VMIR is a standalone library written in C that can parse and execute:
- WebAssembly
.wasmfiles - LLVM Bitcode
.bcfiles
Optionally it can generate machine code (JIT) to speed up execution significantly. JIT is currently only supported on 32 bit ARM.
VMIR is licensed under the MIT license. See LICENSE.
To build VMIR just type:
$ make
... and you will end up with a VMIR binary in very same directory. The library is compiled from a single file src/vmir.c which in turn include other files to keep the code somewhat separated.
Let's create a small program and run it. Type the following well known snippet into a file called helloworld.c
int main(void)
{
printf("Hello world\n");
return 0;
}
Then compile it
clang -emit-llvm -target le32-unknown-nacl -c helloworld.c -o helloworld.bc
And finally, run it:
$ ./vmir helloworld.bc
Hello world
Compiling to WebAssembly requires a bit more work than a single line. However,
there is a pre-built version of sha1sum included in the source repo.
$ echo hello | ./vmir examples/prebuilt/sha1sum.wasm
f572d396fae9206628714fb2ce00f72e94f2258f -
If you're on Linux and want to go all crazy you can use VMIR to execute Bitcode and WebAssembly directly from the command line by installing a kernel binfmt pointing to the VMIR executable:
echo ":vmirwasm:M::\x00\x61\x73\x6d\x01::${PWD}/vmir:" | sudo tee /proc/sys/fs/binfmt_misc/register
echo ":vmirbc:M::\x42\x43\xc0\xde::${PWD}/vmir:" | sudo tee /proc/sys/fs/binfmt_misc/register
And then you just simply just do:
$ echo hello | examples/prebuilt/sha1sum.wasm
f572d396fae9206628714fb2ce00f72e94f2258f -
Interpretation is about 10x slower (on x86) than the same binary compiled as native code. Still it's a lot faster than LLVM's own interpreter (which by all means is not intended to run code fast in any way)
Example run of test/misc/src/sha1test.c over 64MB of random data
| Environment | (Core i7 3.2GHz) | ARMv7 BCM2709 (Rpi2) |
|---|---|---|
| Native | 0.39s | 3.54s |
| VMIR JIT | n/a | 17.5s |
| VMIR | 4.8s | 1m 42s |
| LLVM LLI | 7m 39s | n/a |
VMIR currently passes the gcc torture test suite on optimization level 0, 1 and 2. Those tests can be found in test/gcc-torture. Use make && ./runtest to run the tests.
- The built-in libc is lacking a lot of functions and features. This is where most work needs to be done.
- No support for vector types (Ie, code must be compiled with
-fno-vectorize -fno-slp-vectorize). - Not all instructions classes / value types are JITed.
- No C++ STL solution. Ideas welcome...
VMIR uses the same target as Google NativeClient. There are small examples in test/misc.
When building bigger projects consisting of multiple files you must llvm-link to combine the .bc files into a single file.
Building for WebAssembly is a bit more involved atm. There is a document here: docs/wasm.md, that shows how to setup LLVM + Binaryen and the WebAssembly Binary Toolkit. Once you have that in place there are some small examples in examples/wasm that could get you started.
Including VMIR in your own project is pretty straight forward. Just copy the files from src/ to your project but only compile vmir.c (it will include all other .c -files on its own). The API is defined in vmir.h. See src/main.c for example how to load and execute binaries.
VMIR's libc also offers an option to use TLSF for memory allocation. The default built-in allocator is a very simple linear search first-fit algorithm.
You might ask yourself what the purpose of VMIR actually is and why it even exists?
As with many of these kind of project I just wanted to scratch and itch but also be able to ship plugins written in C and C++ for another project of mine.
Now with the rise of WebAssembly I intend to focus more on that as the primary input to VMIR mostly because it's more stable than LLVM's Bitcode, which is not really meant to be used as a shippable object code.
Given enough time I also hope to improve the JIT engine to be able to emit code for more architectures (In particular ARMv8 and x86_64).
Follow me on https://twitter.com/andoma