perf: helpers reading several fields of one or two parameters pay a full IC tower per read and boxed arithmetic (cli-table3 layout: 13x Node instructions) #12328

A helper that reads several fields from one or two untyped parameters pays a full inline-cache tower per read: a receiver check, a shape compare and a boxed result. Arithmetic on those values is guarded and generic. Real example: cli-table3's layout code. A verbatim copy of its nine layout functions, laying out a 2,000 × 7 table, costs 69.4 G instructions on Perry vs 5.3 G on Node, with 6.5 s vs 0.6 s wall (main 718ca28).

function cellsConflict(cell1, cell2) {
  let yMin1 = cell1.y;
  let yMax1 = cell1.y - 1 + (cell1.rowSpan || 1);
  let yMin2 = cell2.y;
  let yMax2 = cell2.y - 1 + (cell2.rowSpan || 1);
  let yConflict = !(yMin1 > yMax2 || yMin2 > yMax1);
  let xMin1 = cell1.x;
  let xMax1 = cell1.x - 1 + (cell1.colSpan || 1);
  let xMin2 = cell2.x;
  let xMax2 = cell2.x - 1 + (cell2.colSpan || 1);
  let xConflict = !(xMin1 > xMax2 || xMin2 > xMax1);
  return yConflict && xConflict;
}

What happens

cellsConflict takes 52% of self cycles and its caller conflictExists 35%; the generic IC front takes 7–11%.

• One IC site per read. The 12 reads are independent generic sites (expr/property_get/generic_dispatch.rs). The existing statement-run read regions (expr/region_read_run.rs, stmt/region_read_stmts.rs) need consecutive let a = o.k statements, so operator trees that mix two receivers with || never form a run.
• No field representation. A hit returns a boxed value, because the site word holds (ShapeId, slot) and no field representation. So - 1 + (...), || and > take the guarded_arith / js_is_truthy / dyncmp paths, and each let gets a shadow-root slot.

Proposed direction

Extend the existing read regions, without adding a new mechanism:

• statements whose initializers are pure operator, || and comparison trees over reads of up to 2 receivers, with no calls;
• one shape guard per receiver, covering up to MAX_KEYS keys;
• carry the F64 lane bits from the shape's field representation (shapes: field-representation store check and lineage generalization (step 5, P2a) #11655) in the region word, so numeric leaves skip the tag test (reuse stmt/region_loop/numeric_expression.rs from perf: complete one-shape numeric regions and ConstFn method calls #11680).

Watch the documented regression when a refinable let loses its type (region_read_stmts.rs:84-100).

Related

• perf: reads that resolve on the prototype chain (absent keys, inherited data) are 590–2,100× slower than Node (read IC caches own properties only) #10495: absent reads, such as cell1.rowSpan on an {x, y} literal, always call the IC front. That's planned to be handled through the keyed holder entry.
• cellsConflict is also not inlined into conflictExists: it has 11 statements, above MAX_INLINE_STMTS = 10. This is being measured separately.

Witness: the cli-table3 layout micro (layout-micro.ts in lane perry-wgperf), and walletguide's transactions list --limit 2000 --table (8.1 s vs Node's 1.09 s). Details are in secret-tests scratchpad/shape-audit/AUDIT.md, section 2(a) and step 3.