Julia ASTs

Julia has two representations of code. First there is a surface syntax AST returned by the parser (e.g. the Meta.parse function), and manipulated by macros. It is a structured representation of code as it is written, constructed by julia-parser.scm from a character stream. Next there is a lowered form, or IR (intermediate representation), which is used by type inference and code generation. In the lowered form there are fewer types of nodes, all macros are expanded, and all control flow is converted to explicit branches and sequences of statements. The lowered form is constructed by julia-syntax.scm.

First we will focus on the AST, since it is needed to write macros.

Surface syntax AST

Front end ASTs consist almost entirely of Exprs and atoms (e.g. symbols, numbers). There is generally a different expression head for each visually distinct syntactic form. Examples will be given in s-expression syntax. Each parenthesized list corresponds to an Expr, where the first element is the head. For example (call f x) corresponds to Expr(:call, :f, :x) in Julia.

Calls

do syntax:

f(x) do a,b
    body
end

parses as (do (call f x) (-> (tuple a b) (block body))).

Operators

Most uses of operators are just function calls, so they are parsed with the head call. However some operators are special forms (not necessarily function calls), and in those cases the operator itself is the expression head. In julia-parser.scm these are referred to as "syntactic operators". Some operators (+ and *) use N-ary parsing; chained calls are parsed as a single N-argument call. Finally, chains of comparisons have their own special expression structure.

Bracketed forms

Macros

Strings

Doc string syntax:

"some docs"
f(x) = x

parses as (macrocall (|.| Core '@doc) (line) "some docs" (= (call f x) (block x))).

Imports and such

using has the same representation as import, but with expression head :using instead of :import.

Numbers

Julia supports more number types than many scheme implementations, so not all numbers are represented directly as scheme numbers in the AST.

Block forms

A block of statements is parsed as (block stmt1 stmt2 ...).

If statement:

if a
    b
elseif c
    d
else
    e
end

parses as:

(if a (block (line 2) b)
    (elseif (block (line 3) c) (block (line 4) d)
            (block (line 6 e))))

A while loop parses as (while condition body).

A for loop parses as (for (= var iter) body). If there is more than one iteration specification, they are parsed as a block: (for (block (= v1 iter1) (= v2 iter2)) body).

break and continue are parsed as 0-argument expressions (break) and (continue).

let is parsed as (let (= var val) body) or (let (block (= var1 val1) (= var2 val2) ...) body), like for loops.

A basic function definition is parsed as (function (call f x) body). A more complex example:

function f(x::T; k = 1) where T
    return x+1
end

parses as:

(function (where (call f (parameters (kw k 1))
                       (:: x T))
                 T)
          (block (line 2) (return (call + x 1))))

Type definition:

mutable struct Foo{T<:S}
    x::T
end

parses as:

(struct true (curly Foo (<: T S))
        (block (line 2) (:: x T)))

The first argument is a boolean telling whether the type is mutable.

try blocks parse as (try try_block var catch_block finally_block). If no variable is present after catch, var is #f. If there is no finally clause, then the last argument is not present.

Quote expressions

Julia source syntax forms for code quoting (quote and :( )) support interpolation with $. In Lisp terminology, this means they are actually "backquote" or "quasiquote" forms. Internally, there is also a need for code quoting without interpolation. In Julia's scheme code, non-interpolating quote is represented with the expression head inert.

inert expressions are converted to Julia QuoteNode objects. These objects wrap a single value of any type, and when evaluated simply return that value.

A quote expression whose argument is an atom also gets converted to a QuoteNode.

Line numbers

Source location information is represented as (line line_num file_name) where the third component is optional (and omitted when the current line number, but not file name, changes).

These expressions are represented as LineNumberNodes in Julia.

Macros

Macro hygiene is represented through the expression head pair escape and hygienic-scope. The result of a macro expansion is automatically wrapped in (hygienic-scope block module), to represent the result of the new scope. The user can insert (escape block) inside to interpolate code from the caller.

Lowered form

Lowered form (IR) is more important to the compiler, since it is used for type inference, optimizations like inlining, and code generation. It is also less obvious to the human, since it results from a significant rearrangement of the input syntax.

In addition to Symbols and some number types, the following data types exist in lowered form:

• Expr

  Has a node type indicated by the head field, and an args field which is a Vector{Any} of subexpressions. While almost every part of a surface AST is represented by an Expr, the IR uses only a limited number of Exprs, mostly for calls and some top-level-only forms.

• SlotNumber

  Identifies arguments and local variables by consecutive numbering. It has an integer-valued id field giving the slot index. The types of these slots can be found in the slottypes field of their CodeInfo object.

• Argument

  The same as SlotNumber, but appears only post-optimization. Indicates that the referenced slot is an argument of the enclosing function.

• CodeInfo

  Wraps the IR of a group of statements. Its code field is an array of expressions to execute.

• GotoNode

  Unconditional branch. The argument is the branch target, represented as an index in the code array to jump to.

• GotoIfNot

  Conditional branch. If the cond field evaluates to false, goes to the index identified by the dest field.

• ReturnNode

  Returns its argument (the val field) as the value of the enclosing function. If the val field is undefined, then this represents an unreachable statement.

• QuoteNode

  Wraps an arbitrary value to reference as data. For example, the function f() = :a contains a QuoteNode whose value field is the symbol a, in order to return the symbol itself instead of evaluating it.

• GlobalRef

  Refers to global variable name in module mod.

• SSAValue

  Refers to a consecutively-numbered (starting at 1) static single assignment (SSA) variable inserted by the compiler. The number (id) of an SSAValue is the code array index of the expression whose value it represents.

• NewvarNode

  Marks a point where a variable (slot) is created. This has the effect of resetting a variable to undefined.

Expr types

These symbols appear in the head field of Exprs in lowered form.

• call

  Function call (dynamic dispatch). args[1] is the function to call, args[2:end] are the arguments.

• invoke

  Function call (static dispatch). args[1] is the MethodInstance to call, args[2:end] are the arguments (including the function that is being called, at args[2]).

• static_parameter

  Reference a static parameter by index.

• =

  Assignment. In the IR, the first argument is always a SlotNumber or a GlobalRef.

• method

  Adds a method to a generic function and assigns the result if necessary.

  Has a 1-argument form and a 3-argument form. The 1-argument form arises from the syntax function foo end. In the 1-argument form, the argument is a symbol. If this symbol already names a function in the current scope, nothing happens. If the symbol is undefined, a new function is created and assigned to the identifier specified by the symbol. If the symbol is defined but names a non-function, an error is raised. The definition of "names a function" is that the binding is constant, and refers to an object of singleton type. The rationale for this is that an instance of a singleton type uniquely identifies the type to add the method to. When the type has fields, it wouldn't be clear whether the method was being added to the instance or its type.

  The 3-argument form has the following arguments:

  • args[1]

    A function name, or nothing if unknown or unneeded. If a symbol, then the expression first behaves like the 1-argument form above. This argument is ignored from then on. It can be nothing when methods are added strictly by type, (::T)(x) = x, or when a method is being added to an existing function, MyModule.f(x) = x.

  • args[2]

    A SimpleVector of argument type data. args[2][1] is a SimpleVector of the argument types, and args[2][2] is a SimpleVector of type variables corresponding to the method's static parameters.

  • args[3]

    A CodeInfo of the method itself. For "out of scope" method definitions (adding a method to a function that also has methods defined in different scopes) this is an expression that evaluates to a :lambda expression.

• struct_type

  A 7-argument expression that defines a new struct:

  • args[1]

    The name of the struct

  • args[2]

    A call expression that creates a SimpleVector specifying its parameters

  • args[3]

    A call expression that creates a SimpleVector specifying its fieldnames

  • args[4]

    A Symbol, GlobalRef, or Expr specifying the supertype (e.g., :Integer, GlobalRef(Core, :Any), or :(Core.apply_type(AbstractArray, T, N)))

  • args[5]

    A call expression that creates a SimpleVector specifying its fieldtypes

  • args[6]

    A Bool, true if mutable

  • args[7]

    The number of arguments to initialize. This will be the number of fields, or the minimum number of fields called by an inner constructor's new statement.

• abstract_type

  A 3-argument expression that defines a new abstract type. The arguments are the same as arguments 1, 2, and 4 of struct_type expressions.

• primitive_type

  A 4-argument expression that defines a new primitive type. Arguments 1, 2, and 4 are the same as struct_type. Argument 3 is the number of bits.

  Julia 1.5

  struct_type, abstract_type, and primitive_type were removed in Julia 1.5 and replaced by calls to new builtins.

• global

  Declares a global binding.

• const

  Declares a (global) variable as constant.

• new

  Allocates a new struct-like object. First argument is the type. The new pseudo-function is lowered to this, and the type is always inserted by the compiler. This is very much an internal-only feature, and does no checking. Evaluating arbitrary new expressions can easily segfault.

• splatnew

  Similar to new, except field values are passed as a single tuple. Works similarly to splat(new) if new were a first-class function, hence the name.

• isdefined

  Expr(:isdefined, :x) returns a Bool indicating whether x has already been defined in the current scope.

• the_exception

  Yields the caught exception inside a catch block, as returned by jl_current_exception(ct).

• enter

  Enters an exception handler (setjmp). args[1] is the label of the catch block to jump to on error. Yields a token which is consumed by pop_exception.

• leave

  Pop exception handlers. args[1] is the number of handlers to pop.

• pop_exception

  Pop the stack of current exceptions back to the state at the associated enter when leaving a catch block. args[1] contains the token from the associated enter.

  Julia 1.1

  pop_exception is new in Julia 1.1.

• inbounds

  Controls turning bounds checks on or off. A stack is maintained; if the first argument of this expression is true or false (true means bounds checks are disabled), it is pushed onto the stack. If the first argument is :pop, the stack is popped.

• boundscheck

  Has the value false if inlined into a section of code marked with @inbounds, otherwise has the value true.

• loopinfo

  Marks the end of the a loop. Contains metadata that is passed to LowerSimdLoop to either mark the inner loop of @simd expression, or to propagate information to LLVM loop passes.

• copyast

  Part of the implementation of quasi-quote. The argument is a surface syntax AST that is simply copied recursively and returned at run time.

• meta

  Metadata. args[1] is typically a symbol specifying the kind of metadata, and the rest of the arguments are free-form. The following kinds of metadata are commonly used:

  • :inline and :noinline: Inlining hints.

• foreigncall

  Statically-computed container for ccall information. The fields are:

  • args[1] : name

    The expression that'll be parsed for the foreign function.

  • args[2]::Type : RT

    The (literal) return type, computed statically when the containing method was defined.

  • args[3]::SimpleVector (of Types) : AT

    The (literal) vector of argument types, computed statically when the containing method was defined.

  • args[4]::Int : nreq

    The number of required arguments for a varargs function definition.

  • args[5]::QuoteNode{<:Union{Symbol,Tuple{Symbol,UInt16}, Tuple{Symbol,UInt16,Bool}}: calling convention

    The calling convention for the call, optionally with effects, and gc_safe (safe to execute concurrently to GC.).

  • args[6:5+length(args[3])] : arguments

    The values for all the arguments (with types of each given in args[3]).

  • args[6+length(args[3])+1:end] : gc-roots

    The additional objects that may need to be gc-rooted for the duration of the call. See Working with LLVM for where these are derived from and how they get handled.

• new_opaque_closure

  Constructs a new opaque closure. The fields are:

  • args[1] : signature

    The function signature of the opaque closure. Opaque closures don't participate in dispatch, but the input types can be restricted.

  • args[2] : lb

    Lower bound on the output type. (Defaults to Union{})

  • args[3] : ub

    Upper bound on the output type. (Defaults to Any)

  • args[4] : constprop

    Indicates whether the opaque closure's identity may be used for constant propagation. The @opaque macro enables this by default, but this will cause additional inference which may be undesirable and prevents the code from running during precompile. If args[4] is a method, the argument is considered skipped.

  • args[5] : method

    The actual method as an opaque_closure_method expression.

  • args[6:end] : captures

    The values captured by the opaque closure.

  Julia 1.7

  Opaque closures were added in Julia 1.7

Method

A unique'd container describing the shared metadata for a single method.

• name, module, file, line, sig

  Metadata to uniquely identify the method for the computer and the human.

• ambig

  Cache of other methods that may be ambiguous with this one.

• specializations

  Cache of all MethodInstance ever created for this Method, used to ensure uniqueness. Uniqueness is required for efficiency, especially for incremental precompile and tracking of method invalidation.

• source

  The original source code (if available, usually compressed).

• generator

  A callable object which can be executed to get specialized source for a specific method signature.

• roots

  Pointers to non-AST things that have been interpolated into the AST, required by compression of the AST, type-inference, or the generation of native code.

• nargs, isva, called, is_for_opaque_closure,

  Descriptive bit-fields for the source code of this Method.

• primary_world

  The world age that "owns" this Method.

MethodInstance

A unique'd container describing a single callable signature for a Method. See especially Proper maintenance and care of multi-threading locks for important details on how to modify these fields safely.

• specTypes

  The primary key for this MethodInstance. Uniqueness is guaranteed through a def.specializations lookup.

• def

  The Method that this function describes a specialization of. Or a Module, if this is a top-level Lambda expanded in Module, and which is not part of a Method.

• sparam_vals

  The values of the static parameters in specTypes. For the MethodInstance at Method.unspecialized, this is the empty SimpleVector. But for a runtime MethodInstance from the MethodTable cache, this will always be defined and indexable.

• backedges

  We store the reverse-list of cache dependencies for efficient tracking of incremental reanalysis/recompilation work that may be needed after a new method definitions. This works by keeping a list of the other MethodInstance that have been inferred or optimized to contain a possible call to this MethodInstance. Those optimization results might be stored somewhere in the cache, or it might have been the result of something we didn't want to cache, such as constant propagation. Thus we merge all of those backedges to various cache entries here (there's almost always only the one applicable cache entry with a sentinel value for max_world anyways).

• cache

  Cache of CodeInstance objects that share this template instantiation.

CodeInstance

• def

  The MethodInstance that this cache entry is derived from.

• owner

  A token that represents the owner of this CodeInstance. Will use jl_egal to match.

• rettype/rettype_const

  The inferred return type for the specFunctionObject field, which (in most cases) is also the computed return type for the function in general.

• inferred

  May contain a cache of the inferred source for this function, or it could be set to nothing to just indicate rettype is inferred.

• ftpr

  The generic jlcall entry point.

• jlcall_api

  The ABI to use when calling fptr. Some significant ones include:

  • 0 - Not compiled yet
  • 1 - JL_CALLABLE jl_value_t *(*)(jl_function_t *f, jl_value_t *args[nargs], uint32_t nargs)
  • 2 - Constant (value stored in rettype_const)
  • 3 - With Static-parameters forwarded jl_value_t *(*)(jl_svec_t *sparams, jl_function_t *f, jl_value_t *args[nargs], uint32_t nargs)
  • 4 - Run in interpreter jl_value_t *(*)(jl_method_instance_t *meth, jl_function_t *f, jl_value_t *args[nargs], uint32_t nargs)

• min_world / max_world

  The range of world ages for which this method instance is valid to be called. If max_world is the special token value -1, the value is not yet known. It may continue to be used until we encounter a backedge that requires us to reconsider.

• Timing fields

  • time_infer_total: Total cost of computing inferred originally as wall-time from start to finish.

  • time_infer_cache_saved: The cost saved from time_infer_total by having caching. Adding this to time_infer_total should give a stable estimate for comparing the cost of two implementations or one implementation over time. This is generally an over-estimate of the time to infer something, since the cache is frequently effective at handling repeated work.

  • time_infer_self: Self cost of julia inference for inferred (a portion of time_infer_total). This is simply the incremental cost of compiling this one method, if given a fully populated cache of all call targets, even including constant inference results and LimitedAccuracy results, which generally are not in a cache.

  • time_compile: Self cost of llvm JIT compilation (e.g. of computing invoke from inferred). A total cost estimate can be computed by walking all of the edges contents and summing those, while accounting for cycles and duplicates. (This field currently does not include any measured AOT compile times.)

CodeInfo

A (usually temporary) container for holding lowered (and possibly inferred) source code.

• code

  An Any array of statements

• slotnames

  An array of symbols giving names for each slot (argument or local variable).

• slotflags

  A UInt8 array of slot properties, represented as bit flags:

  • 0x02 - assigned (only false if there are no assignment statements with this var on the left)
  • 0x08 - used (if there is any read or write of the slot)
  • 0x10 - statically assigned once
  • 0x20 - might be used before assigned. This flag is only valid after type inference.

• ssavaluetypes

  Either an array or an Int.

  If an Int, it gives the number of compiler-inserted temporary locations in the function (the length of code array). If an array, specifies a type for each location.

• ssaflags

  Statement-level 32 bits flags for each expression in the function. See the definition of jl_code_info_t in julia.h for more details.

These are only populated after inference (or by generated functions in some cases):

• debuginfo

  An object to retrieve source information for each statements, see How to interpret line numbers in a CodeInfo object.

• rettype

  The inferred return type of the lowered form (IR). Default value is Any. This is mostly present for convenience, as (due to the way OpaqueClosures work) it is not necessarily the rettype used by codegen.

• parent

  The MethodInstance that "owns" this object (if applicable).

• edges

  Forward edges to method instances that must be invalidated.

• min_world/max_world

  The range of world ages for which this code was valid at the time when it had been inferred.

Optional Fields:

• slottypes

  An array of types for the slots.

• method_for_inference_limit_heuristics

  The method_for_inference_heuristics will expand the given method's generator if necessary during inference.

Boolean properties:

• propagate_inbounds

  Whether this should propagate @inbounds when inlined for the purpose of eliding @boundscheck blocks.

UInt8 settings:

• constprop, inlineable

  • 0 = use heuristic
  • 1 = aggressive
  • 2 = none

• purity Constructed from 5 bit flags:

  • 0x01 << 0 = this method is guaranteed to return or terminate consistently (:consistent)
  • 0x01 << 1 = this method is free from externally semantically visible side effects (:effect_free)
  • 0x01 << 2 = this method is guaranteed to not throw an exception (:nothrow)
  • 0x01 << 3 = this method is guaranteed to terminate (:terminates_globally)
  • 0x01 << 4 = the syntactic control flow within this method is guaranteed to terminate (:terminates_locally)

  See the documentation of Base.@assume_effects for more details.

How to interpret line numbers in a CodeInfo object

There are 2 common forms for this data: one used internally that compresses the data somewhat and one used in the compiler. They contain the same basic info, but the compiler version is all mutable while the version used internally is not.

Many consumers may be able to call Base.IRShow.buildLineInfoNode, Base.IRShow.append_scopes!, or Stacktraces.lookup(::InterpreterIP) to avoid needing to (re-)implement these details specifically.

The definitions of each of these are:

struct Core.DebugInfo
    @noinline
    def::Union{Method,MethodInstance,Symbol}
    linetable::Union{Nothing,DebugInfo}
    edges::SimpleVector{DebugInfo}
    codelocs::String # compressed data
end
mutable struct Core.Compiler.DebugInfoStream
    def::Union{Method,MethodInstance,Symbol}
    linetable::Union{Nothing,DebugInfo}
    edges::Vector{DebugInfo}
    firstline::Int32 # the starting line for this block (specified by an index of 0)
    codelocs::Vector{Int32} # for each statement:
        # index into linetable (if defined), else a line number (in the file represented by def)
        # then index into edges
        # then index into edges[linetable]
end

• def : where this DebugInfo was defined (the Method, MethodInstance, or Symbol of file scope, for example)

• linetable

  Another DebugInfo that this was derived from, which contains the actual line numbers, such that this DebugInfo contains only the indexes into it. This avoids making copies, as well as makes it possible to track how each individual statement transformed from source to optimized, not just the separate line numbers. If def is not a Symbol, then that object replaces the current function object for the metadata on what function is conceptually being executed (e.g. think Cassette transforms here). The codelocs values described below also are interpreted as an index into the codelocs in this object, instead of being a line number itself.

• edges : Vector of the unique DebugInfo for every function inlined into this (which recursively have the edges for everything inlined into them).

• firstline (when uncompressed to DebugInfoStream)

  The line number associated with the begin statement (or other keyword such as function or quote) that delineates where this code definition "starts".

• codelocs (when uncompressed to DebugInfoStream)

  A vector of indices, with 3 values for each statement in the IR plus one for the starting point of the block, that describe the stacktrace from that point:

  1. the integer index into the linetable.codelocs field, giving the original location associated with each statement (including its syntactic edges), or zero indicating no change to the line number from the previously executed statement (which is not necessarily syntactic or lexical prior), or the line number itself if the linetable field is nothing.
  2. the integer index into edges, giving the DebugInfo inlined there, or zero if there are no edges.
  3. (if entry 2 is non-zero) the integer index into edges[].codelocs, to interpret recursively for each function in the inlining stack, or zero indicating to use edges[].firstline as the line number.

  Special codes include:

  • (zero, zero, *): no change to the line number or edges from the previous statement (you may choose to interpret this either syntactically or lexically). The inlining depth also might have changed, though most callers should ignore that.
  • (zero, non-zero, *) : no line number, just edges (usually because of macro-expansion into top-level code).