Computer Enhance was never meant to be a place for publishing novel investigations into software performance, but sometimes it happens by accident. We

The Case of the Missing Increment - by Casey Muratori

submited by

Style Pass

2024-09-27 05:00:07

Computer Enhance was never meant to be a place for publishing novel investigations into software performance, but sometimes it happens by accident. We already saw this once when bonus materials for the Performance-Aware Programming course lead us to discover a previously undocumented 16-ahead mapping pattern in the Windows memory manager.

In that same vein, on a specific set of Intel CPUs, some of our tutorial loops hit a mostly undocumented performance anomaly in the Golden Cove microarchitecture. I say “mostly” because — SPOILER WARNING — once I figured out what was happening, I was able to search for and find a single solitary hint about it in a writeup of one of Intel’s Architecture Day presentations.

The scene of this particular whodunnit may seem unlikely: the performance anomaly occurs in some of the simplest machine code we could consider. In fact, you might say it’s oversimplified code, since it doesn’t do any useful work. It’s entire raison d'être was illustrating hardware behavior for educational purposes.

But in my experience, extremely simple machine code is actually one of the most common places to witness something strange. Since you’re handing the CPU a few highly-specific instructions without most of the code that would otherwise be interleaved, you’re pushing the boundaries of what its designers expected to encounter in real use. At this point, you’re more likely — not less — to feel the edges of the microarchitecture than you are in a more typical scenario.