BEAM VM ELI5

Registers and Calls

Being a register VM means that stack is not used to pass arguments between calls, but registers are. They are not same registers as one would expect to see in a CPU architecture, but rather just an array of Term values. There are 1024 registers but functions of such arity are rare, so usually VM only uses first 3 to 10 registers.

For example, when a call to a function of arity 3 is made: registers x0, x1, and x2 are set to function arguments. If we ever need to make a recursive call, we will have to overwrite registers and lose the original values in them. In this situation, we will need to save old values on the stack. There is no dynamic stack allocation other than stack frame, determined by the compiler.

Each process has own set of registers.

Instructions

BEAM instruction set contains approximately 158 basic instructions, most of them can be visible if you dump assembly by calling erlc -S yourfile.erl. Each instruction has 0 or more arguments. Different tricks are used to encode arguments in a portable and compact way. For example locations in code (so-called labels) are encoded by their index, and code loader later translates them to memory addresses.

During BEAM code loading, some combinations of opcodes are replaced with a faster opcode. This is optimization trick called super-instruction. For each opcode, there is a piece of C code in beam_emu.c which has a C label. An array of C labels is stored at the end of the same VM loop routine and is used as the lookup table.

Threaded VM Loop

VM emulator loop in emulator/beam/beam_emu.s contains a lot of small pieces of code, each having a label and handling one BEAM instruction. They all belong to one very long function. A table of labels is stored in the same function which is used as lookup table.

After the loading opcodes are replaced with such label addresses, followed by arguments. For example, for opcode #1, an element with index 1 from labels array is placed in the code memory.

This way it is easy to jump to a location in C code which handles next opcode. Just read a void* pointer and do a goto *p. This feature is an extension to C and C++ compilers. This type of VM loop is called direct-threaded dispatch virtual machine loop.

Other types of VM loops are:

• switch dispatch VM (this one is used by BEAM VM source if the C compiler refuses to support labels extension). It is also 20 to 30% slower than direct threading;

• direct call threading

• and a few more exotic http://www.cs.toronto.edu/~matz/dissertation/matzDissertation-latex2html/node6.html