Reproduction for regression in bench perf timing

1. Run cargo bench a few times, observe the fastest timing column.
2. In lib.rs remove fold_copy() and repeat step 1 to compare timing results.
3. Timing should have decreased, verifying the regression.

Output examples

With regression present

     Running benches/criterion.rs (target/release/deps/criterion-a56b6d24776194ad)
Gnuplot not found, using plotters backend
criterion/into_string_fold
                        time:   [133.56 ns 135.63 ns 137.84 ns]
                        change: [-4.5839% -2.5971% -0.6088%] (p = 0.01 < 0.05)
                        Change within noise threshold.

     Running benches/divan.rs (target/release/deps/divan-e88aec68da77a7cc)
Timer precision: 21 ns
divan                fastest       │ slowest       │ median        │ mean          │ samples │ iters
╰─ into_string_fold  137.2 ns      │ 621.2 ns      │ 159.4 ns      │ 156.2 ns      │ 1000    │ 128000

Without regression

     Running benches/criterion.rs (target/release/deps/criterion-a56b6d24776194ad)
Gnuplot not found, using plotters backend
criterion/into_string_fold
                        time:   [107.28 ns 110.38 ns 113.99 ns]
                        change: [-18.487% -15.598% -12.446%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 6 outliers among 100 measurements (6.00%)
  6 (6.00%) high mild

     Running benches/divan.rs (target/release/deps/divan-e88aec68da77a7cc)
Timer precision: 21 ns
divan                fastest       │ slowest       │ median        │ mean          │ samples │ iters
╰─ into_string_fold  102.1 ns      │ 290.8 ns      │ 107.6 ns      │ 113.4 ns      │ 1000    │ 128000

Description

This is a minimized and isolated reproduction I observed where just having a copy of a function (as dead code) was introducing a consistent bump in timings for the actual benched method.

On my system the difference is approx 100ns vs 135ns, this is consistent for me just by removing the copied function from source.

I suspect it may be environment specific (UPDATE: confirmed reproduction by other users), or vary by rust release with the compiler optimizing differently, so I've put this reproduction together for verification.

The comments in lib.rs show where changes can be made to avoid the regression.

System Details

• Platform: x86_64 GNU/Linux
• CPU: AMD Ryzen 9 7940HS
• Linux Kernel: 5.15.123.1-microsoft-standard-WSL2 (Host: Windows 11 23H2)
• WSL2 OS: Ubuntu 22.04.2 LTS (Jammy Jellyfish)

Reproduced with:

• cargo 1.74.0 (ecb9851af 2023-10-18)
• cargo 1.76.0-nightly (71cd3a926 2023-11-20)