List of future optimizations for Wasm GC performance

[fitzgen]

This is a tracking issue collecting all of our ideas for improving Wasm GC performance and potential optimizations we could implement.

Tasks

Beta Give feedback

1. Speed up getting and setting funcref elements/fields in GC objects #9347
   cranelift:goal:optimize-speed performance wasmtime:ref-types +3
2. Avoid locks for subtyping checks in Wasm GC #9352
   cranelift:goal:optimize-speed performance wasmtime:ref-types +3
3. Reuse linear memories, and our code to emit their bound checks, for Wasm GC heaps #9350
   cranelift:goal:optimize-speed performance wasmtime:ref-types +3
4. Escape analysis for Wasm GC #9349
   cranelift:goal:optimize-speed performance wasmtime:ref-types +3
5. Type-based alias analysis for Wasm GC #9348
   cranelift:goal:optimize-speed performance wasmtime:ref-types +3
6. cranelift/egraphs: allow simplifying trapping arithmetic #5908
   cranelift:goal:optimize-speed cranelift:mid-end +2
7. Const fold through uadd_overflow_trap/trap{z,nz} insts that show up frequently in GC heap accesses. Depends on cranelift/egraphs: allow simplifying trapping arithmetic #5908
8. Also add a mid-end rule like (uadd_overflow_trap (uextend x) (uextend y)) ==> (iadd (uextend x) (uextend y)). Depends on cranelift/egraphs: allow simplifying trapping arithmetic #5908
9. Use bulk memcpy-style APIs to implement array.copy when possible and do GC barriers inline when not possible
10. Expose type registry's supertypes arrays to Wasm, so it can do the O(1) subtype checks inline
11. Propagate trapz backwards from ref.cast into the branch-y instance-of check that it emits. Similar for the jumps inside a br_on_cast[_fail]'s instance-of check. See Implement the ref.cast Wasm GC instruction #9437 (comment) for details. Alternatively, clean this stuff up in the mid-end: cranelift/egraphs: Optimize instructions without results #6106
12. Speed up call_indirect in the face of subtyping. Much of this can be addressed by letting Wasm to the O(1) subtype checks inline, but it is still additional work to be done on every call_indirect vs MVP wasm. We should investigate inline caches and such again.

[github-actions]

Subscribe to Label Action

cc @fitzgen

This issue or pull request has been labeled: "wasmtime:ref-types"

Thus the following users have been cc'd because of the following labels:

• fitzgen: wasmtime:ref-types

To subscribe or unsubscribe from this label, edit the .github/subscribe-to-label.json configuration file.

Learn more.

[fitzgen]

See also WebAssembly/gc#565 for potential benchmarks.