14 KB
Newer Older
# Mu Reference Implementation version 2
Kunshan Wang's avatar
Kunshan Wang committed

Kunshan Wang's avatar
Kunshan Wang committed
3 4
[![Build Status](](

Kunshan Wang's avatar
Kunshan Wang committed
Version 2.2.0
Kunshan Wang's avatar
Kunshan Wang committed

Kunshan Wang's avatar
Kunshan Wang committed
This project is the current reference implementation of Mu, the micro virtual
machine designed by [The Micro Virtual Machine Project](
Kunshan Wang's avatar
Kunshan Wang committed

Kunshan Wang's avatar
Kunshan Wang committed
10 11
Version 2.2.0 implements the current [Mu
Kunshan Wang's avatar
Kunshan Wang committed

## How to compile
Kunshan Wang's avatar
Kunshan Wang committed

Kunshan Wang's avatar
Kunshan Wang committed
15 16
**For the impatient**:

Kunshan Wang's avatar
Kunshan Wang committed
17 18
* Install JDK 8. If you use Mac, download from
Kunshan Wang's avatar
Kunshan Wang committed
19 20 21 22 23
* If you use Mac, install [Homebrew](
* Install [Scala]( 2.11. If you use Mac and Homebrew,
  `brew install scala`.
* Install [sbt]( 0.13. If you use Mac and Homebrew,
  `brew install sbt`.
* Install [Scala IDE]( 4.x (Eclipse with pre-installed
Kunshan Wang's avatar
Kunshan Wang committed
25 26 27 28
  plugins for Scala).
* Clone this repository:

Kunshan Wang's avatar
Kunshan Wang committed
29 30 31 32 33 34 35
git clone

If you do not have SSH access to the ANU GitLab repositories, use the HTTPS URL:

git clone
Kunshan Wang's avatar
Kunshan Wang committed
36 37

Kunshan Wang's avatar
Kunshan Wang committed
* In the directory `mu-impl-ref2`, do the following:
Kunshan Wang's avatar
Kunshan Wang committed
39 40

sbt update genSrc eclipse
Kunshan Wang's avatar
Kunshan Wang committed
42 43 44 45 46 47 48

* Open Scala IDE and import the generated project as "existing project into

**Detailed guide**:

Kunshan Wang's avatar
Kunshan Wang committed
49 50 51 52
The reference implementation is developed and tested with Java VM 8. You need a
JRE to build the Scala/Java part, and a JDK to build the C binding.

You also need [Scala]( 2.11 and
[sbt]( 0.13. It is recommended to install them using
Kunshan Wang's avatar
Kunshan Wang committed
54 55 56 57 58 59 60 61 62 63
the package manager of your operating system or distribution (such as apt-get,
yum, pacman, etc. for GNU/Linux distributions and Homebrew for Mac OS X) if such
packages are available.

For Ubuntu users: Ubuntu 15.10 does not provide sbt in its repository. Please
[download sbt]( from the official sbt web
site, or follow the [official sbt installing guide for
Linux](  If
you experience any "certificate" problems, [this
page]( provides a solution.
Kunshan Wang's avatar
Kunshan Wang committed

Kunshan Wang's avatar
Kunshan Wang committed
65 66
Then after cloning this repository, you can simply invoke `sbt compile` to
compile this project. Or you can do it step by step:
Kunshan Wang's avatar
Kunshan Wang committed

Kunshan Wang's avatar
Kunshan Wang committed
68 69 70 71 72 73 74 75
* To download all dependencies from the Maven central repository, invoke `sbt

* To generate the Mu IR parser from the Antlr grammar, invoke `sbt genSrc`. The
  generated sources will be in the `target/scala-2.11/src_managed` directory.

* To compile, invoke `sbt compile`. This will also generate the Mu IR parser
  using Antlr.
Kunshan Wang's avatar
Kunshan Wang committed

To generate an Eclipse project, install the [sbt-eclipse
plugin]( and invoke `sbt eclipse`.
Kunshan Wang's avatar
Kunshan Wang committed
79 80
Make sure you generate the parser (`sbt genSrc`) before creating the Eclipse
project, so that the generated sources will be on the Eclipse build path.
81 82 83

IntelliJ IDEA has plugins for Scala and SBT. Make sure you don't commit `.idea`
or generated project files into the repository.
Kunshan Wang's avatar
Kunshan Wang committed

85 86 87 88 89 90 91 92
### C binding and Python binding

The C binding is in the `cbinding` directory. Just run `make` inside `cbinding`.

The Python binding is in the `pythonbinding` directory. It depends on the C
binding, so make sure you make the C binding first. The Python binding does not
need to be built.

93 94
See [cbinding/](cbinding/ and
[pythonbinding/](pythonbinding/ for more details.

## How to run

98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131
For the impatient: Execute the following command and see Mu running a factorial

sbt 'set fork:=true' 'test:runMain junks.FactorialFromRPython'

### C API

The reference implementation implements the [Mu Client
API]( which allows C
programs to control the micro VM and construct and deliver bundles for the micro
VM to execute.

See [cbinding/](cbinding/ for more details.

### Scala API

The micro VM itself is implemented in Scala.

- `uvm.refimpl.MicroVM` is the counterpart of the `MuVM` struct in the [Mu
  Client API]( It can
  be instantiated with VMConf options explained below.
- `uvm.refimpl.MuCtx` is the counterpart of the `MuCtx` struct in C.
- `uvm.refimpl.MuValue` and its subclasses implement the `MuValue` handles.

As an implementation detail, the micro VM will not start execution until
`MicroVM.execute()` is called. See implementation details below.

The Scala interface is closer to the Scala's style. For example, the
`MuCtx.dumpKeepalives()` method returns a `Seq[MuValue]` rather than writing the
results into a given array. It also does more static type checking than the C

132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150
There is a sample factorial program (generously provided by @johnjiabinzhang) in
the `src/test` directory. To run the program with all dependencies on the
classpath, you need to run it with sbt. Invoke `sbt` to enter the interactive
shell. Then type:

set fork := true
test:runMain junks.FactorialFromRPython

or directly from the command line:

sbt 'set fork:=true' 'test:runMain junks.FactorialFromRPython'

`fork := true` tells sbt to run the program in a different process than the one
running sbt itself.

151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191
### Boot Image

The reference implementation can create boot images, a package that contains a
Mu IR bundle and a serialised Mu memory, including the global memory and the

Boot images can be created using the standard `make_boot_image` method on the
MuVM object. In this reference implementation, the boot image is a zip file. By
convention, boot images have the file-name extension `.muref`.

Before a boot image can be executed, an entry point needs to be specified. Use
the `tools/` script to set the entry point by ID or name. The entry point
is a Mu function that takes an `int<32>` and a `uptr<uptr<int<8>>>` as
parameters, the same as the `main` function in C.

The `tools/` script can also specify extra libraries to be loaded when the
micro VM loads the boot image. EXTERN constants will be resolved from these
libraries in the order of those libraries.

The `tools/` script runs the micro VM with the given boot image.
Additional arguments are passed to the entry point.

### Micro VM Configuration

There are some parameters that controls the behaviour of the reference

When using the C API, the refimpl-specific
[cbinding/refimpl2-start.h](cbinding/refimpl2-start.h) header provides the
`mu_refimpl_new_ex` function which accepts a C-style string. The options are
encoded as `key=value` pairs, one option per line, with no spaces between the
equal sign.

When using the `tools/` script, the options are specified as
command-line options in the form `--key=value` before the boot image file name.


*Sizes may have suffixes K, M, G or T. 1K = 1024 bytes. sosSize, losSize and
globalSize must be a multiple of 32768 bytes (32K).*

192 193 194 195 196 197 198
- **sosSize**: The size of the small object space in bytes. default: 2M
- **losSize**: The size of the large object space in bytes. default: 2M
- **globalSize**: The size of the large object space in bytes. default: 1M
- **stackSize**: The size of each stack in bytes. default: 60K
- **dumpBundle**: Print out the bundle as text every time a bundle is loaded.
  default: false
- **staticCheck**: Run static checker after each bundle is loaded. default: true
199 200
- **sourceInfo**: Provide line/column info in Mu IR when errors occur. May be
  useful for debugging small Mu IR bundles, but will significantly slow down
  parsing!!!  Enable only if the bundle is small. default: false
202 203 204 205 206
- **automagicReloc**: Allow "automagic" relocation.  If true, `uptr` and
  `ufuncptr` fields will also be traced during boot image building.  If a `uptr`
  field points to a global cell field, it will still point to the same field
  after boot image loading; if a `ufuncptr` points to a native function, it will
  point to the same function after boot image loading.  default: false
207 208
- **extraLibs**: Extra libraries to load when starting the micro VM. This is a
  colon-separated list of libraries. Each library has the same syntax of the
209 210 211 212
  `path` argument of the `dlopen` system function. By default, it does not load
  any extra libraries.
- **bootImg**: The path to the boot image. Only useful in the C API. By default,
  it does not load any boot image.
213 214 215 216 217
- **uPtrHack**: When true, it will allow memory locations of general reference
  types to be accessed by `uptr<T>`. By default, such fields can only be
  accessed by `iref<T>`, but this hack is necessary for the current
  [mu-client-pypy]( project to
  work. default: false
218 219 220

*Log levels can be: ALL, TRACE, DEBUG, INFO, WARN, ERROR, OFF. Case-insensitive.
Setting to WARN should get rid of most logging information, except the serious
ones.* The default log level is DEBUG.

223 224 225 226
- **vmLog**: The log level of the micro VM (the "uvm" package)
- **gcLog**: The log level of the garbage collector (the "uvm.refimpl.mem"
  package). If vmLog is set but gcLog is not, it will use the log level of

228 229 230 231 232 233 234 235
## Implementation details

This reference implementation aims to be easy to work with, but does not have
high performance. It may be used by client writers to evaluate the Mu micro VM,
and may also be used by Mu micro VM implementers as a reference to compare with.

The micro VM is implemented as an interpreter written in Scala. The main class
is `uvm.refimpl.MicroVM`, which implements the `MuVM` struct specified by the
Kunshan Wang's avatar
Kunshan Wang committed
236 237 238 239 240 241
[client API](, but is
more Scala-like. The client interacts with the micro VM via `uvm.refimpl.MuCtx`
instances created by the `MicroVM` instance, which corresponds to the `MuCtx`
struct in the spec. `uvm.refimpl.MuValue` and its subclasses implement the
`MuValue` handles, but has a real Scala type hierarchy and does extra type
checking when converting, which is not required by the spec.

243 244
The client can also be written in C, Python or other languages that can
interface with C.
245 246 247 248 249 250 251 252 253 254 255 256 257 258

### Threading model

It uses green threads to execute multiple Mu threads and uses a round-robin
scheduler: the interpreter iterates over all active threads, executes one
instruction for each active thread, then repeat this process. However, the whole
Scala-based program itself is **not thread safe**. Do not run multiple JVM or
native threads. This means, you can still experiment with concurrent Mu
programs, but there are some corner cases that do not work in this refimpl. For

- Waiting for other Mu threads in the trap handler. The trap handler is executed
  by the same thread executing the Mu IR. During trap handler, no Mu program is
  executed. So if you want to use
Kunshan Wang's avatar
Kunshan Wang committed
260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294
  to wait for certain Mu thread to come to a certain rendezvous point (a common
  optimisation trick), you should either wait within Mu IR (not in trap
  handlers) or try the high-performance Mu implementation which is still being

- Synchronising with concurrent native programs via pointers, atomic memory
  access and futex. This is the realistic way for Mu to synchronise with
  native programs or foreign languages, but this refimpl implements atomic
  memory access as not-atomic (since it uses green thread) and implements futex
  in Scala (since it has its own scheduler).

The MicroVM instance will not start executing unless its `execute()` method is
called. This method is specific to this implementation, and is not defined in
the specification. This also means the *client cannot run concurrently with the
MicroVM*, i.e. once started, the client can only intervene in the execution in
**trap handlers**. So a common use pattern is:

val microVM = new MicroVM()

val uir = myCompiler.compile(sourceCode)
val ctx = microVM.newContext()

microVM.setTrapHandler(theTrapHandler)  // Set the trap handler so the client
                                        // can talk with the VM when trapped.

val stack = ctx.newStack(theMainFunction)
val thread = ctx.newThread(stack, Seq(params, to, the, main, function))

microVM.execute() // The current JVM thread will run on behalf of the MicroVM.
                  // This blocks until all Mu threads stop.
                  // However, MicroVM will call theTrapHandler.

295 296
The refimpl implements the text-based IR and HAIL as well as the IR-builder API
to construct Mu IR ASTs programmatically. 
297 298 299 300 301 302 303 304 305 306 307 308 309

### Garbage collection

This reference implementation has an exact tracing garbage collector with a
mark-region small object space and a mark-sweep large object space.

### IR implementation-specific details

- Many undefined behaviours in the specification will raise
  `UvmRuntimeException`, such as division by zero, going below the last frame of
  a stack, accessing a NULL reference, etc. But this behaviour is not

310 311 312 313
- `int<n>` for n = 1 to 128 are implemented. `vec<T n>` is implemented for all T
  that are int, float or double, and all n >= 1. However, only 8, 16, 32, 64,
  128-bit integers, float, double, `vec<int<32> 4>`, `vec<float 4>` and
  `vec<double 2>` can be loaded or stored from the memory.
314 315 316 317 318 319 320 321 322 323

- The tagged reference type `tagref64` is fully implemented.

- Out-of-memory errors will terminate the VM rather than letting the Mu IR
  handle such failures via the exception clauses.

### Native interface

This reference implementation assumes it is running on x86-64 on either Linux or
OSX. It implements the [AMD64 Unix Native
Kunshan Wang's avatar
Kunshan Wang committed
325 326 327 328 329 330 331
of the specification. It can call native functions from Mu IR and let native
programs call back to Mu IR. 

It does not support throwing Mu exceptions into native programs, or handing
C++-based exceptions.

## Author and Copyright
Kunshan Wang's avatar
Kunshan Wang committed
332 333 334 335 336 337

This project is created by Kunshan Wang, Yi Lin, Steve Blackburn, Antony
Hosking, Michael Norrish.

This project is released under the CC-BY-SA license. See `LICENSE`.

## Contact
Kunshan Wang's avatar
Kunshan Wang committed

340 341 342 343 344
Kunshan Wang <>

vim: tw=80