[gitmo/Moose.git] / lib / Moose / Manual / Delta.pod

=pod

=head1 NAME

Moose::Manual::Delta - Important Changes in Moose

=head1 DESCRIPTION

This documents any important or noteworthy changes in Moose, with a
focus on backwards. This does duplicate data from the F<Changes> file,
but aims to provide more details and when possible workarounds.

Besides helping keep up with changes, you can also use this document
for finding the lowest version of Moose that supported a given
feature.  If you encounter a problem and have a solution but don't see
it documented here, or think we missed an important feature, please
send us a patch.

=head1 NEXT

=over 4

=item Native Traits

In previous versions of Moose, the Native delegations were created as
closures. The generated code was often quite slow compared to doing the same
thing by hand. For example, the Array's push delegation ended up doing
something like this:

  push @{ $self->$reader() }, @_;

If the attribute was created without a reader, the C<$reader> sub reference
followed a very slow code path. Even with a reader, this is still slower than
it needs to be.

Native delegations are now generated as inline code, just like other
accessors, so we can access the slot directly.

In addition, native traits now do proper constraint checking in all cases. In
particular, constraint checking has been improved for array and hash
references. Previously, only the I<contained> type (the C<Str> in
C<HashRef[Str]>) would be checked when a new value was added to the
collection. However, if there was a constraint that applied to the whole
value, this was never checked.

In addition, coercions are now called on the whole value.

The delegation methods now do more argument checking. All of the methods check
that a valid number of arguments were passed to the method. In addition, the
delegation methods check that the arguments are sane (array indexes, hash
keys, numbers, etc.) when applicable. We have tried to emulate the behavior of
Perl builtins as much as possible.

Finally, triggers are called whenever the value of the attribute is changed by
a Native delegation.

These changes are only likely to break code in a few cases. If you have a
typed arrayref or hashref attribute where the type enforces a constraint on
the whole collection, this constraint will now be checked. It's possible that
code which previously ran without errors will now cause the constraint to
fail. However, presumably this is a good thing ;)

If you are passing invalid arguments to a delegation which were previously
being ignore, these will now fail.

The other issue is the use of a trigger. If your code relied on the trigger
only being called for a regular writer, that may cause problems.

As always, you are encouraged to test before deploying the latest version of
Moose to production.

=back

=head1 1.09

=over 4

=item All deprecated features now warn

Previously, deprecation mostly consisted of simply saying "X is deprecated" in
the Changes file. We were not very consistent about actually warning. Now, all
deprecated features still present in Moose actually give a warning. The
warning is issued once per calling package. See L<Moose::Deprecated> for more
details.

=item You cannot pass C<< coerce => 1 >> unless the attribute's type constraint has a coercion

Previously, this was accepted, and it sort of worked, except that if you
attempted to set the attribute after the object was created, you would get a
runtime error.

Now you will get a warning when you attempt to define the attribute.

=item C<no Moose>, C<no Moose::Role>, and C<no Moose::Exporter> no longer unimport strict and warnings

This change was made in 1.05, and has now been reverted. We don't know if the
user has explicitly loaded strict or warnings on their own, and unimporting
them is just broken in that case.

=item Reversed logic when defining which options can be changed

L<Moose::Meta::Attribute> now allows all options to be changed in an
overridden attribute. The previous behaviour required each option to be
whitelisted using the C<legal_options_for_inheritance> method. This method has
been removed, and there is a new method, C<illegal_options_for_inheritance>,
which can now be used to prevent certain options from being changeable.

In addition, we only throw an error if the illegal option is actually
changed. If the superclass didn't specify this option at all when defining the
attribute, the subclass version can still add it as an option.

Example of overriding this in an attribute trait:

  package Bar::Meta::Attribute;
  use Moose::Role;

  has 'my_illegal_option' => (
      isa => 'CodeRef',
      is  => 'rw',
  );

  around illegal_options_for_inheritance => sub {
      return ( shift->(@_), qw/my_illegal_option/ );
  };

=back

=head1 1.05

=over 4

=item L<Moose::Object/BUILD> methods are now called when calling C<new_object>

Previously, C<BUILD> methods would only be called from C<Moose::Object::new>,
but now they are also called when constructing an object via
C<Moose::Meta::Class::new_object>. C<BUILD> methods are an inherent part of the
object construction process, and this should make C<< $meta->new_object >>
actually usable without forcing people to use C<< $meta->name->new >>.

=item C<no Moose>, C<no Moose::Role>, and C<no Moose::Exporter> now unimport strict and warnings

In the interest of having C<no Moose> clean up everything that C<use Moose>
does in the calling scope, C<no Moose> (as well as all other
L<Moose::Exporter>-using modules) now unimports strict and warnings.

=item Metaclass compatibility checking and fixing should be much more robust

The L<metaclass compatibility|Moose/METACLASS COMPATIBILITY AND MOOSE> checking
and fixing algorithms have been completely rewritten, in both Class::MOP and
Moose. This should resolve many confusing errors when dealing with non-Moose
inheritance and with custom metaclasses for things like attributes,
constructors, etc. For correct code, the only thing that should require a
change is that custom error metaclasses must now inherit from
L<Moose::Error::Default>.

=back

=head1 1.02

=over 4

=item Moose::Meta::TypeConstraint::Class is_subtype_of behavior

Earlier versions of L<is_subtype_of|Moose::Meta::TypeConstraint::Class/is_subtype_of>
would incorrectly return true when called with itself, its own TC name or
its class name as an argument. (i.e. $foo_tc->is_subtype_of('Foo') == 1) This
behavior was a caused by C<isa> being checked before the class name. The old
behavior can be accessed with L<is_type_of|Moose::Meta::TypeConstraint::Class/is_type_of>

=back

=head1 1.00

=over 4

=item Moose::Meta::Attribute::Native::Trait::Code no longer creates reader methods by default

Earlier versions of L<Moose::Meta::Attribute::Native::Trait::Code> created
read-only accessors for the attributes it's been applied to, even if you didn't
ask for it with C<< is => 'ro' >>. This incorrect behaviour has now been fixed.

=back

=head1 0.95

=over 4

=item Moose::Util add_method_modifier behavior

add_method_modifier (and subsequently the sugar functions Moose::before,
Moose::after, and Moose::around) can now accept arrayrefs, with the same
behavior as lists. Types other than arrayref and regexp result in an error.

=back

=head1 0.93_01 and 0.94

=over 4

=item Moose::Util::MetaRole API has changed

The C<apply_metaclass_roles> function is now called C<apply_metaroles>. The
way arguments are supplied has been changed to force you to distinguish
between metaroles applied to L<Moose::Meta::Class> (and helpers) versus
L<Moose::Meta::Role>.

The old API still works, but will warn in a future release, and eventually be
removed.

=item Moose::Meta::Role has real attributes

The attributes returned by L<Moose::Meta::Role> are now instances of the
L<Moose::Meta::Role::Attribute> class, instead of bare hash references.

=item "no Moose" now removes C<blessed> and C<confess>

Moose is now smart enough to know exactly what it exported, even when it
re-exports functions from other packages. When you unimport Moose, it will
remove these functions from your namespace unless you I<also> imported them
directly from their respective packages.

If you have a C<no Moose> in your code I<before> you call C<blessed> or
C<confess>, your code will break. You can either move the C<no Moose> call
later in your code, or explicitly import the relevant functions from the
packages that provide them.

=item L<Moose::Exporter> is smarter about unimporting re-exports

The change above comes from a general improvement to L<Moose::Exporter>. It
will now unimport any function it exports, even if that function is a
re-export from another package.

=item Attributes in roles can no longer override class attributes with "+foo"

Previously, this worked more or less accidentally, because role attributes
weren't objects. This was never documented, but a few MooseX modules took
advantage of this.

=item The composition_class_roles attribute in L<Moose::Meta::Role> is now a method

This was done to make it possible for roles to alter the the list of
composition class roles by applying a method modifiers. Previously, this was
an attribute and MooseX modules override it. Since that no longer works, this
was made a method.

This I<should> be an attribute, so this may switch back to being an attribute
in the future if we can figure out how to make this work.

=back

=head1 0.93

=over 4

=item Calling $object->new() is no longer deprecated

We decided to undeprecate this. Now it just works.

=item Both C<get_method_map> and C<get_attribute_map> is deprecated

These metaclass methods were never meant to be public, and they are both now
deprecated. The work around if you still need the functionality they provided
is to iterate over the list of names manually.

    my %fields = map { $_ => $meta->get_attribute($_) } $meta->get_attribute_list;

This was actually a change in L<Class::MOP>, but this version of Moose
requires a version of L<Class::MOP> that includes said change.

=back

=head1 0.90

=over 4

=item Added Native delegation for Code refs

See L<Moose::Meta::Attribute::Native::Trait::Code> for details.

=item Calling $object->new() is deprecated

Moose has long supported this, but it's never really been documented, and we
don't think this is a good practice. If you want to construct an object from
an existing object, you should provide some sort of alternate constructor like
C<< $object->clone >>.

Calling C<< $object->new >> now issues a warning, and will be an error in a
future release.

=item Moose no longer warns if you call C<make_immutable> for a class with mutable ancestors

While in theory this is a good thing to warn about, we found so many
exceptions to this that doing this properly became quite problematic.

=back

=head1 0.89_02

=over 4

=item New Native delegation methods from L<List::Util> and L<List::MoreUtils>

In particular, we now have C<reduce>, C<shuffle>, C<uniq>, and C<natatime>.

=item The Moose::Exporter with_caller feature is now deprecated

Use C<with_meta> instead. The C<with_caller> option will start warning in a
future release.

=item Moose now warns if you call C<make_immutable> for a class with mutable ancestors

This is dangerous because modifying a class after a subclass has been
immutabilized will lead to incorrect results in the subclass, due to inlining,
caching, etc. This occasionally happens accidentally, when a class loads one
of its subclasses in the middle of its class definition, so pointing out that
this may cause issues should be helpful. Metaclasses (classes that inherit
from L<Class::MOP::Object>) are currently exempt from this check, since at the
moment we aren't very consistent about which metaclasses we immutabilize.

=item C<enum> and C<duck_type> now take arrayrefs for all forms

Previously, calling these functions with a list would take the first element of
the list as the type constraint name, and use the remainder as the enum values
or method names. This makes the interface inconsistent with the anon-type forms
of these functions (which must take an arrayref), and a free-form list where
the first value is sometimes special is hard to validate (and harder to give
reasonable error messages for). These functions have been changed to take
arrayrefs in all their forms - so, C<< enum 'My::Type' => [qw(foo bar)] >> is
now the preferred way to create an enum type constraint. The old syntax still
works for now, but it will hopefully be deprecated and removed in a future
release.

=back

=head1 0.89_01

L<Moose::Meta::Attribute::Native> has been moved into the Moose core from
L<MooseX::AttributeHelpers>.  Major changes include:

=over 4

=item C<traits>, not C<metaclass>

Method providers are only available via traits.

=item C<handles>, not C<provides> or C<curries>

The C<provides> syntax was like core Moose C<< handles => HASHREF >>
syntax, but with the keys and values reversed.  This was confusing,
and AttributeHelpers now uses C<< handles => HASHREF >> in a way that
should be intuitive to anyone already familiar with how it is used for
other attributes.

The C<curries> functionality provided by AttributeHelpers has been
generalized to apply to all cases of C<< handles => HASHREF >>, though
not every piece of functionality has been ported (currying with a
CODEREF is not supported).

=item C<empty> is now C<is_empty>, and means empty, not non-empty

Previously, the C<empty> method provided by Arrays and Hashes returned true if
the attribute was B<not> empty (no elements).  Now it returns true if the
attribute B<is> empty. It was also renamed to C<is_empty>, to reflect this.

=item C<find> was renamed to C<first>, and C<first> and C<last> were removed

L<List::Util> refers to the functionality that we used to provide under C<find>
as L<first|List::Util/first>, so that will likely be more familiar (and will
fit in better if we decide to add more List::Util functions). C<first> and
C<last> were removed, since their functionality is easily duplicated with
curries of C<get>.

=item Helpers that take a coderef of one argument now use C<$_>

Subroutines passed as the first argument to C<first>, C<map>, and C<grep> now
receive their argument in C<$_> rather than as a parameter to the subroutine.
Helpers that take a coderef of two or more arguments remain using the argument
list (there are technical limitations to using C<$a> and C<$b> like C<sort>
does).

See L<Moose::Meta::Attribute::Native> for the new documentation.

=back

The C<alias> and C<excludes> role parameters have been renamed to C<-alias>
and C<-excludes>. The old names still work, but new code should use the new
names, and eventually the old ones will be deprecated and removed.

=head1 0.89

C<< use Moose -metaclass => 'Foo' >> now does alias resolution, just like
C<-traits> (and the C<metaclass> and C<traits> options to C<has>).

Added two functions C<meta_class_alias> and C<meta_attribute_alias> to
L<Moose::Util>, to simplify aliasing metaclasses and metatraits. This is
a wrapper around the old

  package Moose::Meta::Class::Custom::Trait::FooTrait;
  sub register_implementation { 'My::Meta::Trait' }

way of doing this.

=head1 0.84

When an attribute generates I<no> accessors, we now warn. This is to help
users who forget the C<is> option. If you really do not want any accessors,
you can use C<< is => 'bare' >>. You can maintain back compat with older
versions of Moose by using something like:

    ($Moose::VERSION >= 0.84 ? is => 'bare' : ())

When an accessor overwrites an existing method, we now warn. To work around
this warning (if you really must have this behavior), you can explicitly
remove the method before creating it as an accessor:

    sub foo {}

    __PACKAGE__->meta->remove_method('foo');

    has foo => (
        is => 'ro',
    );

When an unknown option is passed to C<has>, we now warn. You can silence
the warning by fixing your code. :)

The C<Role> type has been deprecated. On its own, it was useless,
since it just checked C<< $object->can('does') >>. If you were using
it as a parent type, just call C<role_type('Role::Name')> to create an
appropriate type instead.

=head1 0.78

C<use Moose::Exporter;> now imports C<strict> and C<warnings> into packages
that use it.

=head1 0.77

C<DEMOLISHALL> and C<DEMOLISH> now receive an argument indicating whether or
not we are in global destruction.

=head1 0.76

Type constraints no longer run coercions for a value that already matches the
constraint.  This may affect some (arguably buggy) edge case coercions that
rely on side effects in the C<via> clause.

=head1 0.75

L<Moose::Exporter> now accepts the C<-metaclass> option for easily
overriding the metaclass (without L<metaclass>). This works for classes
and roles.

=head1 0.74

Added a C<duck_type> sugar function to L<Moose::Util::TypeConstraints>
to make integration with non-Moose classes easier. It simply checks if
C<< $obj->can() >> a list of methods.

A number of methods (mostly inherited from L<Class::MOP>) have been
renamed with a leading underscore to indicate their internal-ness. The
old method names will still work for a while, but will warn that the
method has been renamed. In a few cases, the method will be removed
entirely in the future. This may affect MooseX authors who were using
these methods.

=head1 0.73

Calling C<subtype> with a name as the only argument now throws an
exception. If you want an anonymous subtype do:

    my $subtype = subtype as 'Foo';

This is related to the changes in version 0.71_01.

The C<is_needed> method in L<Moose::Meta::Method::Destructor> is now
only usable as a class method. Previously, it worked as a class or
object method, with a different internal implementation for each
version.

The internals of making a class immutable changed a lot in Class::MOP
0.78_02, and Moose's internals have changed along with it. The
external C<< $metaclass->make_immutable >> method still works the same
way.

=head1 0.72

A mutable class accepted C<< Foo->new(undef) >> without complaint,
while an immutable class would blow up with an unhelpful error. Now,
in both cases we throw a helpful error instead.

This "feature" was originally added to allow for cases such as this:

  my $args;

  if ( something() ) {
      $args = {...};
  }

  return My::Class->new($args);

But we decided this is a bad idea and a little too magical, because it
can easily mask real errors.

=head1 0.71_01

Calling C<type> or C<subtype> without the sugar helpers (C<as>,
C<where>, C<message>) is now deprecated.

As a side effect, this meant we ended up using Perl prototypes on
C<as>, and code like this will no longer work:

  use Moose::Util::TypeConstraints;
  use Declare::Constraints::Simple -All;

  subtype 'ArrayOfInts'
      => as 'ArrayRef'
      => IsArrayRef(IsInt);

Instead it must be changed to this:

  subtype(
      'ArrayOfInts' => {
          as    => 'ArrayRef',
          where => IsArrayRef(IsInt)
      }
  );

If you want to maintain backwards compat with older versions of Moose,
you must explicitly test Moose's C<VERSION>:

  if ( Moose->VERSION < 0.71_01 ) {
      subtype 'ArrayOfInts'
          => as 'ArrayRef'
          => IsArrayRef(IsInt);
  }
  else {
      subtype(
          'ArrayOfInts' => {
              as    => 'ArrayRef',
              where => IsArrayRef(IsInt)
          }
      );
  }

=head1 0.70

We no longer pass the meta-attribute object as a final argument to
triggers. This actually changed for inlined code a while back, but the
non-inlined version and the docs were still out of date.

If by some chance you actually used this feature, the workaround is
simple. You fetch the attribute object from out of the C<$self>
that is passed as the first argument to trigger, like so:

  has 'foo' => (
      is      => 'ro',
      isa     => 'Any',
      trigger => sub {
          my ( $self, $value ) = @_;
          my $attr = $self->meta->find_attribute_by_name('foo');

          # ...
      }
  );

=head1 0.66

If you created a subtype and passed a parent that Moose didn't know
about, it simply ignored the parent. Now it automatically creates the
parent as a class type. This may not be what you want, but is less
broken than before.

You could declare a name with subtype such as "Foo!Bar". Moose would
accept this allowed, but if you used it in a parameterized type such
as "ArrayRef[Foo!Bar]" it wouldn't work. We now do some vetting on
names created via the sugar functions, so that they can only contain
alphanumerics, ":", and ".".

=head1 0.65

Methods created via an attribute can now fulfill a C<requires>
declaration for a role. Honestly we don't know why Stevan didn't make
this work originally, he was just insane or something.

Stack traces from inlined code will now report the line and file as
being in your class, as opposed to in Moose guts.

=head1 0.62_02

When a class does not provide all of a role's required methods, the
error thrown now mentions all of the missing methods, as opposed to
just the first missing method.

Moose will no longer inline a constructor for your class unless it
inherits its constructor from Moose::Object, and will warn when it
doesn't inline. If you want to force inlining anyway, pass
C<< replace_constructor => 1 >> to C<make_immutable>.

If you want to get rid of the warning, pass C<< inline_constructor =>
0 >>.

=head1 0.62

Removed the (deprecated) C<make_immutable> keyword.

Removing an attribute from a class now also removes delegation
(C<handles>) methods installed for that attribute. This is correct
behavior, but if you were wrongly relying on it you might get bit.

=head1 0.58

Roles now add methods by calling C<add_method>, not
C<alias_method>. They make sure to always provide a method object,
which will be cloned internally. This means that it is now possible to
track the source of a method provided by a role, and even follow its
history through intermediate roles.  This means that methods added by
a role now show up when looking at a class's method list/map.

Parameter and Union args are now sorted, this makes Int|Str the same
constraint as Str|Int. Also, incoming type constraint strings are
normalized to remove all whitespace differences. This is mostly for
internals and should not affect outside code.

L<Moose::Exporter> will no longer remove a subroutine that the
exporting package re-exports. Moose re-exports the Carp::confess
function, among others. The reasoning is that we cannot know whether
you have also explicitly imported those functions for your own use, so
we err on the safe side and always keep them.

=head1 0.56

C<Moose::init_meta> should now be called as a method.

New modules for extension writers, L<Moose::Exporter> and
L<Moose::Util::MetaRole>.

=head1 0.55_01

Implemented metaclass traits (and wrote a recipe for it):

  use Moose -traits => 'Foo'

This should make writing small Moose extensions a little
easier.

=head1 0.55

Fixed C<coerce> to accept anon types just like C<subtype> can.
So that you can do:

  coerce $some_anon_type => from 'Str' => via { ... };

=head1 0.51

Added C<BUILDARGS>, a new step in C<< Moose::Object->new() >>.

=head1 0.49

Fixed how the C<< is => (ro|rw) >> works with custom defined
C<reader>, C<writer> and C<accessor> options. See the below table for
details:

  is => ro, writer => _foo    # turns into (reader => foo, writer => _foo)
  is => rw, writer => _foo    # turns into (reader => foo, writer => _foo)
  is => rw, accessor => _foo  # turns into (accessor => _foo)
  is => ro, accessor => _foo  # error, accesor is rw

=head1 0.45

The C<before/around/after> method modifiers now support regexp
matching of method names. NOTE: this only works for classes, it is
currently not supported in roles, but, ... patches welcome.

The C<has> keyword for roles now accepts the same array ref form that
L<Moose>.pm does for classes.

A trigger on a read-only attribute is no longer an error, as it's
useful to trigger off of the constructor.

Subtypes of parameterizable types now are parameterizable types
themselves.

=head1 0.44

Fixed issue where C<DEMOLISHALL> was eating the value in C<$@>, and so
not working correctly. It still kind of eats them, but so does vanilla
perl.

=head1 0.41

Inherited attributes may now be extended without restriction on the
type ('isa', 'does').

The entire set of Moose::Meta::TypeConstraint::* classes were
refactored in this release. If you were relying on their internals you
should test your code carefully.

=head1 0.40

Documenting the use of '+name' with attributes that come from recently
composed roles. It makes sense, people are using it, and so why not
just officially support it.

The C<< Moose::Meta::Class->create >> method now supports roles.

It is now possible to make anonymous enum types by passing C<enum> an
array reference instead of the C<< enum $name => @values >>.

=head1 0.37

Added the C<make_immutable> keyword as a shortcut to calling
C<make_immutable> on the meta object. This eventually got removed!

Made C<< init_arg => undef >> work in Moose. This means "do not accept
a constructor parameter for this attribute".

Type errors now use the provided message. Prior to this release they
didn't.

=head1 0.34

Moose is now a postmodern object system :)

The Role system was completely refactored. It is 100% backwards
compat, but the internals were totally changed. If you relied on the
internals then you are advised to test carefully.

Added method exclusion and aliasing for Roles in this release.

Added the L<Moose::Util::TypeConstraints::OptimizedConstraints>
module.

Passing a list of values to an accessor (which is only expecting one
value) used to be silently ignored, now it throws an error.

=head1 0.26

Added parameterized types and did a pretty heavy refactoring of the
type constraint system.

Better framework extendability and better support for "making your own
Moose".

=head1 0.25 or before

Honestly, you shouldn't be using versions of Moose that are this old,
so many bug fixes and speed improvements have been made you would be
crazy to not upgrade.

Also, I am tired of going through the Changelog so I am stopping here,
if anyone would like to continue this please feel free.

=head1 AUTHOR

Stevan Little E<lt>stevan@iinteractive.comE<gt>

=head1 COPYRIGHT AND LICENSE

Copyright 2009 by Infinity Interactive, Inc.

L<http://www.iinteractive.com>

This library is free software; you can redistribute it and/or modify
it under the same terms as Perl itself.

=cut