Asmble is a compiler that compiles WebAssembly code to JVM bytecode. It also contains an interpreter and utilities for working with WASM code from the command line and from JVM languages.
- WASM to JVM bytecode compiler (no runtime required)
- WASM interpreter (instruction-at-a-time steppable)
- Conversion utilities between WASM binary, WASM text, and WASM AST
- Programmatic JVM library for all of the above (written in Kotlin)
- Examples showing how to use other languages on the JVM via WASM (e.g. Rust)
WebAssembly by itself does not have routines for printing to stdout or any external platform features. For this example we'll use the test harness used by the spec. Java 8 must be installed.
Download the latest TAR/ZIP from the releases area and extract it to
asmble/
.
WebAssembly code is either in a binary file (i.e. .wasm
files) or a
text file (i.e. .wast
files). The following code imports the print
function from the test harness. Then it creates a function calling print
for the integer 70 and sets it to be called
on module init:
(module
(import "spectest" "print_i32" (func $print (param i32)))
(func $print70 (call $print (i32.const 70)))
(start $print70)
)
Save this as print-70.wast
. Now to run this, execute:
./asmble/bin/asmble run -testharness print-70.wast
The result will be:
70 : i32
Which is how the test harness prints an integer. See the examples directory for more examples.
Assuming Java 8 is installed, download the latest release and extract it.
The asmble
command is present in the asmble/bin
folder. There are multiple commands in Asmble that can be seen by
executing asmble
with no commands:
Usage:
COMMAND options...
Commands:
compile - Compile WebAssembly to class file
help - Show command help
invoke - Invoke WebAssembly function
run - Run WebAssembly script commands
translate - Translate WebAssembly from one form to another
For detailed command info, use:
help COMMAND
Some of the commands are detailed below.
Running asmble help compile
:
Command: compile
Description: Compile WebAssembly to class file
Usage:
compile <inFile> [-format <inFormat>] <outClass> [-out <outFile>]
Args:
<inFile> - The wast or wasm WebAssembly file name. Can be '--' to read from stdin. Required.
-format <inFormat> - Either 'wast' or 'wasm' to describe format. Optional, default: <use file extension>
-log <logLevel> - One of: trace, debug, info, warn, error, off. Optional, default: warn
<outClass> - The fully qualified class name. Required.
-out <outFile> - The file name to output to. Can be '--' to write to stdout. Optional, default: <outClass.class>
This is used to compile WebAssembly to a class file. See the compilation details for details about how
WebAssembly translates to JVM bytecode. The result will be a .class
file containing JVM bytecode.
NOTE: There is no runtime required with the class files. They are self-contained.
Running asmble help invoke
:
Command: invoke
Description: Invoke WebAssembly function
Usage:
invoke [-in <inFile>]... [-reg <registration>]... [-mod <module>] [<export>] [<arg>]...
Args:
<arg> - Parameter for the export if export is present. Multiple allowed. Optional, default: <empty>
-defmaxmempages <defaultMaxMemPages> - The maximum number of memory pages when a module doesn't say. Optional, default: 5
<export> - The specific export function to invoke. Optional, default: <start-func>
-in <inFile> - Files to add to classpath. Can be wasm, wast, or class file. Named wasm/wast modules here are automatically registered unless -noreg is set. Multiple allowed. Optional, default: <empty>
-log <logLevel> - One of: trace, debug, info, warn, error, off. Optional, default: warn
-mod <module> - The module name to run. If it's a JVM class, it must have a no-arg constructor. Optional, default: <last-in-entry>
-noreg - If set, this will not auto-register modules with names. Optional.
-reg <registration> - Register class name to a module name. Format: modulename=classname. Multiple allowed. Optional, default: <empty>
-res - If there is a result, print it. Optional.
-testharness - If set, registers the spec test harness as 'spectest'. Optional.
This can run WebAssembly code including compiled .class
files. For example, put the following WebAssembly at
add-20.wast
:
(module
(func (export "doAdd") (param $i i32) (result i32)
(i32.add (get_local 0) (i32.const 20))
)
)
This can be invoked via the following with the result shown:
asmble invoke -res -in add-20.wast doAdd 100
That will print 120
. However, it can be compiled first like so:
asmble compile add-20.wast MyClass
Now there is a file called MyClass.class
. Since it has a no-arg constructor because it doesn't import anything (see
compilation details below), it can be invoked as well:
asmble invoke -res -in MyClass.class -reg myMod=MyClass -mod myMod doAdd 100
Note, that any Java class can be registered for the most part. It just needs to have a no-arg consstructor and any referenced functions need to be public, non-static, and with return/param types of only int, long, float, or double.
Running asmble help translate
:
Command: translate
Description: Translate WebAssembly from one form to another
Usage:
translate <inFile> [-in <inFormat>] [<outFile>] [-out <outFormat>]
Args:
-compact - If set for wast out format, will be compacted. Optional.
<inFile> - The wast or wasm WebAssembly file name. Can be '--' to read from stdin. Required.
-in <inFormat> - Either 'wast' or 'wasm' to describe format. Optional, default: <use file extension>
-log <logLevel> - One of: trace, debug, info, warn, error, off. Optional, default: warn
<outFile> - The wast or wasm WebAssembly file name. Can be '--' to write to stdout. Optional, default: --
-out <outFormat> - Either 'wast' or 'wasm' to describe format. Optional, default: <use file extension or wast for stdout>
Asmble can translate .wasm
files to .wast
or vice versa. It can also translate .wast
to .wast
which has value
because it resolves all names and creates a more raw yet deterministic and sometimes more readable .wast
. Technically,
it can translate .wasm
to .wasm
but there is no real benefit.
All Asmble is doing internally here is converting to a common AST regardless of input then writing it out in the desired output.
Asmble is written in Kotlin but since Kotlin is a thin layer over traditional Java, it can be used quite easily in all JVM languages.
The compiler and annotations are deployed to Maven Central. The compiler is written in Kotlin and can be added as a Gradle dependency with:
compile 'com.github.cretz.asmble:asmble-compiler:0.3.0'
This is only needed to compile of course, the compiled code has no runtime requirement. The compiled code does include some annotations (but in Java its ok to have annotations that are not found). If you do want to reflect the annotations, the annotation library can be added as a Gradle dependency with:
compile 'com.github.cretz.asmble:asmble-annotations:0.3.0'
To manually build, clone the repository:
git clone --recursive https://github.com/cretz/asmble
The reason we use recursive is to clone the spec submodule we have embedded at src/test/resources/spec
. Unlike many
Gradle projects, this project chooses not to embed the Gradle runtime library in the repository. To assemble the entire
project with Gradle installed and on the PATH
(tested with 4.6), run:
gradle :compiler:assembleDist
The API documentation is not yet available at this early stage. But as an overview, here are some useful classes and packages:
asmble.ast.Node
- All WebAssembly AST nodes as static inner classes.asmble.cli
- All code for the CLI.asmble.compile.jvm.AstToAsm
- Entry point to go from AST module to ASM ClassNode.asmble.compile.jvm.Mem
- Interface that can be implemented to change how memory is handled. Right nowByteBufferMem
in the same package is the only implementation and it emitsByteBuffer
.FuncBuilder
- Where the bulk of the WASM-instruction-to-JVM-instruction translation happens.asmble.io
- Classes for translating to/from ast nodes, bytes (i.e. wasm), sexprs (i.e. wast), and strings.asmble.run.jvm
- Tools for running WASM code on the JVM. SpecificallyScriptContext
which helps with linking.asmble.run.jvm.interpret
- The interpreter that can run WASM all at once or allow it to be stepped one instruction at a time.
Note, some code is not complete yet (e.g. a linker and javax.script
support) but beginnings of the code still appear
in the repository.
And for those reading code, here are some interesting algorithms:
asmble.compile.jvm.RuntimeHelpers#bootstrapIndirect
(in Java, not Kotlin) - Manipulating arguments to essentially chainMethodHandle
calls for aninvokedynamic
bootstrap. This is actually taken from the compiled Java class and injected as a synthetic method of the module class if needed.asmble.compile.jvm.msplit
(in Java, not Kotlin) - A rudimentary JVM method bytecode splitter for when method sizes exceed the limit allowed by the JVM (embedded from another project).asmble.compile.jvm.InsnReworker#addEagerLocalInitializers
- Backwards navigation up the instruction list to make sure that a local is set before it is get.asmble.compile.jvm.InsnReworker#injectNeededStackVars
- Inject instructions at certain places to make sure we have certain items on the stack when we need them.asmble.io.ByteReader$InputStream
- A simple eof-peekable input stream reader.asmble.run.jvm.interpret.Interpreter
- Full WASM interpreter in a few hundred lines of Kotlin.
Asmble does its best to compile WASM ops to JVM bytecodes with minimal overhead. Below are some details on how each part is done. Every module is represented as a single class. This section assumes familiarity with WebAssembly concepts.
Asmble creates different constructors based on the memory requirements. Each constructor created contains the imports as parameters (see imports below)
If the module does not define memory, a single constructor is created that accepts all other imports. If the module does define memory, two constructors are created: one accepting a memory instance, and an overload that instead accepts an integer value for max memory that is used to create the memory instance before sending to the first one. If the maximum memory is given for the module, a third constructor is created without any memory parameters and just calls the max memory overload w/ the given max memory value. All three of course have other imports as the rest of the parameters.
After all other constructor duties (described in sections below), the module's start function is called if present.
Memory is built or accepted in the constructor and is stored in a field. The current implementation uses a ByteBuffer
.
Since ByteBuffer
s are not dynamically growable, the max memory is an absolute max even though there is a limit which
is adjusted on grow_memory
. Any data for the memory is set in the constructor.
In the WebAssembly MVP a table is just a set of function pointers. This is stored in a field as an array of
MethodHandle
instances. Any elements for the table are set in the constructor.
Globals are stored as fields on the class. A non-import global is simply a field that is final if not mutable. An import
global is a MethodHandle
to the getter and a MethodHandle
to the setter if mutable. Any values for the globals are
set in the constructor.
The constructor accepts all imports as params. Memory is imported via a ByteBuffer
param, then function
imports as MethodHandle
params, then global imports as MethodHandle
params (one for getter and another for setter if
mutable), then a MethodHandle
array param for an imported table. All of these values are set as fields in the
constructor.
Exports are exported as public methods of the class. The export names are mangled to conform to Java identifier requirements. Function exports are as is whereas memory, global, and table exports have the name capitalized and are then prefixed with "get" to match Java getter conventions.
Exports are always separate methods instead of just changing the name of an existing method or field. This encapsulation allows things like many exports for a single item.
WebAssembly has 4 types: i32
, i64
, f32
, and f64
. These translate quite literally to int
, long
, float
, and
double
respectively.
Operations such as unreachable
(which throws) behave mostly as expected. Branching and looping are handled with jumps.
The problem that occurs with jumping is that WebAssembly does not require compiler writers to clean up their own stack.
Therefore, if the WASM ops have extra stack values, we pop it before jumping which has performance implications but not
big ones. For most sane compilers, the stack will be managed stringently and leftover stack items will not be present.
Luckily, br_table
jumps translate literally to JVM table switches which makes them very fast. There is a special set
of code for handling really large tables (because of Java's method limit) but this is unlikely to affect most in
practice.
Normal call
operations do different things depending upon whether it is an import or not. If it is an import, the
MethodHandle
is retrieved from a field and called via invokeExact
. Otherwise, a normal invokevirtual
is done to
call the local method.
A call_indirect
is done via invokedynamic
on the JVM. Specifically, invokedynamic
specifies a synthetic bootstrap
method that we create. It does a one-time call on that bootstrap method to get a MethodHandle
that can be called in
the future. We wouldn't normally have to use invokedynamic
because we could use the index to reference a
MethodHandle
in the array field. However, in WebAssembly, that index is after the parameters of the call and the
stack manipulation we would have to do would be far too expensive.
So we need a MethodHandle that takes the params of the target method, and then the index, to make the call. But we
also need "this" because it is expected at some point in the future that the table field could be changed underneath and
we don't want that field reference to be cached via the one-time bootstrap call. We do this with a synthetic bootstrap
method which uses some MethodHandle
trickery to manipulate it the way we want. This makes indirect calls very fast,
especially on successive invocations.
A drop
translates literally to a pop
. A select translates to a conditional swap, then a pop.
Local variable access translates fairly easily because WebAssembly and the JVM treat the concept of parameters as the initial locals similarly. Granted the JVM form has "this" at slot 0. Also, WebAssembly doesn't treat 64-bit vars as 2 slots like the JVM, so some simple math is done like it is with the stack.
WebAssembly requires all locals the assume they are 0 whereas the JVM requires locals be set before use. An algorithm in Asmble makes sure that locals are set to 0 before they are fetched in any situation where they weren't explicitly set first.
Global variable access depends on whether it's an import or not. Imports call getter MethodHandle
s whereas non-imports
simply do normal field access.
Memory operations are done via ByteBuffer
methods on a little-endian buffer. All operations including unsigned
operations are tailored to use specific existing Java stdlib functions.
As a special optimization, we put the memory instance as a local var if it is accessed a lot in a function. This is cheaper than constantly fetching the field.
Constants are simply ldc
bytecode ops on the JVM. Comparisons are done via specific bytecodes sometimes combined with
JVM calls for things like unsigned comparison. Operators use idiomatic JVM approaches as well.
The WebAssembly spec requires a runtime check of overflow during trunc
calls. This is enabled by default in Asmble. It
defers to an internal synthetic method that does the overflow check. This can be programmatically disabled for better
performance.
Asmble maintains knowledge of types on the stack during compilation and fails compilation for any invalid stack items. This includes the somewhat complicated logic concerning unreachable code.
In several cases, Asmble needs something on the stack that WebAssembly doesn't, such as "this" before the value of a
putfield
call when setting a non-import global. In order to facilitate this, Asmble does a preprocessing of the
instructions. It builds the stack diffs and injects the needed items (e.g. a reference to the memory class for a load)
at the right place in the instruction list to make sure they are present when needed.
As an unintended side effect of this kind of logic, it turns out that Asmble never needs local variables beyond what WebAssembly specifies. No temp variables or anything. It could be argued however that the use of temp locals might make some of the compilation logic less complicated and could even improve runtime performance in places where we overuse the stack (e.g. some places where we do a swap).
Below are some performance and implementation quirks where there is a bit of an impedance mismatch between WebAssembly and the JVM:
- WebAssembly has a nice data section for byte arrays whereas the JVM does not. Right now we use a single-byte-char
string constant (i.e. ISO-8859 charset). This saves class file size, but this means we call
String::getBytes
on init to load bytes from the string constant. Due to the JVM using an unsigned 16-bit int as the string constant length, the maximum byte length is 65536. Since the string constants are stored as UTF-8 constants, they can be up to four bytes a character. Therefore, we populate memory in data chunks no larger than 16300 (nice round number to make sure that even in the worse case of 4 bytes per char in UTF-8 view, we're still under the max). - The JVM makes no guarantees about trailing bits being preserved on NaN floating point representations like WebAssembly does. This causes some mismatch on WebAssembly tests depending on how the JVM "feels" (I haven't dug into why some bit patterns stay and some don't when NaNs are passed through methods).
- The JVM requires strict stack management where the compiler writer is expected to pop off what he doesn't use before performing unconditional jumps. WebAssembly requires the runtime to discard unused stack items before unconditional jump so we have to handle this. This can cause performance issues because essentially we do a "pop-before-jump" which pops all unneeded stack values before jumping. If the target of the jump expects a fresh item on the stack (i.e. a typed block) then it gets worse because we have to pop what we don't need except for the last stack value which leads to a swap-pop-and-swap. Hopefully in real world use, tools that compile to WebAssembly don't have a bunch of these cases. If they do, we may need to look into spilling to temporary local vars.
- Both memory and tables have "max capacity" and "initial capacity". While memory uses a
ByteBuffer
which has these concepts (i.e. "capacity" and "limit"), tables use an array which only has the "initial capacity". This means that tests that check for max capacity on imports at link time do not fail because we don't store max capacity for a table. This is not a real problem for the MVP since the table cannot be grown. But once it can, we may need to consider bringing another int along with us for table max capacity (or at least make it an option). - WebAssembly has a concept of "unset max capacity" which means there can theoretically be an infinite capacity memory
instance.
ByteBuffer
s do not support this, but care is taken to allow link time and runtime max memory setting to give the caller freedom. - WebAssembly requires some trunc calls to do overflow checks, whereas the JVM does not. So for example, WebAssembly
has
i32.trunc_s/f32
which would usually be a simplef2i
JVM instruction, but we have to do an overflow check that the JVM does not do. We do this via a private static synthetic method in the module. There is too much going on to inline it in the method and if several functions need it, it can become hot and JIT'd. This may be an argument for a more global set of runtime helpers, but we aim to be runtime free. Care was taken to allow the overflow checks to be turned off programmatically. - WebAssembly allows unsigned 32 bit int memory indices.
ByteBuffer
only has signed which means the value can overflow. And in order to support even larger sets of memory, WebAssembly supports constant offsets which are added to the runtime indices. Asmble will eagerly fail compilation if an offset is out of range. But at runtime we don't check by default and the overflow can wrap around and access wrong memory. There is an option to do the overflow check when added to the offset which is disabled by default. Other than this there is nothing we can do easily.
Why?
I like writing compilers and I needed a sufficiently large project to learn Kotlin really well to make a reasonable judgement on it. I also wanted to become familiar w/ WebAssembly. I don't really have a business interest for this and therefore I cannot promise it will forever be maintained.
Will it work on Android?
I have not investigated. But I do use invokedynamic
and MethodHandle
so it would need to be a modern version of
Android. I assume, then, that both runtime and compile-time code might run there. Experiment feedback welcome.
What about JVM to WASM?
This is not an immediate goal of this project, at least not until the WASM GC proposal has been accepted. In the meantime, there is https://github.com/konsoletyper/teavm
So I can compile something in C via Emscripten and have it run on the JVM with this?
Yes, but work is required. WebAssembly is lacking any kind of standard library. So Emscripten will either embed it or import it from the platform (not sure which/where, I haven't investigated). It might be a worthwhile project to build a libc-of-sorts as Emscripten knows it for the JVM. Granted it is probably not the most logical approach to run C on the JVM compared with direct LLVM-to-JVM work.
Debugging?
Not yet, once source maps get standardized I may revisit.
- Add "dump" that basically goes from WebAssembly to "javap" like output so details are clear
- Expose the advanced compilation options
- Add "link" command that will build an entire JAR out of several WebAssembly files and glue code between them
- Annotations to make it clear what imports are expected
- Compile some parts to JS and native with Kotlin
- Add
javax.script
support (which can give things like a free repl w/ jrunscript)