Scalar Replacement Research

A small research project exploring Escape Analysis and Scalar Replacement in JVM HotSpot using JMH and perfasm.

The project demonstrates how the JIT compiler can:

• completely eliminate object allocations;
• optimize non-escaping objects into scalar values;
• disable optimizations when object identity or escaping semantics are required.

The research was inspired by Aleksey Shipilëv’s article:

JVM Anatomy Quark #18: Scalar Replacement

Benchmarks

Scalar Replacement

@Benchmark
public int scalarReplacement() {
    Point p = new Point(x, 10);
    return p.y + p.x;
}

The object does not escape and does not require identity semantics.
HotSpot eliminates the allocation completely using Escape Analysis and Scalar Replacement.

Allocation Without Escape Analysis Optimization

@Benchmark
public int allocationWithoutEscape() {
    Point p = new Point(x, 10);
    return System.identityHashCode(p);
}

System.identityHashCode() requires real object identity, forcing HotSpot to materialize the object and preventing Scalar Replacement.

Allocation With Escape

@Benchmark
public Point allocationWithEscape() {
    return new Point(x, 10);
}

The object escapes from the method, so the allocation cannot be eliminated.

Benchmark Results

Benchmark	Result
scalarReplacement	~0.8 ns/op
allocationWithEscape	~3.9 ns/op
allocationWithoutEscape	~56 ns/op

The results show:

• Scalar Replacement can completely eliminate allocations;
• ordinary allocations in HotSpot are very cheap due to TLABs;
• object identity operations are significantly more expensive.

Running

mvn clean install

Run benchmarks:

java -jar target/benchmarks.jar -prof perfasm

Technologies

• Java 21
• JMH
• perfasm
• JUnit 5

Notes

This project is intended for educational and research purposes to better understand JVM internals, JIT optimizations, Escape Analysis, and allocation elimination in HotSpot.