Cranelift: riscv64: C.J veneer range overflow

[s0me0ne-unkn0wn]

Summary

When compiling large functions targeting riscv64gc with compressed instructions (Zca) enabled, emit_veneer() panics because the veneer is placed too far from the original C.J.

The island emission deadline mechanism is supposed to force island emission before any label reference goes out of range, but, presumably, the deadline calculation doesn't account for the very small range of C.J when traps and constants are flushed before veneers — these can push the veneer past the 2 KiB limit.

.clif Test Case

rvcjump_bug_reduced.clif.gz

Steps to Reproduce

$ clif-util test rvcjump_bug_reduced.clif

Expected Results

Success

Actual Results

thread 'worker #0' panicked at cranelift/codegen/src/isa/riscv64/inst/mod.rs:1759:9:
RVCJump offset '2048' use_offset:'7416' label_offset:'9464'  must not exceed max range.
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace
[2026-03-20T19:04:21Z ERROR cranelift_filetests::concurrent] FAIL: panicked in worker #0: RVCJump offset '2048' use_offset:'7416' label_offset:'9464'  must not exceed max range.
FAIL rvcjump_bug_reduced.clif: panicked in worker #0: RVCJump offset '2048' use_offset:'7416' label_offset:'9464'  must not exceed max range.
1 tests
Error: 1 failure

Versions and Environment

Cranelift version or commit: 0.122.0

Operating system: Linux Manjaro

Architecture: x86_64 cross-compiling to riscv64gc

[cfallin]

Thanks for the report! I'll take a look at fixing this early next week. FWIW, it appears that the particular minimized example doesn't hit the issue on main, but does on the corresponding Cranelift versions for Wasmtime 36.0 and 24.0 (LTS releases); I bisected the "started working" point to the upgrade to newest regalloc2, but that's possibly just because of a shift in code size due to different spills/moves and I suspect the root issue is still present in the riscv64 backend.