package Hash::Util::FieldHash;

use strict;
use warnings;
no warnings 'experimental::builtin';
use builtin qw(reftype);

our $VERSION = '1.26';

use Exporter 'import';
our %EXPORT_TAGS = (
    'all' => [ qw(
our @EXPORT_OK = ( @{ $EXPORT_TAGS{'all'} } );

    require XSLoader;
    my %ob_reg; # private object registry
    sub _ob_reg { \ %ob_reg }

sub fieldhash (\%) {
    for ( shift ) {
        return unless ref() && reftype( $_) eq 'HASH';
        return $_ if Hash::Util::FieldHash::_fieldhash( $_, 0);
        return $_ if Hash::Util::FieldHash::_fieldhash( $_, 2) == 2;

sub idhash (\%) {
    for ( shift ) {
        return unless ref() && reftype( $_) eq 'HASH';
        return $_ if Hash::Util::FieldHash::_fieldhash( $_, 0);
        return $_ if Hash::Util::FieldHash::_fieldhash( $_, 1) == 1;

sub fieldhashes { map &fieldhash( $_), @_ }
sub idhashes { map &idhash( $_), @_ }


=head1 NAME

Hash::Util::FieldHash - Support for Inside-Out Classes


  ### Create fieldhashes
  use Hash::Util qw(fieldhash fieldhashes);

  # Create a single field hash
  fieldhash my %foo;

  # Create three at once...
  fieldhashes \ my(%foo, %bar, %baz);
  # ...or any number
  fieldhashes @hashrefs;

  ### Create an idhash and register it for garbage collection
  use Hash::Util::FieldHash qw(idhash register);
  idhash my %name;
  my $object = \ do { my $o };
  # register the idhash for garbage collection with $object
  register($object, \ %name);
  # the following entry will be deleted when $object goes out of scope
  $name{$object} = 'John Doe';

  ### Register an ordinary hash for garbage collection
  use Hash::Util::FieldHash qw(id register);
  my %name;
  my $object = \ do { my $o };
  # register the hash %name for garbage collection of $object's id
  register $object, \ %name;
  # the following entry will be deleted when $object goes out of scope
  $name{id $object} = 'John Doe';


C<Hash::Util::FieldHash> offers a number of functions in support of
L<The Inside-out Technique> of class construction.


=item id


Returns the reference address of a reference $obj.  If $obj is
not a reference, returns $obj.

This function is a stand-in replacement for
that is, it returns
the reference address of its argument as a numeric value.  The only
difference is that C<refaddr()> returns C<undef> when given a
non-reference while C<id()> returns its argument unchanged.

C<id()> also uses a caching technique that makes it faster when
the id of an object is requested often, but slower if it is needed
only once or twice.

=item id_2obj

    $obj = id_2obj($id)

If C<$id> is the id of a registered object (see L</register>), returns
the object, otherwise an undefined value.  For registered objects this
is the inverse function of C<id()>.

=item register

    register($obj, @hashrefs)

In the first form, registers an object to work with for the function
C<id_2obj()>.  In the second form, it additionally marks the given
hashrefs down for garbage collection.  This means that when the object
goes out of scope, any entries in the given hashes under the key of
C<id($obj)> will be deleted from the hashes.

It is a fatal error to register a non-reference $obj.  Any non-hashrefs
among the following arguments are silently ignored.

It is I<not> an error to register the same object multiple times with
varying sets of hashrefs.  Any hashrefs that are not registered yet
will be added, others ignored.

Registry also implies thread support.  When a new thread is created,
all references are replaced with new ones, including all objects.
If a hash uses the reference address of an object as a key, that
connection would be broken.  With a registered object, its id will
be updated in all hashes registered with it.

=item idhash

    idhash my %hash

Makes an idhash from the argument, which must be a hash.

An I<idhash> works like a normal hash, except that it stringifies a
I<reference used as a key> differently.  A reference is stringified
as if the C<id()> function had been invoked on it, that is, its
reference address in decimal is used as the key.

=item idhashes

    idhashes \ my(%hash, %gnash, %trash)
    idhashes \ @hashrefs

Creates many idhashes from its hashref arguments.  Returns those
arguments that could be converted or their number in scalar context.

=item fieldhash

    fieldhash %hash;

Creates a single fieldhash.  The argument must be a hash.  Returns
a reference to the given hash if successful, otherwise nothing.

A I<fieldhash> is, in short, an idhash with auto-registry.  When an
object (or, indeed, any reference) is used as a fieldhash key, the
fieldhash is automatically registered for garbage collection with
the object, as if C<register $obj, \ %fieldhash> had been called.

=item fieldhashes

    fieldhashes @hashrefs;

Creates any number of field hashes.  Arguments must be hash references.
Returns the converted hashrefs in list context, their number in scalar



A word on terminology:  I shall use the term I<field> for a scalar
piece of data that a class associates with an object.  Other terms that
have been used for this concept are "object variable", "(object) property",
"(object) attribute" and more.  Especially "attribute" has some currency
among Perl programmer, but that clashes with the C<attributes> pragma.  The
term "field" also has some currency in this sense and doesn't seem
to conflict with other Perl terminology.

In Perl, an object is a blessed reference.  The standard way of associating
data with an object is to store the data inside the object's body, that is,
the piece of data pointed to by the reference.

In consequence, if two or more classes want to access an object they
I<must> agree on the type of reference and also on the organization of
data within the object body.  Failure to agree on the type results in
immediate death when the wrong method tries to access an object.  Failure
to agree on data organization may lead to one class trampling over the
data of another.

This object model leads to a tight coupling between subclasses.
If one class wants to inherit from another (and both classes access
object data), the classes must agree about implementation details.
Inheritance can only be used among classes that are maintained together,
in a single source or not.

In particular, it is not possible to write general-purpose classes
in this technique, classes that can advertise themselves as "Put me
on your @ISA list and use my methods".  If the other class has different
ideas about how the object body is used, there is trouble.

For reference C<Name_hash> in L</Example 1> shows the standard implementation of
a simple class C<Name> in the well-known hash based way.  It also demonstrates
the predictable failure to construct a common subclass C<NamedFile>
of C<Name> and the class C<IO::File> (whose objects I<must> be globrefs).

Thus, techniques are of interest that store object data I<not> in
the object body but some other place.

=head2 The Inside-out Technique

With I<inside-out> classes, each class declares a (typically lexical)
hash for each field it wants to use.  The reference address of an
object is used as the hash key.  By definition, the reference address
is unique to each object so this guarantees a place for each field that
is private to the class and unique to each object.  See C<Name_id>
in L</Example 1> for a simple example.

In comparison to the standard implementation where the object is a
hash and the fields correspond to hash keys, here the fields correspond
to hashes, and the object determines the hash key.  Thus the hashes
appear to be turned I<inside out>.

The body of an object is never examined by an inside-out class, only
its reference address is used.  This allows for the body of an actual
object to be I<anything at all> while the object methods of the class
still work as designed.  This is a key feature of inside-out classes.

=head2 Problems of Inside-out

Inside-out classes give us freedom of inheritance, but as usual there
is a price.

Most obviously, there is the necessity of retrieving the reference
address of an object for each data access.  It's a minor inconvenience,
but it does clutter the code.

More important (and less obvious) is the necessity of garbage
collection.  When a normal object dies, anything stored in the
object body is garbage-collected by perl.  With inside-out objects,
Perl knows nothing about the data stored in field hashes by a class,
but these must be deleted when the object goes out of scope.  Thus
the class must provide a C<DESTROY> method to take care of that.

In the presence of multiple classes it can be non-trivial
to make sure that every relevant destructor is called for
every object.  Perl calls the first one it finds on the
inheritance tree (if any) and that's it.

A related issue is thread-safety.  When a new thread is created,
the Perl interpreter is cloned, which implies that all reference
addresses in use will be replaced with new ones.  Thus, if a class
tries to access a field of a cloned object its (cloned) data will
still be stored under the now invalid reference address of the
original in the parent thread.  A general C<CLONE> method must
be provided to re-establish the association.

=head2 Solutions

C<Hash::Util::FieldHash> addresses these issues on several

The C<id()> function is provided in addition to the
existing C<Scalar::Util::refaddr()>.  Besides its short name
it can be a little faster under some circumstances (and a
bit slower under others).  Benchmark if it matters.  The
working of C<id()> also allows the use of the class name
as a I<generic object> as described L<further down|/"The Generic Object">.

The C<id()> function is incorporated in I<id hashes> in the sense
that it is called automatically on every key that is used with
the hash.  No explicit call is necessary.

The problems of garbage collection and thread safety are both
addressed by the function C<register()>.  It registers an object
together with any number of hashes.  Registry means that when the
object dies, an entry in any of the hashes under the reference
address of this object will be deleted.  This guarantees garbage
collection in these hashes.  It also means that on thread
cloning the object's entries in registered hashes will be
replaced with updated entries whose key is the cloned object's
reference address.  Thus the object-data association becomes

Object registry is best done when the object is initialized
for use with a class.  That way, garbage collection and thread
safety are established for every object and every field that is

Finally, I<field hashes> incorporate all these functions in one
package.  Besides automatically calling the C<id()> function
on every object used as a key, the object is registered with
the field hash on first use.  Classes based on field hashes
are fully garbage-collected and thread safe without further

=head2 More Problems

Another problem that occurs with inside-out classes is serialization.
Since the object data is not in its usual place, standard routines
like C<Storable::freeze()>, C<Storable::thaw()> and 
C<Data::Dumper::Dumper()> can't deal with it on their own.  Both
C<Data::Dumper> and C<Storable> provide the necessary hooks to
make things work, but the functions or methods used by the hooks
must be provided by each inside-out class.

A general solution to the serialization problem would require another
level of registry, one that associates I<classes> and fields.
So far, the functions of C<Hash::Util::FieldHash> are unaware of
any classes, which I consider a feature.  Therefore C<Hash::Util::FieldHash>
doesn't address the serialization problems.

=head2 The Generic Object

Classes based on the C<id()> function (and hence classes based on
C<idhash()> and C<fieldhash()>) show a peculiar behavior in that
the class name can be used like an object.  Specifically, methods
that set or read data associated with an object continue to work as
class methods, just as if the class name were an object, distinct from
all other objects, with its own data.  This object may be called
the I<generic object> of the class.

This works because field hashes respond to keys that are not references
like a normal hash would and use the string offered as the hash key.
Thus, if a method is called as a class method, the field hash is presented
with the class name instead of an object and blithely uses it as a key.
Since the keys of real objects are decimal numbers, there is no
conflict and the slot in the field hash can be used like any other.
The C<id()> function behaves correspondingly with respect to non-reference

Two possible uses (besides ignoring the property) come to mind.
A singleton class could be implemented this using the generic object.
If necessary, an C<init()> method could die or ignore calls with
actual objects (references), so only the generic object will ever exist.

Another use of the generic object would be as a template.  It is
a convenient place to store class-specific defaults for various
fields to be used in actual object initialization.

Usually, the feature can be entirely ignored.  Calling I<object
methods> as I<class methods> normally leads to an error and isn't used
routinely anywhere.  It may be a problem that this error isn't
indicated by a class with a generic object.

=head2 How to use Field Hashes

Traditionally, the definition of an inside-out class contains a bare
block inside which a number of lexical hashes are declared and the
basic accessor methods defined, usually through C<Scalar::Util::refaddr>.
Further methods may be defined outside this block.  There has to be
a DESTROY method and, for thread support, a CLONE method.

When field hashes are used, the basic structure remains the same.
Each lexical hash will be made a field hash.  The call to C<refaddr>
can be omitted from the accessor methods.  DESTROY and CLONE methods
are not necessary.

If you have an existing inside-out class, simply making all hashes
field hashes with no other change should make no difference.  Through
the calls to C<refaddr> or equivalent, the field hashes never get to
see a reference and work like normal hashes.  Your DESTROY (and
CLONE) methods are still needed.

To make the field hashes kick in, it is easiest to redefine C<refaddr>

    sub refaddr { shift }

instead of importing it from C<Scalar::Util>.  It should now be possible
to disable DESTROY and CLONE.  Note that while it isn't disabled,
DESTROY will be called before the garbage collection of field hashes,
so it will be invoked with a functional object and will continue to

It is not desirable to import the functions C<fieldhash> and/or
C<fieldhashes> into every class that is going to use them.  They
are only used once to set up the class.  When the class is up and running,
these functions serve no more purpose.

If there are only a few field hashes to declare, it is simplest to

    use Hash::Util::FieldHash;

early and call the functions qualified:

    Hash::Util::FieldHash::fieldhash my %foo;

Otherwise, import the functions into a convenient package like
C<HUF> or, more general, C<Aux>

        package Aux;
        use Hash::Util::FieldHash ':all';

and call

    Aux::fieldhash my %foo;

as needed.

=head2 Garbage-Collected Hashes

Garbage collection in a field hash means that entries will "spontaneously"
disappear when the object that created them disappears.  That must be
borne in mind, especially when looping over a field hash.  If anything
you do inside the loop could cause an object to go out of scope, a
random key may be deleted from the hash you are looping over.  That
can throw the loop iterator, so it's best to cache a consistent snapshot
of the keys and/or values and loop over that.  You will still have to
check that a cached entry still exists when you get to it.

Garbage collection can be confusing when keys are created in a field hash
from normal scalars as well as references.  Once a reference is I<used> with
a field hash, the entry will be collected, even if it was later overwritten
with a plain scalar key (every positive integer is a candidate).  This
is true even if the original entry was deleted in the meantime.  In fact,
deletion from a field hash, and also a test for existence constitute
I<use> in this sense and create a liability to delete the entry when
the reference goes out of scope.  If you happen to create an entry
with an identical key from a string or integer, that will be collected
instead.  Thus, mixed use of references and plain scalars as field hash
keys is not entirely supported.


The examples show a very simple class that implements a I<name>, consisting
of a first and last name (no middle initial).  The name class has four


=item * C<init()>

An object method that initializes the first and last name to its
two arguments. If called as a class method, C<init()> creates an
object in the given class and initializes that.

=item * C<first()>

Retrieve the first name

=item * C<last()>

Retrieve the last name

=item * C<name()>

Retrieve the full name, the first and last name joined by a blank.


The examples show this class implemented with different levels of
support by C<Hash::Util::FieldHash>.  All supported combinations
are shown.  The difference between implementations is often quite
small.  The implementations are:


=item * C<Name_hash>

A conventional (not inside-out) implementation where an object is
a hash that stores the field values, without support by
C<Hash::Util::FieldHash>.  This implementation doesn't allow
arbitrary inheritance.

=item * C<Name_id>

Inside-out implementation based on the C<id()> function.  It needs
a C<DESTROY> method.  For thread support a C<CLONE> method (not shown)
would also be needed.  Instead of C<Hash::Util::FieldHash::id()> the
function C<Scalar::Util::refaddr> could be used with very little
functional difference.  This is the basic pattern of an inside-out

=item * C<Name_idhash>

Idhash-based inside-out implementation.  Like C<Name_id> it needs
a C<DESTROY> method and would need C<CLONE> for thread support.

=item * C<Name_id_reg>

Inside-out implementation based on the C<id()> function with explicit
object registry.  No destructor is needed and objects are thread safe.

=item * C<Name_idhash_reg>

Idhash-based inside-out implementation with explicit object registry.
No destructor is needed and objects are thread safe.

=item * C<Name_fieldhash>

FieldHash-based inside-out implementation.  Object registry happens
automatically.  No destructor is needed and objects are thread safe.


These examples are realized in the code below, which could be copied
to a file F<>.

=head2 Example 1

    use strict; use warnings;

        package Name_hash;  # standard implementation: the
                            # object is a hash
        sub init {
            my $obj = shift;
            my ($first, $last) = @_;
            # create an object if called as class method
            $obj = bless {}, $obj unless ref $obj;
            $obj->{ first} = $first;
            $obj->{ last} = $last;

        sub first { shift()->{ first} }
        sub last { shift()->{ last} }

        sub name {
            my $n = shift;
            join ' ' => $n->first, $n->last;


        package Name_id;
        use Hash::Util::FieldHash qw(id);

        my (%first, %last);

        sub init {
            my $obj = shift;
            my ($first, $last) = @_;
            # create an object if called as class method
            $obj = bless \ my $o, $obj unless ref $obj;
            $first{ id $obj} = $first;
            $last{ id $obj} = $last;

        sub first { $first{ id shift()} }
        sub last { $last{ id shift()} }

        sub name {
            my $n = shift;
            join ' ' => $n->first, $n->last;

        sub DESTROY {
            my $id = id shift;
            delete $first{ $id};
            delete $last{ $id};


        package Name_idhash;
        use Hash::Util::FieldHash;

        Hash::Util::FieldHash::idhashes( \ my (%first, %last) );

        sub init {
            my $obj = shift;
            my ($first, $last) = @_;
            # create an object if called as class method
            $obj = bless \ my $o, $obj unless ref $obj;
            $first{ $obj} = $first;
            $last{ $obj} = $last;

        sub first { $first{ shift()} }
        sub last { $last{ shift()} }

        sub name {
            my $n = shift;
            join ' ' => $n->first, $n->last;

        sub DESTROY {
            my $n = shift;
            delete $first{ $n};
            delete $last{ $n};


        package Name_id_reg;
        use Hash::Util::FieldHash qw(id register);

        my (%first, %last);

        sub init {
            my $obj = shift;
            my ($first, $last) = @_;
            # create an object if called as class method
            $obj = bless \ my $o, $obj unless ref $obj;
            register( $obj, \ (%first, %last) );
            $first{ id $obj} = $first;
            $last{ id $obj} = $last;

        sub first { $first{ id shift()} }
        sub last { $last{ id shift()} }

        sub name {
            my $n = shift;
            join ' ' => $n->first, $n->last;

        package Name_idhash_reg;
        use Hash::Util::FieldHash qw(register);

        Hash::Util::FieldHash::idhashes \ my (%first, %last);

        sub init {
            my $obj = shift;
            my ($first, $last) = @_;
            # create an object if called as class method
            $obj = bless \ my $o, $obj unless ref $obj;
            register( $obj, \ (%first, %last) );
            $first{ $obj} = $first;
            $last{ $obj} = $last;

        sub first { $first{ shift()} }
        sub last { $last{ shift()} }

        sub name {
            my $n = shift;
            join ' ' => $n->first, $n->last;

        package Name_fieldhash;
        use Hash::Util::FieldHash;

        Hash::Util::FieldHash::fieldhashes \ my (%first, %last);

        sub init {
            my $obj = shift;
            my ($first, $last) = @_;
            # create an object if called as class method
            $obj = bless \ my $o, $obj unless ref $obj;
            $first{ $obj} = $first;
            $last{ $obj} = $last;

        sub first { $first{ shift()} }
        sub last { $last{ shift()} }

        sub name {
            my $n = shift;
            join ' ' => $n->first, $n->last;


To exercise the various implementations the script L<below|/"Example 2"> can
be used.

It sets up a class C<Name> that is a mirror of one of the implementation
classes C<Name_hash>, C<Name_id>, ..., C<Name_fieldhash>.  That determines
which implementation is run.

The script first verifies the function of the C<Name> class.

In the second step, the free inheritability of the implementation
(or lack thereof) is demonstrated.  For this purpose it constructs
a class called C<NamedFile> which is a common subclass of C<Name> and
the standard class C<IO::File>.  This puts inheritability to the test
because objects of C<IO::File> I<must> be globrefs.  Objects of C<NamedFile>
should behave like a file opened for reading and also support the C<name()>
method.  This class juncture works with exception of the C<Name_hash>
implementation, where object initialization fails because of the
incompatibility of object bodies.

=head2 Example 2

    use strict; use warnings; $| = 1;

    use Example;

        package Name;
        use parent 'Name_id';  # define here which implementation to run

    # Verify that the base package works
    my $n = Name->init(qw(Albert Einstein));
    print $n->name, "\n";
    print "\n";

    # Create a named file handle (See definition below)
    my $nf = NamedFile->init(qw(/tmp/x Filomena File));
    # use as a file handle...
    for ( 1 .. 3 ) {
        my $l = <$nf>;
        print "line $_: $l";
    # ...and as a Name object
    print "...brought to you by ", $nf->name, "\n";

    # Definition of NamedFile
    package NamedFile;
    use parent 'Name';
    use parent 'IO::File';

    sub init {
        my $obj = shift;
        my ($file, $first, $last) = @_;
        $obj = $obj->IO::File::new() unless ref $obj;
        $obj->open($file) or die "Can't read '$file': $!";
        $obj->Name::init($first, $last);

=head1 GUTS

To make C<Hash::Util::FieldHash> work, there were two changes to
F<perl> itself.  C<PERL_MAGIC_uvar> was made available for hashes,
and weak references now call uvar C<get> magic after a weakref has been
cleared.  The first feature is used to make field hashes intercept
their keys upon access.  The second one triggers garbage collection.

=head2 The C<PERL_MAGIC_uvar> interface for hashes

C<PERL_MAGIC_uvar> I<get> magic is called from C<hv_fetch_common> and
C<hv_delete_common> through the function C<hv_magic_uvar_xkey>, which
defines the interface.  The call happens for hashes with "uvar" magic
if the C<ufuncs> structure has equal values in the C<uf_val> and C<uf_set>
fields.  Hashes are unaffected if (and as long as) these fields
hold different values.

Upon the call, the C<mg_obj> field will hold the hash key to be accessed.
Upon return, the C<SV*> value in C<mg_obj> will be used in place of the
original key in the hash access.  The integer index value in the first
parameter will be the C<action> value from C<hv_fetch_common>, or -1
if the call is from C<hv_delete_common>.

This is a template for a function suitable for the C<uf_val> field in
a C<ufuncs> structure for this call.  The C<uf_set> and C<uf_index>
fields are irrelevant.

    IV watch_key(pTHX_ IV action, SV* field) {
        MAGIC* mg = mg_find(field, PERL_MAGIC_uvar);
        SV* keysv = mg->mg_obj;
        /* Do whatever you need to.  If you decide to
           supply a different key newkey, return it like this
        mg->mg_obj = newkey;
        return 0;

=head2 Weakrefs call uvar magic

When a weak reference is stored in an C<SV> that has "uvar" magic, C<set>
magic is called after the reference has gone stale.  This hook can be
used to trigger further garbage-collection activities associated with
the referenced object.

=head2 How field hashes work

The three features of key hashes, I<key replacement>, I<thread support>,
and I<garbage collection> are supported by a data structure called
the I<object registry>.  This is a private hash where every object
is stored.  An "object" in this sense is any reference (blessed or
unblessed) that has been used as a field hash key.

The object registry keeps track of references that have been used as
field hash keys.  The keys are generated from the reference address
like in a field hash (though the registry isn't a field hash).  Each
value is a weak copy of the original reference, stored in an C<SV> that
is itself magical (C<PERL_MAGIC_uvar> again).  The magical structure
holds a list (another hash, really) of field hashes that the reference
has been used with.  When the weakref becomes stale, the magic is
activated and uses the list to delete the reference from all field
hashes it has been used with.  After that, the entry is removed from
the object registry itself.  Implicitly, that frees the magic structure
and the storage it has been using.

Whenever a reference is used as a field hash key, the object registry
is checked and a new entry is made if necessary.  The field hash is
then added to the list of fields this reference has used.

The object registry is also used to repair a field hash after thread
cloning.  Here, the entire object registry is processed.  For every
reference found there, the field hashes it has used are visited and
the entry is updated.

=head2 Internal function Hash::Util::FieldHash::_fieldhash

    # test if %hash is a field hash
    my $result = _fieldhash \ %hash, 0;

    # make %hash a field hash
    my $result = _fieldhash \ %hash, 1;

C<_fieldhash> is the internal function used to create field hashes.
It takes two arguments, a hashref and a mode.  If the mode is boolean
false, the hash is not changed but tested if it is a field hash.  If
the hash isn't a field hash the return value is boolean false.  If it
is, the return value indicates the mode of field hash.  When called with
a boolean true mode, it turns the given hash into a field hash of this
mode, returning the mode of the created field hash.  C<_fieldhash>
does not erase the given hash.

Currently there is only one type of field hash, and only the boolean
value of the mode makes a difference, but that may change.

=head1 AUTHOR

Anno Siegel (ANNO) wrote the xs code and the changes in perl proper
Jerry Hedden (JDHEDDEN) made it faster


Copyright (C) 2006-2007 by (Anno Siegel)

This library is free software; you can redistribute it and/or modify
it under the same terms as Perl itself, either Perl version 5.8.7 or,
at your option, any later version of Perl 5 you may have available.