
[...]
real	1m9,515s
user	1m42,364s
sys	0m8,968s

time cabal run -w ghc-9.4.5 -- clash-prelude:doctests
[...]
real	0m25,501s
user	0m57,237s
sys	0m6,458s

The slowdown seems to be mostly in the example discovery phase.
With 9.6 after starting the program there is a long pause, then it prints any warnings found in the source code, and then it starts testing all the examples.
This pause with no output is ~45s using 9.6.2, and only ~6s using 9.4.5.

[leonschoorl]

Some more data

GHC	8.10.7	9.0.2	9.4.5	9.6.2
real	0m16,230s	0m12,355s	0m25,501s	1m9,515s
user	1m10,319s	0m56,553s	0m57,237s	1m42,364s
sys	0m6,790s	0m5,586s	0m6,458s	0m8,968s

There is also a ~2x slowdown between 9.0 and 9.4

[christiaanb]

Is the executable statically or dynamically linked? If the former, does linking it dynamically instead make it faster?

[martijnbastiaan]

It seems to have been fixed again in 9.6.3. I'll publish some numbers soonish.

Edit: Maybe https://gitlab.haskell.org/ghc/ghc/-/issues/23464

[martijnbastiaan]

GHC Version	Error	Time
9.0.2	0	5.11
9.2.8	0	3.66
9.4.1	0	3.42
9.4.2	0	3.34
9.4.3	0	3.41
9.4.4	0	3.89
9.4.5	0	3.42
9.4.6	0	3.34
9.4.7	0	4.14
9.4.8	0	55.65
9.6.1	0	60.06
9.6.2	0	56.42
9.6.3	0	6.37
9.6.4	0	7.74

So it broke 9.4.7 → 9.4.8 and was fixed 9.6.2 → 9.6.3.