search

mozilla / rhino

Rhino is an open-source implementation of JavaScript written entirely in Java
https://rhino.github.io
Other
4.18k stars 851 forks source link

Begin to use invokedynamic in the bytecode #1645

Closed gbrail closed 1 month ago

gbrail commented 1 month ago

This begins to use invokedynamic instructions in Rhino's bytecode.

It replaces the operations that call the ScriptRuntime operations for common property operations like setting and getting properties and elements, and common context operations like setting and getting variables, with invokedynamic instructions.

It also adds components that will wire up those invokedynamic instructions to the appropriate ScriptRuntime operations using the Dynalink package.

The result should be that Rhino behaves exactly the same and performs the same as well.

However, once this is implemented we can begin to create additional dynalink "linkers" that do specific things to optimize performance based on what is happening at runtime.

gbrail commented 1 month ago

Unless there are no objections, I'd like to merge this. Just so that you all understand the scope:

Downsides:

  1. It adds a dependency on the jdk.dynalink module. That should not be a problem for anyone I can think of as long as we make sure that this code is never activated on an Android platform (and I think that it will be OK)
  2. It adds an extra layer of abstraction to the bytecode which requires understanding more stuff to understand what it does. I am going to experiment more with the structure of the code. I fear in the current state it will be hard to understand.

Upsides:

  1. The bytecode is smaller and IMO easier to read.
  2. I have an optimization waiting in the wings that will make constants faster. However it only works for "real constants," and sadly existing JavaScript benchmarks tend not to actually use "const." (I also tried a more aggressive optimization that assumed things were const until they aren't, but that needs more work.)
  3. I think that further optimizations are possible for the various math operations.
  4. I have spent lots and lots of time trying to make property and variable accesses more efficient, and so far nothing makes a significant difference, but I will keep working on it.
p-bakker commented 1 month ago

No objections here!

Wrt to the downsides: dunno how feasible it is, but some documentation on the bytecode generation may help? Got the idea that only a few people understand how that works right now. I for one don't at all 🙂