Sunday, January 08, 2012

The invokedynamic API

I thought I should write a bit about the invokedynamic API, since the API is very powerful and flexible, but sometimes needs a bit getting used to it. I won't write about everything or even in detail, just some hints for thinking on a few elements - the ones I used for Groovy mostly and are most probably the ones you will use as well.

Let us say we are in the situation, that we have the method we want to call as handle already and we have the target type of our call site. I call that method handle SM (for selected method) and the target type TT - for obvious reasons.

Now normally in Groovy you would do the following: you transform the arguments into a form you can use for calling the method. If you for example have an int, but want to call an long taking method, then you have to transform the int into a long. You normally do this by using a general type transforming method, that takes the argument and a goal type, inspects the argument and then goes through quite big code parts for all the transformations that are allowed. We do this not during method selection itself, but for each call of course. What we do is kind of like wrapping the method object into something new, that we can call in our call site and that for each call will test if it has to apply the big standard transformation - and well, apply it too.

If you read that big transformation code, then you will have a direct transformation of the arguments to whatever the SM needs. In invokedynamic we work a slightly bit different. Instead of having only one transformation we have many small ones, that we combine into a specialized one. We do this by taking our SM, apply transformers to it and use the result for our method calls. Since the main work is now no longer correcting the big transformer code, but the combination of the many small transformers the workflow feels kind of reversed to me. You now want to apply transformers to make out of your SM, one that accepts the target type TT.

Some notes for better understanding:

  • One thing to remember when working with MethodHandles for example is that it has no receiver. In a method call, we say normally that foo is the receiver (well the class of foo), bar the method name and x the argument. For a MethodHandle it itself is the receiver of course. So if we work on its arguments, then the first argument is the receiver of our method call. A MethodHandle for above would maybe have this MethodType: (Foo,X)Object - taking the foo receiver and an x argument, returning an Object.
  • The classes you work with mostly... You have MethodType to describe a method's parameters, including return type and receiver. There is MethodHandle of course, the core of it all, representing your SM. And there is a collection of helper methods in the class MethodHandles. Note the additional s. Well, and SwitchPoint, but I haven't used it really yet... that is still to come soon.
  • Type matching... your SM has to be changed by the usage of the transforms to fit the requested target type of your call site. For this you use for example asType

But without further delay...

MethodHandles#dropArguments is one quite useful method, that at the same time shows very much the reversed thinking that needs to be applied. This is a transform, that will drop arguments, it does not do it right away. In the old thinking we have for example arguments a,b,c and if we would drop there, we would for example get a,b. This transform *will* do exact this, but what we look at gives a different impression. We have a handle that takes A,B (A being type of a and B being type of b) and dropping then means we will take that and produce a new handle that can take A,B,C. This new handle will then ignore the argument c of type C and in doing so, it does exactly what I described above, but if we debug the code producing the combinations of transformers we see a method handle that gets the dropArguments transformation applied and now takes one more argument instead of one less. So again.... assuming you want to get rid of an argument you start with the handle, that is without it and have to transform it into a handle that takes it... for me that is like reversed thinking and really sometimes produces problems for me. You may want to apply this one if your SM takes less arguments than provided.

MethodHandle#bindTo is also one I use often. Assuming we have a handle taking A,B,C and A will be the same for each call, we can bind it to produce a new handle taking B,C... but only for the first argument. So for example if you have a method call that will be done through the meta class system, then this results mostly in calling a method MetaClass#invokedMethod. But the receiver there is not the receiver from my callsite, so I bind the first argument, the meta class. A changed meta class would require a new method selection so it is kind of constant for this selection. You may want to apply this one if your SM has one more argument than the ones provided and it is the first one

But what to do if we have more than one extra argument our SM would like to take? In this case you use MethodHandles#insertArguments. For example your SM would like to take A,B,C, your TT is A,B and you want to bind a c. Then you can use newHandle = insertArguments(oldHandle,2,c).If it where A,B,C,D, and you want to bind for C and D, then it is insertArguments(oldHandle,2,c,d). If you have A,B,C,D and you want to bind B and D, then it is for example (there are two ways): newHandle = insertArguments(oldHandle,1,b); newHandle = insertArguments(newHandle,2,d). You may have noticed the chaining of transforms in this one. The first makes A,B,C,D into A,C,D, moving D from position 3 to position 2. The second makes A,C then. You may want to apply this one if your SM has more arguments than the ones provided.

MethodHandle#asCollector is also a nice one. There are several invokeMethod methods in Groovy that take an Object[] as argument to contain all the arguments and then delegate calls to builders or do other things. The callsite you start with though may not provide the Object[], but the arguments one by one. So your TT may look like this: A,B,C,D and your SM may accept this: A,Object[]. Meaning we somehow have to wrap B,C,D into an Object[], if thought from the perspective of the arguments. And that is why the method is called asCollector. What you see on your method handle is that with handle = handle.asCollector(Object[].class, 3), your A,Object[] handle is now a A,Object,Object,Object handle. You will need asType to get to the final form. Should you want to collect the arguments at a different place than the last one, you may have to permute the arguments. The opposite way is asSpreader, but I didn't use that one yet. I may do so when I extend the implementation to include the spread operator. You may want to apply this one if your SM has arguments you want to collect into an array

Sometimes we have to make an transformation, that changes the runtime class of a reference type. This is of course not possible, because the JVM is strong typed - but we can create a new object with the desired result. For example Groovy has GString, which is a kind of String. A method call done with an GString argument is supposed to be able to call a method that takes a String. Now GString and String are not in a subclass relation, meaning that we will have to do a real type transformation here and MethodHandles#filterArguments will help us with that. A filter takes one argument and returns the transformed value. For our case it should take a GString and return a String. Let us say our SM takes A,String,String,B; TT be A,GString,GString,B and our filter MethodHandle takes GSTRING. Then we can simply do newHandle = MethodHandles.filterArguments(oldHandle, 1, GSTRING, GSTRING) to produce a handle for A,GString,GString,B. Again I had here some problems with the reversed thinking: The filter's return type must match the type in the SM and the argument type the type of our TT. If you think about it, it is quite clear and obvious, but when I work on a program I always have to stop here and rethink. Anyway... You may want to apply this one if your SM argument differs form the provided argument in a way that boxing and casting alone don't do the transformation requried.

And the last one I want to mention: MethodHandles#guardWithTest. In many situations you have to handle the invalidation of your call site to some extend. One way is for example (if you use a MutableCallsite) to have a guard causing exchanging the target for your callsite by a new method selection. Let us assume we have a call foo(a,x) and the methods foo(Object,Object) and foo(Object,String). Now x may be a String and we want to call foo(Object,String) then, or we want to call foo(Object,Object) if it is not String. Let us assume there will be three calls: The first done with an Integer, the second with String and the last one with an Integer again. We will have to exchange the method each time. A guard takes a number of arguments (0-N) and returns a boolean, which will be used to determine if we call our normal target or a fallback handle. For this case here I have used a handle called SAME_CLASS. It takes a class and an Object argument and tests if argument.getClass() is equal to the provided class. Since the first argument, the class, is fixed I bind it upfront, leaving me a guard handle which takes an Object argument only. guardWithTest does not support any position, but we want to check the second argument to the method, not the first... and there is the receiver too. If you did not jump to this section, then you may find we are in a situation in which we have less arguments than given... only it is not the SM, but our guard now. But that doesn't matter. Still we drop the first two to get a handle Object,Object,Object, which will ignore the 0 position argument, the receiver, and the argument at position 1, but test the one at position 2. Then it is simply newHandle = MethodHandles.guardWithTest(guard, oldHandle, fallback) and be done with it. If you want to have multiple guards, you continue to apply this kind of transform. I haven't found a way to combine guards themselves to have only one guardWithTest call. But maybe that isn't needed too. You may apply this one if your call site needs to be invalidated based on the given arguments.

This is just a short overview and by far nowhere near complete or a replacement for reading the javadoc. Just a little guide that may be of help.