trs runcore: zero-copy sidecar strings + window-image move — the boot-alloc rung

[nanavati]

Why

The -m 1 boot/loop split showed the last two Verilator deficits are RunCore boot mass, not per-cycle code: CFL's boot was 66% of its run (6.1ms of 9.2), BRAM0Test's 56% — while both steady loops alone already beat Verilator's entire runs. The boot profile put the compute share in allocations: parse_sidecar built a String per entry of the full design string table (CFL: thousands of rule names — the dominant push/malloc source, with heap-consolidation churn), and the boot then cloned the whole post-window arena image.

What

• Zero-copy string/path tables: they parse to (offset, len) ranges into the sidecar bytes — take_range UTF-8-validates each entry once and allocates nothing; the buffer moves into RunCore and table_str slices it unchecked under that parse-time invariant (documented). The two runcore_prim_cb name lookups slice field-precisely (a &self accessor would conflict with the live &mut rc.prims borrow). Dyn strings, warn names, prim seeds, and load rows keep owned storage — tiny counts, and runcore_restore is untouched.
• Window-image move: Boot's post-window arena image is take()n instead of cloned — the clone re-touched every arena page (alloc + memcpy) and nothing reads it afterward.

Sidecar format unchanged (no version bump — this is a boot-side parse change only).

Measured

Same-session: CFL boot 6.1 → 4.8ms (full 9.2 → 8.3), BRAM0Test full 9.7 → 8.8, outputs byte-identical. Boot compute after the change is 4.1M Ir (a third of it the one-time UTF-8 validation) — the remaining boot is process/dlopen/page-fault overhead, not instructions.

Same-moment stamp (fresh host, min-of-N interleaved, all legs byte-verified — vl_bytes all green under the harness's trailer/prefix/golden rules):

design	trs	verilator	note
CFL	6.04	5.33	gap −1.7 → −0.71ms
BRAM0Test	6.62	4.62	gap −2.8 → −2.0ms
Mesa	35.92	54.40	−4.8ms (same big string table at boot) — 1.51x win
FloatTest	39.42	44.30	win widens to 11%
Dividers / TrafficBRAM / Sudoku / DFT64v1 / DFT64v5 / Long	2.07 / 23.18 / 132.7 / 8.93 / 9.13 / 1060	3.10 / 29.64 / 370.8 / 12.15 / 15.47 / 16218	wins hold

Scoreboard: 8W/2L across 10 timed Verilator legs, 11/11 vs Bluesim, both remaining deficits ≤2.0ms.

Witnesses

• Regress battery 22/22, classic and TRS_RUNCORE=1.
• Dual full-corpus seals: 1003 PASS / 0 DIFF twice (witnessed TRS_RUNCORE_CHECK=1 TRS_SELFCHECK=1, and driver-armed TRS_RUNCORE=1 — the driver-armed sweep exercises the new parser on all 998 aot designs), zero mismatch, zero panics.

---

Generated by Claude Code