Saturday, September 20th, 2008p>
SquirrelFish Extreme: 943.3 ms V8: 1280.6 ms TraceMonkey: 1464.6 ms
What makes it so fast?
SquirrelFish Extreme uses four different technologies to deliver much better performance than the original SquirrelFish: bytecode optimizations, polymorphic inline caching, a lightweight “context threaded” JIT compiler, and a new regular expression engine that uses our JIT infrastructure.
1. Bytecode Optimizations
When we first announced SquirrelFish, we mentioned that we thought that the basic design had lots of room for improvement from optimizations at the bytecode level. Thanks to hard work by Oliver Hunt, Geoff Garen, Cameron Zwarich, myself and others, we implemented lots of effective optimizations at the bytecode level.
In addition, we’ve improved the bytecode instruction set, and built optimizations that take advantage of these improvements. We’ve added combo instructions, peephole optimizations, faster handling of constants and some specialized opcodes for common cases of general operations.
2. Polymorphic Inline Cache
So how exactly do we cheat? We detect when objects actually have the same underlying structure — the same properties in the same order — and associate them with a structure identifier, or StructureID. Whenever a property access is performed, we do the usual hash lookup (using our highly optimized hashtables) the first time, and record the StructureID and the offset where the property was found. Subsequent times, we check for a match on the StructureID – usually the same piece of code will be working on objects of the same structure. If we get a hit, we can use the cached offset to perform the lookup in only a few machine instructions, which is much faster than hashing.
Here is the classic Self paper that describes the original technique. You can look at Geoff’s implementation of the StructureID class in Subversion to see more details of how we did it.
We’ve only taken the first steps on polymorphic inline caching. We have lots of ideas on how to improve the technique to get even more speed. But already, you’ll see a huge difference on performance tests where the bottleneck is object property access.
3. Context Threaded JIT
Another major change we’ve made with SFX is to introduce native code generation. Our starting point is a technique called a “context threaded interpreter”, which is a bit of a misnomer, because this is actually a simple but effective form of JIT compiler. In the original SquirrelFish announcement, we described our use of direct threading, which is about the fastest form of bytecode intepretation short of generating native code. Context threading takes the next step and introduces some native code generation.
The basic idea of context threading is to convert bytecode to native code, one opcode at a time. Complex opcodes are converted to function calls into the language runtime. Simple opcodes, or in some cases the common fast paths of otherwise complex opcodes, are inlined directly into the native code stream. This has two major advantages. First, the control flow between opcodes is directly exposed to the CPU as straight line code, so much dispatch overhead is removed. Second, many branches that were formally between opcodes are now inline, and made highly predictable to the CPU’s branch predictor.
Here is a paper describing the basic idea of context threading. Our initial prototype of context threading was created by Gavin Barraclough. Several of us helped him polish it and tune the performance over the past few weeks.
One of the great things about our lightweight JIT is that there’s only about 4,000 lines of code involved in native code generation. All the other code remains cross platform. It’s also surprisingly hackable. If you thought compiling to native code is rocket science, think again. Besides Gavin, most of us have little prior experience with native codegen, but we were able to jump right in.
Currently the code is limited to x86 32-bit, but we plan to refactor and add support for more CPU architectures. CPUs that are not yet supported by the JIT can still use the interpreter. We also think we can get a lot more speedups out of the JIT through techniques such as type specialization, better register allocation and liveness analysis. The SquirrelFish bytecode is a good representation for making many of these kinds of transforms.
4. Regular Expression JIT
We thought that making regular expressions fast was a better thing to do than changing the benchmark. A lot of real tasks on the web involve a lot of regexp processing. After all, fundamental tasks on the web, like JSON validation and parsing, depend on regular expressions. And emerging technologies — like John Resig’s processing.js library — extend that dependency ever further.
Major kudos to the entire SFX team for pulling this off. Now, to grab a new nightly…
Posted by Dion Almaer at 12:09 am