This is an experimental package to compile Julia code to standalone libraries. A system image is not needed.
Installation is the same as any other registered Julia package
using Pkg
Pkg.add("StaticCompiler")
StaticCompiler.jl provides the functions compile_executable
and compile_shlib
for compiling a Julia function to a native executable or shared library for use from outside of Julia:
julia> using StaticCompiler, StaticTools
julia> hello() = println(c"Hello, world!")
hello (generic function with 1 method)
julia> compile_executable(hello, (), "./")
"/Users/user/hello"
shell> ls -alh hello
-rwxrwxr-x. 1 user user 8.4K Oct 20 20:36 hello
shell> ./hello
Hello, world!
This approach comes with substantial limitations compared to regular julia code, as you cannot rely on julia's runtime, libjulia
(see, e.g., StaticTools.jl for some ways to work around these limitations).
The low-level function StaticCompiler.generate_obj
(not exported) generates object files. This can be used for more control of compilation. This can be used for example, to cross-compile to other targets.
Sometimes, a julia function you want to statically compile will do things (such as throwing errors) that aren't supported natively by StaticCompiler. One tool provided for working around this is the @device_override
macro which lets you swap out a method, but only inside of a StaticCompiler.jl compilation context. For example:
julia> using Libdl, StaticCompiler
julia> f(x) = g(x) + 1;
julia> g(x) = 2x
julia> @device_override g(x::Int) = x - 10
julia> f(1) # Gives the expected answer in regular julia
3
julia> dlopen(compile_shlib(f, (Int,), "./")) do lib
fptr = dlsym(lib, "f")
# Now use the compiled version where g(x) = 2x is replaced with g(x) = x - 10
@ccall $fptr(1::Int)::Int
end
-8
Typically, errors should be overrided and replaced with @print_and_throw
, which is StaticCompiler friendly, i.e.
we define overrides such as
@device_override @noinline Base.Math.throw_complex_domainerror(f::Symbol, x) =
@print_and_throw c"This operation requires a complex input to return a complex result"
If for some reason, you wish to use a different method table (defined with Base.Experimental.@MethodTable
and Base.Experimental.@overlay
) than the default one provided by StaticCompiler.jl, you can provide it to compile_executable
and compile_shlib
via a keyword argument method_table
.
This package uses the GPUCompiler package to generate code.
- GC-tracked allocations and global variables do not work with
compile_executable
orcompile_shlib
. This has some interesting consequences, including that all functions within the function you want to compile must either be inlined or return only native types (otherwise Julia would have to allocate a place to put the results, which will fail). - Since error handling relies on libjulia, you can only throw errors from standalone-compiled (
compile_executable
/compile_shlib
) code if an explicit overload has been defined for that particular error with@device_override
(see quirks.jl). - Type instability. Type unstable code cannot currently be statically compiled via this package.
- Extra experimental on Windows (PRs welcome if you encounter issues). Should work in WSL on Windows 10+.
To enable code to be statically compiled, consider the following:
-
Use type-stable code.
-
Use Tuples, NamedTuples, StaticArrays, and other types where appropriate. These allocate on the stack and don't use Julia's heap allocation.
-
Avoid Julia's internal allocations. That means don't bake in use of Arrays or Strings or Dicts. Types from StaticTools can help, like StaticStrings and MallocArrays.
-
If need be, manage memory manually, using
malloc
andfree
from StaticTools.jl. This works withStaticTools.MallocString
andStaticTools.MallocArray
, or use Bumper.jl. -
Don't use global variables that need to be allocated and initialized. Instead of global variables, use context structures that have an initialization function. It is okay to use global Tuples or NamedTuples as the use of these should be baked into compiled code.
-
Use context variables to store program state, inputs, and outputs. Parameterize these typese as needed, so your code can handle normal types (Arrays) and static-friendly types (StaticArrays, MallocArrays, or StrideArrays). The SciML ecosystem does this well (example). Use of these context variables also enables allocations and initialization to be centralized, so these could be managed by the calling routines in Julia, Python, JavaScript, or other language.
-
Arguments and returned values from
compile_shlib
must be native objects such asInt
,Float64
, orPtr
. They cannot be things likeTuple{Int, Int}
because that is not natively sized. Such objects need to be passed by reference instead of by value. -
If your code needs an array as a workspace, instead of directly creating it, create it as a function argument (where it could default to a standard array creation). That code could be statically compiled if that function argument is changed to a MallocArray or another static-friendly alternative.
If you're trying to statically compile generic code, you may run into issues if that code uses features not supported by StaticCompiler. One option is to change the code you're calling using the tips above. If that is not easy, you may by able to compile it anyway. One option is to use method overlays to change what methods are called.
Cthulhu is a great help in digging into code, finding type instabilities, and finding other sources of code that may break static compilation.
Because Julia objects follow C memory layouts, compiled libraries should be usable from most languages that can interface with C. For example, results should be usable with Python's CFFI package.
For WebAssembly, interface helpers are available at WebAssemblyInterfaces, and users should also see WebAssemblyCompiler for a package more focused on compilation of WebAssebly in general.