Language introspection

What is metaprogramming? A high-level code that writes high-level code by Steven Johnson.

Why do we need metaprogramming?

• In general, we do not need it, as we can do whatever we need without it, but it can help us to remove a boilerplate code.
  • As an example, consider a @show macro, which just prints the name of the variable (or the expression) and its evaluation. This means that instead of writing println("2+exp(4) = ", 2+exp(4)) we can just write @show 2+exp(4).
  • We have seen @time or @benchmark, which is difficult to implement using normal function, since when you pass 2+exp(4) as a function argument, it will be automatically evaluated. You need to pass it as an expression, that can be evaluated within the function.
  • @chain macro from Chain.jl improves over native piping |>
  • @set macro from Accesors.jl which allows to set the value of a field in a non-mutable struct (it is done by constructing new struct and copying the old values).
  • We have seen @forward macro implementing encapsulation.
  • Macros are used to insert compilation directives not accessible through the syntax, e.g. @inbounds.
• A chapter on its own is definition of Domain Specific Languages.

Translation stages from source code to machine code

Julia (as any modern compiler) uses several stages to convert source code to native code.

• Julia's steps consists from ^[1]
  1. parse the source code to abstract syntax tree (AST) –- parse function
  2. expand macros –- macroexpand function
  3. syntax desugaring
  4. stamentize controlflow
  5. resolve scopes
  6. generate intermediate representation (IR) ("goto" form) –- expand or code_lowered functions
  7. top-level evaluation, method sorting –- methods
  8. type inference
  9. inlining and high level optimization –- code_typed
  10. LLVM IR generation –- code_llvm
  11. LLVM optimizer, native code generation –- code_native
• steps 3-6 are done in one inseparable stage
• Julia's IR is in static single assignment form

Example: Fibonacci numbers

Consider for example a function computing the Fibonacci numbers^[1]

function nextfib(n)
	a, b = one(n), one(n)
	while b < n
		a, b = b, a + b
	end
	return b
end

Parsing

The first thing the compiler does is that it will parse the source code (represented as a string) to the abstract syntax tree (AST). We can inspect the results of this stage as

julia> parsed_fib = Meta.parse(
"""
	function nextfib(n)
		a, b = one(n), one(n)
		while b < n
			a, b = b, a + b
		end
		return b
	end""")
:(function nextfib(n)
      #= none:1 =#
      #= none:2 =#
      (a, b) = (one(n), one(n))
      #= none:3 =#
      while b < n
          #= none:4 =#
          (a, b) = (b, a + b)
      end
      #= none:6 =#
      return b
  end)

AST is a tree representation of the source code, where the parser has already identified individual code elements function call, argument blocks, etc. The parsed code is represented by Julia objects, therefore it can be read and modified by Julia from Julia at your wish (this is what is called homo-iconicity of a language the itself being derived from Greek words homo- meaning "the same" and icon meaning "representation"). Using TreeView

using TreeView, TikzPictures
g = tikz_representation(walk_tree(parsed_fib))
TikzPictures.save(SVG("parsed_fib.svg"), g)

We can see that the AST is indeed a tree, with function being the root node (caused by us parsing a function). Each inner node represents a function call with children of the inner node being its arguments. An interesting inner node is the Block representing a sequence of statements, where we can also see information about lines in the source code inserted as comments. Lisp-like S-Expression can be printed using Meta.show_sexpr(parsed_fib).

(:function, (:call, :nextfib, :n), (:block,
    :(#= none:1 =#),
    :(#= none:2 =#),
    (:(=), (:tuple, :a, :b), (:tuple, (:call, :one, :n), (:call, :one, :n))),
    :(#= none:3 =#),
    (:while, (:call, :<, :b, :n), (:block,
        :(#= none:4 =#),
        (:(=), (:tuple, :a, :b), (:tuple, :b, (:call, :+, :a, :b)))
      )),
    :(#= none:6 =#),
    (:return, :b)
  ))

Expanding macros

If we insert a "useless" macro to nextfib, for example @show b, we see that the macro is not expanded and it is left there as-is.

julia> parsed_fib = Meta.parse(
"""
  function nextfib(n)
    a, b = one(n), one(n)
    while b < n
      a, b = b, a + b
    end
    @show b
    return b
  end""")
:(function nextfib(n)
      #= none:1 =#
      #= none:2 =#
      (a, b) = (one(n), one(n))
      #= none:3 =#
      while b < n
          #= none:4 =#
          (a, b) = (b, a + b)
          #= none:5 =#
      end
      #= none:6 =#
      #= none:6 =# @show b
      #= none:7 =#
      return b
  end)

and we can ask for expansion of the macro

julia> macroexpand(Main, parsed_fib)
:(function nextfib(n)
      #= none:1 =#
      #= none:2 =#
      (a, b) = (one(n), one(n))
      #= none:3 =#
      while b < n
          #= none:4 =#
          (a, b) = (b, a + b)
          #= none:5 =#
      end
      #= none:6 =#
      begin
          Base.println("b = ", Base.repr(begin
                      #= show.jl:1047 =#
                      local var"#62#value" = b
                  end))
          var"#62#value"
      end
      #= none:7 =#
      return b
  end)

Lowering

The next stage is lowering, where AST is converted to Static Single Assignment Form (SSA), in which "each variable is assigned exactly once, and every variable is defined before it is used". Loops and conditionals are transformed into gotos and labels using a single unless/goto construct (this is not exposed in user-level Julia).

julia> @code_lowered nextfib(3)
CodeInfo(
1 ─ %1 = Main.one(n)
│   %2 = Main.one(n)
│        a = %1
└──      b = %2
2 ┄ %5 = b < n
└──      goto #4 if not %5
3 ─ %7 = b
│   %8 = a + b
│        a = %7
│        b = %8
└──      goto #2
4 ─      return b
)

or alternatively lowered_fib = Meta.lower(@__MODULE__, parsed_fib).

We can see that

• compiler has introduced a lot of variables
• while (and for) loops has been replaced by a goto, where goto can be conditional

For inserted debugging information, there is an option to pass keyword argument debuginfo=:source.

julia> @code_lowered debuginfo=:source nextfib(3)
CodeInfo(
    @ none:2 within `nextfib'
1 ─ %1 = Main.one(n)
│   %2 = Main.one(n)
│        a = %1
└──      b = %2
    @ none:3 within `nextfib'
2 ┄ %5 = b < n
└──      goto #4 if not %5
    @ none:4 within `nextfib'
3 ─ %7 = b
│   %8 = a + b
│        a = %7
│        b = %8
└──      goto #2
    @ none:6 within `nextfib'
4 ─      return b
)

Code Typing

Code typing is the process in which the compiler attaches types to variables and tries to infer types of objects returned from called functions. If the compiler fails to infer the returned type, it will give the variable type Any, in which case a dynamic dispatch will be used in subsequent operations with the variable. Inspecting typed code is therefore important for detecting type instabilities (the process can be difficult and error prone, fortunately, new tools like Jet.jl may simplify this task). The output of typing can be inspected using @code_typed macro. If you know the types of function arguments, aka function signature, you can call directly function InteractiveUtils.code_typed(nextfib, (typeof(3),)).

julia> @code_typed nextfib(3)
CodeInfo(
1 ─      nothing::Nothing
2 ┄ %2 = φ (#1 => 1, #3 => %6)::Int64
│   %3 = φ (#1 => 1, #3 => %2)::Int64
│   %4 = Base.slt_int(%2, n)::Bool
└──      goto #4 if not %4
3 ─ %6 = Base.add_int(%3, %2)::Int64
└──      goto #2
4 ─      return %2
) => Int64

We can see that

• some calls have been inlined, e.g. one(n) was replaced by 1 and the type was inferred as Int.
• The expression b < n has been replaced with its implementation in terms of the slt_int intrinsic ("signed integer less than") and the result of this has been annotated with return type Bool.
• The expression a + b has been also replaced with its implementation in terms of the add_int intrinsic and its result type annotated as Int64.
• And the return type of the entire function body has been annotated as Int64.
• The phi-instruction %2 = φ (#1 => 1, #3 => %6) is a selector function, which returns the value depending on from which branch do you come from. In this case, variable %2 will have value 1, if the control was transfered from block #1 and it will have value copied from variable %6 if the control was transferreed from block 3 see also. The φ stands from phony variable.

When we have called @code_lower, the role of types of arguments was in selecting - via multiple dispatch - the appropriate function body among different methods. Contrary in @code_typed, the types of parameters determine the choice of inner methods that need to be called (again with multiple dispatch). This process can trigger other optimization, such as inlining, as seen in the case of one(n) being replaced with 1 directly, though here this replacement is hidden in the φ function.

Note that the same view of the code is offered by the @code_warntype macro, which we have seen in the previous lecture. The main difference from @code_typed is that it highlights type instabilities with red color and shows only unoptimized view of the code. You can view the unoptimized code with a keyword argument optimize=false:

julia> @code_typed optimize=false nextfib(3)
CodeInfo(
1 ─ %1 = Main.one(n)::Core.Const(1)
│   %2 = Main.one(n)::Core.Const(1)
│        (a = %1)::Core.Const(1)
└──      (b = %2)::Core.Const(1)
2 ┄ %5 = (b < n)::Bool
└──      goto #4 if not %5
3 ─ %7 = b::Int64
│   %8 = (a + b)::Int64
│        (a = %7)::Int64
│        (b = %8)::Int64
└──      goto #2
4 ─      return b
) => Int64

Lowering to LLVM IR

Julia uses the LLVM compiler framework to generate machine code. LLVM stands for low-level virtual machine and it is basis of many modern compilers (see wiki). We can see the textual form of code lowered to LLVM IR by invoking

julia> @code_llvm debuginfo=:source nextfib(3)
;  @ REPL[10]:1 within `nextfib'
define i64 @julia_nextfib_890(i64 signext %0) {
top:
  br label %L2

L2:                                               ; preds = %L2, %top
  %value_phi  = phi i64 [ 1, %top ], [ %1, %L2 ]
  %value_phi1 = phi i64 [ 1, %top ], [ %value_phi, %L2 ]
;  @ REPL[10]:3 within `nextfib'
; ┌ @ int.jl:83 within `<'
   %.not = icmp slt i64 %value_phi, %0
; └
;  @ REPL[10]:4 within `nextfib'
; ┌ @ int.jl:87 within `+'
   %1 = add i64 %value_phi1, %value_phi
; └
;  @ REPL[10]:3 within `nextfib'
  br i1 %.not, label %L2, label %L8

L8:                                               ; preds = %L2
;  @ REPL[10]:6 within `nextfib'
  ret i64 %value_phi
}

LLVM code can be tricky to understand first, but one gets used to it. Notice references to the source code, which help with orientation. From the code above, we may infer

• code starts by jumping to label L2, from where it reads values of two variables to two "registers" value_phi and value_phi1 (variables in LLVM starts with %).
• Both registers are treated as int64 and initialized by 1.
• [ 1, %top ], [ %value_phi, %L2 ] means that values are initialized as 1 if you come from the label top and as value value_phi if you come from %2. This is the LLVM's selector (phony φ).
• icmp slt i64 %value_phi, %0 compares the variable %value_phi to the content of variable %0. Notice the anotation that we are comparing Int64.
• %1 = add i64 %value_phi1, %value_phi adds two variables %value_phi1 and %value_phi. Note again than we are using Int64 addition.
• br i1 %.not, label %L2, label %L8 implements a conditional jump depending on the content of %.not variable.
• ret i64 %value_phi returns the value indicating it to be an Int64.

It is not expected you will be directly operating on the LLVM code, though there are libraries which does that. For example Enzyme.jl performs automatic differentiation of LLVM code, which has the benefit of being able to take a gradient through setindex!.

Producing the native vode

Native code The last stage is generation of the native code, which Julia executes. The native code depends on the target architecture (e.g. x86, ARM). As in previous cases there is a macro for viewing the compiled code @code_native

julia> @code_native debuginfo=:source nextfib(3)
	.section	__TEXT,__text,regular,pure_instructions
; ┌ @ REPL[10]:1 within `nextfib'
	movl	$1, %ecx
	movl	$1, %eax
	nopw	(%rax,%rax)
L16:
	movq	%rax, %rdx
	movq	%rcx, %rax
; │ @ REPL[10]:4 within `nextfib'
; │┌ @ int.jl:87 within `+'
	addq	%rcx, %rdx
	movq	%rdx, %rcx
; │└
; │ @ REPL[10]:3 within `nextfib'
; │┌ @ int.jl:83 within `<'
	cmpq	%rdi, %rax
; │└
	jl	L16
; │ @ REPL[10]:6 within `nextfib'
	retq
	nopw	%cs:(%rax,%rax)
; └

and the output is used mainly for debugging / inspection.

Looking around

Language introspection is very convenient for investigating, how things are implemented and how they are optimized / compiled to the native code.

julia> Base.infer_effects(nextfib, (Int,))
(+c,+e,+n,!t,+s,+m,+u,+o,+r)

Broadcasting

Broadcasting is not a unique concept in programming languages (Python/Numpy, MATLAB), however its implementation in Julia allows to easily fuse operations. For example

x = randn(100)
sin.(x) .+ 2 .* cos.(x) .+ x

is all computed in a single loop. We can inspect, how this is achieved in the lowered code:

julia> Meta.@lower sin.(x) .+ 2 .* cos.(x) .+ x
:($(Expr(:thunk, CodeInfo(
    @ none within `top-level scope'
1 ─ %1 = Base.broadcasted(sin, x)
│   %2 = Base.broadcasted(cos, x)
│   %3 = Base.broadcasted(*, 2, %2)
│   %4 = Base.broadcasted(+, %1, %3)
│   %5 = Base.broadcasted(+, %4, x)
│   %6 = Base.materialize(%5)
└──      return %6
))))

Notice that we have not used the usual @code_lowered macro, because the statement to be lowered is not a function call. In these cases, we have to use @code_lowered, which can handle more general program statements. On these cases, we cannot use @which either, as that applies to function calls only as well.

Generators

Meta.@lower [x for x in 1:4]
:($(Expr(:thunk, CodeInfo(
    @ none within `top-level scope'
1 ─ %1 = 1:4
│   %2 = Base.Generator(Base.identity, %1)
│   %3 = Base.collect(%2)
└──      return %3
))))

from which we see that the Generator is implemented using the combination of a Base.collect, which is a function collecting items of a sequence and Base.Generator(f,x), which implements an iterator, which applies function f on elements of x over which is being iterated. So an almost magical generators have instantly lost their magic.

Closures

adder(x) = y -> y + x

julia> @code_lowered adder(5)
CodeInfo(
1 ─ %1 = Main.:(var"#8#9")
│   %2 = Core.typeof(x)
│   %3 = Core.apply_type(%1, %2)
│   #8 = %new(%3, x)
└──      return #8
)

Core.apply_type is a one way of constructing an object briefly mentioned in description of object allocation.

The effect of type-instability

struct Wolf
	name::String
	energy::Int
end

struct Sheep
	name::String
	energy::Int
end

sound(wolf::Wolf) = println(wolf.name, " has howled.")
sound(sheep::Sheep) = println(sheep.name, " has baaed.")
stable_pack   = (Wolf("1", 1), Wolf("2", 2), Sheep("3", 3))
unstable_pack = [Wolf("1", 1), Wolf("2", 2), Sheep("3", 3)]
@code_typed map(sound, stable_pack)
@code_typed map(sound, unstable_pack)

using Cthulhu
@descend map(sound, unstable_pack)

General notes on metaprogramming

According to an excellent talk by Steven Johnson, you should use metaprogramming sparingly, because on one hand it's very powerful, but on the other it is generally difficult to read and it can lead to unexpected errors. Julia allows you to interact with the compiler at two different levels.

1. After the code is parsed to AST, you can modify it directly or through macros.
2. When SSA form is being typed, you can create custom functions using the concept of generated functions or directly emit intermediate representation.

More functionalities are coming through the JuliaCompilerPlugins project, but we will not talk about them (yet), as they are not mature yet.

What is Quotation?

When we are doing metaprogramming, we need to somehow tell the compiler that the next block of code is not a normal block of code to be executed, but that it should be interpreted as data and in any sense it should not be evaluated. Quotation refers to exactly this syntactic sugar. In Julia, quotation is achieved either through :(...) or quote ... end.

Notice the difference between

1 + 1 

and

:(1 + 1)

The type returned by the quotation depends on what is quoted. Observe the returned type of the following quoted code

:(1)      |> typeof
:(:x)     |> typeof
:(1 + x)  |> typeof
quote
    1 + x
    x + 1
end       |> typeof

All of these snippets are examples of the quoted code, but only :(1 + x) and the quote block produce objects of type Expr. An interesting return type is the QuoteNode, which allows to insert piece of code which should contain elements that should not be interpolated. Most of the time, quoting returns Expressions.

Expressions

Abstract Syntax Tree, the output of Julia's parser, is expressed using Julia's own datastructures, which means that you can freely manipulate it (and constructed) from the language itself. This property is called homoiconicity. Julia's compiler allows you to intercept compilation just after it has parsed the source code. Before we will take advantage of this power, we should get familiar with the strucute of the AST.

The best way to inspect the AST is through the combination

• Meta.parse, which parses the source code to AST,
• dump which print AST to terminal,
• eval which evaluates the AST within the current module.

Let's start by investigating a very simple statement 1 + 1, whose AST can be constructed either by Meta.parse("1 + 1") or :(1 + 1) or quote 1+1 end (the last one includes also the line information metadata).

julia> p = :(1+1)
:(1 + 1)

julia> typeof(p)
Expr

julia> dump(p)
Expr
  head: Symbol call
  args: Array{Any}((3,))
    1: Symbol +
    2: Int64 1
    3: Int64 1

The parsed code p is of type Expr, which according to Julia's help^[2] is a type representing compound expressions in parsed julia code (ASTs). Each expression consists: of a head Symbol identifying which kind of expression it is (e.g. a call, for loop, conditional statement, etc.), and subexpressions (e.g. the arguments of a call). The subexpressions are stored in a Vector{Any} field called args. If you recall the figure above, where AST was represented as a tree, head gives each node the name name args are either some parameters of the node, or they point to childs of that node. The interpretation of the node depends on the its type stored in head (note that the word type used here is not in the Julia sense).

julia> eval(:foo)
ERROR: foo not defined

julia> foo = "hello"
"hello"

julia> eval(:foo)
"hello"

julia> eval("foo")
"foo"

julia> :+
:+

julia> :function
:function

julia> :call
:call

julia> :x
:x

julia> Symbol(:Very,"_twisted_",:symbol,"_definition")
:Very_twisted_symbol_definition

julia> Symbol("Symbol with blanks")
Symbol("Symbol with blanks")

Construct code from scratch

Since Expr is a Julia structure, we can construct it manually as we can construct any other structure

julia> Expr(:call, :+, 1 , 1) |> dump
Expr
  head: Symbol call
  args: Array{Any}((3,))
    1: Symbol +
    2: Int64 1
    3: Int64 1

yielding to the same structure as we have created above. Expressions can be evaluated using eval, as has been said. to programmatically evaluate our expression, let's do

e = Expr(:call, :+, 1, 1)
eval(e)

We are free to use variables (identified by symbols) inside the expression

e = Expr(:call, :+, :x, 5)
eval(e)

but unless they are not defined within the scope, the expression cannot produce a meaningful result

x = 3
eval(e)

:(1 + sin(x)) == Expr(:call, :+, 1, Expr(:call, :sin, :x))

Since the expression is a Julia structure, we are free to manipulate it. Let's for example substitutue x in e = :(x + 5) with 2x.

e = :(x + 5)
e.args = map(e.args) do a 
	a == :x ?  Expr(:call, :*, 2, :x) : a 
end

or

e = :(x + 5)
e.args = map(e.args) do a 
	a == :x ? :(2*x) : a 
end

and verify that the results are correct.

julia> dump(e)
Expr
  head: Symbol call
  args: Array{Any}((3,))
    1: Symbol +
    2: Expr
      head: Symbol call
      args: Array{Any}((3,))
        1: Symbol *
        2: Int64 2
        3: Symbol x
    3: Int64 5

julia> eval(e)
11

As already mentioned, the manipulation of Expression can be arbitrary. In the above example, we have been operating directly on the arguments. But what if x would be deeper in the expression, as for example in 2(3 + x) + 2(2 - x)? We can implement the substitution using multiple dispatch as we would do when implementing any other function in Julia.

substitue_x(x::Symbol) = x == :x ? :(2*x) : x
substitue_x(e::Expr) = Expr(e.head, map(substitue_x, e.args)...)
substitue_x(u) = u

which works as promised.

julia> e = :(2(3 + 2x) + 2(2 - x))
:(2 * (3 + x) + 2 * (2 - x))
julia> f = substitue_x(e)
:(2 * (3 + 2x) + 2 * (2 - 2x))

or we can replace the sin function

replace_sin(x::Symbol) = x == :sin ? :cos : x
replace_sin(e::Expr) = Expr(e.head, map(replace_sin, e.args)...)
replace_sin(u) = u

replace_sin(:(1 + sin(x)))

Sometimes, we want to operate on a block of code as opposed to single line expressions. Recall that a block of code is defined-quoted with quote ... end. Let us see how replace_x can handle the following example:

e = quote 
	a = x + 3
	b = 2 - x
	2a + 2b
end

julia> replace_x(e) |> Base.remove_linenums!
quote
    a = 2x + 3
    b = 2 - 2x
    2a + 2b
end

julia> replace_x(e) |> eval
10

Brittleness of code manipulation

When we are manipulating the AST or creating new expressions from scratch, there is no syntactic validation performed by the parser. It is therefore very easy to create AST which does not make any sense and cannot be compiled. We have already seen that we can refer to variables that were not defined yet (this makes perfect sense). The same goes with functions (which also makes a lot of sense).

e = :(g() + 5)
eval(e)
g() = 5
eval(e)

But we can also introduce keywords which the language does not know. For example

e = Expr(:my_keyword, 1, 2, 3)
:($(Expr(:my_keyword, 1, 2, 3)))

julia> e.head
:my_keyword

julia> e.args
3-element Vector{Any}:
 1
 2
 3

julia> eval(e)
ERROR: syntax: invalid syntax (my_keyword 1 2 3)
Stacktrace:
 [1] top-level scope
   @ none:1
 [2] eval
   @ ./boot.jl:360 [inlined]
 [3] eval(x::Expr)
   @ Base.MainInclude ./client.jl:446
 [4] top-level scope
   @ REPL[8]:1

notice that error is not related to undefined variable / function, but the invalid syntax. This also demonstrates the role of head in Expr. More on Julia AST can be found in the developer documentation.

Alternative way to look at code

Meta.parse("x[3]") |> dump

We can see a new Symbol ref as a head and the position 3 of variable x.

Meta.parse("(1,2,3)") |> dump

Meta.parse("1/2/3") |> dump

Code generation

Using metaprogramming in inheritance by encapsulation

Recall that Julia (at the moment) does not support inheritance, therefore the only way to adopt functionality of some object and extend it is through encapsulation. Assuming we have some object T, we wrap that object into a new structure. Let's work out a concrete example, where we define the our own matrix.

struct MyMatrix{T} <: AbstractMatrix{T}
	x::Matrix{T}
end

Now, to make it useful, we should define all the usual methods, like size, length, getindex, setindex!, etc. We can list methods defined with Matrix as an argument methodswith(Matrix) (recall this will load methods that are defined with currently loaded libraries). Now, we would like to overload them. To minimize the written code, we can write

import Base: setindex!, getindex, size, length
for f in [:setindex!, :getindex, :size, :length]
	eval(:($(f)(A::MyMatrix, args...) = $(f)(A.x, args...)))
end

which we can verify now that it works as expected

julia> a = MyMatrix([1 2 ; 3 4])
2×2 MyMatrix{Int64}:
 1  2
 3  4

julia> a[4]
4

julia> a[3] = 0
0

julia> a
2×2 MyMatrix{Int64}:
 1  0
 3  4

In this way, Julia acts as its own preprocessor. The above look can be equally written as

for f in [:setindex!, :getindex, :size, :length]
    println("$(f)(A::MyMatrix, args...) = $(f)(A.x, args...)")
end

for f in [:setindex!, :getindex, :size, :length]
	s = "Base.$(f)(A::MyMatrix, args...) = $(f)(A.x, args...)"
	println(s)
	eval(Meta.parse(s))
end

for f in [:setindex!, :getindex, :size, :length]
	@eval $(f)(A::MyMatrix, args...) = $(f)(A.x, args...)
end

Notice that we have just hand-implemented parts of @forward macro from MacroTools, which does exactly this.

Resources

• Introduction to Julia by Jeff Bezanson on first JuliaCon
• Julia's manual on metaprogramming
• David P. Sanders' workshop @ JuliaCon 2021
• Steven Johnson's keynote talk @ JuliaCon 2019
• James Nash's Is Julia Aot or JIT @ JuliaCon 2017
• Andy Ferris's workshop @ JuliaCon 2018
• From Macros to DSL by John Myles White
• Notes on JuliaCompilerPlugin