perf(arm64): reduce loop scans, spill loads, and state copy overhead

[tignear]

Summary

Heliodor's native ARM64 kernel repeatedly scans inactive entries, reloads spilled values, and copies small state arrays. Reduce this work in MIR optimization, register allocation, and machine-code emission:

• Convert eligible bitmap and circular scans to bit scans, remove redundant loop counters, pack invariant byte predicates, and simplify exclusive predicates and empty conditional branches.
• Preserve bit bounds through lowering, collapse packed-bit reconstruction, and recover contiguous copies.
• Rematerialize constant phi inputs at their edges, weight spills by loop frequency, reuse nearby reloads, and coalesce bit-insert destinations with their bases. Keep critical-edge phi sources live across every spill retry.
• Reuse state page addresses, skip empty sparse-worklist groups, specialize commits covered by one dirty word, and pair short vector copies using fixed offsets. Emit direct conditional branches with range-safe fallbacks.

Loop rewrites check bounds, escaped values, and effects on skipped paths. Memory rewrites preserve aliasing, overlap, partial chunks, and both value planes.

Native ARM64 comparison: release runners compiled on the same Neoverse-N2 host; Celox native/O2; Heliodor a78d04730cf2b37c616e039b4a5bd437c1cfd355. Three alternating rounds compare master 1ff7f81e with this source tree and synchronous Veryl-CC. Timed executions exclude IR dumps and sampling.

Linux boot execution	Median, seconds	Range, seconds
Celox master	33.121	33.120–33.126
Celox with this change	20.487	20.478–20.489
Veryl-CC synchronous	19.914	19.909–19.917

Celox execution time falls by 38.1%, bringing the gap to Veryl-CC to 2.9%. Heliodor compilation falls from 13.263 s to 12.413 s (medians; 6.4% lower). Veryl-CC's AOT compilation was cold in the first round (163.864 s) and cached in the later rounds (8.636–8.779 s).

All nine boots matched cy=87cda0 x3=aa pass=1. The benchmarked commit 26c1b8c1 and the PR head have identical Git trees (698273195684401200cad34a2fa1ab2052e9cbbd). An independent image comparison also measured 20.481 s for the selected code versus 19.955 s for synchronous Veryl-CC, with all fifteen boots matching.

The profiling workflow now supports same-host baseline/candidate/Veryl-CC comparisons. Each comparison and profile invocation has its own 1,200-second timeout, addressing the P2 review.

Validation

• 83 ARM64 backend tests passed under QEMU, including differential checks for loop rewrites, register-pressure/phi preservation, copies, sparse commits, shift boundaries, and far branches.
• 98 simulator/native integration tests and 40 host backend tests passed; the full optimized Linux boot also passed under QEMU.
• Host and ARM64 Clippy passed with warnings denied; formatting, workspace checking, and workflow YAML/shell syntax checks passed.
• The selected source completed the native ARM64 compilation/execution comparison and all Linux boots above. CI passed on the final commit, including ARM64 backend, Rust/SystemVerilog/cocotb, NAPI/WASM, JavaScript, and browser tests. Heliodor gates, compile-time benchmarks, and CodeQL also passed.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-09T01:44:07.737235Z	865eb81	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[codspeed-hq]

Merging this PR will not alter performance

✅ 4 untouched benchmarks

---

_{Comparing perf/heliodor-arm64-phase2 (865eb81) with master (1ff7f81)}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 7ef53d2489

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Apply a timeout to each comparison run

When either the downloaded baseline or candidate runner hangs, this direct invocation bypasses the timeout protection in scripts/run-heliodor-bench.sh:1356-1395, so the comparison stalls until the entire 120-minute job timeout and the remaining rounds/profile never run. GNU timeout --help confirms that timeout DURATION COMMAND terminates a command after the specified duration; invoke the runners through the existing run_in_heliodor helper or wrap each invocation with an equivalent per-run timeout.

Useful? React with 👍 / 👎.