I've read about meta-circular interpreters on the web (including SICP) and I've looked into the code of some implementations (such as PyPy and Narcissus).

I've read quite a bit about two languages which made great use of metacircular evaluation, Lisp and Smalltalk. As far as I understood Lisp was the first self-hosting compiler and Smalltalk had the first "true" JIT implementation.

One thing I've not fully understood is how can those interpreters/compilers achieve so good performance or, in other words, why is PyPy faster than CPython? Is it because of reflection?

And also, my Smalltalk research led me to believe that there's a relationship between JIT, virtual machines and reflection. Virtual Machines such as the JVM and CLR allow a great deal of type introspection and I believe they make great use it in Just-in-Time (and AOT, I suppose?) compilation. But as far as I know, Virtual Machines are kind of like CPUs, in that they have a basic instruction set. Are Virtual Machines efficient because they include type and reference information, which would allow language-agnostic reflection?

I ask this because many both interpreted and compiled languages are now using bytecode as a target (LLVM, Parrot, YARV, CPython) and traditional VMs like JVM and CLR have gained incredible boosts in performance. I've been told that it's about JIT, but as far as I know JIT is nothing new since Smalltalk and Sun's own Self have been doing it before Java. I don't remember VMs performing particularly well in the past, there weren't many non-academic ones outside of JVM and .NET and their performance was definitely not as good as it is now (I wish I could source this claim but I speak from personal experience).

Then all of a sudden, in the late 2000s something changed and a lot of VMs started to pop up even for established languages, and with very good performance. Was something discovered about the JIT implementation that allowed pretty much every modern VM to skyrocket in performance? A paper or a book maybe?