Message in a bottle

Synopsis

What does really happen when we call a method? How do the different Ruby implementations actually figure out what code to execute? What plumbing is going on under the hood to get a speedy dispatch? In this talk we will have a look at the internals of the four major Ruby implementations - 1.8, 1.9, JRuby and Rubinius, focusing on their dispatch. From look-up tables and call site caches, to inlining and what on earth is invokedynamic? Expect C, C++, Java, and, of course, Ruby code. But fear not, all will be explained!

Intro

• Ruby 1.8 is slow because it's interpreted
• Surprise! Ruby 1.9 is too!
• Something like 55 Ruby implementations! (Big ones: MRI, JRuby, Rubinius)
• Ruby implementations like bytecode

Speeding it up

• Reduce ops -> speeds things up
• JIT
• Inlining
• Search (methods)

MRI

• rb_call0
  • search (search_method) and execute (vm_call0 and vm_exec -- looks at the bytecode)

Rubinius

• Structure: Module -> MethodTable (Hash-like) -> Bucket -> CompiledMethod

  compiled_method = DeepThought.method_table.lookup(:ultimate_answer?).method compiled_method.decode compiled_method.opcodes compiled_method.literals

• Some messages get special opcodes for optimization purposes

• A call turns into InlineCache (lookup is done so it doesn't have to redo it, unless types change)

• Can add native breakpoints (not slower, like in MRI)

• Rubinius Bytecode -> LLVM IR -> Native Code

• Does some analysis like JVM's Hotspot (at runtime)

JRuby

JVM bytecode

Old: invokevirtual (uses custom runtime)

New: invokedynamic

• Linker
  • CallSite is like Rubinius InlineCache
  • Replaces with your method lookup logic
  • Bootstrap doesn't look up anything, because it doesn't know what the method will be called on
  • No lookup logic -- it only has classes
  • The Fallback is the actual lookup logic

Takeaways

• Moving to faster lookup logic across implementations
• Fewer instructions == better
  • MRI -- what could it do better? It's adding super-instructions
• Late binding
• Symbols aren't garbage collected. You have user generated symbols? They'll take up more and more memory!
  • They're the same object between instances