[lahma]

AOP proxies didn't always undo what they set up when the advised call threw. This fixes both cases.

Pooled and borrowed targets.

• The emitted proxy method (TargetAopProxyMethodBuilder, used by both composition and decorator proxies) called ReleaseTarget without a finally. A throwing call left a SimplePoolTargetSource object checked out, so a small pool ran dry and blocked.
  • The method logic now runs in try { … } finally { TargetSource.ReleaseTarget(target) }.
  • The return value and ref/out results go through locals, and the single ret comes after the block.
• BaseCompositionAopProxy.Equals, GetHashCode and ToString borrowed and released the target the same way. They now use try/finally.
• AbstractAutoProxyCreator.CreateProxy called targetSource.GetTarget() into a local that nothing read. For a pooled source that object was never returned. For a prototype or lazy source it created or initialised a target too early. The borrow is gone.
  • Behaviour note: auto-proxy creation no longer touches the target. A broken target definition behind a custom target source now fails on the first call, not when the container starts.

The exposed proxy.

• With ExposeProxy set, BeginMethod pushed the proxy onto the thread's AopContext stack and EndMethod popped it, again without a finally. Each throwing call left one entry behind, so AopContext.CurrentProxy could return the wrong proxy.
• The pop now runs in GenerateMethod's outer finally for the composition, decorator, inheritance and introduction builders. The push happens just before the try.
• ExposeProxy is read once per call instead of twice, so the pop always matches the push.
• The two finally blocks unwind in reverse order of acquisition: the target is released first, then the proxy is popped.

Compatibility. Public signatures don't change, but the protected contract of the public IL builders does, for anyone who writes their own:

• GenerateMethodLogic now runs inside protected regions. It must leave the stack empty and not emit ret.
• EndMethod no longer pops.

The XML docs state this. There is no BreakingChanges entry, because the change only affects custom IL builders. The happy path costs the same or less per call.

Tests.

• New ReleaseTargetTests covers composition and decorator proxies, each with and without an interceptor:
  • get/release balance with exact counts, for void, reference, int and struct returns, ref/out parameters and generic methods;
  • a throwing target and a throwing advice;
  • a real SimplePoolTargetSource with MaxSize = 1;
  • the composition Equals/GetHashCode/ToString;
  • AopContext balance after throwing calls, including nested proxies and introduced methods.
• InheritanceAopProxyTests gets an exposed-proxy case.
• AutoProxyTargetSourceBorrowTests checks that proxy creation borrows nothing.
• Every case that exercises a throwing path fails without the fix, and all of them pass with it.
• Spring.Aop.Tests: 555 on net10.0 and 573 on net472, 0 failed. Spring.Core, Data, Web, Mvc5 and Testing.NUnit tests are green, apart from the locale-dependent FormatUsingDefaults.
• spring-reviewer checked the IL block structure, stack balance, branches inside the regions and the hot path. It accepted the change with follow-ups, which are in the last commit.

Found during the analysis of #239; that leak itself was fixed in #347.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LGkLQuAuUjvRjrU1AcTsmk