Note that if you malloc that big of a block, it's going to come directly from a MAP_ANONYMOUS mmap, and so it will be zero initialized!
Your uninitialized mode is slower than zeroed, but nowhere near as slow as random.
The difference between initialized and uninitialized may be due to page faults: in the uninitialized case, the search is touching the memory for the first time. Maybe we're seeing TLB misses?
Yeah I was expecting the uninitialized to be the same as the zeroed case for that reason, but was surprised to see a (slightly) different result.
Edit: you are almost certainly right. I added a getrusage call before and after, and we have ~50k soft page faults in the uninitialized case and ~2 in the zeroed case during the memmem.
Your uninitialized mode is slower than zeroed, but nowhere near as slow as random.
The difference between initialized and uninitialized may be due to page faults: in the uninitialized case, the search is touching the memory for the first time. Maybe we're seeing TLB misses?