One of my favorite go-to examples for this is "Producing Wrong Data Without Doing Anything Obviously Wrong!" 4.2 where they demonstrate a reproducible >30% perf

One of my favorite go-to examples for this is "Producing Wrong Data Without Doing Anything Obviously Wrong!" 4.2 where they demonstrate a reproducible >30% performance difference just by "misaligning" stacks with an unused environment variable. That's in the context of unexpected measurement bias, but you won't convince me piles of cold code can't affect memory layout less than that.