The Perl Toolchain Summit needs more sponsors. If your company depends on Perl, please support this very important event.

NAME

Panda::XS - useful features and typemaps for XS modules.

DESCRIPTION

Panda::XS provides some useful features for XS modules. Also adds default configurable typemaps with most commonly used types. Panda::XS makes it possible for other modules (Perl or XS) to inherit from your XS module. To use it you must have a C++ compiler. Of course most (or all) CPAN modules have private implementation visible only via perl interface (function or method calls). But it is a much better approach to implement functionality in C/C++ classes with public API, make an XS interface, to make it usable from perl and also make you C code visible to other XS modules. This makes it possible for other users to use your C code directly from other XS modules, without perl method/function interface and therefore they can achieve much greater speeds. To make your C code visible to other XS modules when your module is installed, see Panda::Install.

SYNOPSIS

Makefile.PL:

    use strict;
    use Panda::Install 'write_makefile';
    
    write_makefile(
        NAME    => 'MyXS',
        CPLUS   => 1,
        DEPENDS => 'Panda::XS',
    );
    

mytypemap.map:

    MyClass*  T_OPTR
    MyClass3* T_OEXT

MyXS.xs:

    #include <xs/xs.h> /* replaces #include perl.h, ppport.h, XSUB.h, etc... */
    
    ...
    
    # C++ class based object
    MyClass*
    MyClass::new ()
    CODE:
        RETVAL = new MyClass();
        ... // scalar-reference based object. IV value is C pointer address.
    
    void
    MyClass::somefunc ()
    PPCODE:
        THIS->somefunc();
        ...
    
    void
    MyClass::DESTROY ()
    
    # hash and C++ class based object at one time
    using xs::sv_payload_attach;
    using xs::rv_payload;
    MODULE=... PACKAGE=MyClass2
    
    SV*
    new (const char* CLASS)
    CODE:
        RETVAL = sv_bless( newRV_noinc((SV*)newHV()), gv_stashpv(CLASS, GV_ADD) );
        rv_payload_attach(RETVAL, new MyClass2());
    OUTPUT:
        RETVAL
        
    void
    somefunc (SV* OBJ)
    PPCODE:
        if (!SvROK(OBJ)) croak("....");
        MyClass2* THIS = (MyClass2*) rv_payload(OBJ);
        assert(THIS);
        ...
        
    void
    DESTROY (SV* OBJ)
    PPCODE:
        if (!SvROK(OBJ)) croak("....");
        MyClass2* THIS = (MyClass2*) rv_payload(OBJ);
        delete THIS;
        
        
    # ANY-SV and C++ class based object at one time
    MyClass3*
    MyClass3::new ()
    CODE:
        RETVAL = new MyClass2();

    void
    MyClass3::somefunc ()
    PPCODE:
        THIS->somefunc();
        ...

    void
    MyClass3::DESTROY ()
    

MyXS.pm:

    package MyXS;
    use Panda::XS;
    
    our $VERSION = '0.1.3';
    require Panda::XSLoader;
    Panda::XSLoader::bootstrap(); # replacement for XSLoader::load()

TYPEMAP

C types

(u)int(8/16/32/64)_t

Mappings for integers

AV*, HV*, CV*, IO*

Array/Hash/Code/IO references.

    AV*
    get_list ()
    CODE:
        RETVAL = newAV();
        // push values to RETVAL
        
    
    void
    merge (HV* h1, HV* h2)
    PPCODE:
        ...
        
OSV*, OAV*, OHV*, OIO*

Scalar/Array/Hash/IO reference based objects.

    OHV*
    OHV::new ()
    CODE:
        RETVAL = newHV();


    int
    OAV::get_count ()
    CODE:
        RETVAL = *av_fetch(THIS, 1, 1);

Typemap classes (to map your custom types/classes to)

T_OPTR
YOUR_TYPE : T_OPTR([basetype=class], [refcnt], [backref], [wrapper=wrapper_class], [static_cast], [prevent_default_destroy])

IV(SCALAR)-based object with C pointer attached.

    TYPEMAP
    MyClass* T_OPTR
    
    # XS
    
    MyClass*
    MyClass::new ()
    CODE:
        RETVAL = new MyClass();
        
    void
    MyClass::somefunc()
    PPCODE:
        // use 'THIS' (MyClass*)
        
    void
    MyClass::DESTROY ()

The main problem of this method is that one cannot fully inherit from your module (no place to store another class' data). However you can still use inheritance XS->XS->... if your child's XS uses C++ class which inherits from parent's C++ class. In another words, if you still need to hold only single pointer. In this case, set 'basetype' to the name of the most parent C++ class. See HOW TO for details.

See T_OEXT for description of params.

T_OEXT
YOUR_TYPE : T_OEXT([basetype=class], [refcnt], [backref], [wrapper=wrapper_class], [static_cast], [on_svdup=funcname], [prevent_default_destroy])

Extendable object with C pointer attached.

    TYPEMAP
    MyClass* T_OEXT
    
    # XS
    
    MyClass*
    MyClass::new ()
    CODE:
        RETVAL = new MyClass();
        
    
    void
    MyClass::somefunc()
    PPCODE:
        // use 'THIS' (MyClass*)

    void
    MyClass::DESTROY ()
    

This method is much more flexible as it allows for inheritance and even XS -> XS inheritance where each one holds its own C++ object. I.e. it can hold an arbitrary number of pointers.

<T_OEXT C type>::new will create scalar-based object (reference to undef)

    my $obj = MyClass->new; # $obj is RV to undef with C++ MyClass attached
    

Use obj2hv / obj2av to change the object's base to desired type.

    Panda::XS::obj2hv($obj); # $obj is RV to HV with C++ MyClass attached

or

    Panda::XS::obj2av($obj); # $obj is RV to AV with C++ MyClass attached

After that you can store data in $obj->{..} or $obj->[..] as if it was a regular perl object

Parameters

basetype [default is $type]

The most parent C++ class in XS hierarchy. Used as a marker(key) for storing C++ pointer in perl object's magic. Also used for auto typecasting from parent class to child class like this 'THIS = dynamic_cast<$type>(($basetype)stored_ptr)' for input typemaps and from child class to parent class in output typemaps 'stored_ptr = static_cast<$basetype>(RETVAL)'. By default $basetype = $type. If $basetype == $type, then casting is not performed. See HOW TO for details.

refcnt [default is disabled]

Enables automatic refcounting for C++ objects (incrementing on OUTPUT and decrementing on DESTROY). Your C++ class must either subclass panda::RefCounted class or define an interface for getting refcnt, incrementing and decrementing refcounter (see xs::refcnt_get, xs::refcnt_inc, xs::refcnt_dec).

Remember that you HAVE to define DESTROY function even if you don't have anything to do there. Otherwise, refcnt won't be decremented!

    void MyClass::DESTROY ()
backref [default is disabled]

On OUTPUT, tries to cast retval to xs::XSBackref and if it succeedes, returns already existing SV (if any) for that object instead of creating new one every time. As a useful effect this preserves any custom perl classes that might subclass your base class as well as any perl data in object (if upgraded to AV/HV). See HOW TO for more detailed explanation.

IMPORTANT: backref stores backreference from C++ object to SV object. To work properly it syncronizes refcounter between C++ object and SV. Therefore it REQUIRES your object to use panda::RefCounted to work out of the box. If another refcounting class is in use, you must call XSBackref::on_retain() / XSBackref::on_release() every time your object is incremented/decremented.

Example using panda::RefCounted.

    // MyClass - C++ class with panda::RefCounted
    // MyClassXS - objects of this class are created from perl
    class MyClassXS : public XSBackref, public MyClass { ... }
    

Example using custom refcounter.

    // MyClass - C++ class with custom refcounter methods - rcnt_inc, rcnt_dec (they must be virtual to work properly)
    class MyClassXS : public XSBackref, public MyClass {
        virtual void rcnt_inc () const {
            XSBackref::on_retain();
            MyClass::rcnt_inc();
        }
        virtual void rcnt_dec () const {
            XSBackref::on_release();
            MyClass::rcnt_dec();
        }
    }
wrapper [default is none]

Wrapper to use when returning C++ object to perl. A typemap must be defined for that class and it must inherit from T_OWRAPPER typemap class. See HOW TO for examples and more details.

static_cast [default is disabled]

If enabled, INPUT typemap will use static_cast instead of dynamic_cast when downcasting objects. You may need this if your object doesn't have any virtual methods (no virtual table). In any other case, using static_cast is not recommended to avoid programmer's mistakes comming from perl (for example, wrong object passed: with dynamic_cast it will croak, with static_cast undefined behaviour might happen). Don't worry about perfomance, Panda::XS uses panda::dyn_cast instead of default operator. It is 10x-20x faster than dynamic_cast.

on_svdup

Callback (function) name to call on threaded perls when a new thread is created and your object is about to be copied. Function signature is:

If you dont use wrapper

    <basetype> <callbackname> (<basetype> arg);
    

If you use wrapper

    <wrapper_basetype> <callbackname> (<wrapper_basetype> arg);

See "THREADED PERLS SUPPORT" for details.

prevent_default_destroy

Prevents T_OEXT/T_OPTR from inserting object deletion code to the end of XS DESTROY function. Use with caution! See "DESTROY behaviour for T_OEXT and T_OPTR" for details.

Special C variables for OUTPUT

CLASS (required)

Define this variable as either 'const char*' or 'SV*' or 'HV*' and set it to a class name or a class stash (HV*) you want your object to be blessed to. It is done automatically for methods 'new', so that you must not define this variable in 'new' methods. However you can change it's value if you want to.

To receive maximum perfomance, follow these rules (in order they appear):

If you already have class stash, set HV* CLASS.

For example, in an object method, which returns another object, like 'clone':

    HV* CLASS = SvSTASH(SvRV(ST(0)))
    

It's the most effective way for Panda::XS to bless your newly created object. However doing this:

    HV* CLASS = gv_stashpv(classname_str, GV_ADD);
    

won't lead to any perfomance gain over setting const char* CLASS.

If not, but you have a class name in SV*, set SV* CLASS:

For example, in a class method, which creates object.

    SV* CLASS = ST(0);
    

In some cases it runs much faster than setting const char* CLASS (when ST(0) is a shared COW string, since perl 5.21.4).

Don't do it for method 'new' because ExtUtils::ParseXS automatically sets "const char* CLASS = SvPV_nolen(ST(0));" which unfortunately is not the most effective way. Of course it happens only if you "typemap"ed your function (MyClass* MyClass::new ())

If you don't have anything of above, set const char* CLASS

For example, in constructors that are called as functions, not as class methods, like 'uri("http://ya.ru")'

    const char* CLASS = "MyFramework::URI";
self

This variable is automatically defined for all T_OEXT outputs and is set to NULL. If you don't change its value, then a new object (reference to undef) will be created and your RETVAL will be attached to it. If you want to attach your RETVAL to an already existing object, set 'self' variable to some SV* value. Value must be a valid RV to a blessed SV, or an SV itself (in this case an RV to specified SV is created and blessed). Useful for calling super and next methods. See HOW TO for details.

T_OEXT_AV, T_OEXT_HV

HV or AV based object with C pointer attached.

    TYPEMAP
    MyClass* T_OEXT_HV
    
    # XS
    
    MyClass*
    MyClass::new ()
    CODE:
        RETVAL = new MyClass();

    void
    MyClass::somefunc()
    PPCODE:
        // use THIS (MyClass*)

    void
    MyClass::DESTROY ()
    
    # in perl code
    my $obj = MyClass->new; # $obj is a blessed HASHREF
    $obj->somefunc;
    

the above works like

    TYPEMAP
    MyClass* T_OEXT
    
    # XS
    
    MyClass*
    MyClass::new ()
    CODE:
        RETVAL = new MyClass();
        
    void
    MyClass::DESTROY ()

    # in perl code
    my $obj = MyClass->new; # $obj is a blessed SCALARREF (ref to undef)
    Panda::XS::obj2hv($obj); # $obj is a blessed HASHREF

and exactly like

    TYPEMAP
    MyClass* MY_TYPE
    
    OUTPUT
    MY_TYPE : T_OEXT
         if (!self && $var) self = (SV*)newHV();
         
    INPUT
    MY_TYPE : T_OEXT
    
    # XS
    
    MyClass*
    MyClass::new ()
    CODE:
        RETVAL = new MyClass();

    # in perl code
    my $obj = MyClass->new; # $obj is a blessed HASHREF

DESTROY behaviour for T_OEXT and T_OPTR

These typemap classes always adds object deletion code to DESTROY function, even if your XS DESTROY function is non empty.

The code is:

If you use wrapper
    delete THIS_wrapper;
    
If you use refcounting
    xs::refcnt_dec(THIS); // which is THIS->release() for panda::RefCounted
    
If you use shared pointers
    // Code to release shared pointer
    
Nothing of above
    delete THIS;

If you don't want T_OEXT / T_OPTR to insert that for you, set prevent_default_destroy param for typemap class.

Keep in mind that in this case you must by any means avoid memory leaks, i.e. do the same or similar.

HOW TO

CREATE AN XS CLASS HOLDING C++ CLASS OR C STRUCTURE

Here and below by 'XS CLASS' i mean not just XS code with functions, i mean class, objects of which hold a C struct or class.

Create your TYPE in typemap using one of existing typemap classes.

Either classic pointer-based object

    MyClass* T_OPTR
    

or more flexible "attached pointer" object

    MyClass* T_OEXT
    
Use your type in XS code
    MyClass*
    MyClass::new ()
    CODE:
        RETVAL = new MyClass();
    OUTPUT:
        RETVAL

    void
    MyClass::somefunc ()
    PPCODE:
        THIS->somefunc(); // THIS is a variable of type MyClass*
        ...

    void
    MyClass::DESTROY ()
Use your XS class from perl
    my $obj = new MyClass();
    $obj->somefunc();
    

Note how your object looks like inside with T_OPTR

    DB<2> x $obj
      0  MyClass=SCALAR(0x8179dc4e0)
         -> 34732484608

or with T_OEXT

    DB<5> x $obj
      0  MyClass=SCALAR(0x8179e16a8)
         -> undef

CREATE AN EXTENDABLE XS CLASS HOLDING C++ CLASS OR C STRUCTURE

'Extendable' means that we can create child class in perl and hold additional data in $self as it was a regular perl object. It's impossible to implement that using T_OPTR as it makes perl object not applicable for holding anything but C pointer. To do this, we must use T_OEXT typemap class. By default, T_OEXT creates reference to undef as an object base. Most of the time we want it to be a hash reference. There are 2 ways we can achieve that - either use T_OEXT_HV as typemap class (it creates RV to HV as an object base instead of RV to undef) or upgrade already created object to RV/HV from perl.

Solution with T_OEXT_HV

In typemap change T_OEXT to T_OEXT_HV:

    MyClass* T_OEXT_HV
    

That's all! Lets see the results

    DB<6> x new MyClass()
      0  MyClass=HASH(0x81794bf48)
           empty hash
    
Solution with object upgrading

Lets do the upgrade in child class.

    package MyClassChild;
    use parent 'MyClass';
    
    sub new {
        my $self = shift->SUPER::new(@_);
        Panda::XS::obj2hv($self);
        return $self;
    }
    

Lets see the results

    DB<5> x new MyClass()
      0  MyClass=SCALAR(0x81794bd80)
         -> undef
    DB<6> x new MyClassChild()
      0  MyClassChild=HASH(0x81794a318)
           empty hash

CREATE XS CLASS HIERARCHY THAT CORRESPONDS TO C++ CLASS HIERARCHY

Image that you need to port a C framework to perl and you want to leave it's classes structure transparent. For example, there is a C framework Foo that we need to port to perl. It has 2 classes ClassA and ClassB.

    class ClassA {
    private:
        int _propA;
    public:
        ClassA (int propA = 0) : _propA(propA) {}
        
        int  propA ()        { return _propA; }
        void propA (int val) { _propA = val; }
        
        virtual ClassA* clone () { return new ClassA(_propA); }
    }

    class ClassB : public ClassA {
    private:
        int _propB;
    public:
        ClassB (int propA = 0, int propB = 0) : ClassA(propA), _propB(propB) {}

        int  propB ()        { return _propB; }
        void propB (int val) { _propB = val; }

        virtual ClassA* clone () { return new ClassB(_propA, _propB); }
    }

We want to create 2 perl classes - Foo::A which holds C ClassA class and Foo::B which holds C ClassB class. Of course Foo::B should inherit from Foo::A. What do we have to do? Almost nothing!

At first, create a custom typemap class for ClassA/ClassB hierarchy, inherit it from T_OEXT and set parameter basetype to the most parent class which will be visible in perl. In our case it is ClassA (visible via Foo::A).

    ClassA* T_FOO
    ClassB* T_FOO
    
    OUTPUT
    T_FOO : T_OEXT(basetype=ClassA*)
    INPUT
    T_FOO : T_OEXT(basetype=ClassA*)
    

Remember that basetype is required in case of C++ inheritance, otherwise you won't be able to call Foo::A methods on a Foo::B object.

Then lets create an XS code.

Foo.xs (Here and after we use extended Panda::Install XS syntax to make XS functions look like C functions)

    MODULE = Foo                PACKAGE = Foo::A
    PROTOTYPES: DISABLE

    ClassA* ClassA::new (int propA) {
        RETVAL = new ClassA(propA);
    }

    int ClassA::propA (SV* newval = NULL) { // 'THIS' is a ClassA* pointer, even if real perl object is of Foo::B class.
        if (newval) THIS->propA(SvIV(newval));
        RETVAL = THIS->propA();
    }
    
    ClassA* ClassA::clone () {
        HV* CLASS = SvSTASH(SvRV(ST(0)));
        RETVAL = THIS->clone();
    }
    
    void ClassA::DESTROY ()

    MODULE = Foo                PACKAGE = Foo::B
    PROTOTYPES: DISABLE

    ClassB* ClassB::new (int propA, int propB) {
        RETVAL = new ClassB(propA, propB);
    }

    int ClassB::propB (SV* newval = NULL) { // 'THIS' is a ClassB* pointer
        if (newval) THIS->propB(SvIV(newval));
        RETVAL = THIS->propB();
    }
    

Somewhere in Perl

    my $a = new Foo::A(123);
    say $a->propA;
    $a->propA(321);
    say ref $a->clone; # Foo::A
    
    my $b = new Foo::B(1,2);
    say $b->propA;
    say $b->propB;
    say ref $b->clone; # Foo::B
    

Using these technics you can port any C++ framework saving its original OOP structure, keeping your XS wrapper as thin as possible and therefore get maximum perfomance.

PASSING C CLASSES/STRUCTURES AS A PARAMETERS TO FUNCTIONS OF OTHER CLASSES.

It works out of the box. Lets add a third class to previous example, ClassC:

    class ClassC {
    public:
        ClassC () {}
        
        int calculate (ClassA* arg) { return arg->propA * arg->propA; }
    }
    

Add it to a typemap

    ClassC* T_OEXT
    

Add XS for ClassC

    MODULE = Foo                PACKAGE = Foo::C
    PROTOTYPES: DISABLE

    ClassC* ClassC::new ()

    int ClassC::calculate (ClassA* arg)
    
    void ClassC::DESTROY ()

Just do it in perl as you would in C++, typemaps will do all the work for you

    my $a = new Foo::A(10);
    my $b = new Foo::B(20, 30);
    my $c = new Foo::C;
    say $c->calculate($a); # 100;
    say $c->calculate($b); # 400;
    

Remember, everywhere you pass/receive Foo::A, Foo::B etc, in XS you receive/return ClassA*, ClassB*, etc

INHERITING XS CLASS FROM ANOTHER WITH DIFFERENT UNDERLYING DATA.

Sometimes source code of parent class is unavailable, and therefore you can't make a C++ child class as described above. However you can still provide such inheritance on perl level. All you need to do is to call SUPER in constructor.

Suppose we have two C++ classes Bow and Milk. And we already have an XS class for Bow. We want to implement XS for Milk and inherit it from Bow. As we want 2 different C data pointers in one perl object, we should either not provide a basetype parameter in typemap, or set it to something different from Bow*. By default (if not provided), basetype is set to the value of final type name. However we need to call SUPER::new, and attach our data to existing SV, rather than creating new one (otherwise Bow's data would be lost). To do this, set 'self' variable to SV object to attach to. Unfortunately, as of current perl (5.21.5), it's impossible to call SUPER/next from XS. To call SUPER/next, use the API provided by Panda::XS C++ interface - call_super.

Typemap:

    Bow*  T_OEXT
    Milk* T_OEXT
    

Milk.pm:

    package Foo::Milk;
    use parent 'Some::Bow';

XS:

    MODULE = Foo                PACKAGE = Foo::Milk
    PROTOTYPES: DISABLE

    Milk* Milk::new (...) {
        self = xs::call_super(cv, &ST(0), items);
        RETVAL = new Milk();
    }

    void Milk::something () {
        // 'THIS' is a Milk* pointer
    }
    
    Milk* Milk::clone () { // cloning just Milk* without Bow* does NOT make sense. So we must call SUPER::clone
        HV* CLASS = SvSTASH(SvRV(ST(0)));
        self = xs::call_super(cv, &ST(0), items);
        RETVAL = THIS->clone();
    } // new Foo::Milk object returned containing cloned Milk* and Bow* classes.
    
    void Milk::DESTROY () {
        xs::call_super(cv, &ST(0), items, G_DISCARD);
    }
    

CREATE AN XS CLASS WHICH IS SUITABLE FOR MULTIPLE IHERITANCE / C3 CLASS MIXIN

All you need to do is to call next::method or maybe::next::method from your construtor. To do this, use call_next_method() or call_next_maybe() from xs:: namespace.

Typemap:

    MyPlugin*  T_OEXT
    

Perl:

    package Foo::MyPlugin;
    use mro 'c3';    
    

XS:

    MODULE = Foo                PACKAGE = Foo::MyPlugin
    PROTOTYPES: DISABLE

    MyPlugin* MyPlugin::new (...) {
        self = xs::call_next_maybe(cv, &ST(0), items);
        RETVAL = new MyPlugin();
    }

    uint32_t MyPlugin::something () {
        // 'THIS' is a MyPlugin* pointer
        RETVAL = THIS->something;
    }
    
    MyPlugin* MyPlugin::clone () {
        HV* CLASS = SvSTASH(SvRV(ST(0)));
        self = xs::call_next_maybe(cv, &ST(0), items);
        RETVAL = THIS->clone();
    } // new Foo::MyPlugin object returned containing cloned MyPlugin* and all data of other classes
    
    void MyPlugin::DESTROY () {
        xs::call_next_maybe(cv, &ST(0), items, G_DISCARD);
    }
    

Perl:

    package MyServer;
    use parent qw/Foo::MyPlugin Foo::Milk/;
    
    my $obj = new MyServer();
    

CREATE AN XS CLASS HOLDING C++ CLASS AND ADDITIONAL C INFO

Sometimes you need to hold additional info in XS class, but you can't or don't want to create additional fields in your C++ class. Let's see an example. Assume we have a 'Driver' C++ class that we want to port. We'll make Foo::Driver perl class.

    #TYPEMAP
    Driver* T_OEXT
    
    #XS
    
    Driver* Driver::new (int arg1, char* arg2) {
        RETVAL = new Driver(arg1, arg2);
    }
    
    int Driver::category () {
        RETVAL = THIS->category;
    }
    
    char* Driver::name () {
        RETVAL = THIS->name;
    }

    void Driver::ban (int why) {
        THIS->ban(why);
    }
    
    HV* Driver::get_history () {
        DriverHistory* history = THIS->get_history();
        RETVAL = newHV();
        // fill RETVAL with data from DriverHistory*
    }

    void
    Driver::DESTROY ()
    

Now imagine that we want to count, how many times a driver has been banned for particular object. But Driver class doesn't have any field to hold this value. Adding it to Driver, even if you can, would be a bad decision as it shouldn't know anything about your XS class behaviour. Also filling up HV every time 'get_history' is called, could be quite expensive, so that we also would like to cache once built result somewhere. How to do that?

Solution 1: create Driver's child class.
    class DriverXS : public Driver {
        public:
        
        int counter;
        HV* cached_history;
        
        DriverXS (int arg1, char* arg2) : counter(0), cached_history(NULL), Driver(arg1, arg2) {}
        
        ~DriverXS () {
            SvREFCNT_dec(cached_history);
        }
    }

Now use your child class instead of Driver

    #TYPEMAP
    DriverXS* T_OEXT
    
    #XS
    
    DriverXS* DriverXS::new (int arg1, char* arg2) {
        RETVAL = new DriverXS(arg1, arg2);
    }
    
    int DriverXS::category () {
        RETVAL = THIS->category;
    }
    
    char* DriverXS::name () {
        RETVAL = THIS->name;
    }

    void DriverXS::ban (int why) {
        THIS->counter++;
        THIS->ban(why);
    }
    
    HV* DriverXS::get_history () {
        if (!THIS->cached_history) {
            DriverHistory* history = THIS->get_history();
            THIS->cached_history = newHV();
            // fill cached_history with data from DriverHistory*
        }
        RETVAL = THIS->cached_history;
    }

    void
    DriverXS::DESTROY ()
    

Although this will work out for our case, this approach may be unacceptable. For example imagine we have DriverStorage in C++ lib and class method get_driver_by_id. How can we port it? Lets create XS class Foo::DriverStorage. And method get_driver_by_id:

    DriverXS* DriverStorage::get_driver_by_id (int id) {
        RETVAL = THIS->get_driver_by_id(id);
    }
    

But wait a minute, our c++ lib's get_driver_by_id returns Driver* of course, as it doesn't know anything about DriverXS. And even if we stored only DriverXS* objects in DriverStorage, some C++ code might have created a simple Driver* object and store it in DriverStorage. So the example above won't even compile. The only way is to create a copy DriverXS* from Driver*.

    DriverXS* DriverStorage::get_driver_by_id (int id) {
        Driver* driver = THIS->get_driver_by_id(id);
        if (!driver) XSRETURN_UNDEF;
        RETVAL = new DriverXS(driver->category, driver->name); // or new DriverXS(driver) if have such constructor
    }
    

But it might be expensive and not every object can be copied (for example, socket handle classes).

Solution 2: create a wrapper, and wrap Driver objects into it.
    class XSDriverWrapper {
        public:
        
        Driver* obj;
        int counter;
        HV* cached_history;
        
        XSDriverWrapper (Driver* driver) : counter(0), cached_history(NULL), obj(driver) {}
        
        ~XSDriverWrapper () {
            SvREFCNT_dec(cached_history);
            delete obj;
        }
    }
    
    #TYPEMAP
    XSDriverWrapper* T_OEXT
    
    #XS
    XSDriverWrapper* XSDriverWrapper::new (int arg1, char* arg2) {
        RETVAL = new XSDriverWrapper(new Driver(arg1, arg2));
    }
    
    int XSDriverWrapper::category () {
        RETVAL = THIS->obj->category;
    }
    
    char* XSDriverWrapper::name () {
        RETVAL = THIS->obj->name;
    }

    void XSDriverWrapper::ban (int why) {
        THIS->counter++;
        THIS->obj->ban(why);
    }
    
    HV* XSDriverWrapper::get_history () {
        if (!THIS->cached_history) {
            DriverHistory* history = THIS->obj->get_history();
            THIS->cached_history = newHV();
            // fill cached_history with data from DriverHistory*
        }
        RETVAL = THIS->cached_history;
    }

    void
    XSDriverWrapper::DESTROY ()

    ...
    XSDriverWrapper* DriverStorage::get_driver_by_id (int id) {
        Driver* driver = THIS->get_driver_by_id(id);
        if (!driver) XSRETURN_UNDEF;
        RETVAL = new XSDriverWrapper(driver);
    }
    

This approach is much more flexible as it allows you to return perl objects even for c++ objects who wasn't created in XS class. But our code is not good enough for now, because it's very annoying to write THIS->obj. Moreover, if you share your module with someone, they will have to know about your wrapper to work with your XS objects. For example, image you've uploaded Foo to CPAN. Someone is using it:

    write_makefile(
        ...
        BIN_DEPS => ['Foo'], # make C headers, and typemaps of Foo visible to my XS/C code
    );
    
    # some user's XS
    
    void
    change_drivers_name (XSDriverWrapper* driver_wrapper, const char* newname)
    PPCODE:
        Driver* driver = driver_wrapper->obj;
        driver->set_name(newname);
        
        
    # perl code
    my $driver = Foo::Driver->new(1, "vasya");
    Bar::change_drivers_name($driver, "petya");
    say $driver->name;
    

No one should know the details of your XS module (about the fact that you wrap real Driver into XSDriverWrapper). To do that, we need to implement all the details in typemap, not in XS functions itself. We will need to create our own typemap class.

    #TYPEMAP
    XSDriverWrapper* XT_FOO_DRIVER_WRAPPER
    Driver*          XT_FOO_DRIVER
    
    OUTPUT
    
    XT_FOO_DRIVER_WRAPPER : T_OWRAPPER
        $arg = new XSDriverWrapper($var);
        
    XT_FOO_DRIVER : T_OEXT(wrapper=XSDriverWrapper*)
        
    INPUT
    XT_FOO_DRIVER_WRAPPER : T_OWRAPPER
        $var = $arg->obj;
    
    XT_FOO_DRIVER : T_OEXT(wrapper=XSDriverWrapper*)
        

Here we use wrapper feature of T_OEXT/T_OPTR typemap classes. It allows us to wrap an object into another and make it fully transparent to XS code. Variables in XS code ('THIS' for example) will point to the object itself, not the wrapper. If you need access to the wrapper, use _wrapper extension, for example THIS_wrapper.

Now our code looks ok:

    Driver* Driver::new (int arg1, char* arg2) {
        RETVAL = new Driver(arg1, arg2);
    }
    
    int Driver::category () {
        RETVAL = THIS->category;
    }
    
    char* Driver::name () {
        RETVAL = THIS->name;
    }

    void Driver::ban (int why) {
        THIS_wrapper->counter++;
        THIS->ban(why);
    }
    
    HV* Driver::get_history () {
        if (!THIS_wrapper->cached_history) {
            DriverHistory* history = THIS->get_history();
            THIS_wrapper->cached_history = newHV();
            // fill cached_history with data from DriverHistory*
        }
        RETVAL = THIS_wrapper->cached_history;
    }

    void Driver::DESTROY ()

    ...
    Driver* DriverStorage::get_driver_by_id (int id) {
        RETVAL = THIS->get_driver_by_id(id);
    }
    

Now the implementation details are hidden from users of your module. And the third-party example above will now look much clearer:

    void
    change_drivers_name (Driver* driver, const char* newname)
    PPCODE:
        driver->set_name(newname);

Users of your module no longer see XSDriverWrapper.

NOTE: when using wrapper feature of T_OEXT/T_OPTR, your wrapper's typemap class must inherit from T_OWRAPPER. Don't use wrapper's typemap in XS code, it won't work!

    void XSDriverWrapper::myfunc (XSDriverWrapper* other) { // WRONG!!! WON'T COMPILE!
        THIS->counter = other->counter;
    }
    

You should always use the object itself and access wrapper via variable ${var}_wrapper.

    void Driver::myfunc (Driver* other) { // CORRECT
        THIS_wrapper->counter = other_wrapper->counter;
    }

REFERENCE COUNTING

While you only have single XS classes without any relationships with each other, everything's ok. Problems begin in the following example (we use simple (without wrapper) Driver* class and XS from previous examples):

    class Insurance {
        private:
        Driver* _first_driver;
        Driver* _second_driver;
        
        public:
        Insurance () {
            _first_driver = new Driver(1, "john");
            _second_driver = new Driver(2, "mike");
        }
        
        Driver* first_driver  ()            { return _first_driver; }
        Driver* second_driver ()            { return _second_driver; }
        void    first_driver  (Driver* val) { _first_driver = val; }
        void    second_driver (Driver* val) { _second_driver = val; }
        
        ~Insurance () {
            delete _first_driver;
            delete _second_driver;
        }
    }
    
    #XS
    
    Insurance* Insurance::new () {
        RETVAL = new Insurance();
    }
    
    Driver* Insurance::first_driver (Driver* newval = NULL) : ALIAS(second_driver=1) {
        if (newval) {
            if (ix == 0) THIS->first_driver(newval);
            else         THIS->second_driver(newval);
            XSRETURN_UNDEF;
        }
        const char* CLASS = "Foo::Driver";
        if (ix == 0) RETVAL = THIS->first_driver();
        else         RETVAL = THIS->second_driver();
    }
    
    void Insurance::DESTROY ()
    

So where is the problem? Right here:

    #perl code
    my $insurance = new Foo::Insurance;
    say $insurance->first_driver->name; # prints "john"
    say $insurance->first_driver->name; # OOPS. core dump or undefined behaviour
    

When first_driver() is called for the first time, it creates perl object and attaches Driver* object to it. After first say(), the temporary perl variable holding Driver perl object no longer needed and destroyed. Destructor (DESTROY) of Driver's XS class destroys Driver* object as well. Pointer to that freed object still remains in Insurance* object and bad things happen. Ok, lets stop deleting Driver* object from Driver::DESTROY(), i.e. remove this function:

    void
    Driver::DESTROY ()
    

In this case we get a memory leak for this type of code:

    for (1..1000) {
        my $driver = new Foo::Driver(1, "myname");
    }
    

How can we avoid deletion in first case and memory leak in second?

We need to add a reference counter to our objects!

    class Driver {
        private:
        mutable int _refcnt;
        ....
        Driver (...) : _refcnt(0) ... { ... }
        void retain () const { _refcnt++; }
        void release  () const { if (--_refcnt <= 0) delete this; }
        virtual ~Driver () {...}
    }
    
    class Insurance {
        ...
        void first_driver  (Driver* val) {
            if (_first_driver) _first_driver->release();
            _first_driver = val;
            if (_first_driver) _first_driver->retain();
        }
    }
    
    #XS
    
    Driver* Driver::new (...) {
        RETVAL = new Driver(...);
        RETVAL->retain();
        ...
    }
    
    void Driver::DESTROY () {
        THIS->release();
    }
    
    Driver* Insurance::first_driver (Driver* newval = NULL) : ALIAS(second_driver=1) {
        ...
        if (ix == 0) RETVAL = THIS->first_driver();
        else         RETVAL = THIS->second_driver();
        if (RETVAL) RETVAL->retain();
    }    

We can implement it in typemap to hide these details from XS code.

    #TYPEMAP
    Driver* XT_FOO_DRIVER
    
    OUTPUT
    
    XT_FOO_DRIVER : T_OEXT
        $var->retain();
        
    INPUT
    
    XT_FOO_DRIVER : T_OEXT
        DESTROY { $var->release(); }
    
    #XS
    
    Driver* Driver::new (...) {
        RETVAL = new Driver(...);
        ...
    }
    
    void Driver::DESTROY ()
    
    Driver* Insurance::first_driver (Driver* newval = NULL) : ALIAS(second_driver=1) {
        ...
        if (ix == 0) RETVAL = THIS->first_driver();
        else         RETVAL = THIS->second_driver();
    }    

Note, however, that you need to retain() and release() objects in C++ classes code itself properly as in any other refcounted system.

T_OEXT/T_OPTR can do all of these automatically if you set refcnt param. In this case your class must either subclass panda::RefCounted class which provides refcnt/retain/release functionality or add an API for typemap.

Here is an example of our case with refcnt feature.

    #TYPEMAP
    Driver* XT_FOO_DRIVER
    
    OUTPUT
    
    XT_FOO_DRIVER : T_OEXT(refcnt)
        
    INPUT
    
    XT_FOO_DRIVER : T_OEXT(refcnt)
    
    # XS and C++ class remains the same
    
    # somewhere in our code we must define an API
    inline int32_t xs::refcnt_get (const Driver* var) { return var->refcnt(); }
    inline void    xs::refcnt_inc (const Driver* var) { var->retain(); }
    inline void    xs::refcnt_dec (const Driver* var) { var->release(); }
    

Instead of adding refcnt functionally to your Driver class as we did, you can just inherit from panda::RefCounted class. In this case you should not define any API functions.

    class Driver: virtual public RefCounted {
        ...
    }
    
    # all code is the same

Remember: if you use wrapper, it should NOT have a refcounter. Make the underlying object itself refcounted. Wrappers are stricly bound to their SV object and should live in sync with it.

USING SHARED POINTERS

The second solution of the problem above is shared pointers. It allows you to solve this problem even if you can't change sources of C++ classes.

You can use either STL's std::shared_ptr or panda::shared_ptr (STL compatible shared pointer which doesn't require C++11).

T_OEXT/T_OPTR automatically detects if you use shared pointers. Everything will work as you expect.

    typedef shared_ptr<Driver> DriverSP;
    
    #typemap
    DriverSP T_OEXT
    
    #code
    
    class Driver { // we don't need to add anything
        private:
        Driver (...) : ... { ... }
        virtual ~Driver () {...}
    }
    
    class Insurance {
        private:
        DriverSP _first_driver;
        public:
        DriverSP first_driver () {
            return _first_driver;
        }
        void first_driver (DriverSP driver) {
            _first_driver = driver;
        }
    }
    
    #XS
    
    DriverSP new (SV* CLASS, ...) {
        RETVAL = DriverSP(new Driver(...));
        ...
    }
    
    void DESTROY (DriverSP THIS)
    
    ...
    
    DriverSP Insurance::first_driver (DriverSP newval = DriverSP()) : ALIAS(second_driver=1) {
        ...
        if (ix == 0) RETVAL = THIS->first_driver();
        else         RETVAL = THIS->second_driver();
    }

You might have noted several things:

DriverSP T_OEXT

Typemap class is bound to shared pointer's type, not to an object's type. Therefore you must set a shared pointer to OUTPUT variable (RETVAL), and you will receive a shared pointer in INPUT args

void DESTROY (DriverSP THIS)

Shared pointers are passed by value, not by pointer. See how we receive an INPUT arg 'THIS':

    const char* name (DriverSP THIS)  // CORRECT
    const char* name (DriverSP* THIS) // WRONG!!!
    const char* DriverSP::name ()     // WRONG!!!
                

The third example won't work because for ExtUtils::ParseXS it's just an alias for

    const char* name (DriverSP* THIS)
    

:((

Using shared pointers introduces slight overheat, as it have to alloc additional memory per object. However, see next paragraph to find out how you can remove this overheat while still using shared pointers.

Remember: if you use wrapper, don't use it via shared pointers, make shared pointers for objects itself and hold it in wrapper.

USING panda::shared_ptr<> FOR CLASSES WHICH have retain/release methods

If you decided to use panda::shared_ptr instead of std::shared_ptr, you can get some benefit if you inherit your class from panda::RefCounted or by any other means implement refcounting via retain/release methods. panda::shared_ptr automatically detects classes with retain/release and works differently.

Firstly, it doesn't malloc anything, as your class already has a refcounter. It just calls retain() and release().

Secondly, as you know, creating second std::shared_ptr for an object is a critical mistake. However, for our case it isn't true. You can freely create second, third, etc shared_ptr for any object with retain/release methods without losing reference counter.

Thirdly, sizeof(panda::shared_ptr<ClassWhichHasRetainRelease)> is just one pointer.

That means you can freely interchange MyClass* with MyClassSP and vice versa.

For example:

    typedef shared_ptr<Driver> DriverSP;
    
    #typemap
    Driver*  T_OEXT(refcnt)
    DriverSP T_OEXT
    
    #code
    
    class Driver : virtual public panda::RefCounted {
        private:
        Driver (...) : ... { ... }
        virtual ~Driver () {...}
    }
    
    class Insurance {
        private:
        DriverSP _first_driver;
        public:
        DriverSP first_driver () {
            return _first_driver;
        }
        void first_driver (DriverSP driver) {
            _first_driver = driver;
        }
    }
    
    #XS
    
    Driver* Driver::new (...) {
        RETVAL = new Driver(...);
        ...
    }
    
    void Driver::DESTROY ()
    
    ...
    
    # See how we receive DriverSP, while Driver object is T_OEXT(refcnt) (i.e. Driver*)
    # It would be a fatal logical error doing this for shared pointers which wraps non-refcounted objects.
    DriverSP Insurance::first_driver (DriverSP newval = DriverSP()) : ALIAS(second_driver=1) {
        ...
        if (ix == 0) RETVAL = THIS->first_driver();
        else         RETVAL = THIS->second_driver();
    }    

Moreover you don't need DriverSP T_OEXT typemap at all, as it would automatically convert it to and from shared pointer.

And mainly remember, do all of the above in this paragraph ONLY if your class implements refcounting. Otherwise you would split the reference counter into many instances and bad things would surely happen.

USING BACKREFERENCES (PRESERVING ORIGINAL PERL CLASS&DATA)

Consider the following example (using our lovely Driver and Insurance classes with refcnt):

    my $insurance = new Foo::Insurance;
    my $driver = new Foo::Driver(1, 'smith');
    $insurance->first_driver($driver);
    my $same_driver = $insurance->first_driver;
    my $same_driver2 = $insurance->first_driver;
    say "SAME" if $driver eq $same_driver;
    say "SAME" if $same_driver eq $same_driver2;
    

Nothing is printed in our example, because every time we return Driver* to perl, typemap creates a new perl object and associates it with C++ object. Of course those 2 objects hold the same C++ object, so most of the time it's not a problem at all.

However if you plan that users of your module will extend your classes, the problem appears:

    package MyDriver;
    use parent 'Foo::Driver';

    sub new {
        my $self = shift->SUPER::new(@_);
        Panda::XS::obj2hv($self);
        $self->{id} = 123;
    }
    
    sub id { $_[0]->{id} }
    
    package main;
    
    my $insurance = new Foo::Insurance;
    my $driver = new MyDriver(1, 'smith');
    $insurance->first_driver($driver);
    my $same_driver = $insurance->first_driver;
    say ref $same_driver; # prints 'Foo::Driver';

Perl object returned is of a default class (without custom data created in MyDriver::new()), because our XS function Driver* Insurance::first_driver says const char* CLASS = "Foo::Driver". And that's generally ok, because it doesn't know anything about your MyDriver perl class which was the owner of the C++ Driver* object, it only holds C++ Driver* objects which are the same for both MyDriver and the default Foo::Driver.

To link C++ object to a particular SV object we must use backref technique. To do this, make your C++ object extend xs::XSBackref class and set backref parameter for typemap class.

    class XSDriver : public Driver, public xs::XSBackref {
        public:
        XSDriver (int arg1, char* arg2) : Driver(arg1, arg2) {}
    }
    
    # TYPEMAP
    Driver* XT_FOO_DRIVER
    
    OUTPUT
    
    XT_FOO_DRIVER : T_OEXT(basetype=Driver*, refcnt, backref)
    
    INPUT

    XT_FOO_DRIVER : T_OEXT(basetype=Driver*, refcnt, backref)
    
    # XS
    Driver* Driver::new (int arg1, char* arg2) {
        RETVAL = new XSDriver(arg1, arg2);
    }
    

That's all! Check it out now:

    my $insurance = new Foo::Insurance;
    my $driver = new MyDriver(1, 'smith');
    $insurance->first_driver($driver);
    my $same_driver = $insurance->first_driver;
    say ref $same_driver; # prints 'MyDriver';
    say $same_driver->id; # prints 10
    

You can even loose all references to perl object from perl:

    my $insurance = new Foo::Insurance;
    my $driver = new MyDriver(1, 'smith');
    $insurance->first_driver($driver);
    undef $driver;
    my $same_driver = $insurance->first_driver;
    say ref $same_driver; # prints 'MyDriver';
    say $same_driver->id; # prints 10

As you can see it won't die until there are any references from perl or C++. That's because XSBackref syncronizes perl's refcounter and C++'s refcounter.

As a side effect, if using backref, performance of such methods increases, because perl object is no longer created every time you return C++ object to perl.

Note how we created XSDriver to add XSBackref to parent classes. We had an option to change Driver class itself and add XSBackref to its parents, however it wouldn't be a good decision, because Driver is a pure C class and should not know anything about perl.

Keep in mind that not every C++ object in our example extends XSBackref. For example, those 2 who are created in Insurance's constructor are just Driver*. Typemap automatically detects objects which don't extend XSBackref and reverts to default behaviour (without backref param). In this case const char* CLASS in XS method is in effect, and perl objects of that class will be returned every time you call method (like it would without backref param). For example:

    my $insurance = new Foo::Insurance; # fills first and second driver in constructor with base Driver* objects
    my $driver = $insurance->first_driver;
    my $same_driver = $insurance->first_driver;
    say ref $driver; # Foo::Driver
    say $driver eq $same_driver; # false
    $driver->name('abcd');
    say $driver->name eq $same_driver->name; # true, as they are still holding the same C++ object

In order to use backref your class MUST use refcounting, because backreferencing is impossible without reference counting. If you use panda::RefCounted, you have nothing to do additionally. If you use other refcounting class, you have to call XSBackref::on_retain() and XSBackref::on_release() every time your object is retained or released. If you use panda::RefCounted, you should use virtual inheritance.

If you use wrapper, the object itself must extend XSBackref, NOT the wrapper because it doesn't make sense.

If you support threaded perls, you must choose either CLONE_SKIP strategy or clone strategy. Sharing C++ object between threads is impossible because backref cannot point to many different SV objects (each for particular interpreter). See "THREADED PERLS SUPPORT" for details.

C++ CLASSES

panda::RefCounted

A base class for refcounting.

SYNOPSIS

    #include <panda/refcnt.h>
    
    class MyClass : virtual public RefCounted { ... }
    
    MyClass* obj = new MyClass(...);
    obj->retain();
    obj->release();
    shared_ptr<MyClass> myobj(obj);

[protected] RefCounted ()

Initializes object with refcnt = 0. You will need to call retain() if you want to hold this object. Call release() if you don't need it anymore.

void retain () const

Increments refcounter.

void release () const

Decrements refcounter and deletes the object if refcounter <= 0.

int32_t refcnt () const

Returns refcount.

virtual void on_retain () const

If you define this method in your class, it will be called on every retain().

virtual void on_release () const

If you define this method in your class, it will be called on every release().

template <class T> panda::shared_ptr<T>

API is the same as for std::shared_ptr.

xs::XSBackref

Mix-class for XS backreferences. See HOW TO. You will also have to enable backref param in typemap.

SV* perl_object

Public property where original perl object is stored. It is not an RV to object, it is a blessed object itself.

void on_retain ()

If you have custom refcounter, call this method when your object is retained.

void on_release ()

If you have custom refcounter, call this method when your object is released.

C FUNCTIONS

Functions marked with [pTHX] must receive aTHX as a first arg.

panda::dynamic_pointer_cast/const_pointer_cast/static_pointer_cast

The same as std::*_pointer_cast for panda::shared_ptr and std::shared_ptr.

panda::dyn_cast<>()

Same as dynamic_cast, but has a constant execution time, MUCH faster than dynamic_cast. All of the Panda::XS typemaps use it.

PAYLOAD FUNCTIONS

void sv_payload_attach (SV* sv, void* ptr, const payload_marker_t* marker = NULL) [pTHX]

Attach payload ptr to sv. Markers allow you to store multiple payloads in a single SV. Marker doesn't need to be initialized somehow, its just a valid memory pointer. Usually its a static global variable. If no marker provided, default one is used (global to the whole program).

bool sv_payload_exists (const SV* sv, const payload_marker_t* marker = NULL) [pTHX]

Returns true if sv has any payload for marker.

void* sv_payload (const SV* sv, const payload_marker_t* marker = NULL) [pTHX]

Returns payload attached to sv.

int sv_payload_detach (SV* sv, const payload_marker_t* marker = NULL) [pTHX]

Removes payload from sv and returns number of payloads removed.

void rv_payload_attach (SV* rv, void* ptr, const payload_marker_t* marker = NULL) [pTHX]

bool rv_payload_exists (SV* rv, const payload_marker_t* marker = NULL) [pTHX]

void* rv_payload (SV* rv, const payload_marker_t* marker = NULL) [pTHX]

int rv_payload_detach (SV* rv, const payload_marker_t* marker = NULL) [pTHX]

Same as sv_* but operates on sv referenced by a given rv. Doesn't check if a given RV is valid.

SVPayloadMarker<T>::marker

Static marker for particular class. Used internally by T_OEXT typemap class.

SVPayloadMarker<T>::get(on_svdup_t svdup_callback = NULL)

Same as above but returns marker which attaches svdup_callback to magic.

svdup_callback is a callback to run on threaded perls when a thread is created. Called for each object and receives magic struct attach to an object with current marker. Return value is ignored. Used to clone or do something other with payload attached to this magic. Note that you don't need to clone SV* payload, as threaded perl duplicates it by itself.

    typedef int (*on_svdup_t) (pTHX_ MAGIC* mg, CLONE_PARAMS* param);

payload_marker_t* sv_payload_marker (const char* key, on_svdup_t svdup_callback = NULL)

Returns a marker in runtime for specified key. If no marker exists for that key, creates and returns it. Performs hash searching, so that in most cases you should put the result into a static variable to prevent calling this function many times.

Note that marker SVPayloadMarker<MyClass*::marker> is NOT the same marker as sv_payload_marker("MyClass*").

SUPER/NEXT METHOD FUNCTIONS

SV* call_super (CV* cv, SV** args, I32 items, I32 flags = 0) [pTHX]

Calls super method from inside XS method. 'cv' is a pointer to currently running XS function (every XS function has this variable). 'args' and 'items' describes perl variables to pass to super method. If you want to passthrough all the arguments your XS function received, pass '&ST(0)' and 'items'. 'flags' is passed to Perl's API call_method / call_sv as is. SUPER is always called in SCALAR context. Returned SV* has its refcounter increased. If you don't need it, decrement it. If you don't need return value at all, pass G_DISCARD in flags.

SV* call_next_method (CV* cv, SV** args, I32 items, I32 flags = 0) [pTHX]

Calls next method (C3 MRO). See "call_super" for argument details.

SV* call_next_maybe (CV* cv, SV** args, I32 items, I32 flags = 0) [pTHX]

Calls next method (C3 MRO). Return undef if no next method. See "call_super" for argument details.

SV* call_next (CV* cv, SV** args, I32 items, next_t type, I32 flags = 0) [pTHX]

Generic form of call_super/call_next_method/call_next_maybe. 'type' is either xs::NEXT_SUPER or xs::NEXT_METHOD or xs::NEXT_MAYBE.

C MACROS

XS_HV_ITER(hv,code)

Iterates through perl hash very fast. On each iteration executes code, current hash entry is accessible via HE* he variable. Use perl API's HeVAL, HeKEY, HeKLEN to extract value and key from hash entry.

hv must be a non-null HV* variable. Doesn't handle tied hashes.

XS_HV_ITER_NU(hv,code)

Same as XS_HV_ITER but skips entries whose values are undefs. _NU stands for 'not undef'.

XS_AV_ITER(av,code)

Iterates through perl array very fast. On each iteration executes code, current array element is accessible via SV* elem variable. Keep in mind that it might be a NULL pointer in case if current slot in array is empty.

For example, in this case

    my @arr;
    $arr[1] = undef;
    $arr[3] = 10;
    

slots 0 and 2 are empty. Slot 1 is not empty but it's value is undefined. Don't confuse empty slot with undefined values, empty slot is an absence of SV* variable, while undefined value is a legal SV* variable for which SvOK(elem) returns false.

av must be a non-null AV* variable. Doesn't handle tied arrays.

XS_AV_ITER_NE(av,code)

Same as XS_AV_ITER but skips empty slots. _NE stands for 'not empty'.

XS_AV_ITER_NU(av,code)

Same as XS_AV_ITER but skips empty slots and undefined emelents. _NU stands for 'not undef'.

PERL FUNCTIONS

obj2hv ($obj)

obj2av ($obj)

Upgrades $obj to HASHREF or ARRAYREF.

$obj must be a valid reference to something lower than array or hash (undef, number, string, etc). Otherwise will croak. If $obj is already of desired type, nothing is done.

sv_payload_attach ($target, $payload)

Attach $payload to $target. $target can be any l-value scalar. If $target has any $payload, it gets removed.

sv_payload ($target)

Returns payload attached to $target if any, otherwise undef.

sv_payload_exists ($target)

Returns true if $target has any payload.

sv_payload_detach ($target)

Removes payload from $target. Returns true if any payload has been removed.

rv_payload_attach ($target, $payload)

rv_payload ($target)

rv_payload_exists ($target)

rv_payload_detach ($target)

Same as sv_* but $target must be reference to perl object.

any_payload_attach ($target, $payload)

any_payload ($target)

any_payload_exists ($target)

any_payload_detach ($target)

Same as sv_* or rv_*, but if $target is ref then payload will be attached/detached from referenced object, otherwise from $target itself.

THREADED PERLS SUPPORT

HOOKING THREAD CREATION FOR XS OBJECTS

If you don't do anything special, on threaded perls after thread creation you will most likely core dump. The reason is that when perl creates a thread it duplicates all perl variables available in current interpreter. So that you will likely get 2 perl variables holding a pointer to the same C/C++ object (refcnt won't help - nobody calls retain()). When these objects are destroyed, double delete might occur on the same C object. That's bad.

You have 4 options to handle that:

Don't support thread creation (however don't core dump).

To do this, just add CLONE_SKIP method to your class and return a true value from it. In this case perl won't clone SV variable holding your object - it will place an unblessed reference to undef instead. Of course users must know about this behaviour, because they won't be able to use objects of your class created before new thread. However they can still create new objects from new thread and everything will be ok.

This is the only option for you if you use T_OPTR. T_OEXT doesn't have this limitation.

Support thread creation by cloning your C++ object (for example, if it's thread-unsafe or doesn't have a refcounter)

To do this, don't define CLONE_SKIP method, instead set on_svdup parameter of T_OEXT class to 'clone'. In this case when a thread creates, method clone() will be called on your C++ object and this new object is stored in duplicated variable. Of course you must have a clone() method with no args in your class. It must be virtual if you support inheritance. For example:

    T_MYCLASS : T_OEXT(basetype=MyClass*, on_svdup=clone)
    

Note that if you use wrapper feature, then clone() will be called on wrapper, not on the object. You must clone your object from wrapper's clone().

If you use backref feature, backreference in your object (which still points to the old SV from the parent thread) will be changed automatically after clone, you don't need to do something special for that (even if you use wrapper).

Support thread creation by incrementing refcnt on your C++ object (for thread-safe objects with refcounters).

To do this, don't define CLONE_SKIP method, instead set on_svdup parameter of T_OEXT class to 'retain'. In this case when a thread creates, method retain() will be called on your C++ object and the same object is stored in duplicated variable. Of course you must have a retain() method with no args in your class (if you subclass panda::RefCounted, you do).

For example:

    T_MYCLASS : T_OEXT(basetype=MyClass*, on_svdup=retain)

Note that if you use wrapper, you should set on_svdup=clone and do retain() of the object in wrapper's clone(). That's because wrappers are strictly bound to their SV (one-to-one) and cannot be shared, because it doesn't make sense.

You CANNOT use backref, because we must use different SV in each thread, however if you retain you object, we would have only single pointer for backreference. So that if you need backref you should clone your objects, not retain.

Implement custom logic

To do this, don't define CLONE_SKIP method, instead set on_svdup parameter of T_OEXT class to the name of your callback which will handle thread creation for objects of your class. Callback must have the following signature:

    <basetype> <callbackname> (pTHX_ <basetype> arg);
    

Callback will be passed current C object. Callback shall return a C object to be set to the duplicated perl variable (it may be a new one or the same object). Keep in mind that if your module shares API for other modules, this callback MUST be externally visible, because when someone uses your class as a return type from XS functions, typemap will insert code, containing your callback name into those XS functions. Don't forget to specify it with full qualified name for the same reason. Example:

    # in .h
    namespace my {
        MyClass* myclass_onsvdup (pTHX_ MyClass* old);
    }
    # in .cc
    MyClass* my::myclass_onsvdup (pTHX_ MyClass* old) {
        return new MyClass(old->data);
    }
    
    # in typemap
    T_MYCLASS : T_OEXT(basetype=MyClass*, on_svdup=my::myclass_onsvdup)
    

Note that if you use wrapper feature, the return and input types of the callback are basetype of wrapper itself, not the object.

If you use backref, then backreference is changed automatically in returned object to point new SV no matter if it is the same object or not (even with wrapper). Keep in mind that if you implement some kind of retain strategy in your callback, you will break backref. Bad thing will surely happen.

aTHX/pTHX politics and helpers

Perl has 2 options for XS when working on threaded perls:

1) Not defining PERL_NO_GET_CONTEXT. Advantage is that you don't have to put pTHX to every function's signature and don't have to pass it to every function which works with perl variables. Disadvatage is - decreased performance on threaded perls.

2) Defining PERL_NO_GET_CONTEXT.

Unfortunatelly, as Panda::Install/Panda::XS concept is all about sharing C code, modules that use Panda::XS cannot have different PERL_NO_GET_CONTEXT state. Because in that case, binary incompability between function signatures would occur.

Panda::XS strictly sets PERL_NO_GET_CONTEXT options, so that you have generally to accept pTHX and pass aTHX to all functions that work with perl variables. In this case you'll achieve maximum performance on threaded perls.

However if you don't want to write pTHX/aTHX every time, you can write

    dTHX;
    

at the top of those functions and actually saying revert to behaviour without PERL_NO_GET_CONTEXT (but you still need to pass aTHX to functions which are defined with pTHX, unless called via macros - like perl's API).

For example, you can't receive pTHX in C++ objects destructors, so the options for you are either write dTHX; on the top of destructor or save perl interpreter object in constructor of your object to property named 'my_perl', or use using xs::my_perl global interpreter.

One more problem is static initialization, for example:

    # my.h
    extern SV* myname;
    # my.cc
    #include <my.h>
    SV* myname = newSVpv("john", 0);
    

This won't work on threaded perls with PERL_NO_GET_CONTEXT because no one defined my_perl. You shouldn't say dTHX; in the global-scoped code. This problem can be solved using using xs::my_perl.

HELPERS

xs::my_perl

This variable holds correct perl interpreter for each thread. Just say using xs::my_perl and you'll get my_perl interpreter for the rest of your code's scope.

Keep in mind that this my_perl variable is get from thread-local storage, like dTHX, so this is actually a convenience tool, reverting to performance as PERL_NO_GET_CONTEXT has not been defined. If you want maximum performance on threaded perls, use pTHX/aTHX everywhere. You could however use mixed style, saying using xs::my_perl and passing pTHX only for performance-critical functions.

This variable should not be used on non-threaded perls (this is not a problem as aTHX is empty in that case). However it still exists on non-threaded perls, so that you can say using xs::my_perl without wrapping it into #ifdef PERL_IMPLICIT_CONTEXT.

mTHX

Another way to achieve maximum perfomance on threaded-perls for C++ objects is to hold my_perl as class variable. In this case you don't have to receive pTHX in every method of your object and don't even have to say dTHX in desctructor.

This helper defines this class member (it defines nothing on non-threaded perls). Example

    class MyXSClass {
        int val;
        mTHX;
        MyXSClass (int val) : val(val) {
            aTHXa(PERL_GET_THX);
        }
        // pTHX is not needed anymore in method's definitions.
    }

mTHXa(interpreter)

Sometimes you have to pass aTHX to constructors of your class members and therefore setting aTHXa in constructor's code is too late.

mTHXa(interpreter) initializes class member my_perl with value interpreter in initializer list. On threaded perls it is my_perl(interpreter),, otherwise empty. Example:

    class MyXSClass {
        int val;
        MySVContainer container;
        mTHX;
        MyXSClass (int val) : val(val), mTHXa(PERL_GET_THX) container(aTHX) {
            ...
        }
    }

Note that you should not write , after mTHXa. Therefore it cannot be the last on the list.

TIP

If you don't care about performance on threaded perls then the most convenient solution is using xs::my_perl on the top of the file. In this case you don't have to put pTHX in every function's definition. However you still have to pass aTHX to functions which defined with pTHX.

CAVEATS

If module A binary depends on module B and module B updates and becomes binary incompatible, undefined behaviour may happen. To solve this problem, one need to reinstall (rebuild) all modules that depend on module B. To help solving this problem, Panda::Install automatically tracks these depedencies, warns and croaks if any binary depedencies became incompatible and prints the list of modules to reinstall (rebuild).

AUTHOR

Pronin Oleg <syber@crazypanda.ru>, Crazy Panda, CP Decision LTD

LICENSE

You may distribute this code under the same terms as Perl itself.