The idea of generating an interpreter from the compiler is really neat.

What I missed in this article is why the Baseline Interpreter is faster than the C++ interpreter. The code snippet for the load zero instruction looks like what a compiler should produce for a straightforward C++ switch case for that instruction. Except that the code uses a push instruction to store the value directly on the system stack, whereas the C++ interpreter would presumably use a more general store instruction into an array (in the heap, maybe) treated as the interpreter stack.

Is that the difference, or am I missing something else?