Repository navigation
compiler-output-regression: native_owned_typed_views misses vectorization for an unexpected control_flow reason (Linux-only, pre-existing, blocking full-suite-gate) #8489
Description
compiler-output-regression fails its "Gate native-ABI proof compiler output" step on one workload:
native_owned_typed_views
vectorization_expectation:
observed_missed_reason_kinds: {"call_instruction": 4, "control_flow": 3,
"generic_not_vectorized": 7, "unknown_trip_count": 1}
unexpected_missed_reason_kinds: ["control_flow"]
min_vectorized_loops: 0, vectorized_count: 0, passed: false
control_flow is not in that workload's allowed_missed_reason_kinds, so three loops now miss vectorization for a reason the gate has not seen before.
Scope / attribution
- Pre-existing, and not from the recent merge batch. It fails identically on the attribution baseline
3627657c7(run 32368807258), which contains none of fix(hir): ambientdeclare function requiremust not shadow the require intrinsic #8452/fix(link): dedup evidence must match what the linker can actually resolve (#8455) #8458/test(parity): ratchet the matrix-trend baseline over the #8271 dark debt #8461/ci: give full-tier jobs timeout margin over observed contention #8462/fix(runtime): extend #8416's C-unwind to the whole closure dispatch family (#8463) #8464, and it still fails after revert: back out #8464's blanket C-unwind conversion (measured +20 crashes, gc-stress) #8484 reverted fix(runtime): extend #8416's C-unwind to the whole closure dispatch family (#8463) #8464 (run 32377162076). So it is not caused by theextern "C-unwind"unwind edges, which was the obvious first suspect. - Not reproducible on macOS. Running
scripts/compiler_output_regression.py suite --suite native-abi-proof --gatelocally against aperry-devbuild reportsfailed_workloads: []. This appears to be specific to the Linux/release configuration the job uses. - The likeliest origin is the recent string-codegen work (refactor(stdlib): centralize string header readers #8448 / perf(codegen): fuse String.concat arguments into concat chain #8450 / fix(ui): replace borrowed runtime strings with owned copies #8453 / perf(codegen): append strings through variable cells #8454), which is the codegen change in the window, but that is not established — attributing it wants a bisect on Linux.
Why this is a low-severity gate failure: this workload sets min_vectorized_loops = 0 with scalar_baseline = "allowed: native-owned typed view fixture gates proof records and fallback reasons". It does not require vectorization; the reason list is a change detector. Sibling workloads (native_abi_packet_control, native_pod_layout_constants) already list control_flow as acceptable.
It is, however, blocking the release: compiler-output-regression is one of full-suite-gate's required jobs.
Proposed: widen this workload's allowed_missed_reason_kinds to include control_flow, matching its siblings, with a comment pointing here — and keep this issue open to attribute the codegen change that introduced it. That trades a precise change-detector for an unblocked release, so it should be reverted once the cause is known and either fixed or accepted deliberately.
Activity
ci: accept control_flow as a missed-vectorization reason for native_o…
ci: accept control_flow as a missed-vectorization reason for native_o…
proggeramlug commentedon Aug 26, 2026
Attributed:
40e2ce86d(#8457, "make-Osthe default native optimization level"). Recorded in #8854.Verdict: legitimate, not a codegen regression. Perry's emitted IR is byte-identical across that commit (same sha256 at
526e0b502,f14a9e262and40e2ce86d) — only the clang args changed,-O3→-Os. The harness's remark pass reuses the compile's own args, so the analysis moved to-Ostoo, where LLVM declines the SimplifyCFG speculation and if-conversion that previously flattened those branches into selects before the vectorizer. The loops genuinely carry control flow, and vectorizing them would grow code size — exactly what-Osasks for. So #8490's widening was the right call; #8854 only replaces its rationale.Three claims in this issue were wrong, and they are why attribution stalled:
cargo build -p perry→--perry target/debug/perry(test.yml:1809/:1821) — a debug build. The macOS run that "did not reproduce" used--profile perry-dev(opt-level 1). That is a different profile, not a different OS. It reproduces on macOS with a CI-matching debug build.control_flow=3already at5876c9cdd, which precedes refactor(stdlib): centralize string header readers #8448/perf(codegen): fuse String.concat arguments into concat chain #8450/fix(ui): replace borrowed runtime strings with owned copies #8453/perf(codegen): append strings through variable cells #8454. Those are exonerated.The circularity worth noting:
40e2ce86dis an ancestor of the3627657c7attribution baseline. Testing against that baseline could only ever report "pre-existing", which is what sent the search toward the wrong window.Separately, this went unnoticed for a reason worth its own fix: the native-ABI step did not execute between Jul 30 and Aug 17, because an earlier step in the same job was failing and a failing
run:aborts the rest. Filed as #8855.Closing — the cause is identified, the workaround in #8490 is confirmed correct, and the rationale now records why.
proggeramlug commentedon Aug 26, 2026
Attributed to 40e2ce8 (#8457); rationale corrected in #8854.
ci: refresh three stale compiler-output-regression expectations (#10784)