I think it's 6 bytes to encode either version. The 3-operand should result in greater retirement throughput - fewer instructions per unit work. More frontend throughput too, when uops are fed from the micro cache. This would depend on exact code layout and where the 32B cache line boundary is.

The why-nots are unconvincing. Synthetic events can be bypassed, vdom is about to get optimal with React Compiler, and you can use React for server-side apps if you care about bundle size/weak client devices.

The simple mental model of a declarative render is so valuable.

Can you make it into a feed like this one? bsky.app/profile/gpu-...

(I don't know how it's done though)

"will stall [...] only when it tries to *use*" - it still gets swapped regularly by the scheduler, no? I thought that's why s_clause exists