They're still the x86-64 implementation for things like strlen on a bunch of platforms. They're not always the fastest, but have code size advantages.

http://agner.org/optimize/optimizing_assembly.pdf

That's a really interesting document - I am going to spend some time studying it, thanks!

Good point about the code size. I imagine there are likely to be cases where that would let some algorithm run faster overall because it fits in the instruction cache, even if the string operation considered on its own is slower.

> They were the fastest way to do string operations on those early CPUs, but I don't think this the case on modern CPUs.

Modern CPUs still have a fast path for string operations, and I recall hearing that they even had some improvements not too long ago (in Sandy Bridge or other recent arch).

CPUs may be smart enough to detect a memcpy done with a loop, but REPxx is the preferred way - even on modern CPUs.

I also remember them being called "software machine gun" by Duntemann in his assembly book. Using REP STOSW could save you 5 instructions.