Some additional context. I was actually writing a benchmarking tool for certain kinds of search algorithm. I spent a long time reducing and controlling for external sources of noise. CPU pinning, doing hundreds of those runs with different data, and then repeating them several times and taking the best of each score with the same data (to control for transient performance issues due to the machine doing other things).

I got the benchmarking tool itself to give reasonably repeatable measurements.

The tool had high precision, but the accuracy of "which algorithm is better" was not reliable just due to these code layout issues.

I basically gave up and shelved the benchmarking project at that point, because it wasn't actually useful to determine which algorithm was better.