Performance regression from GHC 8.10.4 to GHC 9.0.1

Summary

Two of my libraries, libarchive and phash run slower with GHC 9.0.1 as compared to 8.10.4.

Steps to reproduce

vanessa@vanessa-desktop /development/haskell/phash 🌸 cabal run phash-bench -w ghc-9.0.1
benchmarking fileHash/cat.png
time                 34.88 ms   (34.49 ms .. 35.28 ms)
                     0.999 R²   (0.998 R² .. 1.000 R²)
mean                 34.38 ms   (34.12 ms .. 34.65 ms)
std dev              598.0 μs   (483.6 μs .. 769.4 μs)

benchmarking fileHash/frog.jpeg
time                 23.71 ms   (23.48 ms .. 23.93 ms)
                     1.000 R²   (0.999 R² .. 1.000 R²)
mean                 23.37 ms   (23.21 ms .. 23.51 ms)
std dev              366.7 μs   (274.6 μs .. 489.3 μs)

benchmarking fileHash/frog.png
time                 17.64 ms   (17.50 ms .. 17.79 ms)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 17.62 ms   (17.56 ms .. 17.67 ms)
std dev              141.3 μs   (114.9 μs .. 172.2 μs)

vanessa@vanessa-desktop /development/haskell/phash 🌸 cabal run phash-bench -w ghc-8.10.4
benchmarking fileHash/cat.png
time                 23.30 ms   (23.16 ms .. 23.43 ms)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 23.36 ms   (23.27 ms .. 23.47 ms)
std dev              241.2 μs   (164.9 μs .. 363.2 μs)

benchmarking fileHash/frog.jpeg
time                 19.27 ms   (19.16 ms .. 19.37 ms)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 19.10 ms   (18.99 ms .. 19.17 ms)
std dev              209.3 μs   (165.7 μs .. 262.1 μs)

benchmarking fileHash/frog.png
time                 13.35 ms   (13.24 ms .. 13.48 ms)
                     1.000 R²   (0.999 R² .. 1.000 R²)
mean                 13.49 ms   (13.44 ms .. 13.54 ms)
std dev              134.4 μs   (100.1 μs .. 173.2 μs)

vanessa@vanessa-desktop /development/haskell/libarchive 🌸 cabal run libarchive-bench -w ghc-9.0.1
benchmarking roundtrip/libarchive
time                 214.0 μs   (212.2 μs .. 215.5 μs)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 213.1 μs   (212.4 μs .. 213.9 μs)
std dev              2.483 μs   (1.995 μs .. 3.087 μs)

benchmarking roundtrip/tar
time                 301.7 μs   (300.9 μs .. 302.9 μs)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 303.2 μs   (302.3 μs .. 304.4 μs)
std dev              3.475 μs   (2.790 μs .. 4.311 μs)

benchmarking unpack/libarchive (via bytestring)
time                 957.7 μs   (945.2 μs .. 969.1 μs)
                     0.999 R²   (0.998 R² .. 0.999 R²)
mean                 934.2 μs   (928.0 μs .. 943.4 μs)
std dev              26.62 μs   (19.99 μs .. 38.96 μs)
variance introduced by outliers: 18% (moderately inflated)

benchmarking unpack/libarchive (C API)
time                 902.5 μs   (887.9 μs .. 920.4 μs)
                     0.997 R²   (0.994 R² .. 1.000 R²)
mean                 911.5 μs   (903.5 μs .. 925.9 μs)
std dev              35.59 μs   (22.60 μs .. 55.99 μs)
variance introduced by outliers: 29% (moderately inflated)

benchmarking unpack/tar
time                 3.894 ms   (3.328 ms .. 4.650 ms)
                     0.836 R²   (0.756 R² .. 0.969 R²)
mean                 3.796 ms   (3.547 ms .. 4.223 ms)
std dev              1.007 ms   (685.8 μs .. 1.468 ms)
variance introduced by outliers: 93% (severely inflated)

benchmarking unpack/tarConduit
time                 3.541 ms   (3.184 ms .. 4.002 ms)
                     0.718 R²   (0.499 R² .. 0.962 R²)
mean                 3.821 ms   (3.468 ms .. 4.500 ms)
std dev              1.562 ms   (897.4 μs .. 2.804 ms)
variance introduced by outliers: 98% (severely inflated)

vanessa@vanessa-desktop /development/haskell/libarchive 🌸 cabal run libarchive-bench -w ghc-8.10.4
benchmarking roundtrip/libarchive
time                 196.5 μs   (195.9 μs .. 197.3 μs)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 197.2 μs   (196.7 μs .. 197.7 μs)
std dev              1.660 μs   (1.410 μs .. 2.045 μs)

benchmarking roundtrip/tar
time                 294.3 μs   (292.9 μs .. 295.7 μs)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 291.1 μs   (290.0 μs .. 292.3 μs)
std dev              3.735 μs   (3.192 μs .. 4.461 μs)

benchmarking unpack/libarchive (via bytestring)
time                 880.8 μs   (875.7 μs .. 886.3 μs)
                     1.000 R²   (0.999 R² .. 1.000 R²)
mean                 888.9 μs   (884.1 μs .. 894.6 μs)
std dev              17.34 μs   (13.96 μs .. 22.14 μs)

benchmarking unpack/libarchive (C API)
time                 840.6 μs   (834.9 μs .. 847.7 μs)
                     0.999 R²   (0.997 R² .. 1.000 R²)
mean                 846.6 μs   (841.3 μs .. 859.5 μs)
std dev              25.88 μs   (10.95 μs .. 52.64 μs)
variance introduced by outliers: 21% (moderately inflated)

benchmarking unpack/tar
time                 3.674 ms   (3.316 ms .. 4.108 ms)
                     0.842 R²   (0.696 R² .. 0.953 R²)
mean                 3.526 ms   (3.272 ms .. 3.958 ms)
std dev              1.072 ms   (665.0 μs .. 1.820 ms)
variance introduced by outliers: 95% (severely inflated)

benchmarking unpack/tarConduit
time                 3.672 ms   (3.453 ms .. 4.014 ms)
                     0.650 R²   (0.355 R² .. 0.976 R²)
mean                 4.023 ms   (3.632 ms .. 5.466 ms)
std dev              2.048 ms   (598.5 μs .. 4.177 ms)
variance introduced by outliers: 98% (severely inflated)

Expected behavior

Similar or better performance with GHC 9.0.1

Environment

• GHC version used: GHC 8.10.4 and GHC 9.0.1

Optional:

• Operating System: Linux (Ubuntu 18.04.5)
• System Architecture: x86_64