Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 11 additions & 0 deletions changelog.d/7822-dispatch-ic-retry-can-fire.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,11 @@
**One of #7365's two named flaky tests is fixed**: `obj_dispatch_ic_tests` goes from **10 of 12** isolated runs failing to **0 of 20**.

`with_stable_gen` exists because dispatch-IC entries are keyed on `VTABLE_GEN`, and any class registration anywhere retires the cache — including `a_class_registration_invalidates_every_entry` in the *same module*, which calls `test_bump_vtable_generation()` on purpose. So an insert/lookup pair can straddle a bump and miss for a reason that has nothing to do with what is being asserted. The helper was written to retry in exactly that case.

**The retry could never fire.** `body` asserted internally, so a straddled pair panicked on the *setup* assertion ("the entry we just inserted must be findable") before the loop got to re-check the generation. The 64-retry budget was never spent.

`body` now returns whether its observations were valid: `false` means "a bump landed mid-pair, nothing was learned", which is a retry rather than a failure. The assertions the tests exist for — a different name at the same address must miss, a different class id must miss, an over-long name must not alias a truncated key, a prefix must not hit — stay assertions.

**Why this matters beyond one test, and why it looked so strange.** The failure rate *inverted with scope*: running only this module put its five tests — including the deliberate bumper — on threads together and failed **10 of 12**, while the full 2000-test suite spread them apart and failed about **1 in 6**. So the standard triage move, re-running just the failing test, made an intermittent test look reliably broken, and running the suite made it look nearly fine. Both readings were wrong, and #7365 is largely a catalogue of exactly that confusion.

Measured after the fix: **0/20** isolated runs fail. The full suite still fails about 1 in 10, but on a different test — `gc::tests::root_words::bare_address_in_global_root_survives_a_real_collection`, a sibling of the `bare_address_in_shadow_slot_…` case already measured at 2/10 on clean `main`. That family is untouched here and remains #7365's other half; it is a GC-timing intermittency rather than a test-isolation one, so it needs a different fix.
53 changes: 36 additions & 17 deletions crates/perry-runtime/src/object/class_registry/dispatch.rs
Original file line number Diff line number Diff line change
Expand Up @@ -526,18 +526,34 @@ mod obj_dispatch_ic_tests {

const CID: u32 = 61_001;

/// Run `body` on a stable vtable generation.
/// Run `body` on a stable vtable generation, retrying if it straddled a bump.
///
/// Entries are keyed on `VTABLE_GEN`, and the whole point of that key is
/// that ANY class registration anywhere retires the cache. Sibling tests in
/// this crate register classes concurrently, so an insert/lookup pair can
/// straddle a bump and miss for a reason that has nothing to do with what
/// is being asserted. Retry until the pair runs inside one generation.
fn with_stable_gen(body: &dyn Fn()) {
/// that ANY class registration anywhere retires the cache. Sibling tests
/// register classes concurrently — `a_class_registration_invalidates_every_entry`
/// in THIS module calls `test_bump_vtable_generation()` on purpose — so an
/// insert/lookup pair can straddle a bump and miss for a reason that has
/// nothing to do with what is being asserted.
///
/// #7365: the retry existed but could not fire. `body` asserted internally,
/// so a straddled pair **panicked on the setup assertion** before the loop
/// got to re-check the generation, and the retry budget was never spent.
/// That is why the failure rate INVERTED with scope: running just this
/// module (`--lib obj_dispatch_ic_tests`) put its five tests — including
/// the deliberate bumper — on threads together and failed 10 of 12 runs,
/// while the full 2000-test suite spread them out and failed about 1 in 6.
/// A filtered re-run, the standard triage move, therefore made an
/// intermittent test look reliably broken.
///
/// So `body` now REPORTS whether its observations were valid instead of
/// asserting them: `false` means "a bump landed mid-pair, nothing was
/// learned", which is a retry rather than a failure. The assertions the
/// tests actually exist for stay assertions.
fn with_stable_gen(body: &dyn Fn() -> bool) {
for _ in 0..64 {
let before = VTABLE_GEN.load(Ordering::Acquire);
body();
if VTABLE_GEN.load(Ordering::Acquire) == before {
let observed = body();
if observed && VTABLE_GEN.load(Ordering::Acquire) == before {
return;
}
}
Expand All @@ -557,30 +573,31 @@ mod obj_dispatch_ic_tests {
// different name through the SAME backing storage.
let mut scratch = *b"area\0\0\0\0";
obj_dispatch_ic_insert(CID, &scratch[..4], 0xAAAA, 1, false, false);
assert_eq!(
obj_dispatch_ic_lookup(CID, &scratch[..4]),
Some((0xAAAA, 1, false, false)),
"the entry we just inserted must be findable"
);
// Setup, not the subject: a miss here means a sibling bumped the
// generation between insert and lookup. Report it and retry.
if obj_dispatch_ic_lookup(CID, &scratch[..4]) != Some((0xAAAA, 1, false, false)) {
return false;
}

scratch[..4].copy_from_slice(b"perim"[..4].try_into().unwrap());
assert_eq!(
obj_dispatch_ic_lookup(CID, &scratch[..4]),
None,
"a different name at the same address must MISS"
);
true
});
}

#[test]
fn a_hit_requires_the_matching_class_id() {
with_stable_gen(&|| {
obj_dispatch_ic_insert(CID, b"describe", 0xBBBB, 1, false, false);
assert_eq!(
obj_dispatch_ic_lookup(CID, b"describe"),
Some((0xBBBB, 1, false, false))
);
if obj_dispatch_ic_lookup(CID, b"describe") != Some((0xBBBB, 1, false, false)) {
return false;
}
assert_eq!(obj_dispatch_ic_lookup(CID + 1, b"describe"), None);
true
});
}

Expand All @@ -605,6 +622,7 @@ mod obj_dispatch_ic_tests {
"an over-long name must fall through to the tower, not alias a \
truncated key"
);
true
});
}

Expand All @@ -615,6 +633,7 @@ mod obj_dispatch_ic_tests {
with_stable_gen(&|| {
obj_dispatch_ic_insert(CID, b"describe", 0xEEEE, 1, false, false);
assert_eq!(obj_dispatch_ic_lookup(CID, b"describ"), None);
true
});
}
}
Loading