Some bytestring benchmarks got slower in 9.2.0.20210331

Reproduction

Checkout bytestring:

git clone https://github.com/haskell/bytestring
# git checkout ba3046d9f5bf5cac1246c39f519b8be7ac11f49c

Current master is very close to the released bytestring-0.11.1.0, which is bundled with GHC 9.2, but includes some benchmark fixes.

Run benchmarks with GHC 9.0.1 (it will take around five minutes):

cabal bench -w ghc-9.0.1 --benchmark-options '--csv 9.0.1.csv --stdev 2 --hide-successes'

Compare againt benchmarks with GHC 9.2.0.20210331 (it will take around five minutes):

cabal bench -w ghc-9.2.0.20210331 --allow-newer='tagged:template-haskell,splitmix:base' --benchmark-options '--baseline 9.0.1.csv --csv 9.2.0.csv --stdev 2 --fail-if-slower 10 --hide-successes'

Thanks to a featherlight dependency graph, tasty-bench-based benchmark suites can already be compiled by GHC 9.2.

Expected results

No benchmarks became slower.

Actual results

All
  Data.ByteString.Builder
    Small payload
      intHost 1:                                              FAIL (4.99s)
        540 ns ±  12 ns, 42% slower than baseline
      UTF-8 String (naive):                                   FAIL (5.25s)
        577 ns ±  22 ns, 26% slower than baseline
      UTF-8 String:                                           FAIL (38.86s)
        568 ns ± 6.8 ns, 22% slower than baseline
      String (naive):                                         FAIL (5.34s)
        590 ns ±  24 ns, 27% slower than baseline
      String:                                                 FAIL (5.22s)
        579 ns ±  17 ns, 30% slower than baseline
  Data.ByteString.Builder.Prim.ASCII
    int32Dec (10000):                                         FAIL (0.65s)
      149 μs ± 4.0 μs, 12% slower than baseline
    int8HexFixed (10000):                                     FAIL (1.11s)
       16 μs ± 584 ns, 12% slower than baseline
  words
    lorem ipsum:                                              FAIL (0.92s)
      3.4 μs ±  62 ns, 13% slower than baseline
  folds
    foldl'
      1:                                                      FAIL (1.67s)
         24 ns ± 226 ps, 12% slower than baseline
      2:                                                      FAIL (1.11s)
         32 ns ± 690 ps, 11% slower than baseline
      4:                                                      FAIL (0.43s)
         48 ns ± 1.4 ns, 15% slower than baseline
      8:                                                      FAIL (0.74s)
         84 ns ± 1.3 ns, 15% slower than baseline
      64:                                                     FAIL (0.62s)
        568 ns ±  12 ns, 13% slower than baseline
      128:                                                    FAIL (0.59s)
        1.1 μs ±  37 ns, 13% slower than baseline
      256:                                                    FAIL (0.60s)
        2.2 μs ±  65 ns, 14% slower than baseline
      512:                                                    FAIL (1.17s)
        4.3 μs ±  69 ns, 14% slower than baseline
      1024:                                                   FAIL (0.32s)
        8.9 μs ± 334 ns, 16% slower than baseline
      2048:                                                   FAIL (0.60s)
         18 μs ± 426 ns, 16% slower than baseline
      4096:                                                   FAIL (0.59s)
         35 μs ± 900 ns, 14% slower than baseline
      8192:                                                   FAIL (0.59s)
         70 μs ± 2.5 μs, 15% slower than baseline
      16384:                                                  FAIL (0.60s)
        142 μs ± 5.3 μs, 16% slower than baseline
      32768:                                                  FAIL (0.59s)
        282 μs ± 7.1 μs, 15% slower than baseline
      65536:                                                  FAIL (4.71s)
        571 μs ±  15 μs, 17% slower than baseline
    foldr'
      2:                                                      FAIL (0.57s)
         32 ns ± 808 ps, 11% slower than baseline
      4:                                                      FAIL (3.36s)
         48 ns ± 934 ps, 13% slower than baseline
      8:                                                      FAIL (11.36s)
         84 ns ± 1.1 ns, 18% slower than baseline
      64:                                                     FAIL (0.33s)
        567 ns ±  21 ns, 12% slower than baseline
      128:                                                    FAIL (1.18s)
        1.1 μs ±  20 ns, 13% slower than baseline
      256:                                                    FAIL (0.32s)
        2.2 μs ±  85 ns, 13% slower than baseline
      512:                                                    FAIL (0.61s)
        4.4 μs ± 101 ns, 12% slower than baseline
      1024:                                                   FAIL (0.60s)
        8.8 μs ± 211 ns, 11% slower than baseline
      2048:                                                   FAIL (0.60s)
         18 μs ± 366 ns, 14% slower than baseline
      4096:                                                   FAIL (0.61s)
         36 μs ± 978 ns, 16% slower than baseline
      8192:                                                   FAIL (0.32s)
         71 μs ± 2.7 μs, 17% slower than baseline
      16384:                                                  FAIL (1.19s)
        142 μs ± 3.7 μs, 17% slower than baseline
      32768:                                                  FAIL (1.18s)
        281 μs ± 7.2 μs, 15% slower than baseline
      65536:                                                  FAIL (0.60s)
        569 μs ±  22 μs, 16% slower than baseline

37 out of 263 tests failed (371.06s)

On a brighter side, many other benchmarks got 15-20% faster, which is great!

CC @sjakobi