Julia can start more slowly than a short Python or Bash script, then run much faster after critical code has been compiled. That trade-off comes from just-in-time (JIT) compilation: turning code into machine code while the program runs. section 1.7
The timing distinguishes JIT compilation from ahead-of-time compilation, which translates a program before it runs, and interpretation, which processes source instructions at runtime. Julia was designed for JIT compilation, though a language is not inherently limited to one execution method. section 1.1.1
JIT compilation helps Julia combine dynamic typing with high performance. When Julia generates code for a method, it knows the input argument types. Because an object's fields are fixed by its type, the compiler can use that information to generate optimized machine code. Multiple dispatch is another part of this design. section 1.1.1 section 18.4
Listing 18.10 Vec2D defined with abstract type fields to mimic regular dynamic languages
struct Vec2D #1
Δx::Number #2
Δy::Number #2
end
function multiply(u::Vec2D, k::Number)
Vec2D(k * u.Δx, k * u.Δy)
end
Inspect the Vec2D definition and its multiply method. Its fields are declared as Number, providing a concrete starting point for thinking about the type information available during compilation. section 18.4
Listing 18.11 Pseudocode of how multiply would work in a normal dynamic language
function multiply(u::Vec2D, k::Number)
ux = getfield(u, :Δx)
if !isa(ux, Float64)
error("x must be a float")
end
uy = getfield(u, :Δy)
if !isa(uy, Float64)
error("y must be a float")
end
c = convert(Float64, k)
Vec2D(floatmul(c, ux), floatmul(c, uy))
end
Compare the illustrative pseudocode for a conventional dynamic-language approach: it looks up fields and checks their types at runtime. It is not a listing of Julia's generated code. section 18.4
The practical question is whether the work after compilation outweighs the startup delay. JIT compilation makes Julia less attractive for some small, short-running scripts; caching previous compilations and compiling more selectively are cited approaches to reducing that cost. section 1.7
