Julia 語言設計與 JIT 編譯器 - 杜岳華

請從這裡開始

簡報

Multiple dispatch

呼叫函數時比較參數類型，考慮參數類型而呼叫不同函式（類似 C function overloading ）

Generic programming 泛型

與 Templete 意義類似

• Parametric types and parametric methods
• Similar to multiple dispatch with parametric polymorphism.
• All types are first-class: can be dispatched and declared

Metaprogramming

• Macro
• Generated function

Macro

In [1]

macro sayhello(name)
    return :( println("Hello, ", $name) )
end

Out[1]

@sayhello (macro with 1 method)

In[2]

@sayhello "Andy"

Hello, Andy

輸入變數，吐出一段 code

Generated function

@generated function foo(x)
    Core.println(x)
    return :(x * x)
end

Compile time

Compiler compile time

Compiler runtime

Compiler details

Julia compile time

Julia IR

• SSA form
• 類似 LLVM IR

warn_type

• dynamic programming 時無法取得詳細的 type 資訊
• 補充 type 資訊

Static Single Assignment form (SSA)

• Each variable is assigned exactly once.
• Every variable is defined before it used.

Simplify and improve the compiler optimization

• constant propagation
• value range propagation
• sparse conditional constant propagation
• dead code elimination
• global value numbering
• partial redundancy elimination
• strength reduction
• register allocation

Runtime

Functions

Julia functions are generic functions.
generic functions :

• 紀錄 function name，再由 function table 查詢 method

Julia

• function : 只有 function name
• method : 有實做

A generic function is a single function and consists of many methods.

The methods of a function is stored in a method table.

JIT comiliation

Staged programming

Leverage the type in input of code and generate optimized code

@generated function foo(::Array{T,N}, ...) where {T,N}
    ...
    x = Matrix{T}()
    for i = 1:N
        ...
    end
end

Hackable compiler
Python and Numba

​​​​Ref. Effective Extensible Programming: Unleashing Julia on GPUs 

Hackable compiler

Python and Numba

Ref. Effective Extensible Programming: Unleashing Julia on GPUs

Reflection and introspection

tags: COSCUP2019 Julia language IB501