GCC-14 + GitHub-hosted ubuntu-24.04 runner: SIGILL in benchmark inner loops under -O3

[auscaster]

Summary

The benchmarks job in .github/workflows/ci.yml switched from gcc-14 to clang-18 in commit 81cd0e4 because the GCC build of signalbench and packetstreambench was crashing with SIGILL (signal 4, exit code -4 from the Python runner) on the GitHub-hosted ubuntu-24.04 runner. This issue tracks the root cause.

Symptom

Build succeeds, binary runs to its hot loop, then dies with SIGILL.

[100%] Running icey microbenchmarks and writing bench/benchmarks.json
RuntimeError: /home/runner/work/icey/icey/build-bench/base/bench/signalbench failed with exit code -4

Scope

• Affects signalbench and packetstreambench. Both exercise hot loops over Signal::emit / PacketStream emission.
• Compiler-specific: GCC 14.2.0 with -DCMAKE_BUILD_TYPE=Release (-O3 -DNDEBUG).
• Platform-specific: only reproduces on the GitHub-hosted ubuntu-24.04 runner. A local Linux GCC 14 container with the same source and flags runs cleanly. macOS clang (Apple) and clang-18 on the same GitHub runner both pass.
• Not seen in the regular Linux test jobs (g++-14, clang++-18), the address/thread/undefined sanitizer jobs, the WebRTC + FFmpeg job, the macOS jobs, or the Windows MSVC job.
• First failure: 2026-04-21 (commit d3b3fee). Pre-dates all M4 work.

What was tried

• Per-target -O2 -fno-tree-vectorize on signalbench (commit 3a589aa) — fixed signalbench, but packetstreambench then hit the same SIGILL.
• Generalised -O2 -fno-tree-vectorize to all benches via icy_add_benchmark (commit 26a4c27) — packetstreambench still hit SIGILL.
• Switched the bench job from gcc-14 to clang-18 (commit 81cd0e4) — passes.

The library-level GCC fallback in icy_add_benchmark (-O2 -fno-tree-vectorize) stays as defence in depth for anyone building bench binaries with GCC outside CI.

What's needed

Reproduce on a GCC-14 + ubuntu-24.04 runner profile, capture a core dump (ulimit -c unlimited) and objdump the failing instruction. Likely candidates given the symptom:

• GCC's auto-vectorizer or unswitch-loops pass emitting an instruction the runner CPU does not implement.
• A miscompile around the hot member-function-pointer call inside Signal::emit (the recent linked-list-based slot storage in commit 0bed357 changed this hot path).
• A GCC trap (ud2) for some new UB-detection path under -O3 interacting with the bench's tight loop.

Files

• src/base/include/icy/signal.h
• src/base/bench/signalbench.cpp
• src/base/bench/packetstreambench.cpp
• cmake/icey_modules.cmake (icy_add_benchmark, GCC fallback flags)
• .github/workflows/ci.yml (benchmarks job)

Workaround in place

Bench job runs under clang-18. icey itself builds and tests cleanly under gcc-14 in the regular Linux job and under all sanitizers, so consumers compiling against icey with gcc are not affected.