perf: inline FastMaterializeJsonRenderer hot methods

[He-Pin]

Motivation

The std.manifestJsonEx function was 2.11x slower than jrsonnet (Rust implementation) on Scala Native. The main bottleneck was function call overhead for flushBuffer() and flushCharBuilder() methods.

Key Design Decision

Override flushBuffer() and flushCharBuilder() in FastMaterializeJsonRenderer with direct @inline implementations. Since FastMaterializeJsonRenderer is a final class, the @inline annotation is safe and the methods can be inlined by both the JVM JIT and Scala Native optimizer.

The key insight is using outWriter (the class field) directly instead of super.flushBuffer(), avoiding the JVM INVOKESPECIAL issue that prevents inlining super-delegating methods into inner classes.

Modification

Changed sjsonnet/src/sjsonnet/Renderer.scala:

• Override flushBuffer() with direct @inline implementation
• Override flushCharBuilder() with direct @inline implementation
• Use outWriter field directly instead of super delegation

Benchmark Results

Scala Native vs jrsonnet (hyperfine)

Test	Before	After	Improvement
manifestJsonEx	jrsonnet 2.11x faster	jrsonnet 1.56x faster	Gap reduced 26%
manifestTomlEx	jrsonnet 1.94x faster	jrsonnet 1.51x faster	Gap reduced 22%

JMH (JVM)

Baseline JMH benchmarks are stable; the change benefits all platforms.

Analysis

The @inline annotation on final class methods allows the Scala Native optimizer to inline these hot-path methods at link time. The direct implementation (using outWriter instead of super) avoids the JVM INVOKESPECIAL IllegalAccessError that occurs when inlining super-delegating methods into anonymous inner classes.

References

• Scala Native @inline documentation
• JVM INVOKESPECIAL specification

Result

• ✅ All tests pass (./mill __.test)
• ✅ Code formatted (./mill __.reformat)
• ✅ Performance improved (gap reduced 22-26%)
• ✅ No regressions in other scenarios