perf: this.a inside an object-literal method costs ~1,085 instructions per call (read path misses every cache) #11420

Activity

1. proggeramlug commented on Sep 26, 2026

   

   proggeramlug

   on Sep 26, 2026

   ContributorAuthor

   More actions

   Evidence from the package-benchmark attribution work (Phase 3): the harness's bare-loop control does not measure this issue's shape, and on current main it has no ~1,000-instruction floor at all.

   Setup: origin/main 36420d2 (perry 0.5.1654), release compiler, auto-optimize with PERRY_WORKSPACE_ROOT (per-program runtime rebuild), x86-64 (Zen 4), perf stat -e instructions:u, two-N per-iteration method (n = 2M and 8M; startup cancels). Node 26.5.1. Every output was checked equal to Node's.

   1. The current control (benchmarks/packages/control/bare_loop.ts on wip/pkg-bench-phase1 since 21c9f54, loop bounds hoisted) costs 25 instr/iter (Node: 56). No call survives: LLVM inlines op and mulFnv. The 25 are 3 ToInt32 range guards, 2 cvttsd2si, 3 int→double conversions, the arithmetic, and the loop's induction variable kept in both a double and an i32. There is no object-literal method and no this in it.

   2. The historical ~1,040–1,250 was a plain by-name data-property read (it.n in the loop condition), not a call.

   control variant	Perry instr/iter	Node
   original (first harness commit): Math.imul + i < it.n + header(…, it)	1,178	~0
   same, Math.imul replaced by mulFnv	936	55
   i < it.n, it from imported iters(), no header(…, it)	69	55
   i < it.n, it a same-module literal passed to a same-module function	110	8
   current: bounds hoisted	25	56

   So about 870 instructions per iteration come from it.n when it is returned by an imported function and passed to an imported function (header). Chain (DWARF call graph): js_class_field_get_ic → js_typed_feedback_class_field_get_guard → class_field_fast_contract → shape_descriptor_by_id (the guard fails), then js_object_get_field_by_name_f64 → inherited_read_cache_lookup → get_field_by_name_past_inherited_cache → try_data_get_bytes → keys_find_slot_by_bytes_resolved / is_anon_shape_class_id. That is the receiver-provenance read family (#10769 / #10761), not an object-literal method. Math.imul added about 240 per call on top.

   3. This issue's own shape, measured the same way:

   const o = { k: 3, m(x: number): number { return (x * this.k + 1) % 1000003; } };
   for (let i = 0; i < N; i++) acc = o.m(acc + i);

   call shape (same body)	Perry instr/call	Node
   object-literal method + this.k (this issue)	4,267	24
   class method + this.k	51	41
   direct non-inlined function call	86	23
   closure held in a const	298	24
   f.call(o, x)	3,746	113
   function P(){}; P.prototype.m = …; o.m(x)	9,596	34
   private this.#m(x) reading this.#k	1,023	37
   class getter o.k	483	32

   In this microbenchmark the object-literal method costs about 4.3k per call. 76% of that is the dispatcher under js_typed_feedback_native_call_method_by_id (js_native_call_method, js_native_call_value, dispatch_handle, shape_descriptor_by_id). 16% is the this.k read through js_object_get_field_by_name_f64. So the dispatch costs more than the this.a read in this shape. Your ~1,085 may be the read alone; the call adds the rest here.

   Conclusion for the harness: the control is not representative of #11420, and after 21c9f54 it no longer carries any floor: 25 instr/iter, below Node. The Phase-1 interim per-iteration numbers at 2febf42 included about 0.9–1.2k instr/iter of harness overhead per iteration (it.n + Math.imul). That is negligible for the heavy workloads (≥50k/iter) but material for the light ones. The attribution report re-measures everything on the hoisted harness.
2. proggeramlug

   added a commit that references this issue on Sep 27, 2026

   perf(codegen): register anon shapes before minting their ShapeId (#11420

   ...

   3ae9783
3. proggeramlug

   closed this as completedin #11437on Sep 27, 2026