morten morten - 4 months ago 16
Java Question

Will the jit optimize new objects

I created this class for being immutable and hava a fluent api:

public final class Message {
public final String email;
public final String escalationEmail;
public final String assignee;
public final String conversationId;
public final String subject;
public final String userId;

public Message(String email, String escalationEmail, String assignee, String conversationId, String subject, String userId) {
this.email = email;
this.escalationEmail = escalationEmail;
this.assignee = assignee;
this.conversationId = conversationId;
this.subject = subject;
this.userId = userId;
}

public Message() {
email = "";
escalationEmail = "";
assignee = "";
conversationId = "";
subject = "";
userId = "";
}

public Message email(String e) { return new Message(e, escalationEmail, assignee, conversationId, subject, userId); }
public Message escalationEmail(String e) { return new Message(email, e, assignee, conversationId, subject, userId); }
public Message assignee(String a) { return new Message(email, escalationEmail, a, conversationId, subject, userId); }
public Message conversationId(String c) { return new Message(email, escalationEmail, assignee, c, subject, userId); }
public Message subject(String s) { return new Message(email, escalationEmail, assignee, conversationId, s, userId); }
public Message userId(String u) { return new Message(email, escalationEmail, assignee, conversationId, subject, u); }

}


My question is, will the optimizer be able to avoid lots of object creation when a new object is created like this:

Message m = new Message()
.email("foo@bar.com")
.assignee("bar@bax.com")
.subject("subj");


Is there anything to be gained from making a separate mutable builder object instead?

Update 2:
After reading pangins answer my benchmark is invalidated. I'll keep it here for reference of how not to benchmark :)

Update:
I took the liberty of measuring this myself with this code:

public final class Message {
public final String email;
public final String escalationEmail;
public final String assignee;
public final String conversationId;
public final String subject;
public final String userId;

public static final class MessageBuilder {
private String email;
private String escalationEmail;
private String assignee;
private String conversationId;
private String subject;
private String userId;

MessageBuilder email(String e) { email = e; return this; }
MessageBuilder escalationEmail(String e) { escalationEmail = e; return this; }
MessageBuilder assignee(String e) { assignee = e; return this; }
MessageBuilder conversationId(String e) { conversationId = e; return this; }
MessageBuilder subject(String e) { subject = e; return this; }
MessageBuilder userId(String e) { userId = e; return this; }

public Message create() {
return new Message(email, escalationEmail, assignee, conversationId, subject, userId);
}

}

public static MessageBuilder createNew() {
return new MessageBuilder();
}

public Message(String email, String escalationEmail, String assignee, String conversationId, String subject, String userId) {
this.email = email;
this.escalationEmail = escalationEmail;
this.assignee = assignee;
this.conversationId = conversationId;
this.subject = subject;
this.userId = userId;
}

public Message() {
email = "";
escalationEmail = "";
assignee = "";
conversationId = "";
subject = "";
userId = "";
}

public Message email(String e) { return new Message(e, escalationEmail, assignee, conversationId, subject, userId); }
public Message escalationEmail(String e) { return new Message(email, e, assignee, conversationId, subject, userId); }
public Message assignee(String a) { return new Message(email, escalationEmail, a, conversationId, subject, userId); }
public Message conversationId(String c) { return new Message(email, escalationEmail, assignee, c, subject, userId); }
public Message subject(String s) { return new Message(email, escalationEmail, assignee, conversationId, s, userId); }
public Message userId(String u) { return new Message(email, escalationEmail, assignee, conversationId, subject, u); }


static String getString() {
return new String("hello");
// return "hello";
}

public static void main(String[] args) {
int n = 1000000000;

long before1 = System.nanoTime();

for (int i = 0; i < n; ++i) {
Message m = new Message()
.email(getString())
.assignee(getString())
.conversationId(getString())
.escalationEmail(getString())
.subject(getString())
.userId(getString());
}

long after1 = System.nanoTime();

long before2 = System.nanoTime();

for (int i = 0; i < n; ++i) {
Message m = Message.createNew()
.email(getString())
.assignee(getString())
.conversationId(getString())
.escalationEmail(getString())
.subject(getString())
.userId(getString())
.create();
}

long after2 = System.nanoTime();



System.out.println("no builder : " + (after1 - before1)/1000000000.0);
System.out.println("with builder: " + (after2 - before2)/1000000000.0);
}


}


I found the difference to be significant (builder is faster) if the string arguments are not new objects, but all the same (see commented code in getString)

In what I imagine is a more realistic scenario, when all the strings are new objects, the difference is negligible, and the jvm startup would cause the first one to be a tiny bit slower (I tried both ways)

With the "new String" the code was in total many times slower as well (I had to decrease the n), perhaps indicating that there is some optimization of the "new Message" going on, but not of the "new String".

Answer

Yes, HotSpot JIT can eliminate redundant allocations in a local context.

This optimization is provided by the Escape Analysis enabled since JDK 6u23. It is often confused with on-stack allocation, but in fact it is much more powerful, since it allows not only to allocate objects on stack, but to eliminate allocation altogether by replacing object fields with variables (Scalar Replacement) that are subject to further optimizations.

The optimization is controlled by -XX:+EliminateAllocations JVM option which is ON by default.


Thanks to allocation elimination optimization, both your examples of creating a Message object work effectively the same way. They do not allocate intermediate objects; just the final one.

Your benchmark shows misleading results, because it collects many common pitfalls of microbenchmarking:

  • it incorporates several benchmarks in a single method;
  • it measures an OSR stub instead of the final compiled version;
  • it does not do warm-up iterations;
  • it does not consume results, etc.

Let's measure it correctly with JMH. As a bonus, JMH has the allocation profiler (-prof gc) which shows how many bytes are really allocated per iteration. I've added the third test that runs with EliminateAllocations optimization disabled to show the difference.

package bench;

import org.openjdk.jmh.annotations.*;

@State(Scope.Benchmark)
public class MessageBench {

    @Benchmark
    public Message builder() {
        return Message.createNew()
                .email(getString())
                .assignee(getString())
                .conversationId(getString())
                .escalationEmail(getString())
                .subject(getString())
                .userId(getString())
                .create();
    }

    @Benchmark
    public Message immutable() {
        return new Message()
                .email(getString())
                .assignee(getString())
                .conversationId(getString())
                .escalationEmail(getString())
                .subject(getString())
                .userId(getString());
    }

    @Benchmark
    @Fork(jvmArgs = "-XX:-EliminateAllocations")
    public Message immutableNoOpt() {
        return new Message()
                .email(getString())
                .assignee(getString())
                .conversationId(getString())
                .escalationEmail(getString())
                .subject(getString())
                .userId(getString());
    }

    private String getString() {
        return "hello";
    }
}

Here are the results. Both builder and immutable perform equally and allocate just 40 bytes per iteration (exactly the size of one Message object).

Benchmark                                        Mode  Cnt     Score     Error   Units
MessageBench.builder                             avgt   10     6,232 ±   0,111   ns/op
MessageBench.immutable                           avgt   10     6,213 ±   0,087   ns/op
MessageBench.immutableNoOpt                      avgt   10    41,660 ±   2,466   ns/op

MessageBench.builder:·gc.alloc.rate.norm         avgt   10    40,000 ±   0,001    B/op
MessageBench.immutable:·gc.alloc.rate.norm       avgt   10    40,000 ±   0,001    B/op
MessageBench.immutableNoOpt:·gc.alloc.rate.norm  avgt   10   280,000 ±   0,001    B/op