cranelift: thread source span through jit_hd/jit_at/jit_tl

[danieljohnmorris]

Summary

Cranelift JIT runtime errors used to surface with span: None because the JIT_RUNTIME_ERROR TLS cell only stored a bare VmError. The diagnostic renderer had nothing to underline so errors came out with empty labels: [], breaking parity with tree and VM and forcing agents to re-read source code to locate the offending site. Manifesto cost: extra tokens on every cranelift failure.

This PR plumbs the call-site Span through the JIT helper ABI by packing it as a u64 immediate (start << 32 | end) at the cranelift call site and threading it as a trailing extern "C" arg into jit_hd, jit_at, jit_tl. The helpers store the (VmError, Span) pair on error; the entry point decodes the optional span onto the surfaced VmRuntimeError. Zero is the Span::UNKNOWN sentinel and round-trips to None.

Scoped to those three helpers for v1, as agreed. The other erroring helpers (jit_lst, jit_listget, jit_index, jit_jpth, jit_slc, ...) keep their existing permissive-nil semantics; widening them is parked. Each newly-promoted helper now costs +1 u64 arg and +1 iconst at the call site, which is the dominant runtime cost of this approach.

Repro

$ echo 'f>n;hd []' > /tmp/repro.ilo
$ ilo /tmp/repro.ilo --run-cranelift f --diagnostics-json

Before:

"labels":[]

After (matches --run-vm / --run-tree):

"labels":[{"col":7,"end":9,"line":1,"message":"here","primary":true,"start":6}]

What's in the diff

Commit 1, cranelift: thread source span through jit_hd/jit_at/jit_tl

• src/vm/mod.rs: JIT_RUNTIME_ERROR now holds Option<(VmError, Option<Span>)>. New jit_set_runtime_error_with_span(err, span_bits); old jit_set_runtime_error(err) is a 0-bits wrapper for helpers not yet plumbed. jit_take_runtime_error returns the tuple. New decode_span_bits helper. jit_hd, jit_at, jit_tl extern "C" signatures gain a trailing span_bits: u64 and route every error path through the new helper. Inline unit-test calls updated.
• src/vm/jit_cranelift.rs: New pack_span_bits(Span) -> i64. helpers.hd / helpers.at / helpers.tl declared with the new arity. Compile sites for OP_HD / OP_AT / OP_TL pack chunk.spans[ip] into an iconst and append it to the call args. JIT entry point attaches the decoded span to VmRuntimeError.

Commit 2, test + example: cross-engine span parity for cranelift errors

• tests/regression_cranelift_error_span.rs: 9 tests asserting cranelift (start, end) equals VM (start, end) across every v1 helper error path, plus a belt-and-braces "labels not empty" check.
• examples/cranelift-error-span.ilo: positive-path coverage of the extended helper ABI through the cross-engine harness, so a future call site that forgets the immediate crashes loudly on arity mismatch.

Test plan

• cargo build --release --features cranelift
• cargo test --release --features cranelift --test regression_cranelift_error_span (9 passed)
• cargo test --release --features cranelift full suite (4700 passed, 0 failed)
• cargo fmt --check
• cargo clippy --release --features cranelift --all-targets -- -D warnings
• CI green

Follow-ups

• Widen span plumbing to the remaining erroring JIT helpers (jit_lst, jit_listget, jit_index, jit_jpth, jit_slc, the arithmetic helpers, ...). Each is mechanically the same change: +1 u64 arg, +1 iconst, swap to _with_span. Parking until a real diagnostic complaint surfaces from one of those paths.
• Carry the JIT call stack through too. Currently VmRuntimeError::call_stack is always empty for cranelift; the helper TLS would need a Vec<Frame> push at every JIT call site, which is more invasive.

[codecov]

Codecov Report

❌ Patch coverage is 95.45455% with 3 lines in your changes missing coverage. Please review.
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
src/vm/mod.rs	93.02%	3 Missing ⚠️

📢 Thoughts on this report? Let us know!