A super-quick FFI for Node.js.
dyncall
is used to make dynamic calls to native
functions. In order to avoid some cost of translating JavaScript values into raw
C types, a shared buffer is used for both arguments and return values. Writing
values to a buffer turns out to be quite a bit faster than unpacking them in
native code.
sbffi.getNativeFunction(pathToSharedLibrary, functionName, returnType, [argType1, argType2, ...])
All the arguments are strings. The types must be standard C types. See the Types section below for details. When functions take 64-bit types, the parameters must be passed as BigInts. 64-bit return values will also be BigInts.
// adder.c: some C library compiled to libadder.so
uint32_t add(uint32_t a, uint32_t b) {
return a + b;
}
// index.js
const { getNativeFunction } = require('sbffi');
const libPath = '/path/to/libadder.so';
const add = getNativeFunction(libPath, 'add', 'uint32_t', ['uint32_t', 'uint32_t']);
const result = add(23, 34);
// 57
The following types are supported:
(u)int[8|16|32|64]_t
bool
(unsigned) char
(unsigned) short
(unsigned) int
(unsigned) long
(unsigned) long long
float
double
size_t
128-bit types are not yet supported, and while this list may grow over time, for now other types can be used if they're aliases of the above types.
See the section below about pointers.
Pointers are currently assumed to be 64-bit, and can be passed to native
functions by specifying the type as pointer
or referring to any other type
with an asterisk in the string, for example: uint8_t *
.
You can put raw data into a Buffer, and then get a pointer to the start of that buffer with:
const bufferPointer = sbffi.getBufferPointer(buffer);
Arrays and strings must be passed as pointers.
You can use two different styles of callbacks with sbffi
.
-
Simple callbacks are simply passed into the function as normal. They must be called exactly once by the underlying native function.
-
Advanced callbacks must be wrapped with
createCallback
, and.destroy()
must be called on them when the underlying native function will no longer call it. There's a slight performance advantage in using advanced callbacks and re-using them, since simple callbacks create aNapi::ThreadsafeFunction
per invocation. In addition, APIs that may call the same callback multiple times may be used with advanced callbacks, but not with simple callbacks.
In either case, to specify a simple callback, identify it in the arguments array
passed to getNatveFunction()
as [cbReturnType, [cbArgTyp1, cbArgType2, ...]]
.
For advanced callbacks, after specifying them in the native function signature, you can initialize them as follows:
function myCb (result) { /* ... */ }
const advancedCallback = createCallback(myCb, [cbReturnType, [cbArgTyp1, cbArgType2, ...]]);
You can then pass advancedCallback
to a native function that takes in a
callbacks with that signature, just as you would any other callback. When the
callback is no longer needed, you can call advancedCallback.destroy()
.
Failure to call .destroy()
will keep the Node.js process alive.
For now, sbfffi
doesn't have any built-in support for structs. That being
said, there are some helpful libraries like
shared-structs
and
ref-napi
(and its family of
modules). As long as you can build up a C struct into a Buffer, you can pass
pointers to them into C functions. Non-pointer struct arguments or return values
are not supported.
Using a non-release version of sbffi
requires that
cmake
is installed in order to compile the native
addon.
A simple benchmark can be run with npm run bench
. This will test calling a
simple adding function from the test library using the following techniques:
ffi-napi
: A successor tonode-ffi
compatible with modern versions of Node.js.sbffi
: This library.napi-addon
: A very simple/normal Node.js addon using NAPI in C.napi-addon-sb
: A NAPI addon using the same shared-buffer technique assbffi
, but with a hard-coded function call, rather than a dynamic/FFI call.wasm
: The adding function compiled to WebAssembly.js
: Re-implementing the function in plain JavaScript.
Each function will be called 100000 times, in 5 repetitions, timed with
console.time()
. Here are the results on my machine (AMD Ryzen 9 5900X) with
Ubuntu 22.04.1 and Node.js 19.6.0:
ffi-napi ... done!
sbffi ... done!
napi-addon ... done!
napi-addon-sb ... done!
wasm ... done!
js ... done!
---
ffi-napi ... done!
sbffi ... done!
napi-addon ... done!
napi-addon-sb ... done!
wasm ... done!
js ... done!
---
ffi-napi ... done!
sbffi ... done!
napi-addon ... done!
napi-addon-sb ... done!
wasm ... done!
js ... done!
---
ffi-napi ... done!
sbffi ... done!
napi-addon ... done!
napi-addon-sb ... done!
wasm ... done!
js ... done!
---
ffi-napi ... done!
sbffi ... done!
napi-addon ... done!
napi-addon-sb ... done!
wasm ... done!
js ... done!
---
┌───────────────┬──────┬───────────┬────────────┬────────────────────┐
│ (index) │ min │ max │ mean │ stddev │
├───────────────┼──────┼───────────┼────────────┼────────────────────┤
│ ffi-napi │ 3216 │ 122159103 │ 8271.61785 │ 291769.95704110194 │
│ sbffi │ 120 │ 3657727 │ 177.070514 │ 9480.134200345783 │
│ napi-addon │ 120 │ 2877439 │ 181.689724 │ 9197.256781039703 │
│ napi-addon-sb │ 80 │ 3256319 │ 143.028966 │ 10193.011672112762 │
│ wasm │ 60 │ 3317759 │ 170.443644 │ 16172.981925863021 │
│ js │ 70 │ 3395583 │ 141.67375 │ 12967.590912612988 │
└───────────────┴──────┴───────────┴────────────┴────────────────────┘
For this benchmark, I generally see roughly similar performance between sbffi
and a typical NAPI addon. Of course, YMMV.
Please see CONTRIBUTING.md, CODE_OF_CONDUCT.md and TODO.md.
Please see LICENSE.txt.