[
https://issues.apache.org/jira/browse/GROOVY-12161?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18095931#comment-18095931
]
ASF GitHub Bot commented on GROOVY-12161:
-----------------------------------------
github-actions[bot] commented on PR #2706:
URL: https://github.com/apache/groovy/pull/2706#issuecomment-4959673646
### JMH summary — classic (commit `5dfdfba`)
Speedup vs trailing 90-day baseline on gh-pages. Higher = faster.
`1.00` = in line with history. Per-benchmark ratio, geomean within group.
Time-per-op units inverted so direction is consistent. The *calibrated*
column divides out this runner's speed vs the baseline hardware, as
measured by Groovy-independent pure-Java ruler benchmarks.
| Group | Speedup | Calibrated | n |
|--------|---------|------------|---|
| bench | 0.969 × | 0.990 × | 84 |
| core | 1.123 × | 0.947 × | 77 |
| grails | 0.970 × | 0.878 × | 80 |
> ⚠️ Runner speed differs ≥15% from the historical baseline hardware for:
core-hz. Raw speedups are not meaningful for those parts — use the calibrated
column.
<sub>Runner calibration (this run vs baseline hardware): bench 0.98× (26
rulers) · core-ag 0.96× (3 rulers) · core-hz 1.50× (3 rulers) · grails-ad 1.07×
(3 rulers) · grails-ez 1.13× (3 rulers)</sub>
<sub>Baseline: <code>dev/bench/jmh/<part>/classic/data.js</code> on
gh-pages, trailing 90 days. <a
href="https://apache.github.io/groovy/dev/bench/jmh/summary.html";>Daily
dashboard</a> · <a
href="https://apache.github.io/groovy/dev/bench/jmh/";>Per-suite raw
data</a></sub>
<!
> Omit identity catch-all for try/catch without finally and slim exception
> bytecode
> ---------------------------------------------------------------------------------
>
> Key: GROOVY-12161
> URL: https://issues.apache.org/jira/browse/GROOVY-12161
> Project: Groovy
> Issue Type: Improvement
> Reporter: Daniel Sun
> Priority: Major
>
> h2. Summary
> {{StatementWriter.writeTryCatchFinally}} currently emits a catch-all identity
> rethrow ({{astore}} / {{aload}} / {{athrow}}) and an empty shared-finally
> path for *every* try/catch, including plain try/catch with no finally. That
> diverges from javac, bloats the exception table, and adds avoidable
> control-flow noise that is less JIT-friendly.
> h2. Problem
> For a plain try/catch (empty finally), the generator still:
> * registers a catch-all handler that only stores and rethrows the throwable
> * routes normal fall-through through an empty "finally" block and a second
> {{GOTO}} past the catch-all
> * always creates a catch-side {{BlockRecorder}} even when there are no catch
> clauses
> On abrupt exits ({{return}} / {{break}} / {{continue}}),
> {{CompileStack.applyBlockRecorder}} also emits a trailing {{NOP}} after
> inlined finally/synchronized guards, which is unnecessary once the restarted
> range can bind to the next real instruction.
> Non-empty finally must keep a real catch-all so uncaught exceptions still run
> finally then rethrow. Nested try/catch inside try/finally must keep
> {{BlockRecorder}} range splitting so an enclosing finally inlined on
> {{return}} is *not* covered by an inner typed handler (regression risk for
> GROOVY-8229). Stack-map casts for assignments inside try (GROOVY-9805)
> continue to depend on an active {{BlockRecorder}}.
> h2. Proposed approach
> * Keep a {{BlockRecorder}} on try/catch regions always (empty finally ⇒ no-op
> excluded statement) so range splits, nested finally semantics, and
> GROOVY-9805 casts stay correct.
> * When finally is empty: omit the catch-all identity rethrow and the empty
> shared-finally block; fall-through jumps straight to a join label after the
> handlers.
> * When finally is non-empty: keep typed handlers first, then catch-all +
> shared fall-through finally (existing semantics).
> * Only push a catch-side {{BlockRecorder}} when there is at least one
> {{catch}}.
> * In {{applyBlockRecorder}}, keep a leading {{NOP}} so abrupt-exit-only try
> ranges stay non-empty (illegal empty exception ranges), but drop the trailing
> {{NOP}}.
> h2. Expected bytecode shape (illustrative)
> Plain try/catch after change — typed handlers only, no catch-all:
> {code:java}
> int m(int x) {
> try {
> return x
> } catch (RuntimeException e) {
> return -1
> }
> }
> {code}
> Exception table should list only {{RuntimeException}} (possibly multiple
> ranges after return-path range splits), not a {{null}}/any handler, and must
> not contain an adjacent {{ASTORE}}/{{ALOAD}}/{{ATHROW}} identity rethrow.
> Try/finally must still have a catch-all any-handler so finally runs for
> uncaught exceptions.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
