Perl is polymorphism worth it? - perl

There seem to be a number of ways to do polymorphism in Perl but they all feel "hacky" to me. I'm somewhat new to Perl so I could just be interpreting it wrong but I find most examples make your code too illegible and are controversial.
Say you have a widget class that contains some data and methods that all widgets need. There are also a handful of widget types (i.e. calender, schedule, etc). And they need to communicate with each other (via a parent container class).
Would playing with the namespace and making widget a prototype be worth it?
Should I give each widget a reference to an object (one of the types) at instantiation?
Just forget about types being objects and make widget a large class where only a few methods get used per instance based on what type is set.
Something else?
I come from a C/C++ background and I'm finding it difficult to decide on a Perl programming model.
Also, I'm don't have strong type safety or private member requirements. The project is a mid sized web app with a couple developers and portability to other projects isn't a priority. Easy extensibilty without needing to decipher Perl hacks would be useful though.

The "Modern Perl" approach is probably to define Widget as a role. A role can be thought of as similar to a mixin, an interface, or an abstract base class. Do this with Moose::Role or one of its more light-weight alternatives (Moo::Role, Role::Tiny).
{
package Widget;
use Moo::Role;
sub some_common_method {
my $self = shift;
...;
}
sub another_common_method {
my $self = shift;
...;
}
# Here we're indicating that all Widgets must
# have a method called yet_another_common_method,
# but we're not defining how that method should
# be implemented.
requires "yet_another_common_method";
}
Now you can create a class that composes that role:
{
package Calendar;
use Moo;
with "Widget";
# Here's a member variable.
has year => (is => "ro", required => 1);
# Widget requires us to implement this.
sub yet_another_common_method {
my $self = shift;
...;
}
# We can override Widget's implementation
# of this method.
sub some_common_method {
my $self = shift;
...;
}
# We can install hooks ("method modifiers")
# to Widget's methods.
before another_common_method => sub {
my $self = shift;
print STDERR "Calendar ", $self->year, ": another_common_method() was called.\n";
};
}
And another:
{
package Schedule;
use Moo;
with "Widget", "Editable";
sub yet_another_common_method {
my $self = shift;
...;
}
}
And use the classes:
my $calendar = Calendar->new( year => 2014 );
my $schedule = Schedule->new;
my #widgets = ($calendar, $schedule);
for (#widgets) {
$_->some_common_method if $_->does('Widget');
}

Related

Return a base-class object from a derived-class object

I'm aware that Perl is not statically typed when I want to apply this mechanism to a Perl object of a derived class:
Say I have a base class B and a derived class D inheriting from B.
Also I have an object $obj that holds a D object.
A function Bf() is expecting a parameter of type B.
Obviously (by the rules of polymorphism) I can pass $obj to Bf() like Bf($obj), but unlike to a static-typed language Bf() will see the whole D object (and not just the elements of B).
Is there a (rather clean and simple) solution for this problem in Perl? The solution should "hide" the attributes (and methods) a B does not have from D in Bf(), not restricting modifications of the original B (which is D actually).
Adult Programmers only (added 2020-03-06)
OK, people wanted a more concrete description.
Unfortunately (as pointed out) the original program is highly complex and uses reflection-like mechanisms to generate getters, setters and formatters automatically, to I really can't give a minimum working example here, because it would not be minimal.
First I have a class MessageHandler that handle messages (no surprise!).
Then I have a function log_message($$$) that expects (among others) a MessageHandler object as first argument.
Then I have this hierarchy of classes (it's much more complex in reality):
MessageHandler
ControlMessageHandler (ISA: MessageHandler)
ControlMessageResponseHandler (ISA: ControlMessageHandler)
Now if log_message wants a MessageHandler I can pass a ControlMessageResponseHandler as it conforms to MessageHandler.
But doing so exposes all the attributes of ControlMessageResponseHandler to log_message that are non-existent in MessageHandler.
The danger is that log_message might (by mistake) access an attribute of ControlMessageResponseHandler that is not present in MessageHandler. To prevent errors I'd like to prevent that, or at least get some warning (like I would get in a statically-typed language as Eiffel).
Dirty Details inside
Just in case it matters, I'll sketch how my array objects are built (a lot of extra code would be needed for a working example):
First the array indices are allocated automatically like this:
use constant I_VERBOSITY => IS_NEXT->(); # verbosity level
use constant I_TAG => IS_NEXT->(); # additional tag
use constant I_TAG_STACK => IS_NEXT->(); # tag stack
use constant I_MSG_DEBUG => IS_NEXT->(); # handler for debug messages
...
use constant I_LAST => IS_LAST->(); # last index (must be last)
I_LAST is needed for inheritance.
The attributes are defines like this:
use constant ATTRIBUTES => (
['verbosity', I_VERBOSITY, undef],
['tag', I_TAG, \&Class::_format_string],
['tag_stack', I_TAG_STACK, undef],
['msg_debug', I_MSG_DEBUG, \&Class::_format_code],
...
);
The definition contains a hint how to format each attribute.
This information is used to set up formatters to format each attribute like this:
use constant FORMATTERS =>
(map { Class::_attribute_string($_->[0], $_->[1], undef, $_->[2]) }
ATTRIBUTES); # attribute formatters
Getters and setters are automatically defined like this:
BEGIN {
foreach (ATTRIBUTES) {
Class::_assign_gs_ai(__PACKAGE__, $_->[0], $_->[1]);
}
}
The constructor would use the following lines:
my $self = [];
$#$self = I_LAST;
$self->[I_VERBOSITY] = $verbosity;
...
And finally my object print routine goes like this:
sub as_string($)
{
my $self = shift;
my $a_sep = ', ';
return join($a_sep, map { $_->($self, $a_sep) } FORMATTERS);
}
With inheritance it looks like this:
sub as_string($)
{
my $self = shift;
my $a_sep = ', ';
return join($a_sep, $self->SUPER::as_string(),
map { $_->($self, $a_sep) } FORMATTERS);
}
I'm not sure what your problem is, although I think you took the long way to say "I have a function that expects a B object, and I want to pass it a D object."
If you only want objects of a certain exact type, don't accept anything else:
use Carp qw(croak);
sub Bf {
croak "Bad object! I only like B" unless ref $_[0] eq 'B';
...
}
But, that's a bad idea. A derived class should be just as good as the base class. The clean solution is to not care what type you get.
sub Bf {
croak "Bad object! Doesn't respond to foo!" unless $_[0]->can('foo');
...
}
Since this Bf method works with the base class, why would it look for something in some derived class it didn't know about? If the derived class has changed the interface and no longer acts like its parent, then maybe it's isn't a good fit for inheritance. There are many problems like this that are solved by a different architecture.
I think you'll have to come up with a concrete example where the derived class wouldn't work.
It sounds like for some reason you need your D object to behave like a B object, but at the same time not like a D object. As the existing answers and comments indicate, it's a very common to use a sub-class where the base class is expected, and most algorithms shouldn't care whether what you actually passed is D or B. The only reason I can think of why you would want otherwise is that D overrides (redefines) some methods in an incompatible way, and you want the methods from B instead.
package Dog;
sub new {
my ($class, %args) = #_;
return bless \%args, $class;
}
sub bark { print "Bark!\n"; }
package Dingo;
use parent 'Dog';
sub bark { print "...\n"; }
package main;
my $dingo = Dingo->new;
$dingo->bark; # "..."
(n.b., I've left off the recommended use strict; and use warnings; for terseness, they should be used in all packages)
You may be aware from reading perldoc perlootut and perldoc perlobj that an object in Perl is just a blessed reference of some sort; in the example above, we use a hash reference. If you are trying to get the "attributes" that only exist in B, I think you would have to write some sort of translation method. But, if you care about the methods that exist in B, all you have to do is re-bless it into the parent class.
my $dingo = Dingo->new;
$dingo->bark; # "..."
bless $dingo, "Dog";
$dingo->bark; # "Bark!"
Note that bless does not return a new reference, but modifies that reference in-place; if you want it to behave like a Dingo again, you have to bless it back.
Perhaps more conveniently you can define a method to create a copy for you and bless it into the appropriate class:
package Dog;
sub as_dog {
my ($self) = #_;
# The {} below create a shallow copy, i.e., a new reference
return bless { %{$self} }, __PACKAGE__;
}
#...
package main;
my $dingo = Dingo->new;
$dingo->bark; # ...
$dingo->as_dog->bark; # Bark!
$dingo->bark; # ...
While there doesn't seem to be a perfect solution, temporary "re-blessing" the object seems to get quite close to what is asked for:
sub Bf($) # expects a "B" object (or descendant of "B" (like "D"))
{
my $B = shift;
my $type = ref($B); # save original type
die "unexpected type $type" unless ($B->isa('B'));
bless $B, 'B'; # restrict to "B"'s features
$B->whatever(...);
#...
bless $B, $type; # restore original type
}

Perl Moose augment vs around

Trying to understand Moose:
use Modern::Perl;
package FOO {
use Moose;
sub rep { say " <report></report>"; }
sub doc {
say "<document>";
inner();
say "</document>";
}
}
package BAR {
use Moose;
extends 'FOO';
around 'rep' => sub {
my $orig = shift;
my $self = shift;
say "<document>";
$self->$orig(#_);
say "</document>";
};
augment 'doc' => sub {
say " <report></report>";
};
}
package main {
BAR->new->rep;
say "===";
BAR->new->doc;
}
Produces...
<document>
<report></report>
</document>
===
<document>
<report></report>
</document>
... the same result. When desinging the "model (object hierarchy)" - based on what I should decide when to use around and when augment?
Here are probably other (deeper) things what i currently didn't understand yet.
Can please someone provide an "more deep" explanation, because reading tru the Moose/Manual/MethodModifiers obviously not helped enough...
augment and around do rather different things. augment is designed to make this sort of pattern easier:
package Document {
use Moose;
sub make_document {
my $self = shift;
return "<doc>" . $self->_document_innards . "</doc>"
}
# stub; override in child class
sub _document_innards {
my $self = shift;
return "";
}
}
package Invoice {
use Moose;
extends 'Document';
sub _document_innards {
my $self = shift;
return "Give me money!";
}
}
With augment it becomes:
package Document {
use Moose;
sub make_document {
my $self = shift;
return "<doc>" . inner() . "</doc>"
}
}
package Invoice {
use Moose;
extends 'Document';
augment make_document => sub {
my $self = shift;
return "Give me money!";
};
}
On the other hand, around is used as a replacement for doing $self->SUPER::method(#args) because SUPER can't work in roles (the notion of which package to check superclasses for is bound at compile-time, so $self->SUPER::method(#args) would check superclasses of the role (i.e. none) instead of superclasses of the class that consumed the role. If you're not using roles, then SUPER can still be used in Moose classes just fine. TLDR: SUPER is broken by roles, so Moose gives you around as an alternative.
Another thing to compare is override which is a bit like around, but gives you this super() function which is perhaps slightly cleaner than $self->$orig(#_). It also has an "there can be only one" feature. If two roles try to provide an around modifier for the same method, that's fine: they both get to wrap the method (though the order in which they are applied is undefined). If two roles try to provide an override modifier, that's an error.
The implementation of augment is somewhat fragile in my experience, so that in my book is a reason to avoid it. Don't try to replace it with around, because they do rather different things. Instead, replace it with the pattern used in my first example above.
Using around should always be your first instinct. As (Moose creator) Stevan Little says about augment:
Thankfully, only a small percentage of people actually grok this
feature and of those people only a handful of them are crazy enough to
try and use it.

How to handle this situation in object oriented perl

package a::b::c:d
my $res = a::b::c:e->new(); # i am doing like this
# is there any othere to do this
sub new {
...
my $self = {
#result = a::b::c:e->new();
}
}
sub test {
}
sub test2 {
}
1;
package a::b::c:e
sub new {
...
}
sub testresult {
}
1;
My question is:
how to initalize the e module in d in new itself rather creating every function and
how to use that to store some results into e:testresult
There are two strategies -- either examine the symbol table to initialize on creation, or make use of AUTOLOAD and test with can. AUTOLOAD can be messier as you have to deal with the case where the method isn't there:
sub AUTOLOAD {
my $self = shift;
my $method = $AUTOLOAD;
$method =~ s/.*://; # strip package name
if ( $self->{'result'}->can($method) ) {
return $self->{'result'}->$method(#_);
} else {
croak "Unknown method : $method";
}
}
But the symbol table trick is brittle, as if they're using inheritance, you won't see the inherited methods without walking up #ISA, too. (and even if they're not -- they might start using inheritance in the future, which results in things breaking)
...
Typically, when you're trying to copy another module's interface, you've got a case of inheritance, so you might want to ask yourself what the relationship is between ::d and ::e :
a::b::c::d is an a::b::c::e
a::b::c::d uses a::b::c::e
If it's an is-a relationship, it's typically better suited to inheritance (although you might have wrappers around each of the methods, and still need to go through this whole exercise anyway). If it's a uses relationship, odds are you don't want to inherit from every last method they have, and can just hard code a list (although, the list might change if the used class is updated)
foreach my $method ( #list_of_methods_to_copy ) {
*{$method} = sub {
my $self = shift;
return $self->{'results'}->$method(#_);
}
}
Provided that you're not looking for inheritance between the two classes, it looks to me that you might want to use Class::Delegator for composing your class. That way, you can create the delegation routines by putting the following in a::b::c::d.
use Class::Delegator send => 'testresult', to => '{result}';
But you'd need to fix your constructor anyway:
my $self
= bless {
result => a::b::c::e->new()
}, $class_name
;
return $self;
Having done that, you'd have a field '{result}' to delegate to.

Method not found error when inheriting abstract method in Perl OOP

I have a subclass that calls a method from a superclass. The method in the superclass uses a method that is defined in the superclass as abstract (not really abstract) but implemented in the subclass.
For example:
package BaseClass;
sub new
{
}
sub method1 {
return someAbstractMethod();
}
sub someAbtsractMethod
{
die "oops, this is an abstract method that should " .
"be implemented in a subclass" ;
}
1;
package SubClass;
sub new
{
}
sub someAbtsractMethod
{
print "now we implement the asbtract method";
}
1;
Now when I do:
$sub = new SubClass();
$sub->method1();
...it calls the abstract message and I get the specified error message. If I took off the abstract method from the super class and just leave the implementation in the subclass, It does not recognize the method and I get subroutine abstract method not found error.
You haven't set up an IS_A relationship between the parent and child classes.
You can do this with the base pragma as Ivan suggests, or you can manipulate the #ISA array. Or you can use the parent pragma.
#ISA:
package SubClass;
our #ISA = qw( BaseClass );
parent:
package SubClass;
use parent qw( BaseClass );
By the way, don't use the indirect object syntax ever. To call your constructor do:
my $foo = SubClass->new();
Also, it looks like you aren't using the strict and warnings pragmas. Do so. Always.
Finally, if you have multiple packages in one file, it is helpful to enclose each package in a block.
Check out perlboot and perltoot, they are handy OOP tutorials in the perldoc.
Update:
I just noticed that your method calls are broken. You need to find the invoking class or instance in each method.
package BaseClass;
sub new { bless {}, shift; } # horrible constructor - do not reuse.
sub abstract { die "The present situation is abstract"; }
sub method { my $self = shift; $self->abstract; }
package SubClass;
our #ISA = qw( BaseClass );
sub abstract { print "It's alive\n" );
In the script:
my $obj = SubClass->new;
$obj->method;
my $base = BaseClass->new;
$base->method;
Definitely read the tutorials I linked to. They will help you.
There are some problems with your code, you need to tell the subclass what its parent class is. You also probably need to setup your constructors to pass through to the parent class. There were also a few misspellings in there.
package BaseClass;
sub new {
bless {} => shift;
}
sub method1 {
my $self = shift;
return $self->someAbstractMethod();
}
sub someAbstractMethod
{
die "oops, this is an abstract method that should " .
"be implemented in a subclass";
}
package SubClass;
use base 'BaseClass';
sub new {
shift->SUPER::new
}
sub someAbstractMethod
{
print "now we implement the asbtract method\n";
}
package main;
my $obj = BaseClass->new;
eval {$obj->method1(); 1} or warn $#;
my $subobj = SubClass->new;
$subobj->method1();
Your fundamental problem is that you're not using method calls at all. Had you done
sub method1 {
my $self = shift;
$self->someAbstractMethod();
}
in the base class (and assuming inheritance was set up properly) then things would work. But when you write someAbstractMethod() that's not a method call, it's a function call, so it's resolved immediately at compile-time, without any regard for polymorphism (what the type of the object is).
If you are just starting out with Perl OO dont forget to check out Moose. See also Moose::Manual::Unsweetened for a comparison of Moose with regular Perl 5 OOP style.

Is there a point to Perl's object oriented interfaces if they're not creating objects?

I think I read somewhere that some modules only have object oriented interfaces ( though they didn't create objects, they only held utility functions ). Is there a point to that?
First, its important to remember that in Perl, classes are implemented in a weird way, via packages. Packages also serve for general namespace pollution prevention.
package Foo;
sub new {
my ($class) = #_;
my $self = bless {}, $class;
return $self;
}
1;
That is how you make a Foo class in Perl (which can have an objected instantiated by calling Foo->new or new Foo). The use of new is just a convention; it can be anything at all. In fact, that new is what C++ would call a static method call.
You can easily create packages that contain only static method calls, and I suspect this is what you're referring to. The advantage here is that you can still use OO features like inheritance:
package Bar;
sub DoSomething {
my ($class, $arg) = #_;
$class->Compute($arg);
}
sub Compute {
my ($class, $arg) = #_;
$arg * 2;
}
1;
package Baz;
#Baz::ISA = qw(Bar);
sub Compute {
my ($class, $arg) = #_;
$arg * 2 - 1
}
1;
Given that, then
say Bar->DoSomething(3) # 6
say Baz->DoSomething(3) # 5
In fact, you can even use variables for the class name, so these can function very much like singletons:
my $obj = "Baz"; # or Baz->new could just return "Baz"
print $obj->DoSomething(3) # 5
[Code is untested; typos may be present]
I suspect that this is mostly a philosophical choice on the part of authors who prefer OO to imperative programming. Others have mentioned establishing a namespace, but it's the package that does that, not the interface. OO is not required.
Personally, I see little value in creating classes that are never instantiated (i.e. when there's no object in object-oriented). Perl isn't Java; you don't have to write a class for everything. Some modules acknowledge this. For example: File::Spec has an OO interface but also provides a functional interface via File::Spec::Functions.
File::Spec also provides an example of where OO can be useful for uninstantiated "utility" interfaces. Essentially, File::Spec is an abstract base class -- an interface with no implementation. When you load File::Spec it checks which OS you're using and loads the appropriate implementation. As a programmer, you use the interface (e.g. File::Spec->catfile) without having to worry about which version of catfile (Unix, Windows, VMS, etc.) to actually call.
As others have said, inheritance is the big gain if an actual object is not needed. The only thing I have to add here is the advice to name your variables well when writing such interfaces, e.g.:
package Foo;
# just a static method call
sub func
{
my $class = shift;
my (#args) = #_;
# stuff...
}
I named the variable that holds the classname "$class", rather than $this, to make it clear to subsequent maintainers that func() will be called as Foo->func() rather than $foo->func() (with an instantiated Foo object). This helps avoid someone adding this line later to the method:
my $value = $this->{key};
...which will fail, as there is no object to deference to get the "key" key.
If a method might be called either statically or against an instantiated object (for example, when writing a custom AUTOLOAD method), you can write this:
my method
{
my $this = shift;
my $class = ref($this) || $this;
my (#args) = #_;
# stuff...
}
namespacing, mostly. Why not? Everything that improves perl has my full approval.