Repository navigation
private field access is ~1.5× public (residual after #11864/#11913/#11954) #11791
Description
Private brand markers are stored as own keys on instances, so every instance of a class with #private members leaves the class's birth shape. Method calls on such instances then miss the class-shape fast path.
- Cost: a method call on an instance of a class with private members is 7,540 instructions on main. A per-site extra-shape cache (shape-campaign method-call lane) brings it to 1,844, which works around the miss without removing the cause.
- Fix (charter): make the private brand a shape fact, so the class's birth shape carries the brand, instead of an own key. Brand checks (
#x in o, private field/method access) then compare the shape or its brand fact, and instances stay on the birth shape.
This lives in the class-template / birth-shape area that #11780 (template-owned final shapes for class objects and prototypes) builds. It will follow as its own PR on top of #11780.
Activity
perf: private brands and private fields are facts of the object shape
proggeramlug commentedon Oct 4, 2026
Progress in #11864: private brands and private fields are now facts of the object shape (private rows 4–7× faster; tsc −0.06%, Zod +0.215%). Leaving open unless the author confirms this fully meets the issue's target.
proggeramlug commentedon Oct 4, 2026
More progress in #11913: private fields are typed slots of the class layout. private 447 → 331, field_rw 239 → 169, method_call 273 → 188 instructions; tsc flat, Zod −0.055%. The private row is still 1.5× public; leaving open.
perf: private fields are typed slots of the class layout
perf: own-key lookups read the private entry where they find the key
perf: a shape mint pays nothing for private brands it does not carry
proggeramlug commentedon Oct 8, 2026
Narrowing this issue's scope: the stated cause (the brand as an own key) is fixed by #11864 (private brands and fields are shape facts), #11913 (typed slots) and #11954 (a mint pays nothing for brands it doesn't carry). The gap tests test_gap_11791_* are on main.
Remaining: private field access is still about 1.5× the cost of public field access. This issue now tracks that gap.
[-]Private brands as own keys push every instance off the class birth shape (method calls 7.5k instructions)[/-][+]private field access is ~1.5× public (residual after #11864/#11913/#11954)[/+]