Janet 1.12.2-b91fe8b Documentation
The Janet Abstract Machine
The Janet language is implemented on top of an abstract machine (AM). The compiler converts Janet data structures to this bytecode, which can then be efficiently executed from inside a C program. To understand Janet bytecode, it is useful to understand the abstractions used inside the Janet AM, as well as the C types used to implement these features.
The stack = the fiber
A Janet fiber is the type used to represent multiple concurrent processes in
Janet. It is basically a wrapper around the idea of a stack. The stack is
divided into a number of stack frames (
JanetStackFrame * in C), each of
which contains information such as the function that created the stack frame,
the program counter for the stack frame, a pointer to the previous frame, and
the size of the frame. Each stack frame also is paired with a number registers.
X: Slot X X - Stack Top, for next function call. ----- Frame next ----- X X X X X X X - Stack 0 ----- Frame 0 ----- X X X - Stack -1 ----- Frame -1 ----- X X X X X - Stack -2 ----- Frame -2 ----- ... ... ... ----- Bottom of stack
Fibers also have an incomplete stack frame for the next function call on top of
their stacks. Making a function call involves pushing arguments to this
temporary stack, and then invoking either the
instructions. Arguments for the next function call are pushed via the
PUSHA instructions. The
stack of a fiber will grow as large as needed, although by default Janet will
limit the maximum size of a fiber's stack. The maximum stack size can be
modified on a per-fiber basis.
The slots in the stack are exposed as virtual registers to instructions. They can hold any Janet value.
All functions in Janet are closures; they combine some bytecode instructions
with 0 or more environments. In the C source, a closure (hereby the same as a
function) is represented by the type
JanetFunction *. The bytecode
instruction part of the function is represented by
JanetFuncDef *, and a
function environment is represented with
The function definition part of a function (the 'bytecode' part,
JanetFuncDef *), also stores various metadata about the function which is
useful for debugging, as well as constants referenced by the function.
Janet uses C functions to bridge to native code. A C function
JanetCFunction * in C) is a C function pointer that can be called like a
normal Janet closure. From the perspective of the bytecode instruction set,
there is no difference in invoking a C function and invoking a normal Janet
Janet bytecode operates on an array of identical registers that can hold any
Janet value (
Janet * in C). Most instructions have a destination
register, and 1 or 2 source registers. Registers are simply indices into the
stack frame, which can be thought of as a constant sized array.
Each instruction is a 32-bit integer, meaning that the instruction set is a constant width RISC instruction set like MIPS. The opcode of each instruction is the least significant byte of the instruction. The highest bit of this leading byte is reserved for debugging purpose, so there are 128 possible opcodes encodable with this scheme. Not all of these possible opcodes are defined, and undefined opcodes will trap the interpreter and emit a debug signal. Note that this means an unknown opcode is still valid bytecode, it will just put the interpreter into a debug state when executed.
X - Payload bits O - Opcode bits 4 3 2 1 +----+----+----+----+ | XX | XX | XX | OO | +----+----+----+----+
Using 8 bits for the opcode leaves 24 bits for the payload, which may or may not be utilized. There are a few instruction variants that divide these payload bits.
- 0 arg - Used for noops, returning
nil, or other instructions that take no arguments. The payload is essentially ignored.
- 1 arg - All payload bits correspond to a single value, usually a signed or unsigned integer. Used for instructions of 1 argument, like returning a value, yielding a value to the parent fiber, or doing a (relative) jump.
- 2 arg - Payload is split into byte 2 and bytes 3 and 4. The first
argument is the 8-bit value from byte 2, and the second argument is the
16-bit value from bytes 3 and 4 (
instruction >> 16). Used for instructions of two arguments, like move, normal function calls, conditionals, etc.
- 3 arg - Bytes 2, 3, and 4 each correspond to an 8-bit argument. Used for arithmetic operations, emitting a signal, etc.
These instruction variants can be further refined based on the semantics of the arguments. Some instructions may treat an argument as a slot index, while other instructions will treat the argument as a signed integer literal, an index for a constant, an index for an environment, or an unsigned integer. Keeping the bytecode fairly uniform makes verification, compilation, and debugging simpler.
A listing of all opcode values can be found in
janet.h. The Janet
assembly short names can be found in
src/core/asm.c. In this document,
we will refer to the instructions by their short names as presented to the
assembler rather than their numerical values.
Each instruction is also listed with a signature, which are the arguments the instruction expects. There are a handful of instruction signatures, which combine the arity and type of the instruction. The assembler does not do any type-checking per closure, but does prevent jumping to invalid instructions and failure to return or error.
- The '$' prefix indicates that a instruction parameter is acting as a virtual register (slot). If a parameter does not have the '$' suffix in the description, it is acting as some kind of literal (usually an unsigned integer for indexes, and a signed integer for literal integers).
- Some operators in the description have the suffix 'i' or 'r'. These indicate that these operators correspond to integers or real numbers only, respectively. All bit-wise operators and bit shifts only work with integers.
>>>indicates unsigned right shift, as in Java. Because all integers in Janet are signed, we differentiate the two kinds of right bit shift.
- The 'im' suffix in the instruction name is short for immediate.
|$dest = $lhs + $rhs|
|$dest = $lhs + im|
|$dest = $lhs & $rhs|
|$dest = ~$operand|
|$dest = $lhs | $rhs|
|$dest = $lhs ^ $rhs|
|$dest = call($callee, args)|
|$dest = closure(defs[$index])|
|$dest = janet_compare($lhs, $rhs)|
|$dest = $lhs / $rhs|
|$dest = $lhs / im|
|$dest = $lhs == $rhs|
|$dest = $lhs == im|
|$dest = $lhs .== $rhs|
|Throw error $message.|
|$dest = $ds[$key]|
|$dest = $ds[index]|
|$dest = $lhs > $rhs|
|$dest = $lhs .>= $rhs|
|$dest = $lhs .> im|
|$dest = $lhs .> $rhs|
|pc = label, pc += offset|
|if $cond pc = label else pc++|
|if $cond pc++ else pc = label|
|$dest = constants[index]|
|$dest = false|
|$dest = integer|
|$dest = nil|
|$dest = current closure (self)|
|$dest = true|
|$dest = envs[env][index]|
|$dest = length(ds)|
|$dest = $lhs < $rhs|
|$dest = $lhs .<= $rhs|
|$dest = $lhs .< im|
|$dest = $lhs .< $rhs|
|$dest = call(array, args)|
|$dest = call(tuple/brackets, args)|
|$dest = call(buffer, args)|
|$dest = call(string, args)|
|$dest = call(struct, args)|
|$dest = call(table, args)|
|$dest = call(tuple, args)|
|$dest = $src|
|$dest = $src|
|$dest = $lhs * $rhs|
|$dest = $lhs * im|
|Propagate (Re-raise) a signal that has been caught.|
|Push $val on arg|
|Push $val1, $val2 on args|
|Push $val1, $val2, $val3, on args|
|Push values in $array on args|
|$ds[$key] = $val|
|$ds[index] = $val|
|$dest = resume $fiber with $val|
|envs[env][index] = $val|
|$dest = emit $value as sigtype|
|$dest = $lhs << $rhs|
|$dest = $lhs << shamt|
|$dest = $lhs >> $rhs|
|$dest = $lhs >> shamt|
|$dest = $lhs >>> $rhs|
|$dest = $lhs >>> shamt|
|$dest = $lhs - $rhs|
|Return call($callee, args)|
|Assert $slot matches types|