cranelift: route OP_WINDOW / OP_WINDOW_VIEW through extern helpers

[danieljohnmorris]

Summary

Cranelift's JIT used to bail out (JitCallError::NotEligible) on any function containing OP_WINDOW or OP_WINDOW_VIEW, forcing entire hot paths back to the VM dispatcher. The reason: the VM dispatcher's OP_WINDOW emits HeapObj::ListView strides, and Cranelift's inlined LISTGET / FOREACHPREP / FOREACHNEXT fast paths read Vec metadata at fixed offsets that would UB on a view's struct layout.

Fix: emit calls to the existing jit_window / jit_window_view extern helpers in both arms. Both helpers always return owning HeapObj::Lists (never views), so subsequent inlined foreach reads operate on the expected Vec layout and remain sound. We trade the VM dispatcher's per-stride ListView win for the much larger JIT-compiled outer loop, which is exactly the trade bio-canonical-shape flt p (window k xs) needs.

Repro before / after

f xs:L n>L (L n);window 3 xs on --run-cranelift:

• before: silently fell back to VM dispatcher (no error, but the surrounding function never benefited from JIT).
• after: compiles in JIT, full hot path runs native.

What's in the diff

• cranelift: route OP_WINDOW / OP_WINDOW_VIEW through extern helpers (src/vm/jit_cranelift.rs) — replaces the return None bail with calls to helpers.window (3 args) and helpers.window_view (5 args). OP_WINDOW_VIEW is a two-word op; skip_next = true consumes the data word holding the n register, mirroring OP_SLC / OP_LST / OP_MSET / OP_RGXSUB. Stale audit comments referencing the bail in OP_FOREACHPREP / OP_LISTGET remain accurate (no view reaches them under JIT because the helpers materialise to Lists).
• test: cover OP_WINDOW / OP_WINDOW_VIEW on cranelift JIT (tests/regression_window_cranelift.rs, examples/window-cranelift-jit.ilo) — 5 cross-engine regression tests across tree / VM / cranelift covering plain window, foreach-over-window, the fused flt p (window n xs) shape (OP_WINDOW_VIEW), size-1 windows, and n>len. Example file pins the same shapes for the engine harness.

Test plan

• cargo test --release --features cranelift --test regression_window_cranelift — 5/5 pass
• cargo test --release --features cranelift — full suite green
• cargo fmt --check clean
• cargo clippy --release --features cranelift --all-targets -- -D warnings clean
• Smoke: ilo --run-cranelift 'main xs:L n>L n;ws=window 3 xs;hd ws' 1,2,3,4,5 → [1, 2, 3]

Follow-ups

• A future PR can teach the inlined LISTGET / FOREACHPREP / FOREACHNEXT paths to detect HeapObj::ListView via a [ptr+0] == 1 discriminant guard, fall back to the listget helper on mismatch, and then let OP_WINDOW emit views directly. That recovers the per-stride ListView win that the VM dispatcher currently has exclusively, on top of JIT-compiled outer loops. Out of scope here; the helper-routing path is the smaller, safer first step.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ All tests successful. No failed tests found.

📢 Thoughts on this report? Let us know!