[WhyNotHugo]

OTOH, you can remove the runtime dispatching logic entirely if you compile separate binaries for each architecture variant.

Especially the binaries for the newest variant, since they can entirely conditionals/branching for all older variants.

[jeffbee]

That's a lot of surgery. These libraries do not all share one way to do it. For example zstd will switch to static BMI2 dispatch if it was targeting Haswell or later at compile time, but other libraries don't have that property and will need defines.