| perlxs - XS language reference manual |
length(NAME) Keyword
perlxs - XS language reference manual
XS is an interface description file format used to create an extension interface between Perl and C code (or a C library) which one wishes to use with Perl. The XS interface is combined with the library to create a new library which can then be either dynamically loaded or statically linked into perl. The XS interface description is written in the XS language and is the core component of the Perl extension interface.
Before writing XS, read the CAVEATS section below.
An XSUB forms the basic unit of the XS interface. After compilation by the xsubpp compiler, each XSUB amounts to a C function definition which will provide the glue between Perl calling conventions and C calling conventions.
The glue code pulls the arguments from the Perl stack, converts these Perl values to the formats expected by a C function, call this C function, transfers the return values of the C function back to Perl. Return values here may be a conventional C return value or any C function arguments that may serve as output parameters. These return values may be passed back to Perl either by putting them on the Perl stack, or by modifying the arguments supplied from the Perl side.
The above is a somewhat simplified view of what really happens. Since Perl allows more flexible calling conventions than C, XSUBs may do much more in practice, such as checking input parameters for validity, throwing exceptions (or returning undef/empty list) if the return value from the C function indicates failure, calling different C functions based on numbers and types of the arguments, providing an object-oriented interface, etc.
Of course, one could write such glue code directly in C. However, this would be a tedious task, especially if one needs to write glue for multiple C functions, and/or one is not familiar enough with the Perl stack discipline and other such arcana. XS comes to the rescue here: instead of writing this glue C code in long-hand, one can write a more concise short-hand description of what should be done by the glue, and let the XS compiler xsubpp handle the rest.
The XS language allows one to describe the mapping between how the C
routine is used, and how the corresponding Perl routine is used. It
also allows creation of Perl routines which are directly translated to
C code and which are not related to a pre-existing C function. In cases
when the C interface coincides with the Perl interface, the XSUB
declaration is almost identical to a declaration of a C function (in K&R
style). In such circumstances, there is another tool called h2xs
that is able to translate an entire C header file into a corresponding
XS file that will provide glue to the functions/macros described in
the header file.
The XS compiler is called xsubpp. This compiler creates the constructs necessary to let an XSUB manipulate Perl values, and creates the glue necessary to let Perl call the XSUB. The compiler uses typemaps to determine how to map C function parameters and output values to Perl values and back. The default typemap (which comes with Perl) handles many common C types. A supplementary typemap may also be needed to handle any special structures and types for the library being linked. For more information on typemaps, see the perlxstypemap manpage.
A file in XS format starts with a C language section which goes until the
first MODULE = directive. Other XS directives and XSUB definitions
may follow this line. The ``language'' used in this part of the file
is usually referred to as the XS language. xsubpp recognizes and
skips POD (see the perlpod manpage) in both the C and XS language sections, which
allows the XS file to contain embedded documentation.
See the perlxstut manpage for a tutorial on the whole extension creation process.
Note: For some extensions, Dave Beazley's SWIG system may provide a significantly more convenient mechanism for creating the extension glue code. See http://www.swig.org/ for more information.
Many of the examples which follow will concentrate on creating an interface
between Perl and the ONC+ RPC bind library functions. The rpcb_gettime()
function is used to demonstrate many features of the XS language. This
function has two parameters; the first is an input parameter and the second
is an output parameter. The function also returns a status value.
bool_t rpcb_gettime(const char *host, time_t *timep);
From C this function will be called with the following statements.
#include <rpc/rpc.h>
bool_t status;
time_t timep;
status = rpcb_gettime( "localhost", &timep );
If an XSUB is created to offer a direct translation between this function and Perl, then this XSUB will be used from Perl with the following code. The $status and $timep variables will contain the output of the function.
use RPC;
$status = rpcb_gettime( "localhost", $timep );
The following XS file shows an XS subroutine, or XSUB, which
demonstrates one possible interface to the rpcb_gettime()
function. This XSUB represents a direct translation between
C and Perl and so preserves the interface even from Perl.
This XSUB will be invoked from Perl with the usage shown
above. Note that the first three #include statements, for
EXTERN.h, perl.h, and XSUB.h, will always be present at the
beginning of an XS file. This approach and others will be
expanded later in this document. A #define for PERL_NO_GET_CONTEXT
should be present to fetch the interpreter context more efficiently,
see perlguts for details.
#define PERL_NO_GET_CONTEXT
#include "EXTERN.h"
#include "perl.h"
#include "XSUB.h"
#include <rpc/rpc.h>
MODULE = RPC PACKAGE = RPC
bool_t
rpcb_gettime(host,timep)
char *host
time_t &timep
OUTPUT:
timep
Any extension to Perl, including those containing XSUBs,
should have a Perl module to serve as the bootstrap which
pulls the extension into Perl. This module will export the
extension's functions and variables to the Perl program and
will cause the extension's XSUBs to be linked into Perl.
The following module will be used for most of the examples
in this document and should be used from Perl with the use
command as shown earlier. Perl modules are explained in
more detail later in this document.
package RPC;
require Exporter;
require DynaLoader;
@ISA = qw(Exporter DynaLoader);
@EXPORT = qw( rpcb_gettime );
bootstrap RPC;
1;
Throughout this document a variety of interfaces to the rpcb_gettime()
XSUB will be explored. The XSUBs will take their parameters in different
orders or will take different numbers of parameters. In each case the
XSUB is an abstraction between Perl and the real C rpcb_gettime()
function, and the XSUB must always ensure that the real rpcb_gettime()
function is called with the correct parameters. This abstraction will
allow the programmer to create a more Perl-like interface to the C
function.
The simplest XSUBs consist of 3 parts: a description of the return value, the name of the XSUB routine and the names of its arguments, and a description of types or formats of the arguments.
The following XSUB allows a Perl program to access a C library function called sin(). The XSUB will imitate the C function which takes a single argument and returns a single value.
double
sin(x)
double x
Optionally, one can merge the description of types and the list of argument names, rewriting this as
double
sin(double x)
This makes this XSUB look similar to an ANSI C declaration. An optional semicolon is allowed after the argument list, as in
double
sin(double x);
Parameters with C pointer types can have different semantic: C functions with similar declarations
bool string_looks_as_a_number(char *s);
bool make_char_uppercase(char *c);
are used in absolutely incompatible manner. Parameters to these functions could be described xsubpp like this:
char * s
char &c
Both these XS declarations correspond to the char* C type, but they have
different semantics, see The & Unary Operator.
It is convenient to think that the indirection operator
* should be considered as a part of the type and the address operator &
should be considered part of the variable. See the perlxstypemap manpage
for more info about handling qualifiers and unary operators in C types.
The function name and the return type must be placed on separate lines and should be flush left-adjusted.
INCORRECT CORRECT
double sin(x) double
double x sin(x)
double x
The rest of the function description may be indented or left-adjusted. The following example shows a function with its body left-adjusted. Most examples in this document will indent the body for better readability.
CORRECT
double sin(x) double x
More complicated XSUBs may contain many other sections. Each section of an XSUB starts with the corresponding keyword, such as INIT: or CLEANUP:. However, the first two lines of an XSUB always contain the same data: descriptions of the return type and the names of the function and its parameters. Whatever immediately follows these is considered to be an INPUT: section unless explicitly marked with another keyword. (See The INPUT: Keyword.)
An XSUB section continues until another section-start keyword is found.
The Perl argument stack is used to store the values which are sent as parameters to the XSUB and to store the XSUB's return value(s). In reality all Perl functions (including non-XSUB ones) keep their values on this stack all the same time, each limited to its own range of positions on the stack. In this document the first position on that stack which belongs to the active function will be referred to as position 0 for that function.
XSUBs refer to their stack arguments with the macro ST(x), where x refers to a position in this XSUB's part of the stack. Position 0 for that function would be known to the XSUB as ST(0). The XSUB's incoming parameters and outgoing return values always begin at ST(0). For many simple cases the xsubpp compiler will generate the code necessary to handle the argument stack by embedding code fragments found in the typemaps. In more complex cases the programmer must supply the code.
The RETVAL variable is a special C variable that is declared automatically
for you. The C type of RETVAL matches the return type of the C library
function. The xsubpp compiler will declare this variable in each XSUB
with non-void return type. By default the generated C function
will use RETVAL to hold the return value of the C library function being
called. In simple cases the value of RETVAL will be placed in ST(0) of
the argument stack where it can be received by Perl as the return value
of the XSUB.
If the XSUB has a return type of void then the compiler will
not declare a RETVAL variable for that function. When using
a PPCODE: section no manipulation of the RETVAL variable is required, the
section may use direct stack manipulation to place output values on the stack.
If PPCODE: directive is not used, void return value should be used
only for subroutines which do not return a value, even if CODE:
directive is used which sets ST(0) explicitly.
Older versions of this document recommended to use void return
value in such cases. It was discovered that this could lead to
segfaults in cases when XSUB was truly void. This practice is
now deprecated, and may be not supported at some future version. Use
the return value SV * in such cases. (Currently xsubpp contains
some heuristic code which tries to disambiguate between ``truly-void''
and ``old-practice-declared-as-void'' functions. Hence your code is at
mercy of this heuristics unless you use SV * as return value.)
When you're using RETVAL to return an SV *, there's some magic
going on behind the scenes that should be mentioned. When you're
manipulating the argument stack using the ST(x) macro, for example,
you usually have to pay special attention to reference counts. (For
more about reference counts, see the perlguts manpage.) To make your life
easier, the typemap file automatically makes RETVAL mortal when
you're returning an SV *. Thus, the following two XSUBs are more
or less equivalent:
void
alpha()
PPCODE:
ST(0) = newSVpv("Hello World",0);
sv_2mortal(ST(0));
XSRETURN(1);
SV *
beta()
CODE:
RETVAL = newSVpv("Hello World",0);
OUTPUT:
RETVAL
This is quite useful as it usually improves readability. While
this works fine for an SV *, it's unfortunately not as easy
to have AV * or HV * as a return value. You should be
able to write:
AV *
array()
CODE:
RETVAL = newAV();
/* do something with RETVAL */
OUTPUT:
RETVAL
But due to an unfixable bug (fixing it would break lots of existing
CPAN modules) in the typemap file, the reference count of the AV *
is not properly decremented. Thus, the above XSUB would leak memory
whenever it is being called. The same problem exists for HV *,
CV *, and SVREF (which indicates a scalar reference, not
a general SV *).
In XS code on perls starting with perl 5.16, you can override the
typemaps for any of these types with a version that has proper
handling of refcounts. In your TYPEMAP section, do
AV* T_AVREF_REFCOUNT_FIXED
to get the repaired variant. For backward compatibility with older
versions of perl, you can instead decrement the reference count
manually when you're returning one of the aforementioned
types using sv_2mortal:
AV *
array()
CODE:
RETVAL = newAV();
sv_2mortal((SV*)RETVAL);
/* do something with RETVAL */
OUTPUT:
RETVAL
Remember that you don't have to do this for an SV *. The reference
documentation for all core typemaps can be found in the perlxstypemap manpage.
The MODULE keyword is used to start the XS code and to specify the package of the functions which are being defined. All text preceding the first MODULE keyword is considered C code and is passed through to the output with POD stripped, but otherwise untouched. Every XS module will have a bootstrap function which is used to hook the XSUBs into Perl. The package name of this bootstrap function will match the value of the last MODULE statement in the XS source files. The value of MODULE should always remain constant within the same XS file, though this is not required.
The following example will start the XS code and will place all functions in a package named RPC.
MODULE = RPC
When functions within an XS source file must be separated into packages the PACKAGE keyword should be used. This keyword is used with the MODULE keyword and must follow immediately after it when used.
MODULE = RPC PACKAGE = RPC
[ XS code in package RPC ]
MODULE = RPC PACKAGE = RPCB
[ XS code in package RPCB ]
MODULE = RPC PACKAGE = RPC
[ XS code in package RPC ]
The same package name can be used more than once, allowing for non-contiguous code. This is useful if you have a stronger ordering principle than package names.
Although this keyword is optional and in some cases provides redundant information it should always be used. This keyword will ensure that the XSUBs appear in the desired package.
The PREFIX keyword designates prefixes which should be
removed from the Perl function names. If the C function is
rpcb_gettime() and the PREFIX value is rpcb_ then Perl will
see this function as gettime().
This keyword should follow the PACKAGE keyword when used. If PACKAGE is not used then PREFIX should follow the MODULE keyword.
MODULE = RPC PREFIX = rpc_
MODULE = RPC PACKAGE = RPCB PREFIX = rpcb_
The OUTPUT: keyword indicates that certain function parameters should be
updated (new values made visible to Perl) when the XSUB terminates or that
certain values should be returned to the calling Perl function. For
simple functions which have no CODE: or PPCODE: section,
such as the sin() function above, the RETVAL variable is
automatically designated as an output value. For more complex functions
the xsubpp compiler will need help to determine which variables are output
variables.
This keyword will normally be used to complement the CODE: keyword. The RETVAL variable is not recognized as an output variable when the CODE: keyword is present. The OUTPUT: keyword is used in this situation to tell the compiler that RETVAL really is an output variable.
The OUTPUT: keyword can also be used to indicate that function parameters are output variables. This may be necessary when a parameter has been modified within the function and the programmer would like the update to be seen by Perl.
bool_t
rpcb_gettime(host,timep)
char *host
time_t &timep
OUTPUT:
timep
The OUTPUT: keyword will also allow an output parameter to be mapped to a matching piece of code rather than to a typemap.
bool_t
rpcb_gettime(host,timep)
char *host
time_t &timep
OUTPUT:
timep sv_setnv(ST(1), (double)timep);
xsubpp emits an automatic SvSETMAGIC() for all parameters in the
OUTPUT section of the XSUB, except RETVAL. This is the usually desired
behavior, as it takes care of properly invoking 'set' magic on output
parameters (needed for hash or array element parameters that must be
created if they didn't exist). If for some reason, this behavior is
not desired, the OUTPUT section may contain a SETMAGIC: DISABLE line
to disable it for the remainder of the parameters in the OUTPUT section.
Likewise, SETMAGIC: ENABLE can be used to reenable it for the
remainder of the OUTPUT section. See the perlguts manpage for more details
about 'set' magic.
The NO_OUTPUT can be placed as the first token of the XSUB. This keyword
indicates that while the C subroutine we provide an interface to has
a non-void return type, the return value of this C subroutine should not
be returned from the generated Perl subroutine.
With this keyword present The RETVAL Variable is created, and in the generated call to the subroutine this variable is assigned to, but the value of this variable is not going to be used in the auto-generated code.
This keyword makes sense only if RETVAL is going to be accessed by the
user-supplied code. It is especially useful to make a function interface
more Perl-like, especially when the C return value is just an error condition
indicator. For example,
NO_OUTPUT int
delete_file(char *name)
POSTCALL:
if (RETVAL != 0)
croak("Error %d while deleting file '%s'", RETVAL, name);
Here the generated XS function returns nothing on success, and will die()
with a meaningful error message on error.
This keyword is used in more complicated XSUBs which require special handling for the C function. The RETVAL variable is still declared, but it will not be returned unless it is specified in the OUTPUT: section.
The following XSUB is for a C function which requires special handling of its parameters. The Perl usage is given first.
$status = rpcb_gettime( "localhost", $timep );
The XSUB follows.
bool_t
rpcb_gettime(host,timep)
char *host
time_t timep
CODE:
RETVAL = rpcb_gettime( host, &timep );
OUTPUT:
timep
RETVAL
The INIT: keyword allows initialization to be inserted into the XSUB before the compiler generates the call to the C function. Unlike the CODE: keyword above, this keyword does not affect the way the compiler handles RETVAL.
bool_t
rpcb_gettime(host,timep)
char *host
time_t &timep
INIT:
printf("# Host is %s\n", host );
OUTPUT:
timep
Another use for the INIT: section is to check for preconditions before making a call to the C function:
long long
lldiv(a,b)
long long a
long long b
INIT:
if (a == 0 && b == 0)
XSRETURN_UNDEF;
if (b == 0)
croak("lldiv: cannot divide by 0");
The NO_INIT keyword is used to indicate that a function parameter is being used only as an output value. The xsubpp compiler will normally generate code to read the values of all function parameters from the argument stack and assign them to C variables upon entry to the function. NO_INIT will tell the compiler that some parameters will be used for output rather than for input and that they will be handled before the function terminates.
The following example shows a variation of the rpcb_gettime() function.
This function uses the timep variable only as an output variable and does
not care about its initial contents.
bool_t
rpcb_gettime(host,timep)
char *host
time_t &timep = NO_INIT
OUTPUT:
timep
Starting with Perl 5.16, you can embed typemaps into your XS code instead of or in addition to typemaps in a separate file. Multiple such embedded typemaps will be processed in order of appearance in the XS code and like local typemap files take precedence over the default typemap, the embedded typemaps may overwrite previous definitions of TYPEMAP, INPUT, and OUTPUT stanzas. The syntax for embedded typemaps is
TYPEMAP: <<HERE
... your typemap code here ...
HERE
where the TYPEMAP keyword must appear in the first column of a
new line.
Refer to the perlxstypemap manpage for details on writing typemaps.
C function parameters are normally initialized with their values from
the argument stack (which in turn contains the parameters that were
passed to the XSUB from Perl). The typemaps contain the
code segments which are used to translate the Perl values to
the C parameters. The programmer, however, is allowed to
override the typemaps and supply alternate (or additional)
initialization code. Initialization code starts with the first
=, ; or + on a line in the INPUT: section. The only
exception happens if this ; terminates the line, then this ;
is quietly ignored.
The following code demonstrates how to supply initialization code for
function parameters. The initialization code is eval'ed within double
quotes by the compiler before it is added to the output so anything
which should be interpreted literally [mainly $, @, or \\]
must be protected with backslashes. The variables $var, $arg,
and $type can be used as in typemaps.
bool_t
rpcb_gettime(host,timep)
char *host = (char *)SvPV_nolen($arg);
time_t &timep = 0;
OUTPUT:
timep
This should not be used to supply default values for parameters. One would normally use this when a function parameter must be processed by another library function before it can be used. Default parameters are covered in the next section.
If the initialization begins with =, then it is output in
the declaration for the input variable, replacing the initialization
supplied by the typemap. If the initialization
begins with ; or +, then it is performed after
all of the input variables have been declared. In the ;
case the initialization normally supplied by the typemap is not performed.
For the + case, the declaration for the variable will include the
initialization from the typemap. A global
variable, %v, is available for the truly rare case where
information from one initialization is needed in another
initialization.
Here's a truly obscure example:
bool_t
rpcb_gettime(host,timep)
time_t &timep; /* \$v{timep}=@{[$v{timep}=$arg]} */
char *host + SvOK($v{timep}) ? SvPV_nolen($arg) : NULL;
OUTPUT:
timep
The construct \$v{timep}=@{[$v{timep}=$arg]} used in the above
example has a two-fold purpose: first, when this line is processed by
xsubpp, the Perl snippet $v{timep}=$arg is evaluated. Second,
the text of the evaluated snippet is output into the generated C file
(inside a C comment)! During the processing of char *host line,
$arg will evaluate to ST(0), and $v{timep} will evaluate to
ST(1).
Default values for XSUB arguments can be specified by placing an
assignment statement in the parameter list. The default value may
be a number, a string or the special string NO_INIT. Defaults should
always be used on the right-most parameters only.
To allow the XSUB for rpcb_gettime() to have a default host
value the parameters to the XSUB could be rearranged. The
XSUB will then call the real rpcb_gettime() function with
the parameters in the correct order. This XSUB can be called
from Perl with either of the following statements:
$status = rpcb_gettime( $timep, $host );
$status = rpcb_gettime( $timep );
The XSUB will look like the code which follows. A CODE:
block is used to call the real rpcb_gettime() function with
the parameters in the correct order for that function.
bool_t
rpcb_gettime(timep,host="localhost")
char *host
time_t timep = NO_INIT
CODE:
RETVAL = rpcb_gettime( host, &timep );
OUTPUT:
timep
RETVAL
The PREINIT: keyword allows extra variables to be declared immediately before or after the declarations of the parameters from the INPUT: section are emitted.
If a variable is declared inside a CODE: section it will follow any typemap
code that is emitted for the input parameters. This may result in the
declaration ending up after C code, which is C syntax error. Similar
errors may happen with an explicit ;-type or +-type initialization of
parameters is used (see Initializing Function Parameters). Declaring
these variables in an INIT: section will not help.
In such cases, to force an additional variable to be declared together with declarations of other variables, place the declaration into a PREINIT: section. The PREINIT: keyword may be used one or more times within an XSUB.
The following examples are equivalent, but if the code is using complex typemaps then the first example is safer.
bool_t
rpcb_gettime(timep)
time_t timep = NO_INIT
PREINIT:
char *host = "localhost";
CODE:
RETVAL = rpcb_gettime( host, &timep );
OUTPUT:
timep
RETVAL
For this particular case an INIT: keyword would generate the same C code as the PREINIT: keyword. Another correct, but error-prone example:
bool_t
rpcb_gettime(timep)
time_t timep = NO_INIT
CODE:
char *host = "localhost";
RETVAL = rpcb_gettime( host, &timep );
OUTPUT:
timep
RETVAL
Another way to declare host is to use a C block in the CODE: section:
bool_t
rpcb_gettime(timep)
time_t timep = NO_INIT
CODE:
{
char *host = "localhost";
RETVAL = rpcb_gettime( host, &timep );
}
OUTPUT:
timep
RETVAL
The ability to put additional declarations before the typemap entries are processed is very handy in the cases when typemap conversions manipulate some global state:
MyObject
mutate(o)
PREINIT:
MyState st = global_state;
INPUT:
MyObject o;
CLEANUP:
reset_to(global_state, st);
Here we suppose that conversion to MyObject in the INPUT: section and from
MyObject when processing RETVAL will modify a global variable global_state.
After these conversions are performed, we restore the old value of
global_state (to avoid memory leaks, for example).
There is another way to trade clarity for compactness: INPUT sections allow
declaration of C variables which do not appear in the parameter list of
a subroutine. Thus the above code for mutate() can be rewritten as
MyObject
mutate(o)
MyState st = global_state;
MyObject o;
CLEANUP:
reset_to(global_state, st);
and the code for rpcb_gettime() can be rewritten as
bool_t
rpcb_gettime(timep)
time_t timep = NO_INIT
char *host = "localhost";
C_ARGS:
host, &timep
OUTPUT:
timep
RETVAL
The SCOPE: keyword allows scoping to be enabled for a particular XSUB. If enabled, the XSUB will invoke ENTER and LEAVE automatically.
To support potentially complex type mappings, if a typemap entry used
by an XSUB contains a comment like /*scope*/ then scoping will
be automatically enabled for that XSUB.
To enable scoping:
SCOPE: ENABLE
To disable scoping:
SCOPE: DISABLE
The XSUB's parameters are usually evaluated immediately after entering the XSUB. The INPUT: keyword can be used to force those parameters to be evaluated a little later. The INPUT: keyword can be used multiple times within an XSUB and can be used to list one or more input variables. This keyword is used with the PREINIT: keyword.
The following example shows how the input parameter timep can be
evaluated late, after a PREINIT.
bool_t
rpcb_gettime(host,timep)
char *host
PREINIT:
time_t tt;
INPUT:
time_t timep
CODE:
RETVAL = rpcb_gettime( host, &tt );
timep = tt;
OUTPUT:
timep
RETVAL
The next example shows each input parameter evaluated late.
bool_t
rpcb_gettime(host,timep)
PREINIT:
time_t tt;
INPUT:
char *host
PREINIT:
char *h;
INPUT:
time_t timep
CODE:
h = host;
RETVAL = rpcb_gettime( h, &tt );
timep = tt;
OUTPUT:
timep
RETVAL
Since INPUT sections allow declaration of C variables which do not appear in the parameter list of a subroutine, this may be shortened to:
bool_t
rpcb_gettime(host,timep)
time_t tt;
char *host;
char *h = host;
time_t timep;
CODE:
RETVAL = rpcb_gettime( h, &tt );
timep = tt;
OUTPUT:
timep
RETVAL
(We used our knowledge that input conversion for char * is a ``simple'' one,
thus host is initialized on the declaration line, and our assignment
h = host is not performed too early. Otherwise one would need to have the
assignment h = host in a CODE: or INIT: section.)
In the list of parameters for an XSUB, one can precede parameter names
by the IN/OUTLIST/IN_OUTLIST/OUT/IN_OUT keywords.
IN keyword is the default, the other keywords indicate how the Perl
interface should differ from the C interface.
Parameters preceded by OUTLIST/IN_OUTLIST/OUT/IN_OUT
keywords are considered to be used by the C subroutine via
pointers. OUTLIST/OUT keywords indicate that the C subroutine
does not inspect the memory pointed by this parameter, but will write
through this pointer to provide additional return values.
Parameters preceded by OUTLIST keyword do not appear in the usage
signature of the generated Perl function.
Parameters preceded by IN_OUTLIST/IN_OUT/OUT do appear as
parameters to the Perl function. With the exception of
OUT-parameters, these parameters are converted to the corresponding
C type, then pointers to these data are given as arguments to the C
function. It is expected that the C function will write through these
pointers.
The return list of the generated Perl function consists of the C return value
from the function (unless the XSUB is of void return type or
The NO_OUTPUT Keyword was used) followed by all the OUTLIST
and IN_OUTLIST parameters (in the order of appearance). On the
return from the XSUB the IN_OUT/OUT Perl parameter will be
modified to have the values written by the C function.
For example, an XSUB
void
day_month(OUTLIST day, IN unix_time, OUTLIST month)
int day
int unix_time
int month
should be used from Perl as
my ($day, $month) = day_month(time);
The C signature of the corresponding function should be
void day_month(int *day, int unix_time, int *month);
The IN/OUTLIST/IN_OUTLIST/IN_OUT/OUT keywords can be
mixed with ANSI-style declarations, as in
void day_month(OUTLIST int day, int unix_time, OUTLIST int month)
(here the optional IN keyword is omitted).
The IN_OUT parameters are identical with parameters introduced with
The & Unary Operator and put into the OUTPUT: section (see
The OUTPUT: Keyword). The IN_OUTLIST parameters are very similar,
the only difference being that the value C function writes through the
pointer would not modify the Perl parameter, but is put in the output
list.
The OUTLIST/OUT parameter differ from IN_OUTLIST/IN_OUT
parameters only by the initial value of the Perl parameter not
being read (and not being given to the C function - which gets some
garbage instead). For example, the same C function as above can be
interfaced with as
void day_month(OUT int day, int unix_time, OUT int month);
or
void
day_month(day, unix_time, month)
int &day = NO_INIT
int unix_time
int &month = NO_INIT
OUTPUT:
day
month
However, the generated Perl function is called in very C-ish style:
my ($day, $month); day_month($day, time, $month);
length(NAME) KeywordIf one of the input arguments to the C function is the length of a string
argument NAME, one can substitute the name of the length-argument by
length(NAME) in the XSUB declaration. This argument must be omitted when
the generated Perl function is called. E.g.,
void
dump_chars(char *s, short l)
{
short n = 0;
while (n < l) {
printf("s[%d] = \"\\%#03o\"\n", n, (int)s[n]);
n++;
}
}
MODULE = x PACKAGE = x
void dump_chars(char *s, short length(s))
should be called as dump_chars($string).
This directive is supported with ANSI-type function declarations only.
XSUBs can have variable-length parameter lists by specifying an ellipsis
(...) in the parameter list. This use of the ellipsis is similar to that
found in ANSI C. The programmer is able to determine the number of
arguments passed to the XSUB by examining the items variable which the
xsubpp compiler supplies for all XSUBs. By using this mechanism one can
create an XSUB which accepts a list of parameters of unknown length.
The host parameter for the rpcb_gettime() XSUB can be
optional so the ellipsis can be used to indicate that the
XSUB will take a variable number of parameters. Perl should
be able to call this XSUB with either of the following statements.
$status = rpcb_gettime( $timep, $host );
$status = rpcb_gettime( $timep );
The XS code, with ellipsis, follows.
bool_t
rpcb_gettime(timep, ...)
time_t timep = NO_INIT
PREINIT:
char *host = "localhost";
CODE:
if( items > 1 )
host = (char *)SvPV_nolen(ST(1));
RETVAL = rpcb_gettime( host, &timep );
OUTPUT:
timep
RETVAL
The C_ARGS: keyword allows creating of XSUBS which have different calling sequence from Perl than from C, without a need to write CODE: or PPCODE: section. The contents of the C_ARGS: paragraph is put as the argument to the called C function without any change.
For example, suppose that a C function is declared as
symbolic nth_derivative(int n, symbolic function, int flags);
and that the default flags are kept in a global C variable
default_flags. Suppose that you want to create an interface which
is called as
$second_deriv = $function->nth_derivative(2);
To do this, declare the XSUB as
symbolic
nth_derivative(function, n)
symbolic function
int n
C_ARGS:
n, function, default_flags
The PPCODE: keyword is an alternate form of the CODE: keyword and is used to tell the xsubpp compiler that the programmer is supplying the code to control the argument stack for the XSUBs return values. Occasionally one will want an XSUB to return a list of values rather than a single value. In these cases one must use PPCODE: and then explicitly push the list of values on the stack. The PPCODE: and CODE: keywords should not be used together within the same XSUB.
The actual difference between PPCODE: and CODE: sections is in the
initialization of SP macro (which stands for the current Perl
stack pointer), and in the handling of data on the stack when returning
from an XSUB. In CODE: sections SP preserves the value which was on
entry to the XSUB: SP is on the function pointer (which follows the
last parameter). In PPCODE: sections SP is moved backward to the
beginning of the parameter list, which allows PUSH*() macros
to place output values in the place Perl expects them to be when
the XSUB returns back to Perl.
The generated trailer for a CODE: section ensures that the number of return
values Perl will see is either 0 or 1 (depending on the voidness of the
return value of the C function, and heuristics mentioned in
The RETVAL Variable). The trailer generated for a PPCODE: section
is based on the number of return values and on the number of times
SP was updated by [X]PUSH*() macros.
Note that macros ST(i), XST_m*() and XSRETURN*() work equally
well in CODE: sections and PPCODE: sections.
The following XSUB will call the C rpcb_gettime() function
and will return its two output values, timep and status, to
Perl as a single list.
void
rpcb_gettime(host)
char *host
PREINIT:
time_t timep;
bool_t status;
PPCODE:
status = rpcb_gettime( host, &timep );
EXTEND(SP, 2);
PUSHs(sv_2mortal(newSViv(status)));
PUSHs(sv_2mortal(newSViv(timep)));
Notice that the programmer must supply the C code necessary
to have the real rpcb_gettime() function called and to have
the return values properly placed on the argument stack.
The void return type for this function tells the xsubpp compiler that
the RETVAL variable is not needed or used and that it should not be created.
In most scenarios the void return type should be used with the PPCODE:
directive.
The EXTEND() macro is used to make room on the argument
stack for 2 return values. The PPCODE: directive causes the
xsubpp compiler to create a stack pointer available as SP, and it
is this pointer which is being used in the EXTEND() macro.
The values are then pushed onto the stack with the PUSHs()
macro.
Now the rpcb_gettime() function can be used from Perl with
the following statement.
($status, $timep) = rpcb_gettime("localhost");
When handling output parameters with a PPCODE section, be sure to handle 'set' magic properly. See the perlguts manpage for details about 'set' magic.
Occasionally the programmer will want to return simply
undef or an empty list if a function fails rather than a
separate status value. The rpcb_gettime() function offers
just this situation. If the function succeeds we would like
to have it return the time and if it fails we would like to
have undef returned. In the following Perl code the value
of $timep will either be undef or it will be a valid time.
$timep = rpcb_gettime( "localhost" );
The following XSUB uses the SV * return type as a mnemonic only,
and uses a CODE: block to indicate to the compiler
that the programmer has supplied all the necessary code. The
sv_newmortal() call will initialize the return value to undef, making that
the default return value.
SV *
rpcb_gettime(host)
char * host
PREINIT:
time_t timep;
bool_t x;
CODE:
ST(0) = sv_newmortal();
if( rpcb_gettime( host, &timep ) )
sv_setnv( ST(0), (double)timep);
The next example demonstrates how one would place an explicit undef in the return value, should the need arise.
SV *
rpcb_gettime(host)
char * host
PREINIT:
time_t timep;
bool_t x;
CODE:
if( rpcb_gettime( host, &timep ) ){
ST(0) = sv_newmortal();
sv_setnv( ST(0), (double)timep);
}
else{
ST(0) = &PL_sv_undef;
}
To return an empty list one must use a PPCODE: block and then not push return values on the stack.
void
rpcb_gettime(host)
char *host
PREINIT:
time_t timep;
PPCODE:
if( rpcb_gettime( host, &timep ) )
PUSHs(sv_2mortal(newSViv(timep)));
else{
/* Nothing pushed on stack, so an empty
* list is implicitly returned. */
}
Some people may be inclined to include an explicit return in the above
XSUB, rather than letting control fall through to the end. In those
situations XSRETURN_EMPTY should be used, instead. This will ensure that
the XSUB stack is properly adjusted. Consult perlapi for other
XSRETURN macros.
Since XSRETURN_* macros can be used with CODE blocks as well, one can
rewrite this example as:
int
rpcb_gettime(host)
char *host
PREINIT:
time_t timep;
CODE:
RETVAL = rpcb_gettime( host, &timep );
if (RETVAL == 0)
XSRETURN_UNDEF;
OUTPUT:
RETVAL
In fact, one can put this check into a POSTCALL: section as well. Together with PREINIT: simplifications, this leads to:
int
rpcb_gettime(host)
char *host
time_t timep;
POSTCALL:
if (RETVAL == 0)
XSRETURN_UNDEF;
The REQUIRE: keyword is used to indicate the minimum version of the xsubpp compiler needed to compile the XS module. An XS module which contains the following statement will compile with only xsubpp version 1.922 or greater:
REQUIRE: 1.922
This keyword can be used when an XSUB requires special cleanup procedures before it terminates. When the CLEANUP: keyword is used it must follow any CODE:, or OUTPUT: blocks which are present in the XSUB. The code specified for the cleanup block will be added as the last statements in the XSUB.
This keyword can be used when an XSUB requires special procedures executed after the C subroutine call is performed. When the POSTCALL: keyword is used it must precede OUTPUT: and CLEANUP: blocks which are present in the XSUB.
See examples in The NO_OUTPUT Keyword and Returning Undef And Empty Lists.
The POSTCALL: block does not make a lot of sense when the C subroutine call is supplied by user by providing either CODE: or PPCODE: section.
The BOOT: keyword is used to add code to the extension's bootstrap function. The bootstrap function is generated by the xsubpp compiler and normally holds the statements necessary to register any XSUBs with Perl. With the BOOT: keyword the programmer can tell the compiler to add extra statements to the bootstrap function.
This keyword may be used any time after the first MODULE keyword and should appear on a line by itself. The first blank line after the keyword will terminate the code block.
BOOT:
# The following message will be printed when the
# bootstrap function executes.
printf("Hello from the bootstrap!\n");
The VERSIONCHECK: keyword corresponds to xsubpp's -versioncheck and
-noversioncheck options. This keyword overrides the command line
options. Version checking is enabled by default. When version checking is
enabled the XS module will attempt to verify that its version matches the
version of the PM module.
To enable version checking:
VERSIONCHECK: ENABLE
To disable version checking:
VERSIONCHECK: DISABLE
Note that if the version of the PM module is an NV (a floating point number), it will be stringified with a possible loss of precision (currently chopping to nine decimal places) so that it may not match the version of the XS module anymore. Quoting the $VERSION declaration to make it a string is recommended if long version numbers are used.
The PROTOTYPES: keyword corresponds to xsubpp's -prototypes and
-noprototypes options. This keyword overrides the command line options.
Prototypes are disabled by default. When prototypes are enabled, XSUBs will
be given Perl prototypes. This keyword may be used multiple times in an XS
module to enable and disable prototypes for different parts of the module.
Note that xsubpp will nag you if you don't explicitly enable or disable
prototypes, with:
Please specify prototyping behavior for Foo.xs (see perlxs manual)
To enable prototypes:
PROTOTYPES: ENABLE
To disable prototypes:
PROTOTYPES: DISABLE
This keyword is similar to the PROTOTYPES: keyword above but can be used to force xsubpp to use a specific prototype for the XSUB. This keyword overrides all other prototype options and keywords but affects only the current XSUB. Consult Prototypes in the perlsub manpage for information about Perl prototypes.
bool_t
rpcb_gettime(timep, ...)
time_t timep = NO_INIT
PROTOTYPE: $;$
PREINIT:
char *host = "localhost";
CODE:
if( items > 1 )
host = (char *)SvPV_nolen(ST(1));
RETVAL = rpcb_gettime( host, &timep );
OUTPUT:
timep
RETVAL
If the prototypes are enabled, you can disable it locally for a given XSUB as in the following example:
void
rpcb_gettime_noproto()
PROTOTYPE: DISABLE
...
The ALIAS: keyword allows an XSUB to have two or more unique Perl names
and to know which of those names was used when it was invoked. The Perl
names may be fully-qualified with package names. Each alias is given an
index. The compiler will setup a variable called ix which contain the
index of the alias which was used. When the XSUB is called with its
declared name ix will be 0.
The following example will create aliases FOO::gettime() and
BAR::getit() for this function.
bool_t
rpcb_gettime(host,timep)
char *host
time_t &timep
ALIAS:
FOO::gettime = 1
BAR::getit = 2
INIT:
printf("# ix = %d\n", ix );
OUTPUT:
timep
Instead of writing an overloaded interface using pure Perl, you
can also use the OVERLOAD keyword to define additional Perl names
for your functions (like the ALIAS: keyword above). However, the
overloaded functions must be defined in such a way as to accept the number
of parameters supplied by perl's overload system. For most overload
methods, it will be three parameters; for the nomethod function it will
be four. However, the bitwise operators &, |, ^, and ~ may be
called with three or five arguments (see the overload manpage).
If any function has the OVERLOAD: keyword, several additional lines will be defined in the c file generated by xsubpp in order to register with the overload magic.
Since blessed objects are actually stored as RV's, it is useful to use the typemap features to preprocess parameters and extract the actual SV stored within the blessed RV. See the sample for T_PTROBJ_SPECIAL below.
To use the OVERLOAD: keyword, create an XS function which takes three input parameters (or use the C-style '...' definition) like this:
SV *
cmp (lobj, robj, swap)
My_Module_obj lobj
My_Module_obj robj
IV swap
OVERLOAD: cmp <=>
{ /* function defined here */}
In this case, the function will overload both of the three way comparison operators. For all overload operations using non-alpha characters, you must type the parameter without quoting, separating multiple overloads with whitespace. Note that ``'' (the stringify overload) should be entered as \``\'' (i.e. escaped).
Since, as mentioned above, bitwise operators may take extra arguments, you
may want to use something like (lobj, robj, swap, ...) (with
literal ...) as your parameter list.
In addition to the OVERLOAD keyword, if you need to control how Perl autogenerates missing overloaded operators, you can set the FALLBACK keyword in the module header section, like this:
MODULE = RPC PACKAGE = RPC
FALLBACK: TRUE
...
where FALLBACK can take any of the three values TRUE, FALSE, or UNDEF. If you do not set any FALLBACK value when using OVERLOAD, it defaults to UNDEF. FALLBACK is not used except when one or more functions using OVERLOAD have been defined. Please see fallback in the overload manpage for more details.
This keyword declares the current XSUB as a keeper of the given calling signature. If some text follows this keyword, it is considered as a list of functions which have this signature, and should be attached to the current XSUB.
For example, if you have 4 C functions multiply(), divide(), add(),
subtract() all having the signature:
symbolic f(symbolic, symbolic);
you can make them all to use the same XSUB using this:
symbolic
interface_s_ss(arg1, arg2)
symbolic arg1
symbolic arg2
INTERFACE:
multiply divide
add subtract
(This is the complete XSUB code for 4 Perl functions!) Four generated Perl function share names with corresponding C functions.
The advantage of this approach comparing to ALIAS: keyword is that there
is no need to code a switch statement, each Perl function (which shares
the same XSUB) knows which C function it should call. Additionally, one
can attach an extra function remainder() at runtime by using
CV *mycv = newXSproto("Symbolic::remainder",
XS_Symbolic_interface_s_ss, __FILE__, "$$");
XSINTERFACE_FUNC_SET(mycv, remainder);
say, from another XSUB. (This example supposes that there was no
INTERFACE_MACRO: section, otherwise one needs to use something else instead of
XSINTERFACE_FUNC_SET, see the next section.)
This keyword allows one to define an INTERFACE using a different way
to extract a function pointer from an XSUB. The text which follows
this keyword should give the name of macros which would extract/set a
function pointer. The extractor macro is given return type, CV*,
and XSANY.any_dptr for this CV*. The setter macro is given cv,
and the function pointer.
The default value is XSINTERFACE_FUNC and XSINTERFACE_FUNC_SET.
An INTERFACE keyword with an empty list of functions can be omitted if
INTERFACE_MACRO keyword is used.
Suppose that in the previous example functions pointers for
multiply(), divide(), add(), subtract() are kept in a global C array
fp[] with offsets being multiply_off, divide_off, add_off,
subtract_off. Then one can use
#define XSINTERFACE_FUNC_BYOFFSET(ret,cv,f) \
((XSINTERFACE_CVT_ANON(ret))fp[CvXSUBANY(cv).any_i32])
#define XSINTERFACE_FUNC_BYOFFSET_set(cv,f) \
CvXSUBANY(cv).any_i32 = CAT2( f, _off )
in C section,
symbolic
interface_s_ss(arg1, arg2)
symbolic arg1
symbolic arg2
INTERFACE_MACRO:
XSINTERFACE_FUNC_BYOFFSET
XSINTERFACE_FUNC_BYOFFSET_set
INTERFACE:
multiply divide
add subtract
in XSUB section.
This keyword can be used to pull other files into the XS module. The other files may have XS code. INCLUDE: can also be used to run a command to generate the XS code to be pulled into the module.
The file Rpcb1.xsh contains our rpcb_gettime() function:
bool_t
rpcb_gettime(host,timep)
char *host
time_t &timep
OUTPUT:
timep
The XS module can use INCLUDE: to pull that file into it.
INCLUDE: Rpcb1.xsh
If the parameters to the INCLUDE: keyword are followed by a pipe (|) then
the compiler will interpret the parameters as a command. This feature is
mildly deprecated in favour of the INCLUDE_COMMAND: directive, as documented
below.
INCLUDE: cat Rpcb1.xsh |
Do not use this to run perl: INCLUDE: perl | will run the perl that
happens to be the first in your path and not necessarily the same perl that is
used to run xsubpp. See The INCLUDE_COMMAND: Keyword.
Runs the supplied command and includes its output into the current XS
document. INCLUDE_COMMAND assigns special meaning to the $^X token
in that it runs the same perl interpreter that is running xsubpp:
INCLUDE_COMMAND: cat Rpcb1.xsh
INCLUDE_COMMAND: $^X -e ...
The CASE: keyword allows an XSUB to have multiple distinct parts with each part acting as a virtual XSUB. CASE: is greedy and if it is used then all other XS keywords must be contained within a CASE:. This means nothing may precede the first CASE: in the XSUB and anything following the last CASE: is included in that case.
A CASE: might switch via a parameter of the XSUB, via the ix ALIAS:
variable (see The ALIAS: Keyword), or maybe via the items variable
(see Variable-length Parameter Lists). The last CASE: becomes the
default case if it is not associated with a conditional. The following
example shows CASE switched via ix with a function rpcb_gettime()
having an alias x_gettime(). When the function is called as
rpcb_gettime() its parameters are the usual (char *host, time_t *timep),
but when the function is called as x_gettime() its parameters are
reversed, (time_t *timep, char *host).
long
rpcb_gettime(a,b)
CASE: ix == 1
ALIAS:
x_gettime = 1
INPUT:
# 'a' is timep, 'b' is host
char *b
time_t a = NO_INIT
CODE:
RETVAL = rpcb_gettime( b, &a );
OUTPUT:
a
RETVAL
CASE:
# 'a' is host, 'b' is timep
char *a
time_t &b = NO_INIT
OUTPUT:
b
RETVAL
That function can be called with either of the following statements. Note the different argument lists.
$status = rpcb_gettime( $host, $timep );
$status = x_gettime( $timep, $host );
The EXPORT_XSUB_SYMBOLS: keyword is likely something you will never need.
In perl versions earlier than 5.16.0, this keyword does nothing. Starting
with 5.16, XSUB symbols are no longer exported by default. That is, they
are static functions. If you include
EXPORT_XSUB_SYMBOLS: ENABLE
in your XS code, the XSUBs following this line will not be declared static.
You can later disable this with
EXPORT_XSUB_SYMBOLS: DISABLE
which, again, is the default that you should probably never change.
You cannot use this keyword on versions of perl before 5.16 to make
XSUBs static.
The & unary operator in the INPUT: section is used to tell xsubpp
that it should convert a Perl value to/from C using the C type to the left
of &, but provide a pointer to this value when the C function is called.
This is useful to avoid a CODE: block for a C function which takes a parameter
by reference. Typically, the parameter should be not a pointer type (an
int or long but not an int* or long*).
The following XSUB will generate incorrect C code. The xsubpp compiler will
turn this into code which calls rpcb_gettime() with parameters (char
*host, time_t timep), but the real rpcb_gettime() wants the timep
parameter to be of type time_t* rather than time_t.
bool_t
rpcb_gettime(host,timep)
char *host
time_t timep
OUTPUT:
timep
That problem is corrected by using the & operator. The xsubpp compiler
will now turn this into code which calls rpcb_gettime() correctly with
parameters (char *host, time_t *timep). It does this by carrying the
& through, so the function call looks like rpcb_gettime(host, &timep).
bool_t
rpcb_gettime(host,timep)
char *host
time_t &timep
OUTPUT:
timep
C preprocessor directives are allowed within BOOT:, PREINIT: INIT:, CODE:,
PPCODE:, POSTCALL:, and CLEANUP: blocks, as well as outside the functions.
Comments are allowed anywhere after the MODULE keyword. The compiler will
pass the preprocessor directives through untouched and will remove the
commented lines. POD documentation is allowed at any point, both in the
C and XS language sections. POD must be terminated with a =cut command;
xsubpp will exit with an error if it does not. It is very unlikely that
human generated C code will be mistaken for POD, as most indenting styles
result in whitespace in front of any line starting with =. Machine
generated XS files may fall into this trap unless care is taken to
ensure that a space breaks the sequence ``\n=''.
Comments can be added to XSUBs by placing a # as the first
non-whitespace of a line. Care should be taken to avoid making the
comment look like a C preprocessor directive, lest it be interpreted as
such. The simplest way to prevent this is to put whitespace in front of
the #.
If you use preprocessor directives to choose one of two versions of a function, use
#if ... version1
#else /* ... version2 */
#endif
and not
#if ... version1
#endif
#if ... version2
#endif
because otherwise xsubpp will believe that you made a duplicate definition of the function. Also, put a blank line before the #else/#endif so it will not be seen as part of the function body.
If an XSUB name contains ::, it is considered to be a C++ method.
The generated Perl function will assume that
its first argument is an object pointer. The object pointer
will be stored in a variable called THIS. The object should
have been created by C++ with the new() function and should
be blessed by Perl with the sv_setref_pv() macro. The
blessing of the object by Perl can be handled by a typemap. An example
typemap is shown at the end of this section.
If the return type of the XSUB includes static, the method is considered
to be a static method. It will call the C++
function using the class::method() syntax. If the method is not static
the function will be called using the THIS->method() syntax.
The next examples will use the following C++ class.
class color {
public:
color();
~color();
int blue();
void set_blue( int );
private:
int c_blue;
};
The XSUBs for the blue() and set_blue() methods are defined with the class
name but the parameter for the object (THIS, or ``self'') is implicit and is
not listed.
int
color::blue()
void
color::set_blue( val )
int val
Both Perl functions will expect an object as the first parameter. In the
generated C++ code the object is called THIS, and the method call will
be performed on this object. So in the C++ code the blue() and set_blue()
methods will be called as this:
RETVAL = THIS->blue();
THIS->set_blue( val );
You could also write a single get/set method using an optional argument:
int
color::blue( val = NO_INIT )
int val
PROTOTYPE $;$
CODE:
if (items > 1)
THIS->set_blue( val );
RETVAL = THIS->blue();
OUTPUT:
RETVAL
If the function's name is DESTROY then the C++ delete function will be
called and THIS will be given as its parameter. The generated C++ code for
void
color::DESTROY()
will look like this:
color *THIS = ...; // Initialized as in typemap
delete THIS;
If the function's name is new then the C++ new function will be called
to create a dynamic C++ object. The XSUB will expect the class name, which
will be kept in a variable called CLASS, to be given as the first
argument.
color *
color::new()
The generated C++ code will call new.
RETVAL = new color();
The following is an example of a typemap that could be used for this C++ example.
TYPEMAP
color * O_OBJECT
OUTPUT
# The Perl object is blessed into 'CLASS', which should be a
# char* having the name of the package for the blessing.
O_OBJECT
sv_setref_pv( $arg, CLASS, (void*)$var );
INPUT
O_OBJECT
if( sv_isobject($arg) && (SvTYPE(SvRV($arg)) == SVt_PVMG) )
$var = ($type)SvIV((SV*)SvRV( $arg ));
else{
warn("${Package}::$func_name() -- " .
"$var is not a blessed SV reference");
XSRETURN_UNDEF;
}
When designing an interface between Perl and a C library a straight
translation from C to XS (such as created by h2xs -x) is often sufficient.
However, sometimes the interface will look
very C-like and occasionally nonintuitive, especially when the C function
modifies one of its parameters, or returns failure inband (as in ``negative
return values mean failure''). In cases where the programmer wishes to
create a more Perl-like interface the following strategy may help to
identify the more critical parts of the interface.
Identify the C functions with input/output or output parameters. The XSUBs for these functions may be able to return lists to Perl.
Identify the C functions which use some inband info as an indication of failure. They may be candidates to return undef or an empty list in case of failure. If the failure may be detected without a call to the C function, you may want to use an INIT: section to report the failure. For failures detectable after the C function returns one may want to use a POSTCALL: section to process the failure. In more complicated cases use CODE: or PPCODE: sections.
If many functions use the same failure indication based on the return value, you may want to create a special typedef to handle this situation. Put
typedef int negative_is_failure;
near the beginning of XS file, and create an OUTPUT typemap entry
for negative_is_failure which converts negative values to undef, or
maybe croak()s. After this the return value of type negative_is_failure
will create more Perl-like interface.
Identify which values are used by only the C and XSUB functions themselves, say, when a parameter to a function should be a contents of a global variable. If Perl does not need to access the contents of the value then it may not be necessary to provide a translation for that value from C to Perl.
Identify the pointers in the C function parameter lists and return
values. Some pointers may be used to implement input/output or
output parameters, they can be handled in XS with the & unary operator,
and, possibly, using the NO_INIT keyword.
Some others will require handling of types like int *, and one needs
to decide what a useful Perl translation will do in such a case. When
the semantic is clear, it is advisable to put the translation into a typemap
file.
Identify the structures used by the C functions. In many
cases it may be helpful to use the T_PTROBJ typemap for
these structures so they can be manipulated by Perl as
blessed objects. (This is handled automatically by h2xs -x.)
If the same C type is used in several different contexts which require
different translations, typedef several new types mapped to this C type,
and create separate typemap entries for these new types. Use these
types in declarations of return type and parameters to XSUBs.
When dealing with C structures one should select either T_PTROBJ or T_PTRREF for the XS type. Both types are designed to handle pointers to complex objects. The T_PTRREF type will allow the Perl object to be unblessed while the T_PTROBJ type requires that the object be blessed. By using T_PTROBJ one can achieve a form of type-checking because the XSUB will attempt to verify that the Perl object is of the expected type.
The following XS code shows the getnetconfigent() function which is used
with ONC+ TIRPC. The getnetconfigent() function will return a pointer to a
C structure and has the C prototype shown below. The example will
demonstrate how the C pointer will become a Perl reference. Perl will
consider this reference to be a pointer to a blessed object and will
attempt to call a destructor for the object. A destructor will be
provided in the XS source to free the memory used by getnetconfigent().
Destructors in XS can be created by specifying an XSUB function whose name
ends with the word DESTROY. XS destructors can be used to free memory
which may have been malloc'd by another XSUB.
struct netconfig *getnetconfigent(const char *netid);
A typedef will be created for struct netconfig. The Perl
object will be blessed in a class matching the name of the C
type, with the tag Ptr appended, and the name should not
have embedded spaces if it will be a Perl package name. The
destructor will be placed in a class corresponding to the
class of the object and the PREFIX keyword will be used to
trim the name to the word DESTROY as Perl will expect.
typedef struct netconfig Netconfig;
MODULE = RPC PACKAGE = RPC
Netconfig *
getnetconfigent(netid)
char *netid
MODULE = RPC PACKAGE = NetconfigPtr PREFIX = rpcb_
void
rpcb_DESTROY(netconf)
Netconfig *netconf
CODE:
printf("Now in NetconfigPtr::DESTROY\n");
free( netconf );
This example requires the following typemap entry. Consult the perlxstypemap manpage for more information about adding new typemaps for an extension.
TYPEMAP
Netconfig * T_PTROBJ
This example will be used with the following Perl statements.
use RPC;
$netconf = getnetconfigent("udp");
When Perl destroys the object referenced by $netconf it will send the
object to the supplied XSUB DESTROY function. Perl cannot determine, and
does not care, that this object is a C struct and not a Perl object. In
this sense, there is no difference between the object created by the
getnetconfigent() XSUB and an object created by a normal Perl subroutine.
Starting with Perl 5.8, a macro framework has been defined to allow static data to be safely stored in XS modules that will be accessed from a multi-threaded Perl.
Although primarily designed for use with multi-threaded Perl, the macros have been designed so that they will work with non-threaded Perl as well.
It is therefore strongly recommended that these macros be used by all XS modules that make use of static data.
The easiest way to get a template set of macros to use is by specifying
the -g (--global) option with h2xs (see h2xs).
Below is an example module that makes use of the macros.
#define PERL_NO_GET_CONTEXT
#include "EXTERN.h"
#include "perl.h"
#include "XSUB.h"
/* Global Data */
#define MY_CXT_KEY "BlindMice::_guts" XS_VERSION
typedef struct {
int count;
char name[3][100];
} my_cxt_t;
START_MY_CXT
MODULE = BlindMice PACKAGE = BlindMice
BOOT:
{
MY_CXT_INIT;
MY_CXT.count = 0;
strcpy(MY_CXT.name[0], "None");
strcpy(MY_CXT.name[1], "None");
strcpy(MY_CXT.name[2], "None");
}
int
newMouse(char * name)
PREINIT:
dMY_CXT;
CODE:
if (MY_CXT.count >= 3) {
warn("Already have 3 blind mice");
RETVAL = 0;
}
else {
RETVAL = ++ MY_CXT.count;
strcpy(MY_CXT.name[MY_CXT.count - 1], name);
}
OUTPUT:
RETVAL
char *
get_mouse_name(index)
int index
PREINIT:
dMY_CXT;
CODE:
if (index > MY_CXT.count)
croak("There are only 3 blind mice.");
else
RETVAL = MY_CXT.name[index - 1];
OUTPUT:
RETVAL
void
CLONE(...)
CODE:
MY_CXT_CLONE;
#define MY_CXT_KEY "MyModule::_guts" XS_VERSION
my_cxt_t. The other
CXT* macros assume the existence of the my_cxt_t typedef name.
Declare a typedef named my_cxt_t that is a structure that contains
all the data that needs to be interpreter-local.
typedef struct {
int some_value;
} my_cxt_t;
my_cxt_t.
my_cxt_t struct.
It must be called exactly once, typically in a BOOT: section. If you are maintaining multiple interpreters, it should be called once in each interpreter instance, except for interpreters cloned from existing ones. (But see MY_CXT_CLONE below.)
my_cxt_t struct. For
example, if my_cxt_t is
typedef struct {
int index;
} my_cxt_t;
then use this to access the index member
dMY_CXT;
MY_CXT.index = 2;
dMY_CXT may be quite expensive to calculate, and to avoid the overhead
of invoking it in each function it is possible to pass the declaration
onto other functions using the aMY_CXT/pMY_CXT macros, eg
void sub1() {
dMY_CXT;
MY_CXT.index = 1;
sub2(aMY_CXT);
}
void sub2(pMY_CXT) {
MY_CXT.index = 2;
}
Analogously to pTHX, there are equivalent forms for when the macro is the
first or last in multiple arguments, where an underscore represents a
comma, i.e. _aMY_CXT, aMY_CXT_, _pMY_CXT and pMY_CXT_.
threads->create()), both interpreters share the same physical
my_cxt_t structure. Calling MY_CXT_CLONE (typically via the package's
CLONE() function), causes a byte-for-byte copy of the structure to be
taken, and any future dMY_CXT will cause the copy to be accessed instead.
MY_CXT_INIT_INTERP(my_perl)dMY_CXT_INTERP(my_perl)Note that these macros will only work together within the same source file; that is, a dMY_CTX in one source file will access a different structure than a dMY_CTX in another source file.
Starting from Perl 5.8, in C/C++ level Perl knows how to wrap
system/library interfaces that have thread-aware versions
(e.g. getpwent_r()) into frontend macros (e.g. getpwent()) that
correctly handle the multithreaded interaction with the Perl
interpreter. This will happen transparently, the only thing
you need to do is to instantiate a Perl interpreter.
This wrapping happens always when compiling Perl core source (PERL_CORE is defined) or the Perl core extensions (PERL_EXT is defined). When compiling XS code outside of the Perl core, the wrapping does not take place before Perl 5.28. Starting in that release you can
#define PERL_REENTRANT
in your code to enable the wrapping. It is advisable to do so if you
are using such functions, as intermixing the _r-forms (as Perl compiled
for multithreaded operation will do) and the _r-less forms is neither
well-defined (inconsistent results, data corruption, or even crashes
become more likely), nor is it very portable. Unfortunately, not all
systems have all the _r forms, but using this #define gives you
whatever protection that Perl is aware is available on each system.
File RPC.xs: Interface to some ONC+ RPC bind library functions.
#define PERL_NO_GET_CONTEXT
#include "EXTERN.h"
#include "perl.h"
#include "XSUB.h"
#include <rpc/rpc.h>
typedef struct netconfig Netconfig;
MODULE = RPC PACKAGE = RPC
SV *
rpcb_gettime(host="localhost")
char *host
PREINIT:
time_t timep;
CODE:
ST(0) = sv_newmortal();
if( rpcb_gettime( host, &timep ) )
sv_setnv( ST(0), (double)timep );
Netconfig *
getnetconfigent(netid="udp")
char *netid
MODULE = RPC PACKAGE = NetconfigPtr PREFIX = rpcb_
void
rpcb_DESTROY(netconf)
Netconfig *netconf
CODE:
printf("NetconfigPtr::DESTROY\n");
free( netconf );
File typemap: Custom typemap for RPC.xs. (cf. the perlxstypemap manpage)
TYPEMAP
Netconfig * T_PTROBJ
File RPC.pm: Perl module for the RPC extension.
package RPC;
require Exporter;
require DynaLoader;
@ISA = qw(Exporter DynaLoader);
@EXPORT = qw(rpcb_gettime getnetconfigent);
bootstrap RPC;
1;
File rpctest.pl: Perl test program for the RPC extension.
use RPC;
$netconf = getnetconfigent();
$a = rpcb_gettime();
print "time = $a\n";
print "netconf = $netconf\n";
$netconf = getnetconfigent("tcp");
$a = rpcb_gettime("poplar");
print "time = $a\n";
print "netconf = $netconf\n";
XS code has full access to system calls including C library functions. It thus has the capability of interfering with things that the Perl core or other modules have set up, such as signal handlers or file handles. It could mess with the memory, or any number of harmful things. Don't.
Some modules have an event loop, waiting for user-input. It is highly unlikely that two such modules would work adequately together in a single Perl application.
In general, the perl interpreter views itself as the center of the universe as far as the Perl program goes. XS code is viewed as a help-mate, to accomplish things that perl doesn't do, or doesn't do fast enough, but always subservient to perl. The closer XS code adheres to this model, the less likely conflicts will occur.
One area where there has been conflict is in regards to C locales. (See
the perllocale manpage.) perl, with one exception and unless told otherwise,
sets up the underlying locale the program is running in to the locale
passed
into it from the environment. This is an important difference from a
generic C language program, where the underlying locale is the ``C''
locale unless the program changes it. As of v5.20, this underlying
locale is completely hidden from pure Perl code outside the lexical
scope of use locale except for a couple of function calls in the
POSIX module which of necessity use it. But the underlying locale, with
that
one exception is exposed to XS code, affecting all C library routines
whose behavior is locale-dependent. Your XS code better not assume that
the underlying locale is ``C''. The exception is the
LC_NUMERIC
locale category, and the reason it is an exception is that experience
has shown that it can be problematic for XS code, whereas we have not
had reports of problems with the
other locale categories. And the reason
for this one category being problematic is that the character used as a
decimal point can vary. Many European languages use a comma, whereas
English, and hence Perl are expecting a dot (U+002E: FULL STOP). Many
modules can handle only the radix character being a dot, and so perl
attempts to make it so. Up through Perl v5.20, the attempt was merely
to set LC_NUMERIC upon startup to the "C" locale. Any
setlocale() otherwise would change
it; this caused some failures. Therefore, starting in v5.22, perl tries
to keep LC_NUMERIC always set to "C" for XS code.
To summarize, here's what to expect and how to handle locales in XS code:
The current locale is exposed to XS code except possibly LC_NUMERIC
(explained in the next paragraph).
There have not been reports of problems with the other categories.
Perl initializes things on start-up so that the current locale is the
one which is indicated by the user's environment in effect at that time.
See ENVIRONMENT in the perllocale manpage.
However, up through v5.20, Perl initialized things on start-up so that
LC_NUMERIC was set to the ``C'' locale. But if any code anywhere
changed it, it would stay changed. This means that your module can't
count on LC_NUMERIC being something in particular, and you can't
expect floating point numbers (including version strings) to have dots
in them. If you don't allow for a non-dot, your code could break if
anyone anywhere changed the locale. For this reason, v5.22 changed
the behavior so that Perl tries to keep LC_NUMERIC in the ``C'' locale
except around the operations internally where it should be something
else. Misbehaving XS code will always be able to change the locale
anyway, but the most common instance of this is checked for and
handled.
LC_NUMERIC, as perl has
already set the others up. XS code should avoid changing the locale, as
it can adversely affect other, unrelated, code and may not be
thread-safe. To minimize problems, the macros
perlapi/STORE_LC_NUMERIC_SET_TO_NEEDED,
perlapi/STORE_LC_NUMERIC_FORCE_TO_UNDERLYING, and
perlapi/RESTORE_LC_NUMERIC should be used to affect any needed
change.
But, starting with Perl v5.28, locales are thread-safe on platforms that
support this functionality. Windows has this starting with Visual
Studio 2005. Many other modern platforms support the thread-safe POSIX
2008 functions. The C #define USE_THREAD_SAFE_LOCALE will be
defined iff this build is using these. From Perl-space, the read-only
variable ${SAFE_LOCALES} is 1 if either the build is not threaded, or
if USE_THREAD_SAFE_LOCALE is defined; otherwise it is 0.
The way this works under-the-hood is that every thread has a choice of
using a locale specific to it (this is the Windows and POSIX 2008
functionality), or the global locale that is accessible to all threads
(this is the functionality that has always been there). The
implementations for Windows and POSIX are completely different. On
Windows, the runtime can be set up so that the standard
setlocale(3) function either only knows about the global locale or
the locale for this thread. On POSIX, setlocale always deals with
the global locale, and other functions have been created to handle
per-thread locales. Perl makes this transparent to perl-space code. It
continues to use POSIX::setlocale(), and the interpreter translates
that into the per-thread functions.
All other locale-senstive functions automatically use the per-thread
locale, if that is turned on, and failing that, the global locale. Thus
calls to setlocale are ineffective on POSIX systems for the current
thread if that thread is using a per-thread locale. If perl is compiled
for single-thread operation, it does not use the per-thread functions,
so setlocale does work as expected.
If you have loaded the POSIX module you can use the methods given
in the perlcall manpage to call POSIX::setlocale to safely
change or query the locale (on systems where it is safe to do so), or
you can use the new 5.28 function perlapi/Perl_setlocale instead,
which is a drop-in replacement for the system setlocale(3), and
handles single-threaded and multi-threaded applications transparently.
There are some locale-related library calls that still aren't thread-safe because they return data in a buffer global to all threads. In the past, these didn't matter as locales weren't thread-safe at all. But now you have to be aware of them in case your module is called in a multi-threaded application. The known ones are
asctime() ctime() gcvt() [POSIX.1-2001 only (function removed in POSIX.1-2008)] getdate() wcrtomb() if its final argument is NULL wcsrtombs() if its final argument is NULL wcstombs() wctomb()
Some of these shouldn't really be called in a Perl application, and for others there are thread-safe versions of these already implemented:
asctime_r() ctime_r() Perl_langinfo()
The _r forms are automatically used, starting in Perl 5.28, if you
compile your code, with
#define PERL_REENTRANT
See also perlapi/Perl_langinfo. You can use the methods given in the perlcall manpage, to get the best available locale-safe versions of these
POSIX::localeconv() POSIX::wcstombs() POSIX::wctomb()
And note, that some items returned by Localeconv are available
through perlapi/Perl_langinfo.
The others shouldn't be used in a threaded application.
Some modules may call a non-perl library that is locale-aware. This is
fine as long as it doesn't try to query or change the locale using the
system setlocale. But if these do call the system setlocale,
those calls may be ineffective. Instead,
Perl_setlocale works in all circumstances.
Plain setlocale is ineffective on multi-threaded POSIX 2008 systems. It
operates only on the global locale, whereas each thread has its own
locale, paying no attention to the global one. Since converting
these non-Perl libraries to Perl_setlocale is out of the question,
there is a new function in v5.28
switch_to_global_locale that will
switch the thread it is called from so that any system setlocale
calls will have their desired effect. The function
sync_locale must be called before returning to
perl.
This thread can change the locale all it wants and it won't affect any other thread, except any that also have been switched to the global locale. This means that a multi-threaded application can have a single thread using an alien library without a problem; but no more than a single thread can be so-occupied. Bad results likely will happen.
In perls without multi-thread locale support, some alien libraries,
such as Gtk change locales. This can cause problems for the Perl
core and other modules. For these, before control is returned to
perl, starting in v5.20.1, calling the function
sync_locale() from XS should be sufficient to
avoid most of these problems. Prior to this, you need a pure Perl
statement that does this:
POSIX::setlocale(LC_ALL, POSIX::setlocale(LC_ALL));
or use the methods given in the perlcall manpage.
This document covers features supported by ExtUtils::ParseXS
(also known as xsubpp) 3.13_01.
Originally written by Dean Roehrich <roehrich@cray.com>.
Maintained since 1996 by The Perl Porters <perlbug@perl.org>.
| perlxs - XS language reference manual |