> If it's an optimization, can you show a benchmark to validate that it's really faster as expected?

Yes, I did test it. I didn't save the results, but I can redo them if you want. If you plan to reject the issue anyway, there is no point.

> not building with PGO lead to random performances

I would expect that Py_LIKELY/Py_UNLIKELY helps with this. If the compiler doesn't have probability information, it guesses. This can cause unrelated changes to change the heuristics used by the compiler, impacting the performance of some piece of code. With Py_LIKELY/Py_UNLIKELY, the code generated by the compiler should be more stable.