The explanation is not convincing.

My guess is some kind of measurement error or one of the "load bearing nop" phenomena. By that I mean the alignment of instructions (esp branch targets?) can dramatically affect performance and compilers apparently have rather simplistic models for this or don't consider it at all.