Julia 代码的 eval

学习 Julia 语言如何运行代码的最难的一部分是 学习如何让所有的小部分工作协同工作来执行一段代码。

每个代码块通常会通过许多步骤来执行,在转变为期望的结果之前(但愿如此)。并且你可能不熟悉它们的名称,例如(非特定顺序): flisp,AST,C++,LLVM,evaltypeinfmacroexpand,sysimg(或 system image),启动,变异,解析,执行,即时编译器,解释器解释,装箱,拆箱,内部函数,原始函数

!!! sidebar “Definitions”

  1. * REPL
  2. REPL 表示 读取-求值-输出-循环(Read-Eval-Print Loop)。 我们管这个命令行环境的简称就叫REPL
  3. * AST
  4. 抽象语法树(Abstract Syntax Tree)是代码结构的数据表现。在这种表现形式下代码被符号化,因此更加方便操作和执行。

Julia Execution

整个进程的千里之行如下:

  1. 用户打开了 julia
  2. The C function main() from cli/loader_exe.c gets called. This function processes the command line arguments, filling in the jl_options struct and setting the variable ARGS. It then initializes 在 ui/repl.c 中的 C 语言的函数 main() 被调用。这个函数处理命令行参数,填充到 jl_options 结构图并且设置变了 ARGS 。接下来初始化 Julia (通过调用 julia_init in task.c which may load a previously compiled sysimg). Finally, it passes off control to Julia by calling Base._start().
  3. When _start() takes over control, the subsequent sequence of commands depends on the command line arguments given. For example, if a filename was supplied, it will proceed to execute that file. Otherwise, it will start an interactive REPL.
  4. Skipping the details about how the REPL interacts with the user, let’s just say the program ends up with a block of code that it wants to run.
  5. If the block of code to run is in a file, jl_load(char *filename) gets invoked to load the file and parse it. Each fragment of code is then passed to eval to execute.
  6. Each fragment of code (or AST), is handed off to eval() to turn into results.
  7. eval() takes each code fragment and tries to run it in jl_toplevel_eval_flex().
  8. jl_toplevel_eval_flex() decides whether the code is a “toplevel” action (such as using or module), which would be invalid inside a function. If so, it passes off the code to the toplevel interpreter.
  9. jl_toplevel_eval_flex() then expands the code to eliminate any macros and to “lower” the AST to make it simpler to execute.
  10. jl_toplevel_eval_flex() then uses some simple heuristics to decide whether to JIT compiler the AST or to interpret it directly.
  11. The bulk of the work to interpret code is handled by eval in interpreter.c.
  12. If instead, the code is compiled, the bulk of the work is handled by codegen.cpp. Whenever a Julia function is called for the first time with a given set of argument types, type inference will be run on that function. This information is used by the codegen step to generate faster code.
  13. Eventually, the user quits the REPL, or the end of the program is reached, and the _start() method returns.
  14. Just before exiting, main() calls jl_atexit_hook(exit_code). This calls Base._atexit() (which calls any functions registered to atexit() inside Julia). Then it calls jl_gc_run_all_finalizers(). Finally, it gracefully cleans up all libuv handles and waits for them to flush and close.

Parsing

The Julia parser is a small lisp program written in femtolisp, the source-code for which is distributed inside Julia in src/flisp.

The interface functions for this are primarily defined in jlfrontend.scm. The code in ast.c handles this handoff on the Julia side.

The other relevant files at this stage are julia-parser.scm, which handles tokenizing Julia code and turning it into an AST, and julia-syntax.scm, which handles transforming complex AST representations into simpler, “lowered” AST representations which are more suitable for analysis and execution.

If you want to test the parser without re-building Julia in its entirety, you can run the frontend on its own as follows:

  1. $ cd src
  2. $ flisp/flisp
  3. > (load "jlfrontend.scm")
  4. > (jl-parse-file "<filename>")

Macro Expansion

When eval() encounters a macro, it expands that AST node before attempting to evaluate the expression. Macro expansion involves a handoff from eval() (in Julia), to the parser function jl_macroexpand() (written in flisp) to the Julia macro itself (written in - what else - Julia) via fl_invoke_julia_macro(), and back.

Typically, macro expansion is invoked as a first step during a call to Meta.lower()/jl_expand(), although it can also be invoked directly by a call to macroexpand()/jl_macroexpand().

Type Inference

Type inference is implemented in Julia by typeinf() in compiler/typeinfer.jl. Type inference is the process of examining a Julia function and determining bounds for the types of each of its variables, as well as bounds on the type of the return value from the function. This enables many future optimizations, such as unboxing of known immutable values, and compile-time hoisting of various run-time operations such as computing field offsets and function pointers. Type inference may also include other steps such as constant propagation and inlining.

!!! sidebar “More Definitions”

  1. * JIT
  2. Just-In-Time Compilation The process of generating native-machine code into memory right when
  3. it is needed.
  4. * LLVM
  5. Low-Level Virtual Machine (a compiler) The Julia JIT compiler is a program/library called libLLVM.
  6. Codegen in Julia refers both to the process of taking a Julia AST and turning it into LLVM instructions,
  7. and the process of LLVM optimizing that and turning it into native assembly instructions.
  8. * C++
  9. The programming language that LLVM is implemented in, which means that codegen is also implemented
  10. in this language. The rest of Julia's library is implemented in C, in part because its smaller
  11. feature set makes it more usable as a cross-language interface layer.
  12. * box
  13. This term is used to describe the process of taking a value and allocating a wrapper around the
  14. data that is tracked by the garbage collector (gc) and is tagged with the object's type.
  15. * unbox
  16. The reverse of boxing a value. This operation enables more efficient manipulation of data when
  17. the type of that data is fully known at compile-time (through type inference).
  18. * generic function
  19. A Julia function composed of multiple "methods" that are selected for dynamic dispatch based on
  20. the argument type-signature
  21. * anonymous function or "method"
  22. A Julia function without a name and without type-dispatch capabilities
  23. * primitive function
  24. A function implemented in C but exposed in Julia as a named function "method" (albeit without
  25. generic function dispatch capabilities, similar to a anonymous function)
  26. * intrinsic function
  27. A low-level operation exposed as a function in Julia. These pseudo-functions implement operations
  28. on raw bits such as add and sign extend that cannot be expressed directly in any other way. Since
  29. they operate on bits directly, they must be compiled into a function and surrounded by a call
  30. to `Core.Intrinsics.box(T, ...)` to reassign type information to the value.

JIT Code Generation

Codegen is the process of turning a Julia AST into native machine code.

The JIT environment is initialized by an early call to jl_init_codegen in codegen.cpp.

On demand, a Julia method is converted into a native function by the function emit_function(jl_method_instance_t*). (note, when using the MCJIT (in LLVM v3.4+), each function must be JIT into a new module.) This function recursively calls emit_expr() until the entire function has been emitted.

Much of the remaining bulk of this file is devoted to various manual optimizations of specific code patterns. For example, emit_known_call() knows how to inline many of the primitive functions (defined in builtins.c) for various combinations of argument types.

Other parts of codegen are handled by various helper files:

  • debuginfo.cpp

    Handles backtraces for JIT functions

  • ccall.cpp

    Handles the ccall and llvmcall FFI, along with various abi_*.cpp files

  • intrinsics.cpp

    Handles the emission of various low-level intrinsic functions

!!! sidebar “Bootstrapping” The process of creating a new system image is called “bootstrapping”.

  1. The etymology of this word comes from the phrase "pulling oneself up by the bootstraps", and
  2. refers to the idea of starting from a very limited set of available functions and definitions
  3. and ending with the creation of a full-featured environment.

System Image

The system image is a precompiled archive of a set of Julia files. The sys.ji file distributed with Julia is one such system image, generated by executing the file sysimg.jl, and serializing the resulting environment (including Types, Functions, Modules, and all other defined values) into a file. Therefore, it contains a frozen version of the Main, Core, and Base modules (and whatever else was in the environment at the end of bootstrapping). This serializer/deserializer is implemented by jl_save_system_image/jl_restore_system_image in staticdata.c.

If there is no sysimg file (jl_options.image_file == NULL), this also implies that --build was given on the command line, so the final result should be a new sysimg file. During Julia initialization, minimal Core and Main modules are created. Then a file named boot.jl is evaluated from the current directory. Julia then evaluates any file given as a command line argument until it reaches the end. Finally, it saves the resulting environment to a “sysimg” file for use as a starting point for a future Julia run.