So in the case of matrix multiplication, it knows that it's a double-vector of ints and runs the whole thing compiled. So it should, theoretically, be just as fast as a compiled language.
If those stats prove that it is always an array of floats, then machine code gets generated for an array of floats, with a trap handler for when the code fails.
The trap handler, if called, will throw away the code and restart the analysis.
So if you are nice for the compiler and keep your types regular, the machine code will not be thrown away and there are no speed bumps.
I still don't get it... How do you know when to activate the trap handler if you're not checking every time? It's not enough to wait for an error condition (e.g. an invalid pointer). If it's an array of integers, and I sneak in a string in there, what's stopping the JITed version from treating that pointer as an integer?