JRuby interns most Strings that come up in Ruby code just fine in org.jruby.ir.persistence.IRReaderStream#decodeString:

    public String decodeString() {
        int strLength = decodeInt();

        if (strLength == NULL_STRING) return null;

        byte[] bytes = new byte[strLength]; // FIXME: This seems really innefficient
        buf.get(bytes);

        String newString = new String(bytes).intern();

        if (RubyInstanceConfig.IR_READING_DEBUG) System.out.println("STR<" + newString + ">");

        return newString;
    }

, but a few duplicates from more dynamic contexts are not caught. In the red-black benchmark, this adds up to ~200kb of wasted memory (memory wise this is irrelevant obviously, but duplicates make call site cache lookups slower and waste CPU cache => fixing all these spots resulted in a visible ~2% speedup in the red-black benchmark for me over very large sample sizes 2k+ runs).

So I:

• Tracked down spots generating duplicate constants for ASTish code paths
• Did my best to not introduce duplicate calls to intern by interning in constructors wherever possible instead of further upstream
• Made sure not intern calls made it into the hot path by checking that the method isn't called in the hot path (checked red-black benchmark and Logstash after warm up and intern was never called in either => I have a bit of confidence)

Also:

• Removed public synchronized void defineAliases(List<String> aliases, String oldName) since it was one of the cases that looked like they would create duplicate String constants but then realized that it's never used (checked for dynamic invocations too)
• Reduced visibility of public static JavaCallable getMatchingCallable(Ruby runtime, Class<?> javaClass, String methodName, Class<?>[] argumentTypes) since this method was in a path affected by the added intern calls and only used privately => I didn't want to add redundant calls to intern to it, when I can be sure that its only caller interns Strings just fine already