The Inside-out Technique

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 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 anything at all while the object methods of the class still work as designed. This is a key feature of inside-out classes.

Problems of Inside-out

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 "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 "CLONE" method must be provided to re-establish the association.

Solutions

The "id()" function is provided in addition to the existing "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 "id()" also allows the use of the class name as a generic object as described further down.

The "id()" function is incorporated in 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 "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 thread-safe.

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 initialized.

Finally, field hashes incorporate all these functions in one package. Besides automatically calling the "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 measures.

More Problems

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

The Generic Object

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 "id()" function behaves correspondingly with respect to non-reference arguments.

Two possible uses (besides ignoring the property) come to mind. A singleton class could be implemented this using the generic object. If necessary, an "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 object methods as 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.

How to use Field Hashes

When field hashes are used, the basic structure remains the same. Each lexical hash will be made a field hash. The call to "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 "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 "refaddr" as

    sub refaddr { shift }

instead of importing it from "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 function.

It is not desirable to import the functions "fieldhash" and/or "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 "HUF" or, more general, "Aux"

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

and call

    Aux::fieldhash my %foo;

as needed.

Garbage-Collected Hashes

Garbage collection can be confusing when keys are created in a field hash from normal scalars as well as references. Once a reference is 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 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.

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;
            $obj;
        }

        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;
            $obj;
        }

        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;
            $obj;
        }

        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;
            $obj;
        }

        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;
            $obj;
        }

        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;
            $obj;
        }

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

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

    1;

To exercise the various implementations the script below can be used.

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

The script first verifies the function of the "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 "NamedFile" which is a common subclass of "Name" and the standard class "IO::File". This puts inheritability to the test because objects of "IO::File" must be globrefs. Objects of "NamedFile" should behave like a file opened for reading and also support the "name()" method. This class juncture works with exception of the "Name_hash" implementation, where object initialization fails because of the incompatibility of object bodies.

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";
    exit;


    # 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);
    }
    __END__

The "PERL_MAGIC_uvar" interface for hashes

Upon the call, the "mg_obj" field will hold the hash key to be accessed. Upon return, the "SV*" value in "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 "action" value from "hv_fetch_common", or -1 if the call is from "hv_delete_common".

This is a template for a function suitable for the "uf_val" field in a "ufuncs" structure for this call. The "uf_set" and "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
        */
        sv_2mortal(newkey);
        mg->mg_obj = newkey;
        return 0;
    }

Weakrefs call uvar magic

How field hashes work

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 "SV" that is itself magical ("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.

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;

"_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. "_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.