User-defined functions can be written in C (or a language that can be made compatible with C, such as C++). Such functions are compiled into dynamically loadable objects (also called shared libraries) and are loaded by the server on demand. The dynamic loading feature is what distinguishes “C language” functions from “internal” functions — the actual coding conventions are essentially the same for both. (Hence, the standard internal function library is a rich source of coding examples for user-defined C functions.)
    Currently only one calling convention is used for C functions
    (“version 1”). Support for that calling convention is
    indicated by writing a PG_FUNCTION_INFO_V1() macro
    call for the function, as illustrated below.
   
    The first time a user-defined function in a particular
    loadable object file is called in a session,
    the dynamic loader loads that object file into memory so that the
    function can be called.  The CREATE FUNCTION
    for a user-defined C function must therefore specify two pieces of
    information for the function: the name of the loadable
    object file, and the C name (link symbol) of the specific function to call
    within that object file.  If the C name is not explicitly specified then
    it is assumed to be the same as the SQL function name.
   
    The following algorithm is used to locate the shared object file
    based on the name given in the CREATE FUNCTION
    command:
    
If the name is an absolute path, the given file is loaded.
       If the name starts with the string $libdir,
       that part is replaced by the PostgreSQL package
        library directory
       name, which is determined at build time.
      
If the name does not contain a directory part, the file is searched for in the path specified by the configuration variable dynamic_library_path.
Otherwise (the file was not found in the path, or it contains a non-absolute directory part), the dynamic loader will try to take the name as given, which will most likely fail. (It is unreliable to depend on the current working directory.)
    If this sequence does not work, the platform-specific shared
    library file name extension (often .so) is
    appended to the given name and this sequence is tried again.  If
    that fails as well, the load will fail.
   
    It is recommended to locate shared libraries either relative to
    $libdir or through the dynamic library path.
    This simplifies version upgrades if the new installation is at a
    different location.  The actual directory that
    $libdir stands for can be found out with the
    command pg_config --pkglibdir.
   
The user ID the PostgreSQL server runs as must be able to traverse the path to the file you intend to load. Making the file or a higher-level directory not readable and/or not executable by the postgres user is a common mistake.
    In any case, the file name that is given in the
    CREATE FUNCTION command is recorded literally
    in the system catalogs, so if the file needs to be loaded again
    the same procedure is applied.
   
     PostgreSQL will not compile a C function
     automatically.  The object file must be compiled before it is referenced
     in a CREATE
     FUNCTION command.  See Section 37.10.5 for additional
     information.
    
    To ensure that a dynamically loaded object file is not loaded into an
    incompatible server, PostgreSQL checks that the
    file contains a “magic block” with the appropriate contents.
    This allows the server to detect obvious incompatibilities, such as code
    compiled for a different major version of
    PostgreSQL. To include a magic block,
    write this in one (and only one) of the module source files, after having
    included the header fmgr.h:
PG_MODULE_MAGIC;
After it is used for the first time, a dynamically loaded object file is retained in memory. Future calls in the same session to the function(s) in that file will only incur the small overhead of a symbol table lookup. If you need to force a reload of an object file, for example after recompiling it, begin a fresh session.
    Optionally, a dynamically loaded file can contain initialization and
    finalization functions.  If the file includes a function named
    _PG_init, that function will be called immediately after
    loading the file.  The function receives no parameters and should
    return void.  If the file includes a function named
    _PG_fini, that function will be called immediately before
    unloading the file.  Likewise, the function receives no parameters and
    should return void.  Note that _PG_fini will only be called
    during an unload of the file, not during process termination.
    (Presently, unloads are disabled and will never occur, but this may
    change in the future.)
   
To know how to write C-language functions, you need to know how PostgreSQL internally represents base data types and how they can be passed to and from functions. Internally, PostgreSQL regards a base type as a “blob of memory”. The user-defined functions that you define over a type in turn define the way that PostgreSQL can operate on it. That is, PostgreSQL will only store and retrieve the data from disk and use your user-defined functions to input, process, and output the data.
Base types can have one of three internal formats:
pass by value, fixed-length
pass by reference, fixed-length
pass by reference, variable-length
     By-value  types  can  only be 1, 2, or 4 bytes in length
     (also 8 bytes, if sizeof(Datum) is 8 on your machine).
     You should be careful to define your types such that they will be the
     same size (in bytes) on all architectures.  For example, the
     long type is dangerous because it is 4 bytes on some
     machines and 8 bytes on others, whereas int type is 4 bytes
     on most Unix machines.  A reasonable implementation of the
     int4 type on Unix machines might be:
/* 4-byte integer, passed by value */ typedef int int4;
     (The actual PostgreSQL C code calls this type int32, because
     it is a convention in C that int
     means XXXX bits.  Note
     therefore also that the C type int8 is 1 byte in size.  The
     SQL type int8 is called int64 in C.  See also
     Table 37.1.)
    
On the other hand, fixed-length types of any size can be passed by-reference. For example, here is a sample implementation of a PostgreSQL type:
/* 16-byte structure, passed by reference */
typedef struct
{
    double  x, y;
} Point;
     Only  pointers  to  such types can be used when passing
     them in and out of PostgreSQL functions.
     To return a value of such a type, allocate the right amount of
     memory with palloc, fill in the allocated memory,
     and return a pointer to it.  (Also, if you just want to return the
     same value as one of your input arguments that's of the same data type,
     you can skip the extra palloc and just return the
     pointer to the input value.)
    
     Finally, all variable-length types must also be  passed
     by  reference.   All  variable-length  types must begin
     with an opaque length field of exactly 4 bytes, which will be set
     by SET_VARSIZE; never set this field directly! All data to
     be  stored within that type must be located in the memory
     immediately  following  that  length  field.   The
     length field contains the total length of the structure,
     that is,  it  includes  the  size  of  the  length  field
     itself.
    
Another important point is to avoid leaving any uninitialized bits within data type values; for example, take care to zero out any alignment padding bytes that might be present in structs. Without this, logically-equivalent constants of your data type might be seen as unequal by the planner, leading to inefficient (though not incorrect) plans.
Never modify the contents of a pass-by-reference input value. If you do so you are likely to corrupt on-disk data, since the pointer you are given might point directly into a disk buffer. The sole exception to this rule is explained in Section 37.12.
     As an example, we can define the type text as
     follows:
typedef struct {
    int32 length;
    char data[FLEXIBLE_ARRAY_MEMBER];
} text;
     The [FLEXIBLE_ARRAY_MEMBER] notation means that the actual
     length of the data part is not specified by this declaration.
    
     When manipulating
     variable-length types, we must  be  careful  to  allocate
     the  correct amount  of memory and set the length field correctly.
     For example, if we wanted to  store  40  bytes  in  a text
     structure, we might use a code fragment like this:
#include "postgres.h" ... char buffer[40]; /* our source data */ ... text *destination = (text *) palloc(VARHDRSZ + 40); SET_VARSIZE(destination, VARHDRSZ + 40); memcpy(destination->data, buffer, 40); ...
     VARHDRSZ is the same as sizeof(int32), but
     it's considered good style to use the macro VARHDRSZ
     to refer to the size of the overhead for a variable-length type.
     Also, the length field must be set using the
     SET_VARSIZE macro, not by simple assignment.
    
     Table 37.1 shows the C types
     corresponding to many of the built-in SQL data types
     of PostgreSQL.
     The “Defined In” column gives the header file that
     needs to be included to get the type definition.  (The actual
     definition might be in a different file that is included by the
     listed file.  It is recommended that users stick to the defined
     interface.)  Note that you should always include
     postgres.h first in any source file of server
     code, because it declares a number of things that you will need
     anyway, and because including other headers first can cause
     portability issues.
    
Table 37.1. Equivalent C Types for Built-in SQL Types
| SQL Type | C Type | Defined In | 
|---|---|---|
| boolean | bool | postgres.h(maybe compiler built-in) | 
| box | BOX* | utils/geo_decls.h | 
| bytea | bytea* | postgres.h | 
| "char" | char | (compiler built-in) | 
| character | BpChar* | postgres.h | 
| cid | CommandId | postgres.h | 
| date | DateADT | utils/date.h | 
| float4(real) | float4 | postgres.h | 
| float8(double precision) | float8 | postgres.h | 
| int2(smallint) | int16 | postgres.h | 
| int4(integer) | int32 | postgres.h | 
| int8(bigint) | int64 | postgres.h | 
| interval | Interval* | datatype/timestamp.h | 
| lseg | LSEG* | utils/geo_decls.h | 
| name | Name | postgres.h | 
| numeric | Numeric | utils/numeric.h | 
| oid | Oid | postgres.h | 
| oidvector | oidvector* | postgres.h | 
| path | PATH* | utils/geo_decls.h | 
| point | POINT* | utils/geo_decls.h | 
| regproc | RegProcedure | postgres.h | 
| text | text* | postgres.h | 
| tid | ItemPointer | storage/itemptr.h | 
| time | TimeADT | utils/date.h | 
| time with time zone | TimeTzADT | utils/date.h | 
| timestamp | Timestamp | datatype/timestamp.h | 
| timestamp with time zone | TimestampTz | datatype/timestamp.h | 
| varchar | VarChar* | postgres.h | 
| xid | TransactionId | postgres.h | 
Now that we've gone over all of the possible structures for base types, we can show some examples of real functions.
The version-1 calling convention relies on macros to suppress most of the complexity of passing arguments and results. The C declaration of a version-1 function is always:
Datum funcname(PG_FUNCTION_ARGS)
In addition, the macro call:
PG_FUNCTION_INFO_V1(funcname);
     must appear in the same source file.  (Conventionally, it's
     written just before the function itself.)  This macro call is not
     needed for internal-language functions, since
     PostgreSQL assumes that all internal functions
     use the version-1 convention.  It is, however, required for
     dynamically-loaded functions.
    
     In a version-1 function, each actual argument is fetched using a
     PG_GETARG_
     macro that corresponds to the argument's data type.  (In non-strict
     functions there needs to be a previous check about argument null-ness
     using xxx()PG_ARGISNULL(); see below.)
     The result is returned using a
     PG_RETURN_
     macro for the return type.
     xxx()PG_GETARG_
     takes as its argument the number of the function argument to
     fetch, where the count starts at 0.
     xxx()PG_RETURN_
     takes as its argument the actual value to return.
    xxx()
Here are some examples using the version-1 calling convention:
#include "postgres.h"
#include <string.h>
#include "fmgr.h"
#include "utils/geo_decls.h"
PG_MODULE_MAGIC;
/* by value */
PG_FUNCTION_INFO_V1(add_one);
Datum
add_one(PG_FUNCTION_ARGS)
{
    int32   arg = PG_GETARG_INT32(0);
    PG_RETURN_INT32(arg + 1);
}
/* by reference, fixed length */
PG_FUNCTION_INFO_V1(add_one_float8);
Datum
add_one_float8(PG_FUNCTION_ARGS)
{
    /* The macros for FLOAT8 hide its pass-by-reference nature. */
    float8   arg = PG_GETARG_FLOAT8(0);
    PG_RETURN_FLOAT8(arg + 1.0);
}
PG_FUNCTION_INFO_V1(makepoint);
Datum
makepoint(PG_FUNCTION_ARGS)
{
    /* Here, the pass-by-reference nature of Point is not hidden. */
    Point     *pointx = PG_GETARG_POINT_P(0);
    Point     *pointy = PG_GETARG_POINT_P(1);
    Point     *new_point = (Point *) palloc(sizeof(Point));
    new_point->x = pointx->x;
    new_point->y = pointy->y;
    PG_RETURN_POINT_P(new_point);
}
/* by reference, variable length */
PG_FUNCTION_INFO_V1(copytext);
Datum
copytext(PG_FUNCTION_ARGS)
{
    text     *t = PG_GETARG_TEXT_PP(0);
    /*
     * VARSIZE_ANY_EXHDR is the size of the struct in bytes, minus the
     * VARHDRSZ or VARHDRSZ_SHORT of its header.  Construct the copy with a
     * full-length header.
     */
    text     *new_t = (text *) palloc(VARSIZE_ANY_EXHDR(t) + VARHDRSZ);
    SET_VARSIZE(new_t, VARSIZE_ANY_EXHDR(t) + VARHDRSZ);
    /*
     * VARDATA is a pointer to the data region of the new struct.  The source
     * could be a short datum, so retrieve its data through VARDATA_ANY.
     */
    memcpy((void *) VARDATA(new_t), /* destination */
           (void *) VARDATA_ANY(t), /* source */
           VARSIZE_ANY_EXHDR(t));   /* how many bytes */
    PG_RETURN_TEXT_P(new_t);
}
PG_FUNCTION_INFO_V1(concat_text);
Datum
concat_text(PG_FUNCTION_ARGS)
{
    text  *arg1 = PG_GETARG_TEXT_PP(0);
    text  *arg2 = PG_GETARG_TEXT_PP(1);
    int32 arg1_size = VARSIZE_ANY_EXHDR(arg1);
    int32 arg2_size = VARSIZE_ANY_EXHDR(arg2);
    int32 new_text_size = arg1_size + arg2_size + VARHDRSZ;
    text *new_text = (text *) palloc(new_text_size);
    SET_VARSIZE(new_text, new_text_size);
    memcpy(VARDATA(new_text), VARDATA_ANY(arg1), arg1_size);
    memcpy(VARDATA(new_text) + arg1_size, VARDATA_ANY(arg2), arg2_size);
    PG_RETURN_TEXT_P(new_text);
}
     Supposing that the above code has been prepared in file
     funcs.c and compiled into a shared object,
     we could define the functions to PostgreSQL
     with commands like this:
    
CREATE FUNCTION add_one(integer) RETURNS integer
     AS 'DIRECTORY/funcs', 'add_one'
     LANGUAGE C STRICT;
-- note overloading of SQL function name "add_one"
CREATE FUNCTION add_one(double precision) RETURNS double precision
     AS 'DIRECTORY/funcs', 'add_one_float8'
     LANGUAGE C STRICT;
CREATE FUNCTION makepoint(point, point) RETURNS point
     AS 'DIRECTORY/funcs', 'makepoint'
     LANGUAGE C STRICT;
CREATE FUNCTION copytext(text) RETURNS text
     AS 'DIRECTORY/funcs', 'copytext'
     LANGUAGE C STRICT;
CREATE FUNCTION concat_text(text, text) RETURNS text
     AS 'DIRECTORY/funcs', 'concat_text'
     LANGUAGE C STRICT;
     Here, DIRECTORY stands for the
     directory of the shared library file (for instance the
     PostgreSQL tutorial directory, which
     contains the code for the examples used in this section).
     (Better style would be to use just 'funcs' in the
     AS clause, after having added
     DIRECTORY to the search path.  In any
     case, we can omit the system-specific extension for a shared
     library, commonly .so.)
    
     Notice that we have specified the functions as “strict”,
     meaning that
     the system should automatically assume a null result if any input
     value is null.  By doing this, we avoid having to check for null inputs
     in the function code.  Without this, we'd have to check for null values
     explicitly, using PG_ARGISNULL().
    
     The macro PG_ARGISNULL(
     allows a function to test whether each input is null.  (Of course, doing
     this is only necessary in functions not declared “strict”.)
     As with the
     n)PG_GETARG_ macros,
     the input arguments are counted beginning at zero.  Note that one
     should refrain from executing
     xxx()PG_GETARG_ until
     one has verified that the argument isn't null.
     To return a null result, execute xxx()PG_RETURN_NULL();
     this works in both strict and nonstrict functions.
    
     At first glance, the version-1 coding conventions might appear
     to be just pointless obscurantism, compared to using
     plain C calling conventions.  They do however allow
     us to deal with NULLable arguments/return values,
     and “toasted” (compressed or out-of-line) values.
    
     Other options provided by the version-1 interface are two
     variants of the
     PG_GETARG_
     macros. The first of these,
     xxx()PG_GETARG_,
     guarantees to return a copy of the specified argument that is
     safe for writing into. (The normal macros will sometimes return a
     pointer to a value that is physically stored in a table, which
     must not be written to. Using the
     xxx_COPY()PG_GETARG_
     macros guarantees a writable result.)
    The second variant consists of the
    xxx_COPY()PG_GETARG_
    macros which take three arguments. The first is the number of the
    function argument (as above). The second and third are the offset and
    length of the segment to be returned. Offsets are counted from
    zero, and a negative length requests that the remainder of the
    value be returned. These macros provide more efficient access to
    parts of large values in the case where they have storage type
    “external”. (The storage type of a column can be specified using
    xxx_SLICE()ALTER TABLE . tablename ALTER
    COLUMN colname SET STORAGE
    storagetypestoragetype is one of
    plain, external, extended,
     or main.)
    
     Finally, the version-1 function call conventions make it possible
     to return set results (Section 37.10.8) and
     implement trigger functions (Chapter 38) and
     procedural-language call handlers (Chapter 55).  For more details
     see src/backend/utils/fmgr/README in the
     source distribution.
    
Before we turn to the more advanced topics, we should discuss some coding rules for PostgreSQL C-language functions. While it might be possible to load functions written in languages other than C into PostgreSQL, this is usually difficult (when it is possible at all) because other languages, such as C++, FORTRAN, or Pascal often do not follow the same calling convention as C. That is, other languages do not pass argument and return values between functions in the same way. For this reason, we will assume that your C-language functions are actually written in C.
The basic rules for writing and building C functions are as follows:
        Use pg_config
        --includedir-server
        to find out where the PostgreSQL server header
        files are installed on your system (or the system that your
        users will be running on).
       
Compiling and linking your code so that it can be dynamically loaded into PostgreSQL always requires special flags. See Section 37.10.5 for a detailed explanation of how to do it for your particular operating system.
Remember to define a “magic block” for your shared library, as described in Section 37.10.1.
        When allocating memory, use the
        PostgreSQL functions
        palloc and pfree
        instead of the corresponding C library functions
        malloc and free.
        The memory allocated by palloc will be
        freed automatically at the end of each transaction, preventing
        memory leaks.
       
        Always zero the bytes of your structures using memset
        (or allocate them with palloc0 in the first place).
        Even if you assign to each field of your structure, there might be
        alignment padding (holes in the structure) that contain
        garbage values.  Without this, it's difficult to
        support hash indexes or hash joins, as you must pick out only
        the significant bits of your data structure to compute a hash.
        The planner also sometimes relies on comparing constants via
        bitwise equality, so you can get undesirable planning results if
        logically-equivalent values aren't bitwise equal.
       
        Most of the internal PostgreSQL
        types are declared in postgres.h, while
        the function manager interfaces
        (PG_FUNCTION_ARGS, etc.)  are in
        fmgr.h, so you will need to include at
        least these two files.  For portability reasons it's best to
        include postgres.h first,
        before any other system or user header files.  Including
        postgres.h will also include
        elog.h and palloc.h
        for you.
       
Symbol names defined within object files must not conflict with each other or with symbols defined in the PostgreSQL server executable. You will have to rename your functions or variables if you get error messages to this effect.
Before you are able to use your PostgreSQL extension functions written in C, they must be compiled and linked in a special way to produce a file that can be dynamically loaded by the server. To be precise, a shared library needs to be created.
  For information beyond what is contained in this section
  you should read the documentation of your
  operating system, in particular the manual pages for the C compiler,
  cc, and the link editor, ld.
  In addition, the PostgreSQL source code
  contains several working examples in the
  contrib directory.  If you rely on these
  examples you will make your modules dependent on the availability
  of the PostgreSQL source code, however.
 
Creating shared libraries is generally analogous to linking executables: first the source files are compiled into object files, then the object files are linked together. The object files need to be created as position-independent code (PIC), which conceptually means that they can be placed at an arbitrary location in memory when they are loaded by the executable. (Object files intended for executables are usually not compiled that way.) The command to link a shared library contains special flags to distinguish it from linking an executable (at least in theory — on some systems the practice is much uglier).
  In the following examples we assume that your source code is in a
  file foo.c and we will create a shared library
  foo.so.  The intermediate object file will be
  called foo.o unless otherwise noted.  A shared
  library can contain more than one object file, but we only use one
  here.
 
      The compiler flag to create PIC is
      -fPIC.  To create shared libraries the compiler
      flag is -shared.
gcc -fPIC -c foo.c gcc -shared -o foo.so foo.o
This is applicable as of version 3.0 of FreeBSD.
      The compiler flag of the system compiler to create
      PIC is +z.  When using
      GCC it's -fPIC. The
      linker flag for shared libraries is -b.  So:
cc +z -c foo.c
or:
gcc -fPIC -c foo.c
and then:
ld -b -o foo.sl foo.o
      HP-UX uses the extension
      .sl for shared libraries, unlike most other
      systems.
     
      The compiler flag to create PIC is
      -fPIC.
      The compiler flag to create a shared library is
      -shared.  A complete example looks like this:
cc -fPIC -c foo.c cc -shared -o foo.so foo.o
Here is an example. It assumes the developer tools are installed.
cc -c foo.c cc -bundle -flat_namespace -undefined suppress -o foo.so foo.o
      The compiler flag to create PIC is
      -fPIC.  For ELF systems, the
      compiler with the flag -shared is used to link
      shared libraries.  On the older non-ELF systems, ld
      -Bshareable is used.
gcc -fPIC -c foo.c gcc -shared -o foo.so foo.o
      The compiler flag to create PIC is
      -fPIC.  ld -Bshareable is
      used to link shared libraries.
gcc -fPIC -c foo.c ld -Bshareable -o foo.so foo.o
      The compiler flag to create PIC is
      -KPIC with the Sun compiler and
      -fPIC with GCC.  To
      link shared libraries, the compiler option is
      -G with either compiler or alternatively
      -shared with GCC.
cc -KPIC -c foo.c cc -G -o foo.so foo.o
or
gcc -fPIC -c foo.c gcc -G -o foo.so foo.o
If this is too complicated for you, you should consider using GNU Libtool, which hides the platform differences behind a uniform interface.
  The resulting shared library file can then be loaded into
  PostgreSQL.  When specifying the file name
  to the CREATE FUNCTION command, one must give it
  the name of the shared library file, not the intermediate object file.
  Note that the system's standard shared-library extension (usually
  .so or .sl) can be omitted from
  the CREATE FUNCTION command, and normally should
  be omitted for best portability.
 
Refer back to Section 37.10.1 about where the server expects to find the shared library files.
Composite types do not have a fixed layout like C structures. Instances of a composite type can contain null fields. In addition, composite types that are part of an inheritance hierarchy can have different fields than other members of the same inheritance hierarchy. Therefore, PostgreSQL provides a function interface for accessing fields of composite types from C.
Suppose we want to write a function to answer the query:
SELECT name, c_overpaid(emp, 1500) AS overpaid
    FROM emp
    WHERE name = 'Bill' OR name = 'Sam';
     Using the version-1 calling conventions, we can define
     c_overpaid as:
#include "postgres.h"
#include "executor/executor.h"  /* for GetAttributeByName() */
PG_MODULE_MAGIC;
PG_FUNCTION_INFO_V1(c_overpaid);
Datum
c_overpaid(PG_FUNCTION_ARGS)
{
    HeapTupleHeader  t = PG_GETARG_HEAPTUPLEHEADER(0);
    int32            limit = PG_GETARG_INT32(1);
    bool isnull;
    Datum salary;
    salary = GetAttributeByName(t, "salary", &isnull);
    if (isnull)
        PG_RETURN_BOOL(false);
    /* Alternatively, we might prefer to do PG_RETURN_NULL() for null salary. */
    PG_RETURN_BOOL(DatumGetInt32(salary) > limit);
}
     GetAttributeByName is the
     PostgreSQL system function that
     returns attributes out of the specified row.  It has
     three arguments: the argument of type HeapTupleHeader passed
     into
     the  function, the name of the desired attribute, and a
     return parameter that tells whether  the  attribute
     is  null.   GetAttributeByName returns a Datum
     value that you can convert to the proper data type by using the
     appropriate DatumGet
     macro.  Note that the return value is meaningless if the null flag is
     set; always check the null flag before trying to do anything with the
     result.
    XXX()
     There is also GetAttributeByNum, which selects
     the target attribute by column number instead of name.
    
     The following command declares the function
     c_overpaid in SQL:
CREATE FUNCTION c_overpaid(emp, integer) RETURNS boolean
    AS 'DIRECTORY/funcs', 'c_overpaid'
    LANGUAGE C STRICT;
     Notice we have used STRICT so that we did not have to
     check whether the input arguments were NULL.
    
To return a row or composite-type value from a C-language function, you can use a special API that provides macros and functions to hide most of the complexity of building composite data types. To use this API, the source file must include:
#include "funcapi.h"
     There are two ways you can build a composite data value (henceforth
     a “tuple”): you can build it from an array of Datum values,
     or from an array of C strings that can be passed to the input
     conversion functions of the tuple's column data types.  In either
     case, you first need to obtain or construct a TupleDesc
     descriptor for the tuple structure.  When working with Datums, you
     pass the TupleDesc to BlessTupleDesc,
     and then call heap_form_tuple for each row.  When working
     with C strings, you pass the TupleDesc to
     TupleDescGetAttInMetadata, and then call
     BuildTupleFromCStrings for each row.  In the case of a
     function returning a set of tuples, the setup steps can all be done
     once during the first call of the function.
    
     Several helper functions are available for setting up the needed
     TupleDesc.  The recommended way to do this in most
     functions returning composite values is to call:
TypeFuncClass get_call_result_type(FunctionCallInfo fcinfo,
                                   Oid *resultTypeId,
                                   TupleDesc *resultTupleDesc)
     passing the same fcinfo struct passed to the calling function
     itself.  (This of course requires that you use the version-1
     calling conventions.)  resultTypeId can be specified
     as NULL or as the address of a local variable to receive the
     function's result type OID.  resultTupleDesc should be the
     address of a local TupleDesc variable.  Check that the
     result is TYPEFUNC_COMPOSITE; if so,
     resultTupleDesc has been filled with the needed
     TupleDesc.  (If it is not, you can report an error along
     the lines of “function returning record called in context that
     cannot accept type record”.)
    
      get_call_result_type can resolve the actual type of a
      polymorphic function result; so it is useful in functions that return
      scalar polymorphic results, not only functions that return composites.
      The resultTypeId output is primarily useful for functions
      returning polymorphic scalars.
     
      get_call_result_type has a sibling
      get_expr_result_type, which can be used to resolve the
      expected output type for a function call represented by an expression
      tree.  This can be used when trying to determine the result type from
      outside the function itself.  There is also
      get_func_result_type, which can be used when only the
      function's OID is available.  However these functions are not able
      to deal with functions declared to return record, and
      get_func_result_type cannot resolve polymorphic types,
      so you should preferentially use get_call_result_type.
     
     Older, now-deprecated functions for obtaining
     TupleDescs are:
TupleDesc RelationNameGetTupleDesc(const char *relname)
     to get a TupleDesc for the row type of a named relation,
     and:
TupleDesc TypeGetTupleDesc(Oid typeoid, List *colaliases)
     to get a TupleDesc based on a type OID. This can
     be used to get a TupleDesc for a base or
     composite type.  It will not work for a function that returns
     record, however, and it cannot resolve polymorphic
     types.
    
     Once you have a TupleDesc, call:
TupleDesc BlessTupleDesc(TupleDesc tupdesc)
if you plan to work with Datums, or:
AttInMetadata *TupleDescGetAttInMetadata(TupleDesc tupdesc)
     if you plan to work with C strings.  If you are writing a function
     returning set, you can save the results of these functions in the
     FuncCallContext structure — use the
     tuple_desc or attinmeta field
     respectively.
    
When working with Datums, use:
HeapTuple heap_form_tuple(TupleDesc tupdesc, Datum *values, bool *isnull)
     to build a HeapTuple given user data in Datum form.
    
When working with C strings, use:
HeapTuple BuildTupleFromCStrings(AttInMetadata *attinmeta, char **values)
     to build a HeapTuple given user data
     in C string form.  values is an array of C strings,
     one for each attribute of the return row. Each C string should be in
     the form expected by the input function of the attribute data
     type. In order to return a null value for one of the attributes,
     the corresponding pointer in the values array
     should be set to NULL.  This function will need to
     be called again for each row you return.
    
     Once you have built a tuple to return from your function, it
     must be converted into a Datum. Use:
HeapTupleGetDatum(HeapTuple tuple)
     to convert a HeapTuple into a valid Datum.  This
     Datum can be returned directly if you intend to return
     just a single row, or it can be used as the current return value
     in a set-returning function.
    
An example appears in the next section.
C-language functions have two options for returning sets (multiple rows). In one method, called ValuePerCall mode, a set-returning function is called repeatedly (passing the same arguments each time) and it returns one new row on each call, until it has no more rows to return and signals that by returning NULL. The set-returning function (SRF) must therefore save enough state across calls to remember what it was doing and return the correct next item on each call. In the other method, called Materialize mode, a SRF fills and returns a tuplestore object containing its entire result; then only one call occurs for the whole result, and no inter-call state is needed.
     When using ValuePerCall mode, it is important to remember that the
     query is not guaranteed to be run to completion; that is, due to
     options such as LIMIT, the executor might stop
     making calls to the set-returning function before all rows have been
     fetched.  This means it is not safe to perform cleanup activities in
     the last call, because that might not ever happen.  It's recommended
     to use Materialize mode for functions that need access to external
     resources, such as file descriptors.
    
     The remainder of this section documents a set of helper macros that
     are commonly used (though not required to be used) for SRFs using
     ValuePerCall mode.  Additional details about Materialize mode can be
     found in src/backend/utils/fmgr/README.  Also,
     the contrib modules in
     the PostgreSQL source distribution contain
     many examples of SRFs using both ValuePerCall and Materialize mode.
    
     To use the ValuePerCall support macros described here,
     include funcapi.h.  These macros work with a
     structure FuncCallContext that contains the
     state that needs to be saved across calls.  Within the calling
     SRF, fcinfo->flinfo->fn_extra is used to
     hold a pointer to FuncCallContext across
     calls.  The macros automatically fill that field on first use,
     and expect to find the same pointer there on subsequent uses.
typedef struct FuncCallContext
{
    /*
     * Number of times we've been called before
     *
     * call_cntr is initialized to 0 for you by SRF_FIRSTCALL_INIT(), and
     * incremented for you every time SRF_RETURN_NEXT() is called.
     */
    uint64 call_cntr;
    /*
     * OPTIONAL maximum number of calls
     *
     * max_calls is here for convenience only and setting it is optional.
     * If not set, you must provide alternative means to know when the
     * function is done.
     */
    uint64 max_calls;
    /*
     * OPTIONAL pointer to miscellaneous user-provided context information
     *
     * user_fctx is for use as a pointer to your own data to retain
     * arbitrary context information between calls of your function.
     */
    void *user_fctx;
    /*
     * OPTIONAL pointer to struct containing attribute type input metadata
     *
     * attinmeta is for use when returning tuples (i.e., composite data types)
     * and is not used when returning base data types. It is only needed
     * if you intend to use BuildTupleFromCStrings() to create the return
     * tuple.
     */
    AttInMetadata *attinmeta;
    /*
     * memory context used for structures that must live for multiple calls
     *
     * multi_call_memory_ctx is set by SRF_FIRSTCALL_INIT() for you, and used
     * by SRF_RETURN_DONE() for cleanup. It is the most appropriate memory
     * context for any memory that is to be reused across multiple calls
     * of the SRF.
     */
    MemoryContext multi_call_memory_ctx;
    /*
     * OPTIONAL pointer to struct containing tuple description
     *
     * tuple_desc is for use when returning tuples (i.e., composite data types)
     * and is only needed if you are going to build the tuples with
     * heap_form_tuple() rather than with BuildTupleFromCStrings().  Note that
     * the TupleDesc pointer stored here should usually have been run through
     * BlessTupleDesc() first.
     */
    TupleDesc tuple_desc;
} FuncCallContext;
The macros to be used by an SRF using this infrastructure are:
SRF_IS_FIRSTCALL()
Use this to determine if your function is being called for the first or a subsequent time. On the first call (only), call:
SRF_FIRSTCALL_INIT()
     to initialize the FuncCallContext. On every function call,
     including the first, call:
SRF_PERCALL_SETUP()
     to set up for using the FuncCallContext.
    
If your function has data to return in the current call, use:
SRF_RETURN_NEXT(funcctx, result)
     to return it to the caller.  (result must be of type
     Datum, either a single value or a tuple prepared as
     described above.)  Finally, when your function is finished
     returning data, use:
SRF_RETURN_DONE(funcctx)
to clean up and end the SRF.
     The memory context that is current when the SRF is called is
     a transient context that will be cleared between calls.  This means
     that you do not need to call pfree on everything
     you allocated using palloc; it will go away anyway.  However, if you want to allocate
     any data structures to live across calls, you need to put them somewhere
     else.  The memory context referenced by
     multi_call_memory_ctx is a suitable location for any
     data that needs to survive until the SRF is finished running.  In most
     cases, this means that you should switch into
     multi_call_memory_ctx while doing the
     first-call setup.
     Use funcctx->user_fctx to hold a pointer to
     any such cross-call data structures.
     (Data you allocate
     in multi_call_memory_ctx will go away
     automatically when the query ends, so it is not necessary to free
     that data manually, either.)
    
      While the actual arguments to the function remain unchanged between
      calls, if you detoast the argument values (which is normally done
      transparently by the
      PG_GETARG_ macro)
      in the transient context then the detoasted copies will be freed on
      each cycle. Accordingly, if you keep references to such values in
      your xxxuser_fctx, you must either copy them into the
      multi_call_memory_ctx after detoasting, or ensure
      that you detoast the values only in that context.
     
A complete pseudo-code example looks like the following:
Datum
my_set_returning_function(PG_FUNCTION_ARGS)
{
    FuncCallContext  *funcctx;
    Datum             result;
    further declarations as needed
    if (SRF_IS_FIRSTCALL())
    {
        MemoryContext oldcontext;
        funcctx = SRF_FIRSTCALL_INIT();
        oldcontext = MemoryContextSwitchTo(funcctx->multi_call_memory_ctx);
        /* One-time setup code appears here: */
        user code
        if returning composite
            build TupleDesc, and perhaps AttInMetadata
        endif returning composite
        user code
        MemoryContextSwitchTo(oldcontext);
    }
    /* Each-time setup code appears here: */
    user code
    funcctx = SRF_PERCALL_SETUP();
    user code
    /* this is just one way we might test whether we are done: */
    if (funcctx->call_cntr < funcctx->max_calls)
    {
        /* Here we want to return another item: */
        user code
        obtain result Datum
        SRF_RETURN_NEXT(funcctx, result);
    }
    else
    {
        /* Here we are done returning items, so just report that fact. */
        /* (Resist the temptation to put cleanup code here.) */
        SRF_RETURN_DONE(funcctx);
    }
}
A complete example of a simple SRF returning a composite type looks like:
PG_FUNCTION_INFO_V1(retcomposite);
Datum
retcomposite(PG_FUNCTION_ARGS)
{
    FuncCallContext     *funcctx;
    int                  call_cntr;
    int                  max_calls;
    TupleDesc            tupdesc;
    AttInMetadata       *attinmeta;
    /* stuff done only on the first call of the function */
    if (SRF_IS_FIRSTCALL())
    {
        MemoryContext   oldcontext;
        /* create a function context for cross-call persistence */
        funcctx = SRF_FIRSTCALL_INIT();
        /* switch to memory context appropriate for multiple function calls */
        oldcontext = MemoryContextSwitchTo(funcctx->multi_call_memory_ctx);
        /* total number of tuples to be returned */
        funcctx->max_calls = PG_GETARG_UINT32(0);
        /* Build a tuple descriptor for our result type */
        if (get_call_result_type(fcinfo, NULL, &tupdesc) != TYPEFUNC_COMPOSITE)
            ereport(ERROR,
                    (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
                     errmsg("function returning record called in context "
                            "that cannot accept type record")));
        /*
         * generate attribute metadata needed later to produce tuples from raw
         * C strings
         */
        attinmeta = TupleDescGetAttInMetadata(tupdesc);
        funcctx->attinmeta = attinmeta;
        MemoryContextSwitchTo(oldcontext);
    }
    /* stuff done on every call of the function */
    funcctx = SRF_PERCALL_SETUP();
    call_cntr = funcctx->call_cntr;
    max_calls = funcctx->max_calls;
    attinmeta = funcctx->attinmeta;
    if (call_cntr < max_calls)    /* do when there is more left to send */
    {
        char       **values;
        HeapTuple    tuple;
        Datum        result;
        /*
         * Prepare a values array for building the returned tuple.
         * This should be an array of C strings which will
         * be processed later by the type input functions.
         */
        values = (char **) palloc(3 * sizeof(char *));
        values[0] = (char *) palloc(16 * sizeof(char));
        values[1] = (char *) palloc(16 * sizeof(char));
        values[2] = (char *) palloc(16 * sizeof(char));
        snprintf(values[0], 16, "%d", 1 * PG_GETARG_INT32(1));
        snprintf(values[1], 16, "%d", 2 * PG_GETARG_INT32(1));
        snprintf(values[2], 16, "%d", 3 * PG_GETARG_INT32(1));
        /* build a tuple */
        tuple = BuildTupleFromCStrings(attinmeta, values);
        /* make the tuple into a datum */
        result = HeapTupleGetDatum(tuple);
        /* clean up (this is not really necessary) */
        pfree(values[0]);
        pfree(values[1]);
        pfree(values[2]);
        pfree(values);
        SRF_RETURN_NEXT(funcctx, result);
    }
    else    /* do when there is no more left */
    {
        SRF_RETURN_DONE(funcctx);
    }
}
One way to declare this function in SQL is:
CREATE TYPE __retcomposite AS (f1 integer, f2 integer, f3 integer);
CREATE OR REPLACE FUNCTION retcomposite(integer, integer)
    RETURNS SETOF __retcomposite
    AS 'filename', 'retcomposite'
    LANGUAGE C IMMUTABLE STRICT;
A different way is to use OUT parameters:
CREATE OR REPLACE FUNCTION retcomposite(IN integer, IN integer,
    OUT f1 integer, OUT f2 integer, OUT f3 integer)
    RETURNS SETOF record
    AS 'filename', 'retcomposite'
    LANGUAGE C IMMUTABLE STRICT;
     Notice that in this method the output type of the function is formally
     an anonymous record type.
    
     C-language functions can be declared to accept and
     return the polymorphic types
     anyelement, anyarray, anynonarray,
     anyenum, and anyrange.
     See Section 37.2.5 for a more detailed explanation
     of polymorphic functions. When function arguments or return types
     are defined as polymorphic types, the function author cannot know
     in advance what data type it will be called with, or
     need to return. There are two routines provided in fmgr.h
     to allow a version-1 C function to discover the actual data types
     of its arguments and the type it is expected to return. The routines are
     called get_fn_expr_rettype(FmgrInfo *flinfo) and
     get_fn_expr_argtype(FmgrInfo *flinfo, int argnum).
     They return the result or argument type OID, or InvalidOid if the
     information is not available.
     The structure flinfo is normally accessed as
     fcinfo->flinfo. The parameter argnum
     is zero based.  get_call_result_type can also be used
     as an alternative to get_fn_expr_rettype.
     There is also get_fn_expr_variadic, which can be used to
     find out whether variadic arguments have been merged into an array.
     This is primarily useful for VARIADIC "any" functions,
     since such merging will always have occurred for variadic functions
     taking ordinary array types.
    
For example, suppose we want to write a function to accept a single element of any type, and return a one-dimensional array of that type:
PG_FUNCTION_INFO_V1(make_array);
Datum
make_array(PG_FUNCTION_ARGS)
{
    ArrayType  *result;
    Oid         element_type = get_fn_expr_argtype(fcinfo->flinfo, 0);
    Datum       element;
    bool        isnull;
    int16       typlen;
    bool        typbyval;
    char        typalign;
    int         ndims;
    int         dims[MAXDIM];
    int         lbs[MAXDIM];
    if (!OidIsValid(element_type))
        elog(ERROR, "could not determine data type of input");
    /* get the provided element, being careful in case it's NULL */
    isnull = PG_ARGISNULL(0);
    if (isnull)
        element = (Datum) 0;
    else
        element = PG_GETARG_DATUM(0);
    /* we have one dimension */
    ndims = 1;
    /* and one element */
    dims[0] = 1;
    /* and lower bound is 1 */
    lbs[0] = 1;
    /* get required info about the element type */
    get_typlenbyvalalign(element_type, &typlen, &typbyval, &typalign);
    /* now build the array */
    result = construct_md_array(&element, &isnull, ndims, dims, lbs,
                                element_type, typlen, typbyval, typalign);
    PG_RETURN_ARRAYTYPE_P(result);
}
     The following command declares the function
     make_array in SQL:
CREATE FUNCTION make_array(anyelement) RETURNS anyarray
    AS 'DIRECTORY/funcs', 'make_array'
    LANGUAGE C IMMUTABLE;
     There is a variant of polymorphism that is only available to C-language
     functions: they can be declared to take parameters of type
     "any".  (Note that this type name must be double-quoted,
     since it's also a SQL reserved word.)  This works like
     anyelement except that it does not constrain different
     "any" arguments to be the same type, nor do they help
     determine the function's result type.  A C-language function can also
     declare its final parameter to be VARIADIC "any".  This will
     match one or more actual arguments of any type (not necessarily the same
     type).  These arguments will not be gathered into an array
     as happens with normal variadic functions; they will just be passed to
     the function separately.  The PG_NARGS() macro and the
     methods described above must be used to determine the number of actual
     arguments and their types when using this feature.  Also, users of such
     a function might wish to use the VARIADIC keyword in their
     function call, with the expectation that the function would treat the
     array elements as separate arguments.  The function itself must implement
     that behavior if wanted, after using get_fn_expr_variadic to
     detect that the actual argument was marked with VARIADIC.
    
Add-ins can reserve LWLocks and an allocation of shared memory on server startup. The add-in's shared library must be preloaded by specifying it in shared_preload_libraries. Shared memory is reserved by calling:
void RequestAddinShmemSpace(int size)
     from your _PG_init function.
    
LWLocks are reserved by calling:
void RequestNamedLWLockTranche(const char *tranche_name, int num_lwlocks)
     from _PG_init.  This will ensure that an array of
     num_lwlocks LWLocks is available under the name
     tranche_name.  Use GetNamedLWLockTranche
     to get a pointer to this array.
    
     To avoid possible race-conditions, each backend should use the LWLock
     AddinShmemInitLock when connecting to and initializing
     its allocation of shared memory, as shown here:
static mystruct *ptr = NULL;
if (!ptr)
{
        bool    found;
        LWLockAcquire(AddinShmemInitLock, LW_EXCLUSIVE);
        ptr = ShmemInitStruct("my struct name", size, &found);
        if (!found)
        {
                initialize contents of shmem area;
                acquire any requested LWLocks using:
                ptr->locks = GetNamedLWLockTranche("my tranche name");
        }
        LWLockRelease(AddinShmemInitLock);
}
Although the PostgreSQL backend is written in C, it is possible to write extensions in C++ if these guidelines are followed:
         All functions accessed by the backend must present a C interface
         to the backend;  these C functions can then call C++ functions.
         For example, extern C linkage is required for
         backend-accessed functions.  This is also necessary for any
         functions that are passed as pointers between the backend and
         C++ code.
       
        Free memory using the appropriate deallocation method.  For example,
        most backend memory is allocated using palloc(), so use
        pfree() to free it.  Using C++
        delete in such cases will fail.
       
        Prevent exceptions from propagating into the C code (use a catch-all
        block at the top level of all extern C functions).  This
        is necessary even if the C++ code does not explicitly throw any
        exceptions, because events like out-of-memory can still throw
        exceptions.  Any exceptions must be caught and appropriate errors
        passed back to the C interface.  If possible, compile C++ with
        -fno-exceptions to eliminate exceptions entirely; in such
        cases, you must check for failures in your C++ code, e.g.,  check for
        NULL returned by new().
       
        If calling backend functions from C++ code, be sure that the
        C++ call stack contains only plain old data structures
        (POD).  This is necessary because backend errors
        generate a distant longjmp() that does not properly
        unroll a C++ call stack with non-POD objects.
       
     In summary, it is best to place C++ code behind a wall of
     extern C functions that interface to the backend,
     and avoid exception, memory, and call stack leakage.