[clang-21] Possible performance regression with hackbench?

[torvic9]

Hi,

I usually run a series of benchmarks when a new major kernel version is released, and the hackbench benchmark results caught my attention:

When using a 6.16 kernel built with clang-20, I get the following results (average of 5 runs):

1. hackbench -g 10 -l 2000 (process mode): 5.844 seconds
2. hackbench -g 10 -l 2000 -T (thread mode): 5.704 seconds

Now, when using a kernel built with clang-21 (ceteris paribus):

1. process mode: 5.890 seconds
2. thread mode: 5.891 seconds

When I also add ThinLTO to the clang-21 kernel, it gets worse:

1. process mode: 5.969 seconds
2. thread mode: 5.980 seconds

In all the other benchmarks I've run so far, there is basically no difference (or just noise) between clang-20 and clang-21, but not so much with hackbench.

Not sure what to make of this, but I thought I should report it.

---

N.B.: I used the prebuilt clangs from kernel.org, versions 20.1.8 and 21.1.0-rc2, on an up-to-date Arch Linux installation, running on a Skylake laptop.