3. Benchmarking and Assembly compilation at runtime

How you can see the assembler code that the JIT compiler generates to be run on the actual CPU, explored by way of some examples of benchmarking.

We like our code to run fast, and the just-in-time compiler helps this to happen with a lot of optimisations.

Sometimes we want to measure how fast our code runs, or compare different possible implementations looking for the fastest. This type of measurement is usually done on small scales – microbenchmarking is the practice of isolating small blocks of code to measure their runtime. By removing complexity and only benchmarking a small portion of code we aim to improve measurement, but the fact is that we’re now running code in a very different context to where it will eventually run.

The relative simplicity of code not running in a production environment means that the compiler can find more optimisations to perform when compiling the code. In doing this the compiler can make changes that vastly change what the code is doing, to the extent that we’re no longer measuring anything useful. Great care must be taken to ensure that you’re measuring what you think you’re measuring.

Compiler optimisations can be a huge benefit at runtime; they can be a huge problem in benchmarking.

In this third section we’ll look at just-in-time compilers. These run in the JVM, and transform the intermediate classfiles/bytecode into actual CPU instructions. I’m going to introduce some of the compiler optimisations that can be performed by looking at benchmarking, in particular with the Java Microbenchmarking Harness, JMH.

Benchmarking is an instructive way to learn about JIT in the JVM because you need a certain understanding of what’s going in order to convince yourself that your benchmarks are valid.

It’s easy to write your own benchmarks without any framework, consisting of a timing loop repeatedly running your code. But there’s so much complexity as a result of possible compiler optimisations that it’s very hard to write it yourself and be satisfied that your benchmarks are valid. JMH is a benchmarking tool that aims to handle some of the complexity around measuring correctly. It times invocations of your code, taking averages from many runs; it handles requirements like warming up the JVM; and it provides help for preventing certain optimisations that would spoil your measurements.

There are a lot of articles around about JMH, and the advice presented in them can be often similar to the answers given in this StackOverflow question. It boils down to:

Benchmarking is for comparing implementation choices. It’s not correct to conclude that a benchmark showing that some code runs in 20 nanoseconds in a synthetic benchmark will run in that time in production, surrounded by all manner of other context and under a different pattern of load.

Writing benchmarks

class BenchmarkSomething {
  def wellHelloThere() {
    // code here is timed

Writing the code for a benchmark is pretty straightforward – you put the code you want to run in a method marked with @Benchmark. With sbt it’s also easy to run, with sbt jmh:run. (See: sbt-jmh set-up instructions.)

One practical note: JMH isn’t set up to run inside another project, in the way that you would do for unit tests. Rather than having benchmarks in a folder inside your project, it’s usual to keep them external and either depend on the code you want to benchmark or write it as a one-off exploratory project.

When you run this, it generates the benchmark code for you, and there’s quite a lot of it. Indeed it can be pretty instructive to have a look at it for yourself.

1. Is compiling worth it?

def target_default: Unit = ()
def target_interpreted: Unit = ()

def compiled: Unit = target_default
def interpreted: Unit = target_interpreted

JMH allows control over all sorts of JVM configuration. One JVM option exposed is compiler flags, which include the ability to selectively exclude methods from compilation, forcing them to be interpreted instead.

Benchmark time/invocation
compiled 2.3ns
interpreted 89.5ns

The example here is the most trivial one, compiling or interpreting a no-op method. We can see the difference clearly; the compiled method runs around 40 times faster than the interpreted method.

Yes, compilation is worth it!

2. What’s the cost of method invocation?

def target_dontInline: Unit = ()
def target_inline: Unit = ()

def dontInlineBench: Unit = target_dontInline
def inlineBench: Unit = target_inline

We can force the compiler to inline, or not to inline, a given method. In this benchmark we can test the overhead of invoking a method.

Again the result is clear, and in this example the overhead of the method call is around 2ns per invocation.

The JVM has a default maximum size for a method to be allowed to be inlined, which defaults to 35 bytes. You can change this, but it’s not recommended without a really good reason and some evidence. Why shouldn’t everything be inlined? Inlining makes for bigger methods that take longer to be compiled, for smaller benefit given that in larger methods the time spent invoking other methods is relatively smaller.

Benchmark time/invocation
not inlined 2.3ns
inlined 0.3ns

Note that inlining of methods that you get on the JVM usually only happens in JIT, not in scalac or javac. This allows stack traces to include the full call stack, for example. A notable exception to this is the Scala @inline annotation, which “requests that the compiler should try especially hard to inline the annotated method”.

3. Dead code elimination

var x = Math.PI
def measureLogWrong: Unit = { Math.log(x) }
def measureLogRight: Double = Math.log(x)

Code that produces a value that isn’t used is code that’s just burning CPU unnecessarily. At runtime, it’s great for unused code to be removed, because removing it makes our code better. In a benchmark, sometimes you have to persuade the compiler that your code isn’t useless, otherwise your benchmarks would give the wrong results.

Suppose that we want to measure a utility function like Math.log. If the compiler thinks that the value returned is not used then it is free to remove the call. It’s easy to construct a benchmark method that “suffers” from this optimisation. JMH has support for preventing in the form of a type called BlackHole – stuff goes in and doesn’t come out. Any value returned from a benchmark method is consumed by a BlackHole and it’s possible to access one for use inside the method as well. BlackHole goes to great lengths to ensure that the compiler cannot prove that the value is unused, while doing this quickly and with predictable performance.

Benchmark time/invocation
code eliminated 0.3ns
code runs 21.8ns

Note that the first example here doesn’t return the value – it’s a Unit method and the compiler can easily show that the value is not used. (It also needs to know that there are no side-effects, which is likely easy to do with a native operation like log but may be harder with others.) The second version returns the value, and JMH will consume it for us.

The first version runs much faster, a result of DCE, but the second is the measurement we’re actually interested in.

4. Parameter specialisation

var x = Math.PI
def measureLogConst: Double = Math.log(Math.PI)
def measureLogParam: Double = Math.log(x)

If the compiler can prove that the input to a function is always the same, then it can use this fact to assist in optimisations. If the parameter to a function is always the same value then this value can be loaded directly onto the stack without having to read its value from a field on each invocation.

Benchmark time/invocation
constant param 3.1ns
varying param 21.8ns

We can see that here, where in the first example the value is provably constant. In the second example, having the value come from a var is enough to force the compiler not to make this optimisation. The difference here is indeed so great that we might suspect more is going on that only eliminating a load of a value from a field.

5. Class hierarchy analysis

In a hierarchy of inheriting types with methods that override other methods, actual code for operations is distributed around in different types. Given an instance of some interface or trait, how do you find which code you need to run for a method call?

Quick intro: runtime types

To answer this fully we need a basic understanding of how the JVM represents objects and types at runtime. Any object has a concrete type, and that type is represented by a type descriptor in memory describing its methods, its one super-class and any super-interfaces, among other things. Each object has a pointer in memory to its type descriptor. This type descriptor is never an interface, it’s always a class; you can’t instantiate an interface, you always need some concrete type.

You can see and manipulate representations of these types using reflection; in Java, <your-type>.class or <your-object>.getClass() gives a Class object that exposes details of hierarchy, methods and fields available. (Fun fact: if you look closely enough in the JVM source you’ll discover that these things are called klass, to disambiguate from the Java-level keyword.)

Consider an example type hierarchy with an interface List and three implementations ArrayList, LinkedList and ImmutableList. In code you might have a reference to an object of type List, and polymorphism being a pillar of something means that the object will be one of the three sub-types, but can be treated generically.

// bytecode for method call list.size()
aload_1            // load list onto stack
invokevirtual #28  // Method List.size:()I
// now the integer size is on the stack

Suppose that we want to call a method .size() on our list. By the time this is compiled to bytecode it will look something like this. The method being called is on the interface, but the object on the stack has a concrete type and the code for the method has to be chosen based on the actual object, not the interface.

Doing this requires a few hops in memory:

Note we traverse up the class hierarchy, not down. This is because each object is an instance of a single concrete type, and tracing up instead of down means we find method overrides.

Do we need all of these pointer dereferencing steps? In the usual case, the answer is yes. The code is compiled to handle calling methods polymorphically, so must look up the object header first.

This isn’t the case for static methods, where the location of the code is known at compile time because there’s no overriding (compare this to static vs dynamic linking of libraries). This is the difference between the invokevirtual and invokestatic instructions – the latter doesn’t need to dereference an object in order to be invoked.

Now consider the case where there’s only one implementation of an interface method in the JVM – when the method is monomorphic, not polymorphic. The just-in-time compiler is able to optimise method calls where the code of the method is statically known, using class hierarchy analysis. In our example, if only one implementation of List existed in the JVM (i.e. if its class had been loaded) then the object dereference can be skipped and the method call be made effectively static.

The benchmark

trait ThingA { def get: Int }
trait ThingB { def get: Int }

class ThingA1(value: Int) extends ThingA {
  def get = value
} // ... and other copies
class ThingB1(value: Int) extends ThingB {
  def get = value
// all different types
val a1: ThingA = new ThingA1(1)
val a2: ThingA = new ThingA2(2)
val a3: ThingA = new ThingA3(3)
val a4: ThingA = new ThingA4(4)
// all same type
val b1: ThingB = new ThingB1(1)
val b2: ThingB = new ThingB1(2)
val b3: ThingB = new ThingB1(3)
val b4: ThingB = new ThingB1(4)

def multipleClasses = a1.get + a2.get + a3.get + a4.get
def singleClass     = b1.get + b2.get + b3.get + b4.get

Can we observe this effect in practice? Certainly! List is a poor choice for this because many implementations of it will be loaded just by accident. So we’ll construct a synthetic hierarchy to test this scientifically.

The code for both traits and all implementing classes is the same, so that we can test the difference when the only change is the number of implementations.

We’ve got four instances of ThingA of four different concrete types, and four instances of ThingB all of the same type. In the benchmark we’ll load an integer from each of the four instances and add them together. This ensures that none of method calls are removed by the optimiser.

Benchmark time/invocation
multiple classes 5.3ns
single class 3.6ns

The effect is not insignificant – about a one-third performance boost by collapsing our four types into one. In many settings it may not be possible to do this, where the code is doing different things in different implementation, but effect has been observed in the wild. The ImmutableList implementation in Google Guava, a set of core library extensions for Java, has specialised implementations for small list sizes, and there has been a lot of discussion about how best to balance performance here. This GitHub issue has a lot of the context.

You can imagine why it could have been easy to miss this kind of performance difference in testing. In a typical benchmark you’d test each implementation (empty, one item, two items, …) separately, and CHA would optimise method calls for you. In production code where you’d expect all of these types to be loaded, the opportunity for the JIT to perform these optimisations would much less.

See the generated assembler

Why believe me?

So far I’ve offered explanations for the performance differences in our benchmark examples without providing a lot of evidence. JMH provides tools for inspecting the code generated by the JIT compiler.

Assuming you can get everything installed correctly, it’s actually pretty easy to gather the compiled output. JMH selects the hottest regions encountered in each benchmark, and displays them as below with sbt jmh:run -prof perfasm.

(Handy tip: for troubleshooting, you can see which profilers JMH thinks are available using sbt jmh:run -lprof.)

result.startTime = System.nanoTime();
do {
} while(!control.isDone);
result.stopTime = System.nanoTime();

As the simplest possible example, we’ll look here at the inlining benchmark.

Here is some of the code generated for the benchmark, calling into our inlineBench method; here we see the core measuring loop. This is pretty similar to what would likely do naively: start a timer, run some loops, count iterations, and wait until a boolean flag says we’re done. More about the details about that later.

Running the benchmark, the assembler looks something like this:

cycles   instrs    instr address       instruction                 source instruction
                   0x00007f24446351b7: nopw   0x0(%rax,%rax,1)     ;*invokevirtual inlineBench
23.92%   21.41%   0x00007f24446351c0: movzbl 0x94(%r11),%r10d     ;*getfield isDone
          0.05%   0x00007f24446351c8: add    $0x1,%r14
          0.05%   0x00007f24446351cc: test   %eax,0xa627e2e(%rip) ;*ifeq
                                                                  ;   {poll}
21.38%   23.76%   0x00007f24446351d2: test   %r10d,%r10d
                  0x00007f24446351d5: je     0x00007f24446351c0   ;*ifeq
                   0x00007f24446351d7: mov    $0x7f244d7ee2b0,%r10
                   0x00007f24446351e1: callq  *%r10                ;*invokestatic nanoTime
                   0x00007f24446351e4: mov    %rax,0x30(%r13)      ;*putfield stopTime

Notice the arrow indicating the jump source and destination – this is the measuring loop.

The first item here is the invocation of our benchmark, which has been compiled to a no-op. Even better, it’s been hoisted out of the loop so that it’s only invoked once!

In the loop:

What’s my clock speed?

Recall that many of these benchmarks, especially the ones where the compiler managed to eliminate all of our code, ran in 0.3ns per invocation. This is suspiciously close to three billion invocations per second, or 3GHz – this is my processor speed, of course. What we see is that each measurement loop here executes in a single clock cycle.

How can this be the case given that there are multiple operations to perform on each iteration?

Unfortunately we go beyond what JMH can tell us here, and I don’t know of any instrumentation that is able to expose any further information. Although it’s difficult to prove more, we can speculate, though, that we’re seeing examples here of hardware optimisation happening in the CPU itself, beyond even the JIT compiler’s reach.

part of the Just In Time series