Summary

Implements #153: an entryLoop trace plan now hoists one loop-invariant container's derivation out of the loop body into a per-entry prologue.

• Detection (interp/jit_plan.go): hoistable picks the most-accessed ref local whose recorded ARRAY_GET/ARRAY_SET steps agree on one primitive typed-array itab, provided no block writes that local and the plan contains no CALL/RETURN_CALL (a BL would clobber the hoisted registers). Container provenance comes from a conservative per-block marker stack; anything unknown declines.
• Prologue (interp/jit_arm64.go, arm64Lowerer.hoist): once per native entry — load the container local, tag + itab guard against an empty-snapshot exit resuming at the header, load dataPtr/len into registers that stay live for the whole body. Safe with no pinning because a call-free trace loop callable has only forward control flow.
• Access fast paths: matching ARRAY_GET keeps only its bounds exit, guardIndex, and the element load (shared tail; only the (dataPtr, n) source branches). A hoisted primitive ARRAY_SET keeps bounds + store and stops being a state barrier (no flush(flushSnapshot) + local-cache clear). Non-matching containers, ref arrays, globals/upvals, and terminal stores use the existing paths byte-for-byte.
• Livelock fix (interp/interp.go): a loop fallback exit that resumes at the header itself made no progress, and the header dispatch slot holds the native stub — redispatching would livelock. The loop wrapper now runs the shadowed threaded handler (i.exits[root]) once. The hoist prologue's shape guard is the first exit that can resume at the header; covered by a regression test that alternates the container between an array and null across entries.

Measured effect (honest)

• Sieve's two loop callables shrink ~40%: vm_jit_entry_bytes_total 179,992 → 107,168 (inner) and 209,804 → 124,644 (scan).
• Sieve(256) wall time is unchanged within noise (interleaved A/B on M4 Pro: ~1–2%). The removed per-access loads sat off the out-of-order critical path, so the issue's projected 2.5–3 µs did not materialize.
• TypedArraySum(256) and AllocationGraph(128) unchanged (interleaved A/B).
• Profiling shows where the remaining Sieve gap actually lives, filed as follow-ups: loop-kind branch legs are never folded (one native exit/re-entry per prime in the scan loop), and the 256-copy forward-chained body pays a per-iteration operand flush/reload/budget chain (a pure fill loop runs slower native than threaded today).

Test plan

• go test -race ./... (root and benchmarks/ modules)
• New TestARM64_HoistedContainerLoop: native execution with only loop-exit exits, prologue shape deopt with exact refcounts, mid-loop bounds deopt with identical error, LOCAL_SET-in-loop rejection, mixed hoisted + slow-path containers — all against a threaded oracle comparing results and rc[1:]
• New TestHoistable plan-level unit tests
• make fuzz, make lint
• make coverage-check — fails from the pre-existing goenv go1.25/go1.26 toolchain split, unrelated to this diff

Closes #153.

Summary by CodeRabbit

• Performance

  • Improved ARM64 JIT performance for eligible loops by hoisting loop-invariant array container metadata and reusing derived slice-header information, reducing redundant per-access checks/loads (notably shrinking loop-callable size by ~40%).
  • Added faster hoisted array get/set paths within supported loop traces.

• Bug Fixes

  • Prevented potential loop livelocks when deoptimization resumes at the same loop header position.

• Tests

  • Added ARM64 coverage for hoisted-container loop correctness and profiling expectations.
  • Expanded unit coverage for loop-hoisting eligibility.

• Documentation

  • Documented loop-invariant container hoisting rules and updated benchmark notes accordingly.