Reproducer for #14062 is 11% slower with GHC 9.0.1 than with 8.10.4

$ ghc-8.10.4 -O2 T14062.hs && ./T14062
Loaded package environment from /home/simon/.ghc/x86_64-linux-8.10.4/environments/default
[1 of 1] Compiling Main             ( T14062.hs, T14062.o )
Linking T14062 ...
benchmarking monad transformers overhead/test1
time                 45.52 ms   (45.49 ms .. 45.57 ms)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 45.60 ms   (45.54 ms .. 45.76 ms)
std dev              200.0 μs   (47.44 μs .. 352.6 μs)

benchmarking monad transformers overhead/test2
time                 45.54 ms   (45.49 ms .. 45.61 ms)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 45.55 ms   (45.52 ms .. 45.57 ms)
std dev              49.30 μs   (40.24 μs .. 60.50 μs)

$ ghc-9.0.1 -O2 T14062.hs && ./T14062
Loaded package environment from /home/simon/.ghc/x86_64-linux-9.0.1/environments/default
[1 of 1] Compiling Main             ( T14062.hs, T14062.o )
Linking T14062 ...
benchmarking monad transformers overhead/test1
time                 50.57 ms   (50.54 ms .. 50.61 ms)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 50.66 ms   (50.61 ms .. 50.90 ms)
std dev              177.3 μs   (44.23 μs .. 318.0 μs)

benchmarking monad transformers overhead/test2
time                 50.58 ms   (50.55 ms .. 50.61 ms)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 50.63 ms   (50.59 ms .. 50.74 ms)
std dev              121.8 μs   (36.08 μs .. 211.5 μs)

I'm using mtl-2.2.2, transformers-0.5.6.2 and criterion-1.5.9.0 in both cases.

Adding -fproc-alignment=64 and +RTS -A32m doesn't make the difference go away.

(-fproc-alignment=64 actually results in a warnings:

/usr/bin/ld.gold: warning: T14062.o: section .rodata.str contains incorrectly aligned strings; the alignment of those strings won't be preserved

)