System Image Building

Building the Julia system image

Julia ships with a preparsed system image containing the contents of the Base module, named sys.ji. This file is also precompiled into a shared library called sys.{so,dll,dylib} on as many platforms as possible, so as to give vastly improved startup times. On systems that do not ship with a precompiled system image file, one can be generated from the source files shipped in Julia's DATAROOTDIR/julia/base folder.

Julia will by default generate its system image on half of the available system threads. This may be controlled by the JULIA_IMAGE_THREADS environment variable.

This operation is useful for multiple reasons. A user may:

  • Build a precompiled shared library system image on a platform that did not ship with one, thereby improving startup times.
  • Modify Base, rebuild the system image and use the new Base next time Julia is started.
  • Include a userimg.jl file that includes packages into the system image, thereby creating a system image that has packages embedded into the startup environment.

The PackageCompiler.jl package contains convenient wrapper functions to automate this process.

System image optimized for multiple microarchitectures

The system image can be compiled simultaneously for multiple CPU microarchitectures under the same instruction set architecture (ISA). Multiple versions of the same function may be created with minimum dispatch point inserted into shared functions in order to take advantage of different ISA extensions or other microarchitecture features. The version that offers the best performance will be selected automatically at runtime based on available CPU features.

Specifying multiple system image targets

A multi-microarchitecture system image can be enabled by passing multiple targets during system image compilation. This can be done either with the JULIA_CPU_TARGET make option or with the -C command line option when running the compilation command manually. Multiple targets are separated by ; in the option string. The syntax for each target is a CPU name followed by multiple features separated by ,. All features supported by LLVM are supported and a feature can be disabled with a - prefix. (+ prefix is also allowed and ignored to be consistent with LLVM syntax). Additionally, a few special features are supported to control the function cloning behavior.

Note

It is good practice to specify either clone_all or base(<n>) for every target apart from the first one. This makes it explicit which targets have all functions cloned, and which targets are based on other targets. If this is not done, the default behavior is to not clone every function, and to use the first target's function definition as the fallback when not cloning a function.

  1. clone_all

    By default, only functions that are the most likely to benefit from the microarchitecture features will be cloned. When clone_all is specified for a target, however, all functions in the system image will be cloned for the target. The negative form -clone_all can be used to prevent the built-in heuristic from cloning all functions.

  2. base(<n>)

    Where <n> is a placeholder for a non-negative number (e.g. base(0), base(1)). By default, a partially cloned (i.e. not clone_all) target will use functions from the default target (first one specified) if a function is not cloned. This behavior can be changed by specifying a different base with the base(<n>) option. The nth target (0-based) will be used as the base target instead of the default (0th) one. The base target has to be either 0 or another clone_all target. Specifying a non-clone_all target as the base target will cause an error.

  3. opt_size

    This causes the function for the target to be optimized for size when there isn't a significant runtime performance impact. This corresponds to -Os GCC and Clang option.

  4. min_size

    This causes the function for the target to be optimized for size that might have a significant runtime performance impact. This corresponds to -Oz Clang option.

As an example, at the time of this writing, the following string is used in the creation of the official x86_64 Julia binaries downloadable from julialang.org:

generic;sandybridge,-xsaveopt,clone_all;haswell,-rdrnd,base(1)

This creates a system image with three separate targets; one for a generic x86_64 processor, one with a sandybridge ISA (explicitly excluding xsaveopt) that explicitly clones all functions, and one targeting the haswell ISA, based off of the sandybridge sysimg version, and also excluding rdrnd. When a Julia implementation loads the generated sysimg, it will check the host processor for matching CPU capability flags, enabling the highest ISA level possible. Note that the base level (generic) requires the cx16 instruction, which is disabled in some virtualization software and must be enabled for the generic target to be loaded. Alternatively, a sysimg could be generated with the target generic,-cx16 for greater compatibility, however note that this may cause performance and stability problems in some code.

Implementation overview

This is a brief overview of different part involved in the implementation. See code comments for each components for more implementation details.

  1. System image compilation

    The parsing and cloning decision are done in src/processor*. We currently support cloning of function based on the present of loops, simd instructions, or other math operations (e.g. fastmath, fma, muladd). This information is passed on to src/llvm-multiversioning.cpp which does the actual cloning. In addition to doing the cloning and insert dispatch slots (see comments in MultiVersioning::runOnModule for how this is done), the pass also generates metadata so that the runtime can load and initialize the system image correctly. A detailed description of the metadata is available in src/processor.h.

  2. System image loading

    The loading and initialization of the system image is done in src/processor* by parsing the metadata saved during system image generation. Host feature detection and selection decision are done in src/processor_*.cpp depending on the ISA. The target selection will prefer exact CPU name match, larger vector register size, and larger number of features. An overview of this process is in src/processor.cpp.

Trimming

System images are typically quite large, since Base includes a lot of functionality, and by default system images also include several packages such as LinearAlgebra for convenience and backwards compatibility. Most programs will use only a fraction of the functions in these packages. Therefore it makes sense to build binaries that exclude unused functions to save space, referred to as "trimming".

While the basic idea of trimming is sound, Julia has dynamic and reflective features that make it difficult (or impossible) to know in general which functions are unused. As an extreme example, consider code like

getglobal(Base, Symbol(readchomp(stdin)))(1)

This code reads a function name from stdin and calls the named function from Base on the value 1. In this case it is impossible to predict which function will be called, so no functions can reliably be considered "unused". With some noteworthy exceptions (Julia's own REPL being one of them), most real-world programs do not do things like this.

Less extreme cases occur, for example, when there are type instabilities that make it impossible for the compiler to predict which method will be called. However, if code is well-typed and does not use reflection, a complete and (hopefully) relatively small set of needed methods can be determined, and the rest can be removed. The --trim command-line option requests this kind of compilation.

When --trim is specified in a command used to build a system image, the compiler begins tracing calls starting at methods marked using Base.Experimental.entrypoint. If a call is too dynamic to reasonably narrow down the possible call targets, an error is given at compile time showing the location of the call. For testing purposes, it is possible to skip these errors by specifying --trim=unsafe or --trim=unsafe-warn. Then you will get a system image built, but it may crash at run time if needed code is not present.

It typically makes sense to specify --strip-ir along with --trim, since trimmed binaries are fully compiled and therefore don't need Julia IR. At some point we may make --trim imply --strip-ir, but for now we have kept them orthogonal.

To get the smallest possible binary, it will also help to specify --strip-metadata and run the Unix strip utility. However, those steps remove Julia-specific and native (DWARF format) debug info, respectively, and so will make debugging more difficult.

Common problems

  • The Base global variables stdin, stdout, and stderr are non-constant and so their types are not known. All printing should use a specific IO object with a known type. The easiest substitution is to use print(Core.stdout, x) instead of print(x) or print(stdout, x).
  • Use tools like JET.jl, Cthulhu.jl, and/or SnoopCompile to identify failures of type-inference, and follow our Performance Tips to fix them.

Compatibility concerns

We have identified many small changes to Base that significantly increase the set of programs that can be reliably trimmed. Unfortunately some of those changes would be considered breaking, and so are only applied when trimming is requested (this is done by an external build script, currently maintained inside the test suite as contrib/juliac-buildscript.jl). Therefore in many cases trimming will require you to opt in to new variants of Base and some standard libraries.

If you want to use trimming, it is important to set up continuous integration testing that performs a trimmed build and fully tests the resulting program. Fortunately, if your program successfully compiles with --trim then it is very likely to work the same as it did before. However, CI is needed to ensure that your program continues to build with trimming as you develop it.

Package authors may wish to test that their package is "trimming safe", however this is impossible in general. Trimming is only expected to work given concrete entry points such as main() and library entry points meant to be called from outside Julia. For generic packages, existing tests for type stability like @inferred and JET.@report_call are about as close as you can get to checking trim compatibility.

Trimming also introduces new compatibility issues between minor versions of Julia. At this time, we are not able to guarantee that a program that can be trimmed in one version of Julia can also be trimmed in all future versions of Julia. However, breakage of that kind is expected to be rare. We also plan to try to increase the set of programs that can be trimmed over time.