Compile WebAssembly to JVM and other WASM tools
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.
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
WebAssembly code is either in a binary file (i.e.
.wasmfiles) or a text file (i.e.
.wastfiles). The following code imports the
(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
asmblecommand is present in the
asmble/binfolder. There are multiple commands in Asmble that can be seen by executing
asmblewith 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.
asmble help compile:
Command: compile Description: Compile WebAssembly to class file Usage: compile [-format ] [-out ]
Args: - The wast or wasm WebAssembly file name. Can be '--' to read from stdin. Required. -format - Either 'wast' or 'wasm' to describe format. Optional, default: -log - One of: trace, debug, info, warn, error, off. Optional, default: warn - The fully qualified class name. Required. -out - The file name to output to. Can be '--' to write to stdout. Optional, default: <outclass.class> </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
.classfile containing JVM bytecode.
NOTE: There is no runtime required with the class files. They are self-contained.
asmble help invoke:
Command: invoke Description: Invoke WebAssembly function Usage: invoke [-in ]... [-reg ]... [-mod ]  ...
Args: - Parameter for the export if export is present. Multiple allowed. Optional, default: -defmaxmempages - The maximum number of memory pages when a module doesn't say. Optional, default: 5 - The specific export function to invoke. Optional, default: -in - 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: -log - One of: trace, debug, info, warn, error, off. Optional, default: warn -mod - The module name to run. If it's a JVM class, it must have a no-arg constructor. Optional, default: -noreg - If set, this will not auto-register modules with names. Optional. -reg - Register class name to a module name. Format: modulename=classname. Multiple allowed. Optional, default: -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
.classfiles. For example, put the following WebAssembly at
(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.
asmble help translate:
Command: translate Description: Translate WebAssembly from one form to another Usage: translate [-in ]  [-out ]
Args: -compact - If set for wast out format, will be compacted. Optional. - The wast or wasm WebAssembly file name. Can be '--' to read from stdin. Required. -in - Either 'wast' or 'wasm' to describe format. Optional, default: -log - One of: trace, debug, info, warn, error, off. Optional, default: warn - The wast or wasm WebAssembly file name. Can be '--' to write to stdout. Optional, default: -- -out - Either 'wast' or 'wasm' to describe format. Optional, default:
Asmble can translate
.wastor vice versa. It can also translate
.wastwhich has value because it resolves all names and creates a more raw yet deterministic and sometimes more readable
.wast. Technically, it can translate
.wasmbut 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:
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:
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:
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 now
ByteBufferMemin the same package is the only implementation and it emits
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. Specifically
ScriptContextwhich 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.scriptsupport) 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 chain
MethodHandlecalls for an
invokedynamicbootstrap. 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
ByteBuffers 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
MethodHandleinstances. 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
MethodHandleto the getter and a
MethodHandleto 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
ByteBufferparam, then function imports as
MethodHandleparams, then global imports as
MethodHandleparams (one for getter and another for setter if mutable), then a
MethodHandlearray 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:
f64. These translate quite literally to
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.
br_tablejumps 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.
calloperations do different things depending upon whether it is an import or not. If it is an import, the
MethodHandleis retrieved from a field and called via
invokeExact. Otherwise, a normal
invokevirtualis done to call the local method.
call_indirectis done via
invokedynamicon the JVM. Specifically,
invokedynamicspecifies a synthetic bootstrap method that we create. It does a one-time call on that bootstrap method to get a
MethodHandlethat can be called in the future. We wouldn't normally have to use
invokedynamicbecause we could use the index to reference a
MethodHandlein 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
MethodHandletrickery to manipulate it the way we want. This makes indirect calls very fast, especially on successive invocations.
droptranslates 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
MethodHandles whereas non-imports simply do normal field access.
Memory operations are done via
ByteBuffermethods 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
ldcbytecode 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
trunccalls. 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
putfieldcall 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:
String::getByteson 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).
ByteBufferwhich 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).
ByteBuffers do not support this, but care is taken to allow link time and runtime max memory setting to give the caller freedom.
i32.trunc_s/f32which would usually be a simple
f2iJVM 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.
ByteBufferonly 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.
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
MethodHandleso 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.
Not yet, once source maps get standardized I may revisit.
javax.scriptsupport (which can give things like a free repl w/ jrunscript)