turbopack: Add NEEDS_RESOLVE const to reduce binary size from monomorphization

[lukesandberg]

Summary

• Binary size analysis reports for turbopack's native binary (153MB unstripped / ~103MB stripped / ~34MB gzipped), breaking down overhead by crate, section, and monomorphization category.
• TaskInput::NEEDS_RESOLVE const — a compile-time constant on the TaskInput trait that indicates whether a type may contain unresolved Vcs. Only Vc<T> sets it to true; compound types OR their elements' values.
• trivial_clone_resolve — a simpler resolve functor used when NEEDS_RESOLVE is false, replacing the full async resolve_functor_impl state machine for ~60% of turbo-tasks functions whose argument tuples contain no Vc values.

Key findings from analysis

Subsystem	.text size	% of code
turbo-tasks infrastructure	~8.4 MiB	~12-15% of stripped binary
functor() closures	~3.5 MiB	Unique per fn-item type
conditional_update closures	1.36 MiB	Per-type equality comparison
resolve_functor_impl	688 KiB	Per unique arg-tuple type

This PR addresses

resolve_functor_impl monomorphization: ~527 of 877 instances contain no Vc arguments and compile to trivially-different async state machines. By branching on the new NEEDS_RESOLVE const at compile time, these instances use trivial_clone_resolve instead — a simpler function body that the compiler can better deduplicate via ICF. Estimated savings: ~200-350 KiB of .text.

Design notes

• The derive(TaskInput) macro computes NEEDS_RESOLVE at macro-expansion time via syntactic Vc detection rather than const-eval of field type NEEDS_RESOLVE. This avoids const-evaluation cycles with recursive types (e.g. ChunkGroupEntry which has Box of ChunkGroupEntry fields).
• The if T::NEEDS_RESOLVE branch in ArgMeta::new / with_filter_trait_call is evaluated in a const fn context — both arms are bare function pointers, so this is valid.

Future opportunities identified in the reports

Proposal	Estimated savings
Feature-gate ValueDebug (dbg/dbg_depth)	~2.5-3.5 MiB
Type-erase conditional_update comparison	~700-900 KiB
Eliminate duplicate dbg() turbo-task	~350-700 KiB
Enable --icf=all linker flag	~500-1,500 KiB

Test plan

• cargo check -p turbo-tasks -p turbopack-core -p turbopack-ecmascript -p turbopack -p next-api — all compile cleanly
• cargo test -p turbo-tasks --lib --tests — 50/50 tests pass
• cargo fmt -- --check — no formatting issues

[lukesandberg]

is false the right default?

[vercel]

Suggested change

	let needs_resolve_impl = if any_field_needs_resolve {
	quote! { true }
	} else {
	quote! { false }
	// Collect the idents of generic type parameters so we can OR in their NEEDS_RESOLVE.
	// This is safe from const-eval cycles because `T::NEEDS_RESOLVE` refers to the concrete
	// substitution at monomorphization time, not `Self`.
	let generic_param_idents: Vec<_> = generic_params.iter().map(|p| &p.ident).collect();
	let needs_resolve_impl = if any_field_needs_resolve {
	quote! { true }
	} else {
	quote! { false #(|| <#generic_param_idents as turbo_tasks::TaskInput>::NEEDS_RESOLVE)* }

The TaskInput derive macro always emits NEEDS_RESOLVE = false for generic types, ignoring that generic type parameters could be Vc<T> at monomorphization time.

[codspeed-hq]

Merging this PR will not alter performance

✅ 17 untouched benchmarks
⏩ 3 skipped benchmarks¹

---

_{Comparing turbopack_binary_size (dfa08f7) with canary (db0fdc3)}

Footnotes

1. 3 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩