ci(jit): an instrument for the crash in #726, before anyone edits the emitter - #731
Open
gHashTag wants to merge 1 commit into
Open
ci(jit): an instrument for the crash in #726, before anyone edits the emitter#731gHashTag wants to merge 1 commit into
gHashTag wants to merge 1 commit into
Conversation
… emitter `jit.test.JitCompiler dot product correctness` does not fail an assertion. It kills the test process: zig prints "the following test command failed" with no error name and no stack trace, because a fault inside emitted machine code never reaches zig's panic handler. Five sibling failures in the same log each carry a trace; this one carries nothing. So there is nothing to reason from, and the tests guard themselves against being reasoned about anyway -- each JIT test skips unless `builtin.cpu.arch == .x86_64`, which means the emitted code runs on the CI runner and nowhere else. It cannot be reproduced on an arm64 machine, and the pinned Zig 0.15.2 does not link there. This workflow produces the missing observation -- signal, faulting address, backtrace -- rather than a patch written blind. src/jit.zig imports the `vsa` module and so cannot be built standalone with `zig test`; the suite runs whole and the core is recovered afterwards. The runner pipes cores to apport by default, which drops them, so core_pattern is set to a plain path first. Its contract is deliberately about the instrument, not the subject: it fails when the process dies and no core was written -- an instrument that failed -- and reports plainly when the crash does not reproduce. Manual dispatch plus the two paths that can change the answer, so it never destabilises the main gate.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Produces the evidence #726 asks for. Closes nothing yet.
Why an instrument and not a fix
jit.test.JitCompiler dot product correctnessdoes not fail an assertion — it kills the test process. zig printsthe following test command failedwith no error name and no stack trace, unlike the five sibling failures in the same log which all carry one. A fault inside emitted machine code never reaches zig's panic handler.It also cannot be reproduced anywhere convenient. Each JIT test guards itself with
if (builtin.cpu.arch != .x86_64) return error.SkipZigTest(src/jit.zig:571, 596, 630, 667).ubuntu-latestis x86-64, so the guard does not fire there and the emitted code genuinely runs — on the CI runner and nowhere else. This machine is arm64, where it self-skips, and the pinned Zig 0.15.2 does not link on this macOS.Guessing at a segfault with no instrument is how a session becomes twenty hours of repairing hardware that was never broken.
What it does
src/jit.zigimports thevsamodule, sozig testcannot build it standalone — and recovers the core afterwards.kernel.core_patternto a plain path first: the runner pipes cores to apport, which discards them for a process it does not recognise.gdb -batchforbt full,rip/rsp/rbp, threads andx/8i $rip, with the exec path read from the core's own notes so the trace symbolises.test.logand every core.Its contract
It gates the instrument, not the JIT: it fails when the process dies and no core was written, because that is the instrument failing rather than the subject. When the crash does not reproduce it says so plainly instead of reporting green.
workflow_dispatchplus the two paths that can change the answer, so it never destabilisesCodegen Validation— which only just reached green for the first time.