Nim Experimental Features

  Source   Edit

Authors:Andreas Rumpf

About this document

This document describes features of Nim that are to be considered experimental. Some of these are not covered by the .experimental pragma or --experimental switch because they are already behind a special syntax and one may want to use Nim libraries using these features without using them oneself.

Note: Unless otherwise indicated, these features are not to be removed, but refined and overhauled.

Package level objects

Every Nim module resides in a (nimble) package. An object type can be attached to the package it resides in. If that is done, the type can be referenced from other modules as an incomplete object type. This feature allows to break up recursive type dependencies across module boundaries. Incomplete object types are always passed byref and can only be used in pointer like contexts (var/ref/ptr IncompleteObject) in general since the compiler does not yet know the size of the object. To complete an incomplete object the package pragma has to be used. package implies byref.

As long as a type T is incomplete, neither sizeof(T) nor runtime type information for T is available.


# module A (in an arbitrary package)
  Pack.SomeObject = object ## declare as incomplete object of package 'Pack'
  Triple = object
    a, b, c: ref SomeObject ## pointers to incomplete objects are allowed

## Incomplete objects can be used as parameters:
proc myproc(x: SomeObject) = discard
# module B (in package "Pack")
  SomeObject* {.package.} = object ## Use 'package' to complete the object
    s, t: string
    x, y: int

Void type

The void type denotes the absence of any type. Parameters of type void are treated as non-existent, void as a return type means that the procedure does not return a value:

proc nothing(x, y: void): void =
  echo "ha"

nothing() # writes "ha" to stdout

The void type is particularly useful for generic code:

proc callProc[T](p: proc (x: T), x: T) =
  when T is void:

proc intProc(x: int) = discard
proc emptyProc() = discard

callProc[int](intProc, 12)

However, a void type cannot be inferred in generic code:

# Error: type mismatch: got (proc ())
# but expected one of:
# callProc(p: proc (T), x: T)

The void type is only valid for parameters and return types; other symbols cannot have the type void.

Automatic dereferencing

Automatic dereferencing is performed for the first argument of a routine call. This feature has to be enabled via {.experimental: "implicitDeref".}:

{.experimental: "implicitDeref".}

proc depth(x: NodeObj): int = ...

  n: Node
echo n.depth
# no need to write n[].depth either

Code reordering

The code reordering feature can implicitly rearrange procedure, template, and macro definitions along with variable declarations and initializations at the top level scope so that, to a large extent, a programmer should not have to worry about ordering definitions correctly or be forced to use forward declarations to preface definitions inside a module.


{.experimental: "codeReordering".}

proc foo(x: int) =

proc bar(x: int) =


Variables can also be reordered as well. Variables that are initialized (i.e. variables that have their declaration and assignment combined in a single statement) can have their entire initialization statement reordered. Be wary of what code is executed at the top level:

{.experimental: "codeReordering".}

proc a() =

var foo = 5

a() # outputs: "5"

It is important to note that reordering only works for symbols at top level scope. Therefore, the following will fail to compile:

{.experimental: "codeReordering".}

proc a() =
  proc b() =


Named argument overloading

Routines with the same type signature can be called differently if a parameter has different names. This does not need an experimental switch, but is an unstable feature.

proc foo(x: int) =
  echo "Using x: ", x
proc foo(y: int) =
  echo "Using y: ", y

foo(x = 2)
# Using x: 2
foo(y = 2)
# Using y: 2

Do notation

As a special more convenient notation, proc expressions involved in procedure calls can use the do keyword:

sort(cities) do (x,y: string) -> int:
  cmp(x.len, y.len)

# Less parenthesis using the method plus command syntax:
cities = do (x:string) -> string:
  "City of " & x

# In macros, the do notation is often used for quasi-quoting
macroResults.add quote do:
  if not `ex`:
    echo `info`, ": Check failed: ", `expString`

do is written after the parentheses enclosing the regular proc params. The proc expression represented by the do block is appended to them. In calls using the command syntax, the do block will bind to the immediately preceding expression, transforming it in a call.

do with parentheses is an anonymous proc; however a do without parentheses is just a block of code. The do notation can be used to pass multiple blocks to a macro:

macro performWithUndo(task, undo: untyped) = ...

performWithUndo do:
  # multiple-line block of code
  # to perform the task
  # code to undo it

Special Operators

dot operators

Note: Dot operators are still experimental and so need to be enabled via {.experimental: "dotOperators".}.

Nim offers a special family of dot operators that can be used to intercept and rewrite proc call and field access attempts, referring to previously undeclared symbol names. They can be used to provide a fluent interface to objects lying outside the static confines of the type system such as values from dynamic scripting languages or dynamic file formats such as JSON or XML.

When Nim encounters an expression that cannot be resolved by the standard overload resolution rules, the current scope will be searched for a dot operator that can be matched against a re-written form of the expression, where the unknown field or proc name is passed to an untyped parameter:

a.b # becomes `.`(a, b)
a.b(c, d) # becomes `.`(a, b, c, d)

The matched dot operators can be symbols of any callable kind (procs, templates and macros), depending on the desired effect:

template `.`(js: PJsonNode, field: untyped): JSON = js[astToStr(field)]

var js = parseJson("{ x: 1, y: 2}")
echo js.x # outputs 1
echo js.y # outputs 2

The following dot operators are available:

operator .

This operator will be matched against both field accesses and method calls.

operator .()

This operator will be matched exclusively against method calls. It has higher precedence than the . operator and this allows one to handle expressions like x.y and x.y() differently if one is interfacing with a scripting language for example.

operator .=

This operator will be matched against assignments to missing fields.

a.b = c # becomes `.=`(a, b, c)

Call operator

The call operator, (), matches all kinds of unresolved calls and takes precedence over dot operators, however it does not match missing overloads for existing routines. The experimental callOperator switch must be enabled to use this operator.

{.experimental: "callOperator".}

template `()`(a: int, b: float): untyped = $(a, b)

  let a = 1.0
  let b = 2
  doAssert b(a) == `()`(b, a)
  doAssert a.b == `()`(b, a)

  let a = 1.0
  proc b(): int = 2
  doAssert not compiles(b(a))
  doAssert not compiles(a.b) # `()` not called

  let a = 1.0
  proc b(x: float): int = int(x + 1)
  let c = 3.0
  doAssert not compiles(a.b(c)) # gives a type mismatch error same as b(a, c)
  doAssert (a.b)(c) == `()`(a.b, c)

Not nil annotation

Note: This is an experimental feature. It can be enabled with {.experimental: "notnil".}.

All types for which nil is a valid value can be annotated with the not nil annotation to exclude nil as a valid value:

{.experimental: "notnil".}

  PObject = ref TObj not nil
  TProc = (proc (x, y: int)) not nil

proc p(x: PObject) =
  echo "not nil"

# compiler catches this:

# and also this:
var x: PObject

The compiler ensures that every code path initializes variables which contain non-nilable pointers. The details of this analysis are still to be specified here.

Strict not nil checking

Note: This feature is experimental, you need to enable it with

{.experimental: "strictNotNil".}


nim c --experimental:strictNotNil <program>

In the second case it would check builtin and imported modules as well.

It checks the nilability of ref-like types and makes dereferencing safer based on flow typing and not nil annotations.

Its implementation is different than the notnil one: defined under strictNotNil. Keep in mind the difference in option names, be careful with distinguishing them.

We check several kinds of types for nilability:

  • ref types
  • pointer types
  • proc types
  • cstrings


The default kind of nilability types is the nilable kind: they can have the value nil. If you have a non-nilable type T, you can use T nil to get a nilable type for it.

not nil

You can annotate a type where nil isn't a valid value with not nil.

  NilableObject = ref object
    a: int
  Object = NilableObject not nil
  Proc = (proc (x, y: int))

proc p(x: Object) =
  echo x.a # ensured to dereference without an error
# compiler catches this:
# and also this:
var x: NilableObject
if x.isNil:
  p(x) # ok

If a type can include nil as a valid value, dereferencing values of the type is checked by the compiler: if a value which might be nil is derefenced, this produces a warning by default, you can turn this into an error using the compiler options --warningAsError:strictNotNil.

If a type is nilable, you should dereference its values only after a isNil or equivalent check.

local turn on/off

You can still turn off nil checking on function/module level by using a {.strictNotNil: off.} pragma. Note: test that/TODO for code/manual.

nilability state

Currently a nilable value can be Safe, MaybeNil or Nil : we use internally Parent and Unreachable but this is an implementation detail(a parent layer has the actual nilability).

  • Safe means it shouldn't be nil at that point: e.g. after assignment to a non-nil value or not a.isNil check
  • MaybeNil means it might be nil, but it might not be nil: e.g. an argument, a call argument or a value after an if and else.
  • Nil means it should be nil at that point; e.g. after an assignment to nil or a .isNil check.
  • Unreachable means it shouldn't be possible to access this in this branch: so we do generate a warning as well.

We show an error for each dereference ([], .field, [index] () etc) which is of a tracked expression which is in MaybeNil or Nil state.

type nilability

Types are either nilable or non-nilable. When you pass a param or a default value, we use the type : for nilable types we return MaybeNil and for non-nilable Safe.

TODO: fix the manual here. (This is not great, as default values for non-nilables and nilables are usually actually nil , so we should think a bit more about this section.)

params rules

Param's nilability is detected based on type nilability. We use the type of the argument to detect the nilability.

assignment rules

Let's say we have left = right.

When we assign, we pass the right's nilability to the left's expression. There should be special handling of aliasing and compound expressions which we specify in their sections. (Assignment is a possible alias move or move out).

call args rules

When we call with arguments, we have two cases when we might change the nilability.


Here callByVar can re-assign a, so this might change a's nilability, so we change it to MaybeNil. This is also a possible aliasing move out (moving out of a current alias set).


Here call can change a field or element of a, so if we have a dependant expression of a : e.g. a.field. Dependats become MaybeNil.

branches rules

Branches are the reason we do nil checking like this: with flow checking. Sources of brancing are if, while, for, and, or, case, try and combinations with return, break, continue and raise

We create a new layer/"scope" for each branch where we map expressions to nilability. This happens when we "fork": usually on the beginning of a construct. When branches "join" we usually unify their expression maps or/and nilabilities.

Merging usually merges maps and alias sets: nilabilities are merged like this:

template union(l: Nilability, r: Nilability): Nilability =
  ## unify two states
  if l == r:

Special handling is for .isNil and == nil, also for not, and and or.

not reverses the nilability, and is similar to "forking" : the right expression is checked in the layer resulting from the left one and or is similar to "merging": the right and left expression should be both checked in the original layer.

isNil, == nil make expressions Nil. If there is a not or != nil, they make them Safe. We also reverse the nilability in the opposite branch: e.g. else.

compound expressions: field, index expressions

We want to track also field(dot) and index(bracket) expressions.

We track some of those compound expressions which might be nilable as dependants of their bases: a.field is changed if a is moved (re-assigned), similarly a[index] is dependent on a and a.field.field on a.field.

When we move the base, we update dependants to MaybeNil. Otherwise we usually start with type nilability.

When we call args, we update the nilability of their dependants to MaybeNil as the calls usually can change them. We might need to check for strictFuncs pure funcs and not do that then.

For field expressions a.field, we calculate an integer value based on a hash of the tree and just accept equivalent trees as equivalent expressions.

For item expression a[index], we also calculate an integer value based on a hash of the tree and accept equivalent trees as equivalent expressions: for static values only. For now we support only constant indices: we dont track expression with no-const indices. For those we just report a warning even if they are safe for now: one can use a local variable to workaround. For loops this might be annoying: so one should be able to turn off locally the warning using the {.warning[StrictCheckNotNil]:off}..

For bracket expressions, in the future we might count a[<any>] as the same general expression. This means we should should the index but otherwise handle it the same for assign (maybe "aliasing" all the non-static elements) and differentiate only for static: e.g. a[0] and a[1].

element tracking

When we assign an object construction, we should track the fields as well:

var a = Nilable(field: Nilable()) # a : Safe, a.field: Safe

Usually we just track the result of an expression: probably this should apply for elements in other cases as well. Also related to tracking initialization of expressions/fields.

unstructured control flow rules

Unstructured control flow keywords as return, break, continue, raise mean that we jump from a branch out. This means that if there is code after the finishing of the branch, it would be ran if one hasn't hit the direct parent branch of those: so it is similar to an else. In those cases we should use the reverse nilabilities for the local to the condition expressions. E.g.

for a in c:
  if not a.isNil:
  code # here a: Nil , because if not, we would have breaked


We support alias detection for local expressions.

We track sets of aliased expressions. We start with all nilable local expressions in separate sets. Assignments and other changes to nilability can move / move out expressions of sets.

move: Moving left to right means we remove left from its current set and unify it with the right's set. This means it stops being aliased with its previous aliases.

var left = b
left = right # moving left to right

move out: Moving out left might remove it from the current set and ensure that it's in its own set as a single element. e.g.

var left = b
left = nil # moving out

initialization of non nilable and nilable values


warnings and errors

We show an error for each dereference ([], .field, [index] () etc) which is of a tracked expression which is in MaybeNil or Nil state.

We might also show a history of the transitions and the reasons for them that might change the nilability of the expression.


Concepts, also known as "user-defined type classes", are used to specify an arbitrary set of requirements that the matched type must satisfy.

Concepts are written in the following form:

  Comparable = concept x, y
    (x < y) is bool
  Stack[T] = concept s, var v
    s.pop() is T
    s.len is Ordinal
    for value in s:
      value is T

The concept is a match if:

  1. all of the expressions within the body can be compiled for the tested type
  2. all statically evaluable boolean expressions in the body must be true

The identifiers following the concept keyword represent instances of the currently matched type. You can apply any of the standard type modifiers such as var, ref, ptr and static to denote a more specific type of instance. You can also apply the type modifier to create a named instance of the type itself:

  MyConcept = concept x, var v, ref r, ptr p, static s, type T

Within the concept body, types can appear in positions where ordinary values and parameters are expected. This provides a more convenient way to check for the presence of callable symbols with specific signatures:

  OutputStream = concept var s

In order to check for symbols accepting type params, you must prefix the type with the explicit type modifier. The named instance of the type, following the concept keyword is also considered to have the explicit modifier and will be matched only as a type.

  # Let's imagine a user-defined casting framework with operators
  # such as `` and ``. We can test
  # for these with the following concept:
  MyCastables = concept x string) JSonValue)
  # Let's define a couple of concepts, known from Algebra:
  AdditiveMonoid* = concept x, y, type T
    x + y is T is T # require a proc such as `` or ''
  AdditiveGroup* = concept x, y, type T
    x is AdditiveMonoid
    -x is T
    x - y is T

Please note that the is operator allows one to easily verify the precise type signatures of the required operations, but since type inference and default parameters are still applied in the concept body, it's also possible to describe usage protocols that do not reveal implementation details.

Much like generics, concepts are instantiated exactly once for each tested type and any static code included within the body is executed only once.

Concept diagnostics

By default, the compiler will report the matching errors in concepts only when no other overload can be selected and a normal compilation error is produced. When you need to understand why the compiler is not matching a particular concept and, as a result, a wrong overload is selected, you can apply the explain pragma to either the concept body or a particular call-site.

  MyConcept {.explain.} = concept ...

overloadedProc(x, y, z) {.explain.}

This will provide Hints in the compiler output either every time the concept is not matched or only on the particular call-site.

Generic concepts and type binding rules

The concept types can be parametric just like the regular generic types:

### matrixalgo.nim

import std/typetraits

  AnyMatrix*[R, C: static int; T] = concept m, var mvar, type M
    M.ValueType is T
    M.Rows == R
    M.Cols == C
    m[int, int] is T
    mvar[int, int] = T
    type TransposedType = stripGenericParams(M)[C, R, T]
  AnySquareMatrix*[N: static int, T] = AnyMatrix[N, N, T]
  AnyTransform3D* = AnyMatrix[4, 4, float]

proc transposed*(m: AnyMatrix): m.TransposedType =
  for r in 0 ..< m.R:
    for c in 0 ..< m.C:
      result[r, c] = m[c, r]

proc determinant*(m: AnySquareMatrix): int =

proc setPerspectiveProjection*(m: AnyTransform3D) =

### matrix.nim

  Matrix*[M, N: static int; T] = object
    data: array[M*N, T]

proc `[]`*(M: Matrix; m, n: int): M.T =[m * M.N + n]

proc `[]=`*(M: var Matrix; m, n: int; v: M.T) =[m * M.N + n] = v

# Adapt the Matrix type to the concept's requirements
template Rows*(M: typedesc[Matrix]): int = M.M
template Cols*(M: typedesc[Matrix]): int = M.N
template ValueType*(M: typedesc[Matrix]): typedesc = M.T

### usage.nim

import matrix, matrixalgo

  m: Matrix[3, 3, int]
  projectionMatrix: Matrix[4, 4, float]

echo m.transposed.determinant
setPerspectiveProjection projectionMatrix

When the concept type is matched against a concrete type, the unbound type parameters are inferred from the body of the concept in a way that closely resembles the way generic parameters of callable symbols are inferred on call sites.

Unbound types can appear both as params to calls such as s.push(T) and on the right-hand side of the is operator in cases such as x.pop is T and is seq[T].

Unbound static params will be inferred from expressions involving the == operator and also when types dependent on them are being matched:

  MatrixReducer[M, N: static int; T] = concept x
    x.reduce(SquareMatrix[N, T]) is array[M, int]

The Nim compiler includes a simple linear equation solver, allowing it to infer static params in some situations where integer arithmetic is involved.

Just like in regular type classes, Nim discriminates between bind once and bind many types when matching the concept. You can add the distinct modifier to any of the otherwise inferable types to get a type that will be matched without permanently inferring it. This may be useful when you need to match several procs accepting the same wide class of types:

  Enumerable[T] = concept e
    for v in e:
      v is T

  MyConcept = concept o
    # this could be inferred to a type such as Enumerable[int] is distinct Enumerable
    # this could be inferred to a different type such as Enumerable[float] is distinct Enumerable
    # it's also possible to give an alias name to a `bind many` type class
    type Enum = distinct Enumerable
    o.baz is Enum

On the other hand, using bind once types allows you to test for equivalent types used in multiple signatures, without actually requiring any concrete types, thus allowing you to encode implementation-defined types:

  MyConcept = concept x
    type T1 = auto # both procs must accept the same type
    type T2 = seq[SomeNumber]
    x.alpha(T2) # both procs must accept the same type
                # and it must be a numeric sequence

As seen in the previous examples, you can refer to generic concepts such as Enumerable[T] just by their short name. Much like the regular generic types, the concept will be automatically instantiated with the bind once auto type in the place of each missing generic param.

Please note that generic concepts such as Enumerable[T] can be matched against concrete types such as string. Nim doesn't require the concept type to have the same number of parameters as the type being matched. If you wish to express a requirement towards the generic parameters of the matched type, you can use a type mapping operator such as genericHead or stripGenericParams within the body of the concept to obtain the uninstantiated version of the type, which you can then try to instantiate in any required way. For example, here is how one might define the classic Functor concept from Haskell and then demonstrate that Nim's Option[T] type is an instance of it:

import std/[sugar, typetraits]

  Functor[A] = concept f
    type MatchedGenericType = genericHead(typeof(f))
      # `f` will be a value of a type such as `Option[T]`
      # `MatchedGenericType` will become the `Option` type
    f.val is A
      # The Functor should provide a way to obtain
      # a value stored inside it
    type T = auto
    map(f, A -> T) is MatchedGenericType[T]
      # And it should provide a way to map one instance of
      # the Functor to a instance of a different type, given
      # a suitable `map` operation for the enclosed values

import std/options
echo Option[int] is Functor # prints true

Concept derived values

All top level constants or types appearing within the concept body are accessible through the dot operator in procs where the concept was successfully matched to a concrete type:

  DateTime = concept t1, t2, type T
    const Min = T.MinDate
    T.Now is T
    t1 < t2 is bool
    type TimeSpan = typeof(t1 - t2)
    TimeSpan * int is TimeSpan
    TimeSpan + TimeSpan is TimeSpan
    t1 + TimeSpan is T

proc eventsJitter(events: Enumerable[DateTime]): float =
    # this variable will have the inferred TimeSpan type for
    # the concrete Date-like value the proc was called with:
    averageInterval: DateTime.TimeSpan
    deviation: float

Concept refinement

When the matched type within a concept is directly tested against a different concept, we say that the outer concept is a refinement of the inner concept and thus it is more-specific. When both concepts are matched in a call during overload resolution, Nim will assign a higher precedence to the most specific one. As an alternative way of defining concept refinements, you can use the object inheritance syntax involving the of keyword:

  Graph = concept g, type G of EquallyComparable, Copyable
      VertexType = G.VertexType
      EdgeType = G.EdgeType
    VertexType is Copyable
    EdgeType is Copyable
      v: VertexType
      e: EdgeType
  IncidendeGraph = concept of Graph
    # symbols such as variables and types from the refined
    # concept are automatically in scope:
    g.source(e) is VertexType is VertexType
    g.outgoingEdges(v) is Enumerable[EdgeType]
  BidirectionalGraph = concept g, type G
    # The following will also turn the concept into a refinement when it
    # comes to overload resolution, but it doesn't provide the convenient
    # symbol inheritance
    g is IncidendeGraph
    g.incomingEdges(G.VertexType) is Enumerable[G.EdgeType]

proc f(g: IncidendeGraph)
proc f(g: BidirectionalGraph) # this one will be preferred if we pass a type
                              # matching the BidirectionalGraph concept

Type bound operations

There are 4 operations that are bound to a type:

  1. Assignment
  2. Moves
  3. Destruction
  4. Deep copying for communication between threads

These operations can be overridden instead of overloaded. This means the implementation is automatically lifted to structured types. For instance if type T has an overridden assignment operator = this operator is also used for assignments of the type seq[T]. Since these operations are bound to a type they have to be bound to a nominal type for reasons of simplicity of implementation: This means an overridden deepCopy for ref T is really bound to T and not to ref T. This also means that one cannot override deepCopy for both ptr T and ref T at the same time; instead a helper distinct or object type has to be used for one pointer type.

Assignments, moves and destruction are specified in the destructors document.


=deepCopy is a builtin that is invoked whenever data is passed to a spawn'ed proc to ensure memory safety. The programmer can override its behaviour for a specific ref or ptr type T. (Later versions of the language may weaken this restriction.)

The signature has to be:

proc `=deepCopy`(x: T): T

This mechanism will be used by most data structures that support shared memory like channels to implement thread safe automatic memory management.

The builtin deepCopy can even clone closures and their environments. See the documentation of spawn for details.

Case statement macros

Macros named case can rewrite case statements for certain types in order to implement pattern matching. The following example implements a simplistic form of pattern matching for tuples, leveraging the existing equality operator for tuples (as provided in system.==):

{.experimental: "caseStmtMacros".}

import std/macros

macro `case`(n: tuple): untyped =
  result = newTree(nnkIfStmt)
  let selector = n[0]
  for i in 1 ..< n.len:
    let it = n[i]
    case it.kind
    of nnkElse, nnkElifBranch, nnkElifExpr, nnkElseExpr:
      result.add it
    of nnkOfBranch:
      for j in
        let cond = newCall("==", selector, it[j])
        result.add newTree(nnkElifBranch, cond, it[^1])
      error "custom 'case' for tuple cannot handle this node", it

case ("foo", 78)
of ("foo", 78): echo "yes"
of ("bar", 88): echo "no"
else: discard

Currently case statement macros must be enabled explicitly via {.experimental: "caseStmtMacros".}.

case macros are subject to overload resolution. The type of the case statement's selector expression is matched against the type of the first argument of the case macro. Then the complete case statement is passed in place of the argument and the macro is evaluated.

In other words, the macro needs to transform the full case statement but only the statement's selector expression is used to determine which macro to call.

Term rewriting macros

Term rewriting macros are macros or templates that have not only a name but also a pattern that is searched for after the semantic checking phase of the compiler: This means they provide an easy way to enhance the compilation pipeline with user defined optimizations:

template optMul{`*`(a, 2)}(a: int): int = a+a

let x = 3
echo x * 2

The compiler now rewrites x * 2 as x + x. The code inside the curlies is the pattern to match against. The operators *, **, |, ~ have a special meaning in patterns if they are written in infix notation, so to match verbatim against * the ordinary function call syntax needs to be used.

Term rewriting macro are applied recursively, up to a limit. This means that if the result of a term rewriting macro is eligible for another rewriting, the compiler will try to perform it, and so on, until no more optimizations are applicable. To avoid putting the compiler into an infinite loop, there is a hard limit on how many times a single term rewriting macro can be applied. Once this limit has been passed, the term rewriting macro will be ignored.

Unfortunately optimizations are hard to get right and even the tiny example is wrong:

template optMul{`*`(a, 2)}(a: int): int = a+a

proc f(): int =
  echo "side effect!"
  result = 55

echo f() * 2

We cannot duplicate 'a' if it denotes an expression that has a side effect! Fortunately Nim supports side effect analysis:

template optMul{`*`(a, 2)}(a: int{noSideEffect}): int = a+a

proc f(): int =
  echo "side effect!"
  result = 55

echo f() * 2 # not optimized ;-)

You can make one overload matching with a constraint and one without, and the one with a constraint will have precedence, and so you can handle both cases differently.

So what about 2 * a? We should tell the compiler * is commutative. We cannot really do that however as the following code only swaps arguments blindly:

template mulIsCommutative{`*`(a, b)}(a, b: int): int = b*a

What optimizers really need to do is a canonicalization:

template canonMul{`*`(a, b)}(a: int{lit}, b: int): int = b*a

The int{lit} parameter pattern matches against an expression of type int, but only if it's a literal.

Parameter constraints

The parameter constraint expression can use the operators | (or), & (and) and ~ (not) and the following predicates:

atomThe matching node has no children.
litThe matching node is a literal like "abc", 12.
symThe matching node must be a symbol (a bound identifier).
identThe matching node must be an identifier (an unbound identifier).
callThe matching AST must be a call/apply expression.
lvalueThe matching AST must be an lvalue.
sideeffectThe matching AST must have a side effect.
nosideeffectThe matching AST must have no side effect.
paramA symbol which is a parameter.
genericparamA symbol which is a generic parameter.
moduleA symbol which is a module.
typeA symbol which is a type.
varA symbol which is a variable.
letA symbol which is a let variable.
constA symbol which is a constant.
resultThe special result variable.
procA symbol which is a proc.
methodA symbol which is a method.
iteratorA symbol which is an iterator.
converterA symbol which is a converter.
macroA symbol which is a macro.
templateA symbol which is a template.
fieldA symbol which is a field in a tuple or an object.
enumfieldA symbol which is a field in an enumeration.
forvarA for loop variable.
labelA label (used in block statements).
nk*The matching AST must have the specified kind. (Example: nkIfStmt denotes an if statement.)
aliasStates that the marked parameter needs to alias with some other parameter.
noaliasStates that every other parameter must not alias with the marked parameter.

Predicates that share their name with a keyword have to be escaped with backticks. The alias and noalias predicates refer not only to the matching AST, but also to every other bound parameter; syntactically they need to occur after the ordinary AST predicates:

template ex{a = b + c}(a: int{noalias}, b, c: int) =
  # this transformation is only valid if 'b' and 'c' do not alias 'a':
  a = b
  inc a, c

Another example:

proc somefunc(s: string)                 = assert s == "variable"
proc somefunc(s: string{nkStrLit})       = assert s == "literal"
proc somefunc(s: string{nkRStrLit})      = assert s == r"raw"
proc somefunc(s: string{nkTripleStrLit}) = assert s == """triple"""
proc somefunc(s: static[string])         = assert s == "constant"

# Use parameter constraints to provide overloads based on both the input parameter type and form.
var variable = "variable"
const constant = "constant"

Pattern operators

The operators *, **, |, ~ have a special meaning in patterns if they are written in infix notation.

The | operator

The | operator if used as infix operator creates an ordered choice:

template t{0|1}(): untyped = 3
let a = 1
# outputs 3:
echo a

The matching is performed after the compiler performed some optimizations like constant folding, so the following does not work:

template t{0|1}(): untyped = 3
# outputs 1:
echo 1

The reason is that the compiler already transformed the 1 into "1" for the echo statement. However, a term rewriting macro should not change the semantics anyway. In fact they can be deactivated with the --patterns:off command line option or temporarily with the patterns pragma.

The {} operator

A pattern expression can be bound to a pattern parameter via the expr{param} notation:

template t{(0|1|2){x}}(x: untyped): untyped = x+1
let a = 1
# outputs 2:
echo a

The ~ operator

The ~ operator is the not operator in patterns:

template t{x = (~x){y} and (~x){z}}(x, y, z: bool) =
  x = y
  if x: x = z

  a = false
  b = true
  c = false
a = b and c
echo a

The * operator

The * operator can flatten a nested binary expression like a & b & c to &(a, b, c):

  calls = 0

proc `&&`(s: varargs[string]): string =
  result = s[0]
  for i in 1..len(s)-1: result.add s[i]
  inc calls

template optConc{ `&&` * a }(a: string): untyped = &&a

let space = " "
echo "my" && (space & "awe" && "some " ) && "concat"

# check that it's been optimized properly:
doAssert calls == 1

The second operator of * must be a parameter; it is used to gather all the arguments. The expression "my" && (space & "awe" && "some " ) && "concat" is passed to optConc in a as a special list (of kind nkArgList) which is flattened into a call expression; thus the invocation of optConc produces:

`&&`("my", space & "awe", "some ", "concat")

The ** operator

The ** is much like the * operator, except that it gathers not only all the arguments, but also the matched operators in reverse polish notation:

import std/macros

  Matrix = object
    dummy: int

proc `*`(a, b: Matrix): Matrix = discard
proc `+`(a, b: Matrix): Matrix = discard
proc `-`(a, b: Matrix): Matrix = discard
proc `$`(a: Matrix): string = result = $a.dummy
proc mat21(): Matrix =
  result.dummy = 21

macro optM{ (`+`|`-`|`*`) ** a }(a: Matrix): untyped =
  echo treeRepr(a)
  result = newCall(bindSym"mat21")

var x, y, z: Matrix

echo x + y * z - x

This passes the expression x + y * z - x to the optM macro as an nnkArgList node containing:

  Sym "x"
  Sym "y"
  Sym "z"
  Sym "*"
  Sym "+"
  Sym "x"
  Sym "-"

(Which is the reverse polish notation of x + y * z - x.)


Parameters in a pattern are type checked in the matching process. If a parameter is of the type varargs it is treated specially and it can match 0 or more arguments in the AST to be matched against:

template optWrite{
  write(f, x)
  ((write|writeLine){w})(f, y)
}(x, y: varargs[untyped], f: File, w: untyped) =
  w(f, x, y)

Example: Partial evaluation

The following example shows how some simple partial evaluation can be implemented with term rewriting:

proc p(x, y: int; cond: bool): int =
  result = if cond: x + y else: x - y

template optP1{p(x, y, true)}(x, y: untyped): untyped = x + y
template optP2{p(x, y, false)}(x, y: untyped): untyped = x - y

Example: Hoisting

The following example shows how some form of hoisting can be implemented:

import std/pegs

template optPeg{peg(pattern)}(pattern: string{lit}): Peg =
  var gl {.global, gensym.} = peg(pattern)

for i in 0 .. 3:
  echo match("(a b c)", peg"'(' @ ')'")
  echo match("W_HI_Le", peg"\y 'while'")

The optPeg template optimizes the case of a peg constructor with a string literal, so that the pattern will only be parsed once at program startup and stored in a global gl which is then re-used. This optimization is called hoisting because it is comparable to classical loop hoisting.

AST based overloading

Parameter constraints can also be used for ordinary routine parameters; these constraints affect ordinary overloading resolution then:

proc optLit(a: string{lit|`const`}) =
  echo "string literal"
proc optLit(a: string) =
  echo "no string literal"

  constant = "abc"

  variable = "xyz"


However, the constraints alias and noalias are not available in ordinary routines.

Parallel & Spawn

Nim has two flavors of parallelism:

  1. Structured parallelism via the parallel statement.
  2. Unstructured parallelism via the standalone spawn statement.

Nim has a builtin thread pool that can be used for CPU intensive tasks. For IO intensive tasks the async and await features should be used instead. Both parallel and spawn need the threadpool module to work.

Somewhat confusingly, spawn is also used in the parallel statement with slightly different semantics. spawn always takes a call expression of the form f(a, ...). Let T be f's return type. If T is void then spawn's return type is also void otherwise it is FlowVar[T].

Within a parallel section sometimes the FlowVar[T] is eliminated to T. This happens when T does not contain any GC'ed memory. The compiler can ensure the location in location = spawn f(...) is not read prematurely within a parallel section and so there is no need for the overhead of an indirection via FlowVar[T] to ensure correctness.

Note: Currently exceptions are not propagated between spawn'ed tasks!

Spawn statement

spawn can be used to pass a task to the thread pool:

import std/threadpool

proc processLine(line: string) =
  discard "do some heavy lifting here"

for x in lines("myinput.txt"):
  spawn processLine(x)

For reasons of type safety and implementation simplicity the expression that spawn takes is restricted:

  • It must be a call expression f(a, ...).
  • f must be gcsafe.
  • f must not have the calling convention closure.
  • f's parameters may not be of type var. This means one has to use raw ptr's for data passing reminding the programmer to be careful.
  • ref parameters are deeply copied which is a subtle semantic change and can cause performance problems but ensures memory safety. This deep copy is performed via system.deepCopy and so can be overridden.
  • For safe data exchange between f and the caller a global TChannel needs to be used. However, since spawn can return a result, often no further communication is required.

spawn executes the passed expression on the thread pool and returns a data flow variable FlowVar[T] that can be read from. The reading with the ^ operator is blocking. However, one can use blockUntilAny to wait on multiple flow variables at the same time:

import std/threadpool, ...

# wait until 2 out of 3 servers received the update:
proc main =
  var responses = newSeq[FlowVarBase](3)
  for i in 0..2:
    responses[i] = spawn tellServer(Update, "key", "value")
  var index = blockUntilAny(responses)
  assert index >= 0
  discard blockUntilAny(responses)

Data flow variables ensure that no data races are possible. Due to technical limitations not every type T is possible in a data flow variable: T has to be of the type ref, string, seq or of a type that doesn't contain a type that is garbage collected. This restriction is not hard to work-around in practice.

Parallel statement


# Compute PI in an inefficient way
import std/[strutils, math, threadpool]
{.experimental: "parallel".}

proc term(k: float): float = 4 * math.pow(-1, k) / (2*k + 1)

proc pi(n: int): float =
  var ch = newSeq[float](n+1)
    for k in
      ch[k] = spawn term(float(k))
  for k in
    result += ch[k]

echo formatFloat(pi(5000))

The parallel statement is the preferred mechanism to introduce parallelism in a Nim program. A subset of the Nim language is valid within a parallel section. This subset is checked during semantic analysis to be free of data races. A sophisticated disjoint checker ensures that no data races are possible even though shared memory is extensively supported!

The subset is in fact the full language with the following restrictions / changes:

  • spawn within a parallel section has special semantics.
  • Every location of the form a[i] and a[i..j] and dest where dest is part of the pattern dest = spawn f(...) has to be provably disjoint. This is called the disjoint check.
  • Every other complex location loc that is used in a spawned proc (spawn f(loc)) has to be immutable for the duration of the parallel section. This is called the immutability check. Currently it is not specified what exactly "complex location" means. We need to make this an optimization!
  • Every array access has to be provably within bounds. This is called the bounds check.
  • Slices are optimized so that no copy is performed. This optimization is not yet performed for ordinary slices outside of a parallel section.

Guards and locks

Apart from spawn and parallel Nim also provides all the common low level concurrency mechanisms like locks, atomic intrinsics or condition variables.

Nim significantly improves on the safety of these features via additional pragmas:

  1. A guard annotation is introduced to prevent data races.
  2. Every access of a guarded memory location needs to happen in an appropriate locks statement.
  3. Locks and routines can be annotated with lock levels to allow potential deadlocks to be detected during semantic analysis.

Guards and the locks section

Protecting global variables

Object fields and global variables can be annotated via a guard pragma:

var glock: TLock
var gdata {.guard: glock.}: int

The compiler then ensures that every access of gdata is within a locks section:

proc invalid =
  # invalid: unguarded access:
  echo gdata

proc valid =
  # valid access:
  {.locks: [glock].}:
    echo gdata

Top level accesses to gdata are always allowed so that it can be initialized conveniently. It is assumed (but not enforced) that every top level statement is executed before any concurrent action happens.

The locks section deliberately looks ugly because it has no runtime semantics and should not be used directly! It should only be used in templates that also implement some form of locking at runtime:

template lock(a: TLock; body: untyped) =
  {.locks: [a].}:

The guard does not need to be of any particular type. It is flexible enough to model low level lockfree mechanisms:

var dummyLock {.compileTime.}: int
var atomicCounter {.guard: dummyLock.}: int

template atomicRead(x): untyped =
  {.locks: [dummyLock].}:

echo atomicRead(atomicCounter)

The locks pragma takes a list of lock expressions locks: [a, b, ...] in order to support multi lock statements. Why these are essential is explained in the lock levels section.

Protecting general locations

The guard annotation can also be used to protect fields within an object. The guard then needs to be another field within the same object or a global variable.

Since objects can reside on the heap or on the stack this greatly enhances the expressivity of the language:

  ProtectedCounter = object
    v {.guard: L.}: int
    L: TLock

proc incCounters(counters: var openArray[ProtectedCounter]) =
  for i in 0..counters.high:
    lock counters[i].L:
      inc counters[i].v

The access to field x.v is allowed since its guard x.L is active. After template expansion, this amounts to:

proc incCounters(counters: var openArray[ProtectedCounter]) =
  for i in 0..counters.high:
    {.locks: [counters[i].L].}:
        inc counters[i].v

There is an analysis that checks that counters[i].L is the lock that corresponds to the protected location counters[i].v. This analysis is called path analysis because it deals with paths to locations like obj.field[i].fieldB[j].

The path analysis is currently unsound, but that doesn't make it useless. Two paths are considered equivalent if they are syntactically the same.

This means the following compiles (for now) even though it really should not:

{.locks: [a[i].L].}:
  inc i
  access a[i].v

Lock levels

Lock levels are used to enforce a global locking order in order to detect potential deadlocks during semantic analysis. A lock level is an constant integer in the range 0..1_000. Lock level 0 means that no lock is acquired at all.

If a section of code holds a lock of level M than it can also acquire any lock of level N < M. Another lock of level M cannot be acquired. Locks of the same level can only be acquired at the same time within a single locks section:

var a, b: TLock[2]
var x: TLock[1]
# invalid locking order: TLock[1] cannot be acquired before TLock[2]:
{.locks: [x].}:
  {.locks: [a].}:
# valid locking order: TLock[2] acquired before TLock[1]:
{.locks: [a].}:
  {.locks: [x].}:

# invalid locking order: TLock[2] acquired before TLock[2]:
{.locks: [a].}:
  {.locks: [b].}:

# valid locking order, locks of the same level acquired at the same time:
{.locks: [a, b].}:

Here is how a typical multilock statement can be implemented in Nim. Note how the runtime check is required to ensure a global ordering for two locks a and b of the same lock level:

template multilock(a, b: ptr TLock; body: untyped) =
  if cast[ByteAddress](a) < cast[ByteAddress](b):
  {.locks: [a, b].}:

Whole routines can also be annotated with a locks pragma that takes a lock level. This then means that the routine may acquire locks of up to this level. This is essential so that procs can be called within a locks section:

proc p() {.locks: 3.} = discard

var a: TLock[4]
{.locks: [a].}:
  # p's locklevel (3) is strictly less than a's (4) so the call is allowed:

As usual locks is an inferred effect and there is a subtype relation: proc () {.locks: N.} is a subtype of proc () {.locks: M.} iff (M <= N).

The locks pragma can also take the special value "unknown". This is useful in the context of dynamic method dispatching. In the following example, the compiler can infer a lock level of 0 for the base case. However, one of the overloaded methods calls a procvar which is potentially locking. Thus, the lock level of calling g.testMethod cannot be inferred statically, leading to compiler warnings. By using {.locks: "unknown".}, the base method can be marked explicitly as having unknown lock level as well:

type SomeBase* = ref object of RootObj
type SomeDerived* = ref object of SomeBase
  memberProc*: proc ()

method testMethod(g: SomeBase) {.base, locks: "unknown".} = discard
method testMethod(g: SomeDerived) =
  if g.memberProc != nil:

noRewrite pragma

Term rewriting macros and templates are currently greedy and they will rewrite as long as there is a match. There was no way to ensure some rewrite happens only once, e.g. when rewriting term to same term plus extra content.

noRewrite pragma can actually prevent further rewriting on marked code, e.g. with given example echo("ab") will be rewritten just once:

template pwnEcho{echo(x)}(x: untyped) =
  {.noRewrite.}: echo("pwned!")

echo "ab"

noRewrite pragma can be useful to control term-rewriting macros recursion.

Aliasing restrictions in parameter passing

Note: The aliasing restrictions are currently not enforced by the implementation and need to be fleshed out further.

"Aliasing" here means that the underlying storage locations overlap in memory at runtime. An "output parameter" is a parameter of type var T, an input parameter is any parameter that is not of type var.

  1. Two output parameters should never be aliased.
  2. An input and an output parameter should not be aliased.
  3. An output parameter should never be aliased with a global or thread local variable referenced by the called proc.
  4. An input parameter should not be aliased with a global or thread local variable updated by the called proc.

One problem with rules 3 and 4 is that they affect specific global or thread local variables, but Nim's effect tracking only tracks "uses no global variable" via .noSideEffect. The rules 3 and 4 can also be approximated by a different rule:

  1. A global or thread local variable (or a location derived from such a location) can only passed to a parameter of a .noSideEffect proc.

Noalias annotation

Since version 1.4 of the Nim compiler, there is a .noalias annotation for variables and parameters. It is mapped directly to C/C++'s restrict keyword and means that the underlying pointer is pointing to a unique location in memory, no other aliases to this location exist. It is unchecked that this alias restriction is followed, if the restriction is violated, the backend optimizer is free to miscompile the code. This is an unsafe language feature.

Ideally in later versions of the language, the restriction will be enforced at compile time. (Which is also why the name noalias was choosen instead of a more verbose name like unsafeAssumeNoAlias.)

Strict funcs

Since version 1.4 a stricter definition of "side effect" is available. In addition to the existing rule that a side effect is calling a function with side effects the following rule is also enforced:

Any mutation to an object does count as a side effect if that object is reachable via a parameter that is not declared as a var parameter.

For example:

{.experimental: "strictFuncs".}

  Node = ref object
    le, ri: Node
    data: string

func len(n: Node): int =
  # valid: len does not have side effects
  var it = n
  while it != nil:
    inc result
    it = it.ri

func mut(n: Node) =
  let m = n # is the statement that connected the mutation to the parameter = "yeah" # the mutation is here
  # Error: 'mut' can have side effects
  # an object reachable from 'n' is potentially mutated

The algorithm behind this analysis is described in the view types section.

View types

Note: --experimental:views is more effective with --experimental:strictFuncs.

A view type is a type that is or contains one of the following types:

  • lent T (view into T)
  • openArray[T] (pair of (pointer to array of T, size))

For example:

  View1 = openArray[byte]
  View2 = lent string
  View3 = Table[openArray[char], int]

Exceptions to this rule are types constructed via ptr or proc. For example, the following types are not view types:

  NotView1 = proc (x: openArray[int])
  NotView2 = ptr openArray[char]
  NotView3 = ptr array[4, lent int]

The mutability aspect of a view type is not part of the type but part of the locations it's derived from. More on this later.

A view is a symbol (a let, var, const, etc.) that has a view type.

Since version 1.4 Nim allows view types to be used as local variables. This feature needs to be enabled via {.experimental: "views".}.

A local variable of a view type borrows from the locations and it is statically enforced that the view does not outlive the location it was borrowed from.

For example:

{.experimental: "views".}

proc take(a: openArray[int]) =
  echo a.len

proc main(s: seq[int]) =
  var x: openArray[int] = s # 'x' is a view into 's'
  # it is checked that 'x' does not outlive 's' and
  # that 's' is not mutated.
  for i in 0 .. high(x):
    echo x[i]
  take(x.toOpenArray(0, 1)) # slicing remains possible
  let y = x  # create a view from a view
  take y
  # it is checked that 'y' does not outlive 'x' and
  # that 'x' is not mutated as long as 'y' lives.

main(@[11, 22, 33])

A local variable of a view type can borrow from a location derived from a parameter, another local variable, a global const or let symbol or a thread-local var or let.

Let p the proc that is analysed for the correctness of the borrow operation.

Let source be one of:

  • A formal parameter of p. Note that this does not cover parameters of inner procs.
  • The result symbol of p.
  • A local var or let or const of p. Note that this does not cover locals of inner procs.
  • A thread-local var or let.
  • A global let or const.
  • A constant array/seq/object/tuple constructor.

Path expressions

A location derived from source is then defined as a path expression that has source as the owner. A path expression e is defined recursively:

  • source itself is a path expression.
  • Container access like e[i] is a path expression.
  • Tuple access e[0] is a path expression.
  • Object field access e.field is a path expression.
  • system.toOpenArray(e, ...) is a path expression.
  • Pointer dereference e[] is a path expression.
  • An address addr e, unsafeAddr e is a path expression.
  • A type conversion T(e) is a path expression.
  • A cast expression cast[T](e) is a path expression.
  • f(e, ...) is a path expression if f's return type is a view type. Because the view can only have been borrowed from e, we then know that owner of f(e, ...) is e.

If a view type is used as a return type, the location must borrow from a location that is derived from the first parameter that is passed to the proc. See for details about how this is done for var T.

A mutable view can borrow from a mutable location, an immutable view can borrow from both a mutable or an immutable location.

If a view borrows from a mutable location, the view can be used to update the location. Otherwise it cannot be used for mutations.

The duration of a borrow is the span of commands beginning from the assignment to the view and ending with the last usage of the view.

For the duration of the borrow operation, no mutations to the borrowed locations may be performed except via the view that borrowed from the location. The borrowed location is said to be sealed during the borrow.

{.experimental: "views".}

  Obj = object
    field: string

proc dangerous(s: var seq[Obj]) =
  let v: lent Obj = s[0] # seal 's'
  s.setLen 0  # prevented at compile-time because 's' is sealed.
  echo v.field

The scope of the view does not matter:

proc valid(s: var seq[Obj]) =
  let v: lent Obj = s[0]  # begin of borrow
  echo v.field            # end of borrow
  s.setLen 0  # valid because 'v' isn't used afterwards

The analysis requires as much precision about mutations as is reasonably obtainable, so it is more effective with the experimental strict funcs feature. In other words --experimental:views works better with --experimental:strictFuncs.

The analysis is currently control flow insensitive:

proc invalid(s: var seq[Obj]) =
  let v: lent Obj = s[0]
  if false:
    s.setLen 0
  echo v.field

In this example, the compiler assumes that s.setLen 0 invalidates the borrow operation of v even though a human being can easily see that it will never do that at runtime.

Start of a borrow

A borrow starts with one of the following:

  • The assignment of a non-view-type to a view-type.
  • The assignment of a location that is derived from a local parameter to a view-type.

End of a borrow

A borrow operation ends with the last usage of the view variable.


A view v can borrow from multiple different locations. However, the borrow is always the full span of v's lifetime and every location that is borrowed from is sealed during v's lifetime.


The following section is an outline of the algorithm that the current implementation uses. The algorithm performs two traversals over the AST of the procedure or global section of code that uses a view variable. No fixpoint iterations are performed, the complexity of the analysis is O(N) where N is the number of nodes of the AST.

The first pass over the AST computes the lifetime of each local variable based on a notion of an "abstract time", in the implementation it's a simple integer that is incremented for every visited node.

In the second pass information about the underlying object "graphs" is computed. Let v be a parameter or a local variable. Let G(v) be the graph that v belongs to. A graph is defined by the set of variables that belong to the graph. Initially for all v: G(v) = {v}. Every variable can only be part of a single graph.

Assignments like a = b "connect" two variables, both variables end up in the same graph {a, b} = G(a) = G(b). Unfortunately, the pattern to look for is much more complex than that and can involve multiple assignment targets and sources:

f(x, y) = g(a, b)

connects x and y to a and b: G(x) = G(y) = G(a) = G(b) = {x, y, a, b}. A type based alias analysis rules out some of these combinations, for example a string value cannot possibly be connected to a seq[int].

A pattern like v[] = value or v.field = value marks G(v) as mutated. After the second pass a set of disjoint graphs was computed.

For strict functions it is then enforced that there is no graph that is both mutated and has an element that is an immutable parameter (that is a parameter that is not of type var T).

For borrow checking a different set of checks is performed. Let v be the view and b the location that is borrowed from.

  • The lifetime of v must not exceed b's lifetime. Note: The lifetime of a parameter is the complete proc body.
  • If v is used for a mutation, b must be a mutable location too.
  • During v's lifetime, G(b) can only be modified by v (and only if v is a mutable view).
  • If v is result then b has to be a location derived from the first formal parameter or from a constant location.
  • A view cannot be used for a read or a write access before it was assigned to.