• Tracing vs method based
  • Tracing: better best and avg case, and simpler, but worse worst-case
  • Mozilla: TraceMonkey → JagerMonkey (method)
  • Only LuaJIT and PyPy still use tracing
  • Method based likely more complex: On-Stack Replacement (if the current method becomes "hot", it needs to be replaced by a compiled version), Inline Caches, Hidden Classes and Deoptimization
  • HN
• Template interpreter: rather than a switch statement, each bytecode has a sequence of CPU instructions (template), so interpreter just jumps to corresponding template
• Copy and patch: pregenerate snippets beforehand, so less work at JIT time. Just twiddle the snippet holes.
  • Used in new Python JIT

Implementations

• PyPy: meta-tracing approach
• V8
  • 3 tiers
  • Ignition: interpreter
  • Maglev: fast optimizing compiler
  • TurboFan: slow optimizing compiler
  • source
• JVM
  • C1 / HotSpot Client: simple optimizing compiler
  • C2 / HotSpot Server: sophisticated optimizing compiler
  • invokedynamic: flexible way to invoke method handles, you can provide the dispatching logic
  • Object repr: header + fields
    • Header
      • Reference to a class object
      • Monitor (lock word)
      • Identity hashcode
      • GC flags
    • Fields: may be reordered in sake of size optimization, alignment, or target architecture specifics
  • source