[
https://issues.apache.org/jira/browse/GROOVY-11985?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18080616#comment-18080616
]
Paul King commented on GROOVY-11985:
------------------------------------
[~blackdrag] Good catch — the wording in that bullet is ambiguous. "private"
there describes the *callee* (the method being invoked), not the caller. The
transform decision in {{TraitReceiverTransformer.transformMethodCallOnThis}}
only inspects the resolved target's modifiers; the visibility of the enclosing
method doesn't enter.
So your scenario — a private static {{m1}} in the trait calling a public static
{{m2}} that exists only in the implementer — works:
{code:groovy}
trait T {
private static String m1() { m2() } // m1 is private
static String entry() { m1() }
}
class A implements T {
static String m2() { 'A.m2' } // only on the implementer
}
assert A.entry() == 'A.m2'
{code}
The call {{m2()}} inside {{m1}}'s body is rewritten by looking up {{m2}} in
{{T}} — it isn't there, so we fall through to the {{$static$self.m2(args)}}
path and resolve to {{A.m2}} at runtime. The "private static dispatch to the
helper" rule applies to call sites whose *target* is a private static, not to
call sites that happen to live inside one.
I'll reword the bullet to remove the ambiguity, e.g.:
{quote}A call whose target is a *private* trait static is not rewritten through
the implementing class; it dispatches to the trait helper directly. The
override pattern is intentionally unavailable for private statics — they aren't
composed onto the implementer.{quote}
The decision table for {{this.m(args)}}/{{m(args)}} inside a trait body, keyed
on properties of the *callee*:
||#||{{m}} found in trait?||static/instance||callee visibility||call
site||rewritten form||impl override wins?||
|1|yes|static|public|trait method body|{{((Class)$self.getClass()).m(args)}} +
{{DO_DYNAMIC}}|*yes*|
|2|yes|static|public|inside a closure|{{T$Helper.m($static$self, args)}}|no|
|3|yes|static|private|anywhere|{{T$Helper.m($static$self, args)}}|no|
|4|no (only on impl/super)|static|—|trait method body|{{$static$self.m(args)}}
(fallback)|*yes*|
|5|no|static|—|inside a closure|{{$static$self.m(args)}} via {{thisExpr}}|*yes*|
|6|yes|instance|any|trait method body|{{this.m($self, args)}} (existing)|*yes*
at runtime|
Row 1 is the only new behaviour from this PR. Rows 4–6 were already dynamic
pre-PR. Rows 2/3 are the deliberately conservative cells: closures stay on the
helper path so trait closure bodies remain compile-time checkable, and private
statics stay on the helper because they aren't overridable to begin with.
Orthogonal to all of the above is the {{B extends A implements T}} axis from
your earlier question: even when row 1 dispatches dynamically, the dispatch
anchors on the *direct implementer* ({{A}}), not on the receiver class ({{B}}).
So {{B.m1()}} still picks {{A}}'s {{m2}}, not {{B}}'s, unless {{B}} also
redeclares {{m1}} (which forces re-composition).
> Static method override on trait implementer ignored when called via this in
> trait body
> --------------------------------------------------------------------------------------
>
> Key: GROOVY-11985
> URL: https://issues.apache.org/jira/browse/GROOVY-11985
> Project: Groovy
> Issue Type: Bug
> Reporter: Paul King
> Priority: Major
>
> Analysis by AI of the description in the Grails canary build comments yielded
> the following:
> h2. Summary
> In Groovy 4.x, calling {{this.someStaticMethod()}} from inside a trait's
> static method body dispatched dynamically and honoured a static method
> override declared on the implementing class. In Groovy 5.x and
> 6.0.0-SNAPSHOT, the same call is rewritten by {{TraitReceiverTransformer}} to
> dispatch through the trait helper, which can never see the implementing-class
> override. The override is silently lost — no exception, no compile warning,
> the trait's default value just always wins.
> This was discovered as part of the Grails 8 / Groovy 5 migration
> (apache/grails-core PRs
> [#15557|https://github.com/apache/grails-core/pull/15557] and
> [#15558|https://github.com/apache/grails-core/pull/15558]) and motivated a
> reflection-based workaround in
> {{{}Validateable.resolveDefaultNullable(Class){}}}.
> h2. Reproducer
> Standalone repro:
> [https://github.com/jamesfredley/groovy-trait-static-method-override-bug]
> {code:java|title=Validateable.groovy}
> trait Validateable {
> static boolean defaultNullable() {
> false
> }
> static boolean defaultNullableSeenByTrait() {
> // expected to dispatch to the implementing class override
> this.defaultNullable()
> }
> }
> {code}
> {code:java|title=MyNullableValidateable.groovy}
> class MyNullableValidateable implements Validateable {
> static boolean defaultNullable() {
> true
> }
> }
> {code}
> {code:java|title=Driver}
> assert MyNullableValidateable.defaultNullable() // direct call:
> true on every version
> assert MyNullableValidateable.defaultNullableSeenByTrait() // expected true;
> gets false on 5.x / 6.0
> {code}
> h2. Observed behaviour
> ||Groovy version||Direct call||{{this.defaultNullable()}} from trait body||
> |4.0.27|true (PASS)|true (PASS)|
> |5.0.5|true (PASS)|*false (FAIL)*|
> |6.0.0-SNAPSHOT|true (PASS)|*false (FAIL)*|
> h2. Root cause
> The bytecode emitted for the trait helper's
> {{defaultNullableSeenByTrait(Class)}} method changed shape:
> {noformat}
> // Groovy 4.0.27 // Groovy 5.0.5 /
> 6.0.0-SNAPSHOT
> 0: aload_0 0: ldc //
> Validateable$Trait$Helper.class
> 1: invokedynamic invoke: 2: aload_0
> (Ljava/lang/Class;)Ljava/lang/Object; 3: invokedynamic invoke:
>
> (Ljava/lang/Class;Ljava/lang/Class;)Ljava/lang/Object;
> {noformat}
> In 4.x the indy receiver is {{aload_0}} — the implementing class — so the
> dynamic dispatch resolves {{defaultNullable()}} against
> {{MyNullableValidateable}} and finds the override. In 5.x+ the receiver is
> hard-coded {{Validateable$Trait$Helper.class}} (an {{{}ldc{}}}) and the
> implementing class is demoted to an argument, so the dispatch resolves
> {{defaultNullable(Class)}} on the trait helper itself and always lands on the
> lowered trait default.
> The behaviour change was introduced by commit {{0aa78d0a33}} (GROOVY-8854,
> Sep 2023):
> {quote}write {{T.m(p)}} as {{this.m($static$self,p)}} not {{$self.m(p)}}
> {quote}
> That commit rewrote {{TraitReceiverTransformer.transformMethodCallOnThis}} so
> that {{this.someStaticMethod()}} inside a trait body is rewritten as {{(this
> | T$Trait$Helper).m((Class)$self.getClass(), args)}} — routed through the
> trait helper's lowered static, with the implementing class passed as the
> {{$static$self}} argument. The existing trait static method test coverage in
> {{TraitASTTransformationTest.testTraitStaticMethod}} (including the
> GROOVY-8854 case at line 2218) doesn't exercise the
> override-on-implementing-class scenario, so the regression wasn't caught.
> h2. Tradeoff
> Groovy 5's behaviour is arguably closer to Java semantics: static methods are
> not virtual, and {{this}} inside a Java static method doesn't exist, so
> "{{{}this.staticMethod(){}}} virtually dispatches to the implementing class's
> static" was always Groovy-specific magic that relied on MOP. However:
> * The Groovy 4 behaviour was depended on by real code — Grails
> {{Validateable}} is the visible canary; the same idiom likely exists
> elsewhere.
> * The failure mode is silent — no exception, no compile warning, the trait
> default just wins every time.
> * The change surfaced as a side effect of GROOVY-8854 (whose ticket was
> about something else), not as a deliberate "trait statics are no longer
> virtual" decision, and it isn't called out in the Groovy 5 release notes.
> h2. Suggested options
> # Restore Groovy 4 semantics for {{this.staticMethod()}} in trait bodies —
> emit a dynamic lookup whose receiver is {{$static$self}} (the implementing
> class) rather than the trait helper.
> # At minimum, emit a compile-time warning when a trait body calls a
> same-named static and the dispatch will provably not hit any override on the
> implementing class.
> # Document the new contract explicitly in the Groovy 5 release notes and
> trait docs, so consumers can adapt at the trait level instead of debugging
> silent no-ops.
> h2. Workaround (already applied in Grails)
> Bypass {{TraitReceiverTransformer}} via plain Java reflection, which it can't
> see through:
> {code:groovy}
> private static boolean resolveDefaultNullable(Class<?> clazz) {
> try {
> return clazz.getMethod('defaultNullable').invoke(null) as boolean
> } catch (NoSuchMethodException ignored) {
> return false
> }
> }
> {code}
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
