[compiler][poc] Quick experiment with SSR-optimization pass

[josephsavona]

Just a quick poc:

• Inline useState when the initializer is known to not be a function. The heuristic could be improved but will handle a large number of cases already.
• Prune effects
• Prune useRef if the ref is unused, by pruning 'ref' props on primitive components. Then DCE does the rest of the work - with a small change to allow useRef() calls to be dropped since function calls aren't normally eligible for dropping.
• Prune event handlers, by pruning props whose names start w "on" from primitive components. Then DCE removes the functions themselves.

Per the fixture, this gets pretty far.

---

Stack created with Sapling. Best reviewed with ReviewStack.

• [compiler] Deprecate noEmit, add outputMode #35112
• -> [compiler][poc] Quick experiment with SSR-optimization pass #35102

[josephsavona]

This is where the compiler machinery helps: the ref is referenced here, but DCE runs in a fixpoint. By pruning the effect call that makes its callback dead code, which then removes this reference to the ref. Then the OptimizeForSSR pass has already pruned the ref prop, so we're able to remove the useRef.

[josephsavona]

Another thing to add: for custom components we don't know for sure which callbacks are event handlers and which are render props. As implemented that means a lot of event/effect functions can't be pruned. But if we disallow setState during initial mount, then we know that any function which calls setState must be an event handler/effect. Similar for anything that calls startTransition. That should allow identifying a decent percentage of callbacks, though probably still far from all of them.

edit: done

[mofeiZ]

Exciting!

[mofeiZ]

But if we disallow setState during initial mount, then we know that any function which calls setState must be an event handler/effect. Similar for anything that calls startTransition. That should allow identifying a decent percentage of callbacks, though probably still far from all of them.

I'm sure you've already considered this, but it might be a pain to debug if we compile / replace these functions with a no-op -- but they actually get used during render

[josephsavona]

Yeah fair, we could do a follow-up where we replace with a simple function that throws if called.