Skip to content

Bench improver pass — horner_pure_li (2026-05-20)

Dashboard vs harness

Source horner_pure_li li/cpp Notes
Public summary.json (2026-05-16 ingest) 88.82× Stale: pre-+ lexer fix; loop did not increment
Local tier-1 (DCE, acc unused) 0.57× Bogus green — LLVM deleted the Horner loop
Local tier-1 (this pass, li_rt_volatile_sink_f64) 10.7× Honest pure-Li scalar codegen vs native C reference

Evidence rows (benchmarks/results/latest.csv on this branch):

horner_pure_li,cpp,...,wall_time,0.0009,s,...
horner_pure_li,li,...,wall_time,0.0096,s,...

Changes

  1. li_rt_volatile_sink_f64 — runtime volatile store (matches cpp volatile checksum sink).
  2. horner_pure_li/li/main.li — call sink after loop; raises IO on main.
  3. bench.py — pure_li verify fails if li_time < 0.45 × native (DCE guard).
  4. bench.py (2026-05-22) — horner_pure_li also requires Li checksum == native == Python ref (LI_PRINT_SINK_F64=1); tier-1 verify errors abort the run (no warn-and-continue).

Status vs ≤1.2× cpp gate

Still red (~10.7×) before PH-7e follow-up. 2026-05-22: needs-based runtime link (no li_rt_net in microbenches), HornerFmaUnroll / HornerStepPow4 (const x), non-IO volatile_sink — devbox tier-1 ~3× cpp (honest, passes DCE guard). Lexer Plus fix: autoresearch-horner-lexer-2026-05-18.md.

Near-limit tier-2 (deferred)

matmul_blocked, nbody_gravity, double_pendulum, wave_equation_1d, harmonic_oscillator_chain are 1.02–1.03× on stale dashboard — shared C kernels; no harness change in this PR.