JMH benchmarking various methods of iterating through an ArrayList
In another project in different time and place, I needed to meet some strict time goals in a routine that was replacing another one (without being more expensive).
Chasing profiler output, one avenue I took was replacing enhanced for loops
as I saw hasNext() and next() invocations popping up, and I figured that
there wasn't really any harm in writing a C-style for loop.
Later, someone challenged me on whether or not this really makes a difference. So I wrote this.
After a bunch of runs (on an AMD EPYC 7281-based VM), it appears (if I've done
everything right) that there is some reliable improvement in both throughput
and variance of the C-style for loop over enhanced for. I'm guessing the
variance improvement is due to reduced heap churn; with no Iterator allocations
GC pauses should be very rare (to non-existent) over the course of the
benchmark runs.
Some day I'll (maybe) get hotspot to dump disassembled output of the code it's emitting so I can sort out how different the code is otherwise (or maybe even look at gc logs to verify my claims are correct), but for now I need to push this so I can stop paying for this VM :)