The prebuild (provided by Mikael) for the linux toolchain upgrade shows a regression:

-1.2% 2dBench-bimg_misc-G1 on Linux-aarch64

when comparing promo.25-b9 against the same source base built with the new toolchain. The regression is small, but it reproduces.

To be sure, the baseline (promo.25-b9) was run second time. Both results produce the each other's performance level. And the new tool chain build was re-run another 7 times, for which all results produce the same performance level to each other.