[anonymousDan]

I believe the technical term for the correspondence between what it is possible to compile and what is valid bytecode/machine code is fully abstract compilation. It's an interesting concept with many interesting implications (e.g. for security). In the past at least there were various examples of Java programs that were illegal in the language but nonetheless could be created directly as bytecode and would be loaded by the JVM. This obviously becomes a security problem if your program loads bytecode dynamically and makes assumptions about its capabilities at the language level as opposed to the bytecode level.

[drdrey]

If you load untrusted code dynamically, it strikes me as wrong to assume anything about its capabilities. Even more so "at the language level". Untrusted code can do anything unless you sandbox it.

[anonymousDan]

You're right but I guess it's perhaps easy to overlook, e.g. that if you decompile/disassemble a valid bytecode program it might give you a program that is not valid at the source level. Some interesting examples and discussion here: http://lambda-the-ultimate.org/node/5364