The purpose of this thing isn't really performance but fun and exploration, if I'm reading the author's comments right. For examples like yours though, between

The purpose of this thing isn't really performance but fun and exploration, if I'm reading the author's comments right. For examples like yours though, between the static analysis done by the compiler, whatever the JIT does and intrinsics, you can't readily predict what's going to run faster by looking at the Java code or the bytecode. These days, a much broader range of things that are equivalent-looking end up performing the same once the JIT gets involved, including stuff that seems to be naively doing allocations in inner loops. There's a whole genre of SO questions and answers that boil down to throwing a micro-example like yours into JMH and finding out they have identical performance profiles.