Mu Syntax

Here are two valid statements in Mu:

increment x
y <- increment

Understanding when to use one vs the other is the critical idea in Mu. In short, the former increments a value in memory, while the latter increments a value in a register.

Most languages start from some syntax and do what it takes to implement it. Mu, however, is designed as a safe way to program in a regular subset of 32-bit x86 machine code, satisficing rather than optimizing for a clean syntax. To keep the mapping to machine code lightweight, Mu exclusively uses statements. Most statements map to a single instruction of machine code.

Since the x86 instruction set restricts how many memory locations an instruction can use, Mu makes registers explicit as well. Variables must be explicitly mapped to specific registers; otherwise they live in memory. While you have to do your own register allocation, Mu will helpfully point out when you get it wrong.

Statements consist of 3 parts: the operation, optional inouts and optional outputs. Outputs come before the operation name and <-.

Outputs are always registers; memory locations that need to be modified are passed in by reference in inouts.

So Mu programmers need to make two new categories of decisions: whether to define variables in registers or memory, and whether to put variables to the left or right. There’s always exactly one way to write any given operation. In return for this overhead you get a lightweight and future-proof stack. And Mu will provide good error messages to support you.

Further down, this page enumerates all available primitives in Mu, and a separate page describes how each primitive is translated to machine code. There is also a useful list of pre-defined functions (implemented in unsafe machine code) in and

Functions and calls

Zooming out from single statements, here’s a complete sample program in Mu that runs in Linux:

Mu programs are lists of functions. Each function has the following form:

fn _name_ _inout_ ... -> _output_ ... {

Each function has a header line, and some number of statements, each on a separate line. Headers describe inouts and outputs. Inouts can’t be registers, and outputs must be registers (specified using metadata after a /). Outputs can’t take names.

The above program also demonstrates a function call (to the function do-add). Function calls look the same as primitive statements: they can return (multiple) outputs in registers, and modify inouts passed in by reference. In addition, there’s one more constraint: output registers must match the function header. For example:

fn f -> _/eax: int {
fn g {
  a/eax <- f  # ok
  a/ebx <- f  # wrong; `a` must be in register `eax`

You can exit a function at any time with the return instruction. Give it the right number of arguments, and it’ll assign them respectively to the function’s outputs before jumping back to the caller.

Mu encloses multi-word types in parentheses, and types can get quite expressive. For example, you read main’s inout type as “an address to an array of addresses to arrays of bytes.” Since addresses to arrays of bytes are almost always strings in Mu, you’ll quickly learn to mentally shorten this type to “an address to an array of strings”.

Mu currently has no way to name magic constants. Instead, document integer literals using metadata after a /. For example:

var x/eax: int <- copy 3/margin-left

Here we use metadata in two ways: to specify a register for the variable x (checked), and to give a name to the constant 3 (unchecked; purely for documentation).

Variables can’t currently accept unchecked metadata for documentation. (Perhaps this should change.)

The function main is special. It’s where Mu programs start executing. It has a different signature depending on whether a Mu program requires Linux or can run without an OS. On Linux, the signature looks like this:

fn main args: (addr array addr array byte) -> _/ebx: int

It takes an array of strings and returns a status code to Linux in register ebx.

Without an OS, the signature looks like this:

fn main screen: (addr screen), keyboard: (addr keyboard), data-disk: (addr disk)

A screen and keyboard are explicitly passed in. The goal is for all hardware dependencies to always be explicit. However there are currently gaps:


Blocks are useful for grouping related statements. They’re delimited by { and }, each alone on a line.

Blocks can nest:

    _more statements_

Blocks can be named (with the name ending in a : on the same line as the {):

$name: {

Further down we’ll see primitive statements for skipping or repeating blocks. Besides control flow, the other use for blocks is…

Local variables

Functions can define new variables at any time with the keyword var. There are two variants of the var statement, for defining variables in registers or memory.

var name: type
var name/reg: type <- ...

Variables on the stack are never initialized. (They’re always implicitly zeroed out.) Variables in registers are always initialized.

Register variables can go in 6 integer registers (eax, ebx, ecx, edx, esi, edi) or 8 floating-point registers (xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7).

Defining a variable in a register either clobbers the previous variable (if it was defined in the same block) or shadows it temporarily (if it was defined in an outer block).

Variables exist from their definition until the end of their containing block. Register variables may also die earlier if their register is clobbered by a new variable.

Variables on the stack can be of many types (but not byte). Integer registers can only contain 32-bit values: int, byte, boolean, (addr ...). Floating-point registers can only contain values of type float.

Integer primitives

Here is the list of arithmetic primitive operations supported by Mu. The name n indicates a literal integer rather than a variable, and var/reg indicates a variable in a register, though that’s not always valid Mu syntax.

var/reg <- increment
increment var
var/reg <- decrement
decrement var
var1/reg1 <- add var2/reg2
var/reg <- add var2
add-to var1, var2/reg
var/reg <- add n
add-to var, n

var1/reg1 <- subtract var2/reg2
var/reg <- subtract var2
subtract-from var1, var2/reg
var/reg <- subtract n
subtract-from var, n

var1/reg1 <- xor var2/reg2
var/reg <- xor var2
xor-with var1, var2/reg
var/reg <- xor n
xor-with var, n

var1/reg1 <- negate
negate var

var/reg <- copy var2/reg2
copy-to var1, var2/reg
var/reg <- copy var2
var/reg <- copy n
copy-to var, n

compare var1, var2/reg
compare var1/reg, var2
compare var/eax, n
compare var, n

var/reg <- shift-left n
var/reg <- shift-right n
var/reg <- shift-right-signed n
shift-left var, n
shift-right var, n
shift-right-signed var, n

var/reg <- multiply var2

Bitwise operations:

var1/reg1 <- and var2/reg2
var/reg <- and var2
and-with var1, var2/reg
var/reg <- and n
and-with var, n

var1/reg1 <- or var2/reg2
var/reg <- or var2
or-with var1, var2/reg
var/reg <- or n
or-with var, n

var1/reg1 <- not
not var

Any statement above that takes a variable in memory can be replaced with a dereference (*) of an address variable (of type (addr ...)) in a register. You can’t dereference variables in memory. You have to load them into a register first.

Excluding dereferences, the above statements must operate on non-address values with primitive types: int, boolean or byte. (Booleans are really just ints, and Mu assumes any value but 0 is true.) You can copy addresses to int variables, but not the other way around.

Floating-point primitives

These instructions may use the floating-point registers xmm0xmm7 (denoted by /xreg2 or /xrm32). They also use integer values on occasion (/rm32 and /r32).

var/xreg <- add var2/xreg2
var/xreg <- add var2
var/xreg <- add *var2/reg2

var/xreg <- subtract var2/xreg2
var/xreg <- subtract var2
var/xreg <- subtract *var2/reg2

var/xreg <- multiply var2/xreg2
var/xreg <- multiply var2
var/xreg <- multiply *var2/reg2

var/xreg <- divide var2/xreg2
var/xreg <- divide var2
var/xreg <- divide *var2/reg2

var/xreg <- reciprocal var2/xreg2
var/xreg <- reciprocal var2
var/xreg <- reciprocal *var2/reg2

var/xreg <- square-root var2/xreg2
var/xreg <- square-root var2
var/xreg <- square-root *var2/reg2

var/xreg <- inverse-square-root var2/xreg2
var/xreg <- inverse-square-root var2
var/xreg <- inverse-square-root *var2/reg2

var/xreg <- min var2/xreg2
var/xreg <- min var2
var/xreg <- min *var2/reg2

var/xreg <- max var2/xreg2
var/xreg <- max var2
var/xreg <- max *var2/reg2

Remember, when these instructions use indirect mode, they still use an integer register. Floating-point registers can’t hold addresses.

Two instructions in the above list are approximate. According to the Intel manual, reciprocal and inverse-square-root go off the rails around the fourth decimal place. If you need more precision, use divide separately.

Most instructions operate exclusively on integer or floating-point operands. The only exceptions are the instructions for converting between integers and floating-point numbers.

var/xreg <- convert var2/reg2
var/xreg <- convert var2
var/xreg <- convert *var2/reg2

var/reg <- convert var2/xreg2
var/reg <- convert var2
var/reg <- convert *var2/reg2

var/reg <- truncate var2/xreg2
var/reg <- truncate var2
var/reg <- truncate *var2/reg2

There are no instructions accepting floating-point literals. To obtain integer literals in floating-point registers, copy them to general-purpose registers and then convert them to floating-point.

The floating-point instructions above always write to registers. The only instructions that can write floats to memory are copy instructions.

var/xreg <- copy var2/xreg2
copy-to var1, var2/xreg
var/xreg <- copy var2
var/xreg <- copy *var2/reg2

Finally, there are floating-point comparisons. They must always put a register on the left-hand side:

compare var1/xreg1, var2/xreg2
compare var1/xreg1, var2

Operating on individual bytes

A special case is variables of type byte. Mu is a 32-bit platform so for the most part only supports types that are multiples of 32 bits. However, we do want to support strings in ASCII and UTF-8, which will be arrays of 8-bit bytes.

Since most x86 instructions implicitly load 32 bits at a time from memory, variables of type ‘byte’ are only allowed in registers, not on the stack. Here are the possible statements for reading bytes to/from memory:

var/reg <- copy-byte var2/reg2      # var: byte
var/reg <- copy-byte *var2/reg2     # var: byte
copy-byte-to *var1/reg1, var2/reg2  # var1: (addr byte)

In addition, variables of type ‘byte’ are restricted to (the lowest bytes of) just 4 registers: eax, ecx, edx and ebx. As always, this is due to constraints of the x86 instruction set.

Primitive jumps

There are two kinds of jumps, both with many variations: break and loop. break instructions jump to the end of the containing block. loop instructions jump to the beginning of the containing block.

All jumps can take an optional label starting with ‘$’:

loop $foo

This instruction jumps to the beginning of the block called $foo. The corresponding break jumps to the end of the block. Either jump statement must lie somewhere inside such a block. Jumps are only legal to containing blocks. (Use named blocks with restraint; jumps to places far away can get confusing.)

There are two unconditional jumps:

loop label
break label

The remaining jump instructions are all conditional. Conditional jumps rely on the result of the most recently executed compare instruction. (To keep programs easy to read, keep compare instructions close to the jump that uses them.)

break-if-= label
break-if-!= label

Inequalities are similar, but have additional variants for addresses and floats.

break-if-< label
break-if-> label
break-if-<= label
break-if->= label

break-if-addr< label
break-if-addr> label
break-if-addr<= label
break-if-addr>= label

break-if-float< label
break-if-float> label
break-if-float<= label
break-if-float>= label

Similarly, conditional loops:

loop-if-= label
loop-if-!= label

loop-if-< label
loop-if-> label
loop-if-<= label
loop-if->= label

loop-if-addr< label
loop-if-addr> label
loop-if-addr<= label
loop-if-addr>= label

loop-if-float< label
loop-if-float> label
loop-if-float<= label
loop-if-float>= label


Passing objects by reference requires the address operation, which returns an object of type addr.

var/reg: (addr T) <- address var2: T

Here var2 can’t live in a register.

Array operations

Here’s an example definition of a fixed-length array:

var x: (array int 3)

The length (here 3) must be an integer literal. We’ll show how to create dynamically-sized arrays further down.

Arrays can be large; to avoid copying them around on every function call you’ll usually want to manage addrs to them. Here’s an example computing the address of an array.

var n/eax: (addr array int) <- address x

Addresses to arrays don’t include the array length in their type. However, you can obtain the length of an array like this:

var/reg: int <- length arr/reg: (addr array T)

To operate on elements of an array, use the index statement:

var/reg: (addr T) <- index arr/reg: (addr array T), n
var/reg: (addr T) <- index arr: (array T len), n

The index can also be a variable in a register, with a caveat:

var/reg: (addr T) <- index arr/reg: (addr array T), idx/reg: int
var/reg: (addr T) <- index arr: (array T len), idx/reg: int

The caveat: the size of T must be 1, 2, 4 or 8 bytes. The x86 instruction set has complex addressing modes that can index into an array in a single instruction in these situations.

For other sizes of T you’ll need to split up the work, performing a compute-offset before the index.

var/reg: (offset T) <- compute-offset arr: (addr array T), idx/reg: int     # arr can be in reg or mem
var/reg: (offset T) <- compute-offset arr: (addr array T), idx: int         # arr can be in reg or mem

The compute-offset statement returns a value of type (offset T) after performing any necessary bounds checking. Now the offset can be passed to index as usual:

var/reg: (addr T) <- index arr/reg: (addr array T), idx/reg: (offset T)

Stream operations

A common use for arrays is as buffers. Save a few items to a scratch space and then process them. This pattern is so common (we use it in files) that there’s special support for it with a built-in type: stream.

Streams are like arrays in many ways. You can initialize them with a length:

var x: (stream int 3)

However, streams don’t provide random access with an index instruction. Instead, you write to them sequentially, and read back what you wrote.

read-from-stream s: (addr stream T), out: (addr T)
write-to-stream s: (addr stream T), in: (addr T)
var/eax: boolean <- stream-empty? s: (addr stream)
var/eax: boolean <- stream-full? s: (addr stream)

You can clear streams:

clear-stream f: (addr stream _)

You can also rewind them to reread what’s been written:

rewind-stream f: (addr stream _)

Compound types

Primitive types can be combined together using the type keyword. For example:

type point {
  x: int
  y: int

Mu programs are currently sequences of fn and type definitions.

Compound types can’t include addr types for safety reasons (use handle instead, which is described below). They also can’t currently include array, stream or byte types. Since arrays and streams carry their size with them, supporting them in compound types complicates variable initialization. Instead of defining them inline in a type definition, define a handle to them. Bytes shouldn’t be used for anything but utf-8 strings.

To access within a compound type, use the get instruction. There are two forms. You need either a variable of the type itself (say T) in memory, or a variable of type (addr T) in a register.

var/reg: (addr T_f) <- get var/reg: (addr T), f
var/reg: (addr T_f) <- get var: T, f

The f here is the field name from the type definition, and its type T_f must match the type of f in the type definition. For example, some legal instructions for the definition of point above:

var a/eax: (addr int) <- get p, x
var a/eax: (addr int) <- get p, y

You can clear arbitrary types using the clear-object function:

clear-object var: (addr T)

Don’t clear arrays or streams using clear-object; doing so will irreversibly make their length 0 as well.

You can shallow-copy arbitrary types using the copy-object function:

copy-object src: (addr T), dest: (addr T)

Handles for safe access to the heap

We’ve seen the addr type, but it’s intended to be short-lived. addr values should never escape from functions. Function outputs can’t be addrs, function inouts can’t include addr in their payload type. Finally, you can’t save addr values inside compound types. To do that you need a “fat pointer” called a handle that is safe to keep around for extended periods and ensures it’s used safely without corrupting the heap and causing security issues or hard-to-debug misbehavior.

To actually use a handle, we have to turn it into an addr first using the lookup statement.

var y/reg: (addr T) <- lookup x: (handle T)

Now operate on y as usual, safe in the knowledge that you can later recover any writes to its payload from x.

It’s illegal to continue to use an addr after a function that reclaims heap memory. You have to repeat the lookup from the handle. (Luckily Mu doesn’t implement reclamation yet.)

Having two kinds of addresses takes some getting used to. Do we pass in variables by value, by addr or by handle? In inputs or outputs? Here are 3 rules of thumb:

Try to avoid mixing these use cases.

If you have a variable src of type (handle ...), you can save it inside a compound type like this (provided the types match):

var dest/reg: (addr handle T_f) <- get var: (addr T), f
copy-handle src, dest

Or this:

var dest/reg: (addr handle T) <- index arr: (addr array handle T), n
copy-handle src, dest

To create handles to non-array types, use allocate:

var x: (addr handle T)
... initialize x ...
allocate x

To create handles to array types (of potentially dynamic size), use populate:

var x: (addr handle array T)
... initialize x ...
populate x, 3  # array of 3 T's


I said at the start that most instructions map 1:1 to x86 machine code. To enforce type- and memory-safety, I was forced to carve out a few exceptions:

If you’re curious, the compiler summary page has the complete nitty-gritty on how each instruction is implemented. Including the above exceptions.


Anything not allowed here is forbidden, even if the compiler doesn’t currently detect and complain about it. Please contact me or report issues when you encounter a missing or misleading error message.