======================
Nim Tutorial (Part II)
======================
:Author: Andreas Rumpf
:Version: |nimversion|
.. contents::
Introduction
============
"Repetition renders the ridiculous reasonable." -- Norman Wildberger
This document is a tutorial for the advanced constructs of the *Nim*
programming language. **Note that this document is somewhat obsolete as the**
`manual <manual.html>`_ **contains many more examples of the advanced language
features.**
Pragmas
=======
Pragmas are Nim's method to give the compiler additional information/
commands without introducing a massive number of new keywords. Pragmas are
enclosed in the special ``{.`` and ``.}`` curly dot brackets. This tutorial
does not cover pragmas. See the `manual <manual.html#pragmas>`_ or `user guide
<nimc.html#additional-features>`_ for a description of the available
pragmas.
Object Oriented Programming
===========================
While Nim's support for object oriented programming (OOP) is minimalistic,
powerful OOP techniques can be used. OOP is seen as *one* way to design a
program, not *the only* way. Often a procedural approach leads to simpler
and more efficient code. In particular, preferring composition over inheritance
is often the better design.
Inheritance
-----------
Inheritance in Nim is entirely optional. To enable inheritance with
runtime type information the object needs to inherit from
``RootObj``. This can be done directly, or indirectly by
inheriting from an object that inherits from ``RootObj``. Usually
types with inheritance are also marked as ``ref`` types even though
this isn't strictly enforced. To check at runtime if an object is of a certain
type, the ``of`` operator can be used.
.. code-block:: nim
:test: "nim c $1"
type
Person = ref object of RootObj
name*: string # the * means that `name` is accessible from other modules
age: int # no * means that the field is hidden from other modules
Student = ref object of Person # Student inherits from Person
id: int # with an id field
var
student: Student
person: Person
assert(student of Student) # is true
# object construction:
student = Student(name: "Anton", age: 5, id: 2)
echo student[]
Inheritance is done with the ``object of`` syntax. Multiple inheritance is
currently not supported. If an object type has no suitable ancestor, ``RootObj``
can be used as its ancestor, but this is only a convention. Objects that have
no ancestor are implicitly ``final``. You can use the ``inheritable`` pragma
to introduce new object roots apart from ``system.RootObj``. (This is used
in the GTK wrapper for instance.)
Ref objects should be used whenever inheritance is used. It isn't strictly
necessary, but with non-ref objects assignments such as ``let person: Person =
Student(id: 123)`` will truncate subclass fields.
**Note**: Composition (*has-a* relation) is often preferable to inheritance
(*is-a* relation) for simple code reuse. Since objects are value types in
Nim, composition is as efficient as inheritance.
Mutually recursive types
------------------------
Objects, tuples and references can model quite complex data structures which
depend on each other; they are *mutually recursive*. In Nim
these types can only be declared within a single type section. (Anything else
would require arbitrary symbol lookahead which slows down compilation.)
Example:
.. code-block:: nim
:test: "nim c $1"
type
Node = ref object # a reference to an object with the following field:
le, ri: Node # left and right subtrees
sym: ref Sym # leaves contain a reference to a Sym
Sym = object # a symbol
name: string # the symbol's name
line: int # the line the symbol was declared in
code: Node # the symbol's abstract syntax tree
Type conversions
----------------
Nim distinguishes between `type casts`:idx: and `type conversions`:idx:.
Casts are done with the ``cast`` operator and force the compiler to
interpret a bit pattern to be of another type.
Type conversions are a much more polite way to convert a type into another:
They preserve the abstract *value*, not necessarily the *bit-pattern*. If a
type conversion is not possible, the compiler complains or an exception is
raised.
The syntax for type conversions is ``destination_type(expression_to_convert)``
(like an ordinary call):
.. code-block:: nim
proc getID(x: Person): int =
Student(x).id
The ``InvalidObjectConversionDefect`` exception is raised if ``x`` is not a
``Student``.
Object variants
---------------
Often an object hierarchy is overkill in certain situations where simple
variant types are needed.
An example:
.. code-block:: nim
:test: "nim c $1"
# This is an example how an abstract syntax tree could be modelled in Nim
type
NodeKind = enum # the different node types
nkInt, # a leaf with an integer value
nkFloat, # a leaf with a float value
nkString, # a leaf with a string value
nkAdd, # an addition
nkSub, # a subtraction
nkIf # an if statement
Node = ref object
case kind: NodeKind # the ``kind`` field is the discriminator
of nkInt: intVal: int
of nkFloat: floatVal: float
of nkString: strVal: string
of nkAdd, nkSub:
leftOp, rightOp: Node
of nkIf:
condition, thenPart, elsePart: Node
var n = Node(kind: nkFloat, floatVal: 1.0)
# the following statement raises an `FieldDefect` exception, because
# n.kind's value does not fit:
n.strVal = ""
As can been seen from the example, an advantage to an object hierarchy is that
no conversion between different object types is needed. Yet, access to invalid
object fields raises an exception.
Method call syntax
------------------
There is a syntactic sugar for calling routines:
The syntax ``obj.method(args)`` can be used instead of ``method(obj, args)``.
If there are no remaining arguments, the parentheses can be omitted:
``obj.len`` (instead of ``len(obj)``).
This method call syntax is not restricted to objects, it can be used
for any type:
.. code-block:: nim
:test: "nim c $1"
import std/strutils
echo "abc".len # is the same as echo len("abc")
echo "abc".toUpperAscii()
echo({'a', 'b', 'c'}.card)
stdout.writeLine("Hallo") # the same as writeLine(stdout, "Hallo")
(Another way to look at the method call syntax is that it provides the missing
postfix notation.)
So "pure object oriented" code is easy to write:
.. code-block:: nim
:test: "nim c $1"
import std/[strutils, sequtils]
stdout.writeLine("Give a list of numbers (separated by spaces): ")
stdout.write(stdin.readLine.splitWhitespace.map(parseInt).max.`$`)
stdout.writeLine(" is the maximum!")
Properties
----------
As the above example shows, Nim has no need for *get-properties*:
Ordinary get-procedures that are called with the *method call syntax* achieve
the same. But setting a value is different; for this a special setter syntax
is needed:
.. code-block:: nim
:test: "nim c $1"
type
Socket* = ref object of RootObj
h: int # cannot be accessed from the outside of the module due to missing star
proc `host=`*(s: var Socket, value: int) {.inline.} =
## setter of host address
s.h = value
proc host*(s: Socket): int {.inline.} =
## getter of host address
s.h
var s: Socket
new s
s.host = 34 # same as `host=`(s, 34)
(The example also shows ``inline`` procedures.)
The ``[]`` array access operator can be overloaded to provide
`array properties`:idx:\ :
.. code-block:: nim
:test: "nim c $1"
type
Vector* = object
x, y, z: float
proc `[]=`* (v: var Vector, i: int, value: float) =
# setter
case i
of 0: v.x = value
of 1: v.y = value
of 2: v.z = value
else: assert(false)
proc `[]`* (v: Vector, i: int): float =
# getter
case i
of 0: result = v.x
of 1: result = v.y
of 2: result = v.z
else: assert(false)
The example is silly, since a vector is better modelled by a tuple which
already provides ``v[]`` access.
Dynamic dispatch
----------------
Procedures always use static dispatch. For dynamic dispatch replace the
``proc`` keyword by ``method``:
.. code-block:: nim
:test: "nim c $1"
type
Expression = ref object of RootObj ## abstract base class for an expression
Literal = ref object of Expression
x: int
PlusExpr = ref object of Expression
a, b: Expression
# watch out: 'eval' relies on dynamic binding
method eval(e: Expression): int {.base.} =
# override this base method
quit "to override!"
method eval(e: Literal): int = e.x
method eval(e: PlusExpr): int = eval(e.a) + eval(e.b)
proc newLit(x: int): Literal = Literal(x: x)
proc newPlus(a, b: Expression): PlusExpr = PlusExpr(a: a, b: b)
echo eval(newPlus(newPlus(newLit(1), newLit(2)), newLit(4)))
Note that in the example the constructors ``newLit`` and ``newPlus`` are procs
because it makes more sense for them to use static binding, but ``eval`` is a
method because it requires dynamic binding.
**Note:** Starting from Nim 0.20, to use multi-methods one must explicitly pass
``--multimethods:on`` when compiling.
In a multi-method all parameters that have an object type are used for the
dispatching:
.. code-block:: nim
:test: "nim c --multiMethods:on $1"
type
Thing = ref object of RootObj
Unit = ref object of Thing
x: int
method collide(a, b: Thing) {.inline.} =
quit "to override!"
method collide(a: Thing, b: Unit) {.inline.} =
echo "1"
method collide(a: Unit, b: Thing) {.inline.} =
echo "2"
var a, b: Unit
new a
new b
collide(a, b) # output: 2
As the example demonstrates, invocation of a multi-method cannot be ambiguous:
Collide 2 is preferred over collide 1 because the resolution works from left to
right. Thus ``Unit, Thing`` is preferred over ``Thing, Unit``.
**Performance note**: Nim does not produce a virtual method table, but
generates dispatch trees. This avoids the expensive indirect branch for method
calls and enables inlining. However, other optimizations like compile time
evaluation or dead code elimination do not work with methods.
Exceptions
==========
In Nim exceptions are objects. By convention, exception types are
suffixed with 'Error'. The `system <system.html>`_ module defines an
exception hierarchy that you might want to stick to. Exceptions derive from
``system.Exception``, which provides the common interface.
Exceptions have to be allocated on the heap because their lifetime is unknown.
The compiler will prevent you from raising an exception created on the stack.
All raised exceptions should at least specify the reason for being raised in
the ``msg`` field.
A convention is that exceptions should be raised in *exceptional* cases,
they should not be used as an alternative method of control flow.
Raise statement
---------------
Raising an exception is done with the ``raise`` statement:
.. code-block:: nim
:test: "nim c $1"
var
e: ref OSError
new(e)
e.msg = "the request to the OS failed"
raise e
If the ``raise`` keyword is not followed by an expression, the last exception
is *re-raised*. For the purpose of avoiding repeating this common code pattern,
the template ``newException`` in the ``system`` module can be used:
.. code-block:: nim
raise newException(OSError, "the request to the OS failed")
Try statement
-------------
The ``try`` statement handles exceptions:
.. code-block:: nim
:test: "nim c $1"
from std/strutils import parseInt
# read the first two lines of a text file that should contain numbers
# and tries to add them
var
f: File
if open(f, "numbers.txt"):
try:
let a = readLine(f)
let b = readLine(f)
echo "sum: ", parseInt(a) + parseInt(b)
except OverflowDefect:
echo "overflow!"
except ValueError:
echo "could not convert string to integer"
except IOError:
echo "IO error!"
except:
echo "Unknown exception!"
# reraise the unknown exception:
raise
finally:
close(f)
The statements after the ``try`` are executed unless an exception is
raised. Then the appropriate ``except`` part is executed.
The empty ``except`` part is executed if there is an exception that is
not explicitly listed. It is similar to an ``else`` part in ``if``
statements.
If there is a ``finally`` part, it is always executed after the
exception handlers.
The exception is *consumed* in an ``except`` part. If an exception is not
handled, it is propagated through the call stack. This means that often
the rest of the procedure - that is not within a ``finally`` clause -
is not executed (if an exception occurs).
If you need to *access* the actual exception object or message inside an
``except`` branch you can use the `getCurrentException()
<system.html#getCurrentException>`_ and `getCurrentExceptionMsg()
<system.html#getCurrentExceptionMsg>`_ procs from the `system <system.html>`_
module. Example:
.. code-block:: nim
try:
doSomethingHere()
except:
let
e = getCurrentException()
msg = getCurrentExceptionMsg()
echo "Got exception ", repr(e), " with message ", msg
Annotating procs with raised exceptions
---------------------------------------
Through the use of the optional ``{.raises.}`` pragma you can specify that a
proc is meant to raise a specific set of exceptions, or none at all. If the
``{.raises.}`` pragma is used, the compiler will verify that this is true. For
instance, if you specify that a proc raises ``IOError``, and at some point it
(or one of the procs it calls) starts raising a new exception the compiler will
prevent that proc from compiling. Usage example:
.. code-block:: nim
proc complexProc() {.raises: [IOError, ArithmeticDefect].} =
...
proc simpleProc() {.raises: [].} =
...
Once you have code like this in place, if the list of raised exception changes
the compiler will stop with an error specifying the line of the proc which
stopped validating the pragma and the raised exception not being caught, along
with the file and line where the uncaught exception is being raised, which may
help you locate the offending code which has changed.
If you want to add the ``{.raises.}`` pragma to existing code, the compiler can
also help you. You can add the ``{.effects.}`` pragma statement to your proc and
the compiler will output all inferred effects up to that point (exception
tracking is part of Nim's effect system). Another more roundabout way to
find out the list of exceptions raised by a proc is to use the Nim ``doc``
command which generates documentation for a whole module and decorates all
procs with the list of raised exceptions. You can read more about Nim's
`effect system and related pragmas in the manual <manual.html#effect-system>`_.
Generics
========
Generics are Nim's means to parametrize procs, iterators or types
with `type parameters`:idx:. Generic parameters are written within square
brackets, for example ``Foo[T]``. They are most useful for efficient type safe
containers:
.. code-block:: nim
:test: "nim c $1"
type
BinaryTree*[T] = ref object # BinaryTree is a generic type with
# generic param ``T``
le, ri: BinaryTree[T] # left and right subtrees; may be nil
data: T # the data stored in a node
proc newNode*[T](data: T): BinaryTree[T] =
# constructor for a node
new(result)
result.data = data
proc add*[T](root: var BinaryTree[T], n: BinaryTree[T]) =
# insert a node into the tree
if root == nil:
root = n
else:
var it = root
while it != nil:
# compare the data items; uses the generic ``cmp`` proc
# that works for any type that has a ``==`` and ``<`` operator
var c = cmp(it.data, n.data)
if c < 0:
if it.le == nil:
it.le = n
return
it = it.le
else:
if it.ri == nil:
it.ri = n
return
it = it.ri
proc add*[T](root: var BinaryTree[T], data: T) =
# convenience proc:
add(root, newNode(data))
iterator preorder*[T](root: BinaryTree[T]): T =
# Preorder traversal of a binary tree.
# This uses an explicit stack (which is more efficient than
# a recursive iterator factory).
var stack: seq[BinaryTree[T]] = @[root]
while stack.len > 0:
var n = stack.pop()
while n != nil:
yield n.data
add(stack, n.ri) # push right subtree onto the stack
n = n.le # and follow the left pointer
var
root: BinaryTree[string] # instantiate a BinaryTree with ``string``
add(root, newNode("hello")) # instantiates ``newNode`` and ``add``
add(root, "world") # instantiates the second ``add`` proc
for str in preorder(root):
stdout.writeLine(str)
The example shows a generic binary tree. Depending on context, the brackets are
used either to introduce type parameters or to instantiate a generic proc,
iterator or type. As the example shows, generics work with overloading: the
best match of ``add`` is used. The built-in ``add`` procedure for sequences
is not hidden and is used in the ``preorder`` iterator.
There is a special ``[:T]`` syntax when using generics with the method call syntax:
.. code-block:: nim
:test: "nim c $1"
proc foo[T](i: T) =
discard
var i: int
# i.foo[int]() # Error: expression 'foo(i)' has no type (or is ambiguous)
i.foo[:int]() # Success
Templates
=========
Templates are a simple substitution mechanism that operates on Nim's
abstract syntax trees. Templates are processed in the semantic pass of the
compiler. They integrate well with the rest of the language and share none
of C's preprocessor macros flaws.
To *invoke* a template, call it like a procedure.
Example:
.. code-block:: nim
template `!=` (a, b: untyped): untyped =
# this definition exists in the System module
not (a == b)
assert(5 != 6) # the compiler rewrites that to: assert(not (5 == 6))
The ``!=``, ``>``, ``>=``, ``in``, ``notin``, ``isnot`` operators are in fact
templates: this has the benefit that if you overload the ``==`` operator,
the ``!=`` operator is available automatically and does the right thing. (Except
for IEEE floating point numbers - NaN breaks basic boolean logic.)
``a > b`` is transformed into ``b < a``.
``a in b`` is transformed into ``contains(b, a)``.
``notin`` and ``isnot`` have the obvious meanings.
Templates are especially useful for lazy evaluation purposes. Consider a
simple proc for logging:
.. code-block:: nim
:test: "nim c $1"
const
debug = true
proc log(msg: string) {.inline.} =
if debug: stdout.writeLine(msg)
var
x = 4
log("x has the value: " & $x)
This code has a shortcoming: if ``debug`` is set to false someday, the quite
expensive ``$`` and ``&`` operations are still performed! (The argument
evaluation for procedures is *eager*).
Turning the ``log`` proc into a template solves this problem:
.. code-block:: nim
:test: "nim c $1"
const
debug = true
template log(msg: string) =
if debug: stdout.writeLine(msg)
var
x = 4
log("x has the value: " & $x)
The parameters' types can be ordinary types or the meta types ``untyped``,
``typed``, or ``type``. ``type`` suggests that only a type symbol may be given
as an argument, and ``untyped`` means symbol lookups and type resolution is not
performed before the expression is passed to the template.
If the template has no explicit return type,
``void`` is used for consistency with procs and methods.
To pass a block of statements to a template, use ``untyped`` for the last parameter:
.. code-block:: nim
:test: "nim c $1"
template withFile(f: untyped, filename: string, mode: FileMode,
body: untyped) =
let fn = filename
var f: File
if open(f, fn, mode):
try:
body
finally:
close(f)
else:
quit("cannot open: " & fn)
withFile(txt, "ttempl3.txt", fmWrite):
txt.writeLine("line 1")
txt.writeLine("line 2")
In the example the two ``writeLine`` statements are bound to the ``body``
parameter. The ``withFile`` template contains boilerplate code and helps to
avoid a common bug: to forget to close the file. Note how the
``let fn = filename`` statement ensures that ``filename`` is evaluated only
once.
Example: Lifting Procs
----------------------
.. code-block:: nim
:test: "nim c $1"
import std/math
template liftScalarProc(fname) =
## Lift a proc taking one scalar parameter and returning a
## scalar value (eg ``proc sssss[T](x: T): float``),
## to provide templated procs that can handle a single
## parameter of seq[T] or nested seq[seq[]] or the same type
##
## .. code-block:: Nim
## liftScalarProc(abs)
## # now abs(@[@[1,-2], @[-2,-3]]) == @[@[1,2], @[2,3]]
proc fname[T](x: openarray[T]): auto =
var temp: T
type outType = typeof(fname(temp))
result = newSeq[outType](x.len)
for i in 0..<x.len:
result[i] = fname(x[i])
liftScalarProc(sqrt) # make sqrt() work for sequences
echo sqrt(@[4.0, 16.0, 25.0, 36.0]) # => @[2.0, 4.0, 5.0, 6.0]
Compilation to JavaScript
=========================
Nim code can be compiled to JavaScript. However in order to write
JavaScript-compatible code you should remember the following:
- ``addr`` and ``ptr`` have slightly different semantic meaning in JavaScript.
It is recommended to avoid those if you're not sure how they are translated
to JavaScript.
- ``cast[T](x)`` in JavaScript is translated to ``(x)``, except for casting
between signed/unsigned ints, in which case it behaves as static cast in
C language.
- ``cstring`` in JavaScript means JavaScript string. It is a good practice to
use ``cstring`` only when it is semantically appropriate. E.g. don't use
``cstring`` as a binary data buffer.
Part 3
======
The next part is entirely about metaprogramming via macros: `Part III <tut3.html>`_