Repository navigation
Investigate macOS arm64 NativeAOT RTCP argument corruption #9
Description
A local osx-arm64 NativeAOT run of the existing whole-root smoke fails in RtcpTests.Schedule(). Managed Release passes (63 protocol and 321 network cases); the independent final Pion peer suite passes 450/450 cases. Linux NativeAOT is covered separately in CI.
Reproduced with SDK 10.0.301 and an isolated official SHA-512-verified SDK 10.0.402. The failure is independent of the Pion dependency update: RtcpTransmissionSchedule.cs and RtcpTests.cs are unchanged by PR #8. A standalone project linking only the scheduler and scheduling test also reproduces it.
The failing source statement is:
var delayedClock = new ClockSource();
var delayed = new RtcpTransmissionSchedule(100, 200, true, delayedClock, () => .5);Constructor instrumentation reports bandwidthBytesPerSecond == 0 at this call, after earlier constructors receive 100 and 1000 correctly. The valid source argument appears corrupted during native compilation/execution; the constructor correctly rejects the received zero. The same minimal project under NativeAOT size optimization also fails. This points to an AOT/toolchain issue but needs a smaller independent upstream reproducer and confirmation; do not weaken the scheduler guard or remove its tests.
Reproduction: publish src/tests/WebRTC.AotSmoke with -c Release -r osx-arm64 -p:PublishAot=true, then execute WebRTC.AotSmoke. The crash follows the simultaneous SCTP INIT checks and precedes the RTCP success message. Native publish completes with no IL trimming warnings; Apple static-library module-debug-cache warnings also appear and are a separate debug-symbol limitation.
Next: minimize the scheduler/test sequence for dotnet/runtime, compare on another physical arm64 Mac, and add a macOS native execution lane once the cause is understood. Preserve Linux x64/arm64 native execution and all protocol guards meanwhile.
Activity
HavenDV commentedon Oct 17, 2026
Fixed the reproduced repository failure in 57ebc00.
Native disassembly and LLDB tracing narrowed the failure to a saved floating-point register: Schedule loaded 100 into d8, the expected regular-send rejection overwrote d8 with zero, and the later constructor reused that register. Both original SDK 10.0.301 and isolated official 10.0.402 reproduced the issue. Moving the calculation to a throwing helper or replacing the monitor with System.Threading.Lock did not fix it.
Sent now returns expected admission failures normally from a non-inlined calculation frame, leaves the lock, and then throws the same InvalidOperationException with the same message. Successful sends retain the same lock and state updates; failed sends retain all validation and do not mutate the schedule. No test assertion, guard, trimming warning, or optimization was disabled. This is a targeted workaround for the native exception/unwind behavior; it does not claim to repair the .NET runtime generally.
The unchanged Schedule test now passes in the macOS arm64 NativeAOT executable. Added a permanent macOS arm64 whole-root NativeAOT publish-and-execute CI lane. All nine validation jobs passed, including managed protocols/network tests on macOS/Windows/Linux, Linux x64/arm64 AOT, both Pion peer lanes, Chromium, and the new macOS lane. Exact-source package publication succeeded: https://github.com/tryAGI/WebRTC/actions/runs/39040687587 . Local Release build, 63 protocol tests, and 321 network cases also passed.