perf(runtime): compress array forwarding chains - #8674
Conversation
|
Important Review skippedAuto reviews are disabled on base/target branches other than the default branch. Please check the settings in the CodeRabbit UI or the ⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Plus Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
…ape guards (#8682) Lands three reviewed PRs as one squash: #8674, #8675, #8676. - #8674: compress array forwarding chains (`clean_arr_ptr` multi-hop walk). - #8675: share numeric guards across dynamic add trees. - #8676: pack monomorphic method shape guards. These three were authored as a stack on top of #8672, but their contents touch disjoint files, so they are cherry-picked onto main on their own. #8672 is NOT included: it defines its own `is_bound_native_method_closure_value` (true for any bound native-module export with a non-empty module name), which #8662 superseded on main with the strictly narrower `is_bound_native_constructor_closure_value` (gated on explicit constructor metadata). Those predicates have different truth sets, so the substitution is a behavioural change at every call site and is left to the author to rebase. Also splits `array/tests.rs`, which #8674 pushed over the 2000-line cap, into an `array/forwarding_tests.rs` sibling. Pure relocation. Version bump not included per maintainer policy. Co-authored-by: Ralph Küpper <ralph@skelpo.com>
|
Landed on Your branch was stacked on #8672, but the contents touch disjoint files, so this commit was cherry-picked onto Validated on the merged result: 9 ratchet gates + One adjustment: #8674's new test pushed |
1 participant
Summary
Why
Array growth preserves stale aliases with forwarding stubs. A receiver retained across multiple capacity-crossing grows can therefore point through two or more stubs. Generated indexed-access guards intentionally heal one forwarding edge inline; after that single hop they still see
GC_FLAG_FORWARDED, reject the array guard, and enter generic indexed lookup on every access.clean_arr_ptralready performs the complete bounded, tracked-array-validated walk. Once that walk succeeds, this change points the original retained stub directly at the live head. Subsequent generated accesses then need only their existing one-hop healing path. This is central forwarding metadata maintenance rather than an ECS- or benchmark-specific shortcut.This PR is stacked on #8672. Please review/merge that PR first; the runtime and test change is commit
514e0f553, followed only by its changelog fragment.Correctness
cargo fmt --all -- --checkcargo test -p perry-runtime: 2,644 passed, 0 failed, 4 ignoredPERRY_GC_FORCE_EVACUATE=1 PERRY_GC_VERIFY_EVACUATION=1The complete ECS suite under forced evacuation verification still reports the same pre-existing stale heap-field pointer failure on both the #8672 control and this candidate. Normal ECS execution passes, and the focused forced-evacuation fixture plus runtime GC regression pass.
Performance
Apple M1 Mac mini, AC power, Node 26.5.1. Each comparison followed a 60-sample quiet-host gate (all samples at or below 25% active CPU), used
taskpolicy -t 0 -l 0, alternated process order, and ran 11 pairs.Against the unchanged #8672 binary (
repeat=64):Directly against Node 26.5.1 (
repeat=256):This removes the repeated generic indexed-lookup pathology. It is a large improvement, but it does not claim Node parity; the remaining steady-state generated loop cost is separate follow-up work.