perl/pod/perlclassguts.pod

=head1 NAME

perlclassguts - Internals of how C<feature 'class'> and class syntax works

=head1 DESCRIPTION

This document provides in-depth information about the way in which the perl
interpreter implements the C<feature 'class'> syntax and overall behaviour.
It is not intended as an end-user guide on how to use the feature. For that,
see L<perlclass>.

The reader is assumed to be generally familiar with the perl interpreter
internals overall. For a more general overview of these details, see also
L<perlguts>.

=head1 DATA STORAGE

=head2 Classes

A class is fundamentally a package, and exists in the symbol table as an HV
with an aux structure in exactly the same way as a non-class package. It is
distinguished from a non-class package by the fact that the
C<HvSTASH_IS_CLASS()> macro will return true on it.

Extra information relating to it being a class is stored in the
C<struct xpvhv_aux> structure attached to the stash, in the following fields:

    HV          *xhv_class_superclass;
    CV          *xhv_class_initfields_cv;
    AV          *xhv_class_adjust_blocks;
    PADNAMELIST *xhv_class_fields;
    PADOFFSET    xhv_class_next_fieldix;
    HV          *xhv_class_param_map;

=over 4

=item *

C<xhv_class_superclass> will be C<NULL> for a class with no superclass. It
will point directly to the stash of the parent class if one has been set with
the C<:isa()> class attribute.

=item *

C<xhv_class_initfields_cv> will contain a C<CV *> pointing to a function to be
invoked as part of the constructor of this class or any subclass thereof. This
CV is responsible for initializing all the fields defined by this class for a
new instance. This CV will be an anonymous real function - i.e. while it has no
name and no GV, it is I<not> a protosub and may be directly invoked.

=item *

C<xhv_class_adjust_blocks> may point to an AV containing CV pointers to each of
the C<ADJUST> blocks defined on the class. If the class has a superclass, this
array will additionally contain duplicate pointers of the CVs of its parent
class. The AV is created lazily the first time an element is pushed to it; it
is valid for there not to be one, and this pointer will be C<NULL> in that
case.

The CVs are stored directly, not via RVs. Each CV will be an anonymous real
function.

=item *

C<xhv_class_fields> will point to a C<PADNAMELIST> containing C<PADNAME>s,
each being one defined field of the class. They are stored in order of
declaration. Note however, that the index into this array will not necessarily
be equal to the C<fieldix> of each field, because in the case of a subclass,
the array will begin at zero but the index of the first field in it will be
non-zero if its parent class contains any fields at all.

For more information on how individual fields are represented, see L</Fields>.

=item *

C<xhv_class_next_fieldix> gives the field index that will be assigned to the
next field to be added to the class. It is only useful at compile-time.

=item *

C<xhv_class_param_map> may point to an HV which maps field C<:param> attribute
names to the field index of the field with that name. This mapping is copied
from parent classes; each class will contain the sum total of all its parents
in addition to its own.

=back

=head2 Fields

A field is still fundamentally a lexical variable declared in a scope, and
exists in the C<PADNAMELIST> of its corresponding CV. Methods and other
method-like CVs can still capture them exactly as they can with regular
lexicals. A field is distinguished from other kinds of pad entry in that the
C<PadnameIsFIELD()> macro will return true on it.

Extra information relating to it being a field is stored in an additional
structure accessible via the C<PadnameFIELDINFO()> macro on the padname. This
structure has the following fields:

    PADOFFSET  fieldix;
    HV        *fieldstash;
    OP        *defop;
    SV        *paramname;
    bool       def_if_undef;
    bool       def_if_false;

=over 4

=item *

C<fieldix> stores the "field index" of the field; that is, the index into the
instance field array where this field's value will be stored. Note that the
first index in the array is not specially reserved. The first field in a class
will start from field index 0.

=item *

C<fieldstash> stores a pointer to the stash of the class that defined this
field. This is necessary in case there are multiple classes defined within the
same scope; it is used to disambiguate the fields of each.

    {
        class C1; field $x;
        class C2; field $x;
    }

=item *

C<defop> may store a pointer to a defaulting expression optree for this field.
Defaulting expressions are optional; this field may be C<NULL>.

=item *

C<paramname> may point to a regular string SV containing the C<:param> name
attribute given to the field. If none, it will be C<NULL>.

=item *

One of C<def_if_undef> and C<def_if_false> will be true if the defaulting
expression was set using the C<//=> or C<||=> operators respectively.

=back

=head2 Methods

A method is still fundamentally a CV, and has the same basic representation as
one. It has an optree and a pad, and is stored via a GV in the stash of its
containing package. It is distinguished from a non-method CV by the fact that
the C<CvIsMETHOD()> macro will return true on it.

(Note: This macro should not be confused with the one that was previously
called C<CvMETHOD()>. That one does not relate to the class system, and was
renamed to C<CvNOWARN_AMBIGUOUS()> to avoid this confusion.)

There is currently no extra information that needs to be stored about a method
CV, so the structure does not add any new fields.

=head2 Instances

Object instances are represented by an entirely new SV type, whose base type
is C<SVt_PVOBJ>. This should still be blessed into its class stash and wrapped
in an RV in the usual manner for classical object.

As these are their own unique container type, distinct from hashes or arrays,
the core C<builtin::reftype> function returns a new value when asked about
these. That value is C<"OBJECT">.

Internally, such an object is an array of SV pointers whose size is fixed at
creation time (because the number of fields in a class is known after
compilation). An object instance stores the max field index within it (for
basic error-checking on access), and a fixed-size array of SV pointers storing
the individual field values.

Fields of array and hash type directly store AV or HV pointers into the array;
they are not stored via an intervening RV.

=head1 API

The data structures described above are supported by the following API
functions.

=head2 Class Manipulation

=head3 class_setup_stash

    void class_setup_stash(HV *stash);

Called by the parser on encountering the C<class> keyword. It upgrades the
stash into being a class and prepares it for receiving class-specific items
like methods and fields.

=head3 class_seal_stash

    void class_seal_stash(HV *stash);

Called by the parser at the end of a C<class> block, or for unit classes its
containing scope. This function performs various finalisation activities that
are required before instances of the class can be constructed, but could not
have been done until all the information about the members of the class is
known.

Any additions to or modifications of the class under compilation must be
performed between these two function calls. Classes cannot be modified once
they have been sealed.

=head3 class_add_field

    void class_add_field(HV *stash, PADNAME *pn);

Called by F<pad.c> as part of defining a new field name in the current pad.
Note that this function does I<not> create the padname; that must already be
done by F<pad.c>. This API function simply informs the class that the new
field name has been created and is now available for it.

=head3 class_add_ADJUST

    void class_add_ADJUST(HV *stash, CV *cv);

Called by the parser once it has parsed and constructed a CV for a new
C<ADJUST> block. This gets added to the list stored by the class.

=head2 Field Manipulation

=head3 class_prepare_initfield_parse

    void class_prepare_initfield_parse();

Called by the parser just before parsing an initializing expression for a
field variable. This makes use of a suspended compcv to combine all the field
initializing expressions into the same CV.

=head3 class_set_field_defop

    void class_set_field_defop(PADNAME *pn, OPCODE defmode, OP *defop);

Called by the parser after it has parsed an initializing expression for the
field. Sets the defaulting expression and mode of application. C<defmode>
should either be zero, or one of C<OP_ORASSIGN> or C<OP_DORASSIGN> depending
on the defaulting mode.

=head3 padadd_FIELD

    #define padadd_FIELD

This flag constant tells the C<pad_add_name_*> family of functions that the
new name should be added as a field. There is no need to call
C<class_add_field()>; this will be done automatically.

=head2 Method Manipulation

=head3 class_prepare_method_parse

    void class_prepare_method_parse(CV *cv);

Called by the parser after C<start_subparse()> but immediately before doing
anything else. This prepares the C<PL_compcv> for parsing a method; arranging
for the C<CvIsMETHOD> test to be true, adding the C<$self> lexical, and any
other activities that may be required.

=head3 class_wrap_method_body

    OP *class_wrap_method_body(OP *o);

Called by the parser at the end of parsing a method body into an optree but
just before wrapping it in the eventual CV. This function inserts extra ops
into the optree to make the method work correctly.

=head2 Object Instances

=head3 SVt_PVOBJ

    #define SVt_PVOBJ

An SV type constant used for comparison with the C<SvTYPE()> macro.

=head3 ObjectMAXFIELD

    SSize_t ObjectMAXFIELD(sv);

A function-like macro that obtains the maximum valid field index that can be
accessed from the C<ObjectFIELDS> array.

=head3 ObjectFIELDS

    SV **ObjectFIELDS(sv);

A function-like macro that obtains the fields array directly out of an object
instance. Fields can be accessed by their field index, from 0 up to the maximum
valid index given by C<ObjectMAXFIELD>.

=head1 OPCODES

=head2 OP_METHSTART

    newUNOP_AUX(OP_METHSTART, ...);

An C<OP_METHSTART> is an C<UNOP_AUX> which must be present at the start of a
method CV in order to make it work properly. This is inserted by
C<class_wrap_method_body()>, and even appears before any optree fragment
associated with signature argument checking or extraction.

This op is responsible for shifting the value of C<$self> out of the arguments
list and binding any field variables that the method requires access to into
the pad. The AUX vector will contain details of the field/pad index pairings
required.

This op also performs sanity checking on the invocant value. It checks that it
is definitely an object reference of a compatible class type. If not, an
exception is thrown.

If the C<op_private> field includes the C<OPpINITFIELDS> flag, this indicates
that the op begins the special C<xhv_class_initfields_cv> CV. In this case it
should additionally take the second value from the arguments list, which
should be a plain HV pointer (I<directly>, not via RV). and bind it to the
second pad slot, where the generated optree will expect to find it.

=head2 OP_INITFIELD

An C<OP_INITFIELD> is only invoked as part of the C<xhv_class_initfields_cv>
CV during the construction phase of an instance. This is the time that the
individual SVs that make up the mutable fields of the instance (including AVs
and HVs) are actually assigned into the C<ObjectFIELDS> array. The
C<OPpINITFIELD_AV> and C<OPpINITFIELD_HV> private flags indicate whether it is
creating an AV or HV; if neither is set then an SV is created.

If the op has the C<OPf_STACKED> flag it expects to find an initializing value
on the stack. For SVs this is the topmost SV on the data stack. For AVs and
HVs it expects a marked list.

=head1 COMPILE-TIME BEHAVIOUR

=head2 C<ADJUST> Phasers

During compiletime, parsing of an C<ADJUST> phaser is handled in a
fundamentally different way to the existing perl phasers (C<BEGIN>, etc...)

Rather than taking the usual route, the tokenizer recognises that the
C<ADJUST> keyword introduces a phaser block. The parser then parses the body
of this block similarly to how it would parse an (anonymous) method body,
creating a CV that has no name GV. This is then inserted directly into the
class information by calling C<class_add_ADJUST>, entirely bypassing the
symbol table.

=head2 Attributes

During compilation, attributes of both classes and fields are handled in a
different way to existing perl attributes on subroutines and lexical
variables.

The parser still forms an C<OP_LIST> optree of C<OP_CONST> nodes, but these
are passed to the C<class_apply_attributes> or C<class_apply_field_attributes>
functions. Rather than using a class lookup for a method in the class being
parsed, a fixed internal list of known attributes is used to find functions to
apply the attribute to the class or field. In future this may support
user-supplied extension attribute, though at present it only recognises ones
defined by the core itself.

=head2 Field Initializing Expressions

During compilation, the parser makes use of a suspended compcv when parsing
the defaulting expression for a field. All the expressions for all the fields
in the class share the same suspended compcv, which is then compiled up into
the same internal CV called by the constructor to initialize all the fields
provided by that class.

=head1 RUNTIME BEHAVIOUR

=head2 Constructor

The generated constructor for a class itself is an XSUB which performs three
tasks in order: it creates the instance SV itself, invokes the field
initializers, then invokes the ADJUST block CVs. The constructor for any class
is always the same basic shape, regardless of whether the class has a
superclass or not.

The field initializers are collected into a generated optree-based CV called
the field initializer CV. This is the CV which contains all the optree
fragments for the field initializing expressions. When invoked, the field
initializer CV might make a chained call to the superclass initializer if one
exists, before invoking all of the individual field initialization ops. The
field initializer CV is invoked with two items on the stack; being the
instance SV and a direct HV containing the constructor parameters. Note
carefully: this HV is passed I<directly>, not via an RV reference. This is
permitted because both the caller and the callee are directly generated code
and not arbitrary pure-perl subroutines.

The ADJUST block CVs are all collected into a single flat list, merging all of
the ones defined by the superclass as well. They are all invoked in order,
after the field initializer CV.

=head2 C<$self> Access During Methods

When C<class_prepare_method_parse()> is called, it arranges that the pad of
the new CV body will begin with a lexical called C<$self>. Because the pad
should be freshly-created at this point, this will have the pad index of 1.
The function checks this and aborts if that is not true.

Because of this fact, code within the body of a method or method-like CV can
reliably use pad index 1 to obtain the invocant reference. The C<OP_INITFIELD>
opcode also relies on this fact.

In similar fashion, during the C<xhv_class_initfields_cv> the next pad slot is
relied on to store the constructor parameters HV, at pad index 2.

=head1 AUTHORS

Paul Evans

=cut