Avoid Kotlin reflection in MethodParameter for non-suspending functions

[gregjotau]

MethodParameter.getGenericParameterType() and getParameterType() resolve the return type of a Kotlin method (parameter index -1) through KotlinDelegate, which calls ReflectJvmMapping.getKotlinFunction(method) only to check KFunction.isSuspend(). That lookup materializes all members of the declaring KClass through Kotlin reflection (metadata deserialization plus a linear scan comparing javaMethod per member) and is repeated for every method whose return type is resolved.

A suspending function always compiles to a JVM method whose last parameter is kotlin.coroutines.Continuation, which KotlinDetector.isSuspendingFunction(Method) already checks without Kotlin reflection; the rest of the framework (AOP, caching, scheduling, messaging, ConstructorResolver) relies on that check. This change uses it as a guard in KotlinDelegate.getGenericReturnType(...) and getReturnType(...), so regular Kotlin methods fall back to Method.getGenericReturnType() / Method.getReturnType() immediately. Suspending functions keep the existing KFunction based resolution, and a non-suspending function that declares a trailing Continuation parameter explicitly is still verified through KFunction.isSuspend(), so results are unchanged in every case.

Measurements

Cold-JVM micro-benchmark that constructs new MethodParameter(method, -1) and calls getGenericParameterType() and getParameterType() for the 1,418 non-synthetic methods of the 337 Kotlin repository interfaces of a large Spring Boot 4.2.0-SNAPSHOT / Framework 7.1.0-SNAPSHOT application (JDK 27, Apple M5 Max, 5 runs each):

	main	This change
Return type resolution, 1,418 methods	1,584–2,766 ms	57–150 ms

The baseline number includes Kotlin reflection bootstrap, which the patched path avoids entirely; in an application that touches Kotlin reflection elsewhere the saving is correspondingly smaller. In a wall-clock startup profile of that application (async-profiler, 2 ms sampling, extracted Boot layout, 243 Spring Data JPA repositories), MethodParameter$KotlinDelegate.getGenericReturnType / getReturnType accounted for 124 of 8,694 main-thread samples (1.4 % of startup), all below ReflectJvmMapping.getKotlinFunction; with this change the frames no longer appear. Spring Data's KotlinReflectionUtils.isSuspend has the same pattern; I have proposed the equivalent guard there in spring-projects/spring-data-commons#3545.

Verification

MethodParameterKotlinTests gains two tests that pin the return type resolution of a regular function and of a regular function that declares a Continuation parameter explicitly; both pass on main and on this branch. ./gradlew :spring-core:check passes on JDK 25, including the JDK 21 and JDK 24 test suites, checkstyle and architecture checks. Related to #21546.

[gregjotau]

Handing this over to the team: please feel free to take the change from here in whatever form suits you (team commit, different shape, or a different guard). We don't need authorship or ownership of the code; what we care about is the startup cost, so use it as you see fit.

We have tested it thoroughly against reai.no, a large Spring Boot Kotlin accounting application, with the profiles and numbers in the description above. The signed-off commit stays reachable at refs/pull/37299/head after I remove my fork, and the description contains everything needed to reproduce the measurements.