This is a great example of what I'd call "vectorization without vector instructions", though you could also just call it pipelining. You recognize that you're g

This is a great example of what I'd call "vectorization without vector instructions", though you could also just call it pipelining. You recognize that you're going to be operating on big chunks of data at once, so you make allowances to be able to only operate in chunks - like ensuring you can overrun the end of the buffer - or split your main loop out so there is a 'slow path' for the start/end of your operation and a fast path for the rest of it. Then you just operate on as much data as possible in one go and ignore things like branching, early out, etc. You can mask your results at the end or simply discard unused outputs. The '(running & 0x8080808080808080)' example in this article is perfect because it recognizes that in the end you only need to be sure that nothing failed, you don't need to know what and you don't need to know when. Most of the time if you see people doing this it's using vector registers and newer instruction sets, because that's where you get 2-10x speedups by doing it. But even on ARM and x86 without using vector registers, careful pipelining and vectorization can deliver pretty sizable speedups. It increases cache locality, cuts down on branches, and reduces register pressure. I ended up working on a project a while ago where we had to port a game's software rasterizer to multiple targets, including a low-spec ARM device. Instead of trying to hand-code a NEON implementation (I don't remember for sure if it even had NEON...) I started by just vectorizing it to operate on groups of 4 pixels instead of one (because I could fit 4 pixels in a single regular-sized integer register) so that each iteration of the main loop was doing more work and spending less time waiting on memory. Just that alone ended up improving performance considerably at the kind of clock speeds I was dealing with, and on higher-spec x86 cores it was still a slight improvement for about a day of hacking and staring at disassembly.