[sad75]

Summary

Cut the per-call overhead of Execute/Evaluate for small scripts.

Details

Problem

For a small script on a warm engine, the cost of Evaluate is dominated by the work around the script, not the script itself. engine.Evaluate("a + b") allocates 312 B per call on main; 248 B of it is avoidable, and so is part of the time. Four steps run unconditionally:

#	Where	What it does for nothing
1	ExecuteForCompletion	Allocates a closure and a Func<T> delegate (120 B) on every call, to pass four locals to ScriptEvaluation.
2	ScriptEvaluation	Calls Debugger.OnBeforeEvaluate. Reading Debugger creates a DebugHandler on every engine, even when nothing uses debugging.
3	RunAvailableContinuations	Takes the host-call claim and walks the drain after every script, even when there is no job, timer, Atomics.waitAsync waiter or rejection.
4	GlobalDeclarationInstantiation	Allocates two Lists and a HashSet (128 B) before checking whether the script declares anything.

Fix

1. ExecuteForCompletion passes its state in a readonly record struct ScriptEvaluationState to a static callback, through the existing ExecuteWithConstraints<TState, TResult> overload. No allocation.
2. BeforeEvaluate is raised only if (_debugger is not null || _isDebugMode). A host can only subscribe by reading engine.Debugger, which creates the handler, so subscribers are still notified; debug mode still creates it at script entry.
3. RunAvailableContinuations returns first if IsContinuationDrainNoOp() (new, in Engine.Pump.cs): no queued job and no task lane, no Atomics.waitAsync deadline, no pending rejection notification and, on net8.0+, _webApi is null. Otherwise the drain runs as before.
4. GlobalDeclarationInstantiation returns right after MarkScriptEvaluated when there is no function, var or lexical declaration, and no Annex B block function that could bind (sloppy mode only).

Engines with web-API engine state (timers, scheduler, idle callbacks) never take shortcut 3. Module evaluation (Engine.Modules.cs) still reads Debugger directly; left out of scope.

Linked issue

None.

Test plan

• Added or updated unit tests in Jint.Tests
• Ran dotnet test --configuration Release locally
• For ECMAScript spec changes: ran Jint.Tests.Test262 — not a spec change
• For interop changes — not applicable
• For perf changes: included before/after numbers from Jint.Benchmark

New tests, one per behaviour a shortcut must keep:

• EngineTests.BeforeEvaluateFiresForAHandlerSubscribedAfterEarlierEvaluations
• EngineTests.ADrainAfterAnIdleDrainStillRunsNewlyQueuedJobs
• PromiseRejectionNotificationTests.ARejectionFromAScriptThatQueuesNoJobIsReportedOnce
• AnnexBBlockFunctionTests.AScriptWhoseOnlyDeclarationIsABlockFunctionStillGetsTheGlobalAlias (source and Prepared<Script>)
• AnnexBBlockFunctionTests.AStrictScriptWhoseOnlyDeclarationIsABlockFunctionGetsNoGlobalAlias (directive and Options.Strict)

Jint.Tests (39,239 passed) and Jint.Tests.PublicInterface (11,085 passed) are green on net472, net8.0 and net10.0.

Benchmark

New TinyScriptEntryBenchmark: three tiny scripts on a warm engine, plus a FunctionBody control row (Invoke) that none of the changes reach. measure-paired.ps1, 8 alternating rounds, baseline main at 4fd6bb9:

Row	Time Δ (median, 95% CI)	Allocated: main → PR
Expression (a + b)	−29.3 % [−30.4, −27.6]	312 B → 64 B
Comparison (o.x > a && b !== 0)	−19.0 % [−24.9, −16.8]	312 B → 64 B
Declaration (var c = a + b; c)	−13.3 % [−18.4, −10.2]	400 B → 280 B
FunctionBody (control)	+1.7 % [−0.9, +4.4], no change	112 B → 112 B

Fixes 1 and 4 account for the bytes (120 B + 128 B; Declaration declares a var, so it only gets fix 1). Fixes 2 and 3 save time only on a warm engine; fix 2 also saves one DebugHandler per engine, which this benchmark does not measure.

Run on GitHub-hosted ubuntu-latest (AMD EPYC 7763, 2 vCPU), BenchmarkDotNet 0.15.8, .NET 10.0.13. The benchmark file was copied into the baseline tree and BenchmarkDotNet's build timeout raised in both trees; nothing else in the harness changed. The SunSpider/Dromaeo gate was not run: these changes only touch per-entry setup, which a long script crosses once. Happy to run it if wanted.

Breaking change?

No. All changes are internal.