$PostgreSQL: pgsql/src/backend/utils/fmgr/README,v 1.16 2008/10/31 19:37:56 tgl Exp $

Function Manager
================

Proposal For Function-Manager Redesign			19-Nov-2000
--------------------------------------

We know that the existing mechanism for calling Postgres functions needs
to be redesigned.  It has portability problems because it makes
assumptions about parameter passing that violate ANSI C; it fails to
handle NULL arguments and results cleanly; and "function handlers" that
support a class of functions (such as fmgr_pl) can only be done via a
really ugly, non-reentrant kluge.  (Global variable set during every
function call, forsooth.)  Here is a proposal for fixing these problems.

In the past, the major objections to redoing the function-manager
interface have been (a) it'll be quite tedious to implement, since every
built-in function and everyplace that calls such functions will need to
be touched; (b) such wide-ranging changes will be difficult to make in
parallel with other development work; (c) it will break existing
user-written loadable modules that define "C language" functions.  While
I have no solution to the "tedium" aspect, I believe I see an answer to
the other problems: by use of function handlers, we can support both old
and new interfaces in parallel for both callers and callees, at some
small efficiency cost for the old styles.  That way, most of the changes
can be done on an incremental file-by-file basis --- we won't need a
"big bang" where everything changes at once.  Support for callees
written in the old style can be left in place indefinitely, to provide
backward compatibility for user-written C functions.


Changes In pg_proc (System Data About a Function)
-------------------------------------------------

A new column "proisstrict" will be added to the system pg_proc table.
This is a boolean value which will be TRUE if the function is "strict",
that is it always returns NULL when any of its inputs are NULL.  The
function manager will check this field and skip calling the function when
it's TRUE and there are NULL inputs.  This allows us to remove explicit
NULL-value tests from many functions that currently need them (not to
mention fixing many more that need them but don't have them).  A function
that is not marked "strict" is responsible for checking whether its inputs
are NULL or not.  Most builtin functions will be marked "strict".

An optional WITH parameter will be added to CREATE FUNCTION to allow
specification of whether user-defined functions are strict or not.  I am
inclined to make the default be "not strict", since that seems to be the
more useful case for functions expressed in SQL or a PL language, but
am open to arguments for the other choice.


The New Function-Manager Interface
----------------------------------

The core of the new design is revised data structures for representing
the result of a function lookup and for representing the parameters
passed to a specific function invocation.  (We want to keep function
lookup separate from function call, since many parts of the system apply
the same function over and over; the lookup overhead should be paid once
per query, not once per tuple.)


When a function is looked up in pg_proc, the result is represented as

typedef struct
{
    PGFunction  fn_addr;    /* pointer to function or handler to be called */
    Oid         fn_oid;     /* OID of function (NOT of handler, if any) */
    short       fn_nargs;   /* 0..FUNC_MAX_ARGS, or -1 if variable arg count */
    bool        fn_strict;  /* function is "strict" (NULL in => NULL out) */
    bool        fn_retset;  /* function returns a set (over multiple calls) */
    unsigned char fn_stats; /* collect stats if track_functions > this */
    void       *fn_extra;   /* extra space for use by handler */
    MemoryContext fn_mcxt;  /* memory context to store fn_extra in */
    Node       *fn_expr;    /* expression parse tree for call, or NULL */
} FmgrInfo;

For an ordinary built-in function, fn_addr is just the address of the C
routine that implements the function.  Otherwise it is the address of a
handler for the class of functions that includes the target function.
The handler can use the function OID and perhaps also the fn_extra slot
to find the specific code to execute.  (fn_oid = InvalidOid can be used
to denote a not-yet-initialized FmgrInfo struct.  fn_extra will always
be NULL when an FmgrInfo is first filled by the function lookup code, but
a function handler could set it to avoid making repeated lookups of its
own when the same FmgrInfo is used repeatedly during a query.)  fn_nargs
is the number of arguments expected by the function, fn_strict is its
strictness flag, and fn_retset shows whether it returns a set; all of
these values come from the function's pg_proc entry.  fn_stats is also
set up to control whether or not to track runtime statistics for calling
this function.  If the function is being called as part of a SQL expression,
fn_expr will point to the expression parse tree for the function call; this
can be used to extract parse-time knowledge about the actual arguments.

FmgrInfo already exists in the current code, but has fewer fields.  This
change should be transparent at the source-code level.


During a call of a function, the following data structure is created
and passed to the function:

typedef struct
{
    FmgrInfo   *flinfo;         /* ptr to lookup info used for this call */
    Node       *context;        /* pass info about context of call */
    Node       *resultinfo;     /* pass or return extra info about result */
    bool        isnull;         /* function must set true if result is NULL */
    short       nargs;          /* # arguments actually passed */
    Datum       arg[FUNC_MAX_ARGS];  /* Arguments passed to function */
    bool        argnull[FUNC_MAX_ARGS];  /* T if arg[i] is actually NULL */
} FunctionCallInfoData;
typedef FunctionCallInfoData* FunctionCallInfo;

flinfo points to the lookup info used to make the call.  Ordinary functions
will probably ignore this field, but function class handlers will need it
to find out the OID of the specific function being called.

context is NULL for an "ordinary" function call, but may point to additional
info when the function is called in certain contexts.  (For example, the
trigger manager will pass information about the current trigger event here.)
If context is used, it should point to some subtype of Node; the particular
kind of context is indicated by the node type field.  (A callee should
always check the node type before assuming it knows what kind of context is
being passed.)  fmgr itself puts no other restrictions on the use of this
field.

resultinfo is NULL when calling any function from which a simple Datum
result is expected.  It may point to some subtype of Node if the function
returns more than a Datum.  (For example, resultinfo is used when calling a
function that returns a set, as discussed below.)  Like the context field,
resultinfo is a hook for expansion; fmgr itself doesn't constrain the use
of the field.

nargs, arg[], and argnull[] hold the arguments being passed to the function.
Notice that all the arguments passed to a function (as well as its result
value) will now uniformly be of type Datum.  As discussed below, callers
and callees should apply the standard Datum-to-and-from-whatever macros
to convert to the actual argument types of a particular function.  The
value in arg[i] is unspecified when argnull[i] is true.

It is generally the responsibility of the caller to ensure that the
number of arguments passed matches what the callee is expecting; except
for callees that take a variable number of arguments, the callee will
typically ignore the nargs field and just grab values from arg[].

The isnull field will be initialized to "false" before the call.  On
return from the function, isnull is the null flag for the function result:
if it is true the function's result is NULL, regardless of the actual
function return value.  Note that simple "strict" functions can ignore
both isnull and argnull[], since they won't even get called when there
are any TRUE values in argnull[].

FunctionCallInfo replaces FmgrValues plus a bunch of ad-hoc parameter
conventions, global variables (fmgr_pl_finfo and CurrentTriggerData at
least), and other uglinesses.


Callees, whether they be individual functions or function handlers,
shall always have this signature:

Datum function (FunctionCallInfo fcinfo);

which is represented by the typedef

typedef Datum (*PGFunction) (FunctionCallInfo fcinfo);

The function is responsible for setting fcinfo->isnull appropriately
as well as returning a result represented as a Datum.  Note that since
all callees will now have exactly the same signature, and will be called
through a function pointer declared with exactly that signature, we
should have no portability or optimization problems.


Function Coding Conventions
---------------------------

As an example, int4 addition goes from old-style

int32
int4pl(int32 arg1, int32 arg2)
{
    return arg1 + arg2;
}

to new-style

Datum
int4pl(FunctionCallInfo fcinfo)
{
    /* we assume the function is marked "strict", so we can ignore
     * NULL-value handling */

    return Int32GetDatum(DatumGetInt32(fcinfo->arg[0]) +
                         DatumGetInt32(fcinfo->arg[1]));
}

This is, of course, much uglier than the old-style code, but we can
improve matters with some well-chosen macros for the boilerplate parts.
I propose below macros that would make the code look like

Datum
int4pl(PG_FUNCTION_ARGS)
{
    int32   arg1 = PG_GETARG_INT32(0);
    int32   arg2 = PG_GETARG_INT32(1);

    PG_RETURN_INT32( arg1 + arg2 );
}

This is still more code than before, but it's fairly readable, and it's
also amenable to machine processing --- for example, we could probably
write a script that scans code like this and extracts argument and result
type info for comparison to the pg_proc table.

For the standard data types float4, float8, and int8, these macros should hide
whether the types are pass-by-value or pass-by reference, by incorporating
indirection and space allocation if needed.  This will offer a considerable
gain in readability, and it also opens up the opportunity to make these types
be pass-by-value on machines where it's feasible to do so.

Here are the proposed macros and coding conventions:

The definition of an fmgr-callable function will always look like

Datum
function_name(PG_FUNCTION_ARGS)
{
	...
}

"PG_FUNCTION_ARGS" just expands to "FunctionCallInfo fcinfo".  The main
reason for using this macro is to make it easy for scripts to spot function
definitions.  However, if we ever decide to change the calling convention
again, it might come in handy to have this macro in place.

A nonstrict function is responsible for checking whether each individual
argument is null or not, which it can do with PG_ARGISNULL(n) (which is
just "fcinfo->argnull[n]").  It should avoid trying to fetch the value
of any argument that is null.

Both strict and nonstrict functions can return NULL, if needed, with
	PG_RETURN_NULL();
which expands to
	{ fcinfo->isnull = true; return (Datum) 0; }

Argument values are ordinarily fetched using code like
	int32	name = PG_GETARG_INT32(number);

For float4, float8, and int8, the PG_GETARG macros will hide whether the
types are pass-by-value or pass-by-reference.  For example, if float8 is
pass-by-reference then PG_GETARG_FLOAT8 expands to
	(* (float8 *) DatumGetPointer(fcinfo->arg[number]))
and would typically be called like this:
	float8  arg = PG_GETARG_FLOAT8(0);
For what are now historical reasons, the float-related typedefs and macros
express the type width in bytes (4 or 8), whereas we prefer to label the
widths of integer types in bits.

Non-null values are returned with a PG_RETURN_XXX macro of the appropriate
type.  For example, PG_RETURN_INT32 expands to
	return Int32GetDatum(x)
PG_RETURN_FLOAT4, PG_RETURN_FLOAT8, and PG_RETURN_INT64 hide whether their
data types are pass-by-value or pass-by-reference, by doing a palloc if
needed.

fmgr.h will provide PG_GETARG and PG_RETURN macros for all the basic data
types.  Modules or header files that define specialized SQL datatypes
(eg, timestamp) should define appropriate macros for those types, so that
functions manipulating the types can be coded in the standard style.

For non-primitive data types (particularly variable-length types) it won't
be very practical to hide the pass-by-reference nature of the data type,
so the PG_GETARG and PG_RETURN macros for those types won't do much more
than DatumGetPointer/PointerGetDatum plus the appropriate typecast (but see
TOAST discussion, below).  Functions returning such types will need to
palloc() their result space explicitly.  I recommend naming the GETARG and
RETURN macros for such types to end in "_P", as a reminder that they
produce or take a pointer.  For example, PG_GETARG_TEXT_P yields "text *".

When a function needs to access fcinfo->flinfo or one of the other auxiliary
fields of FunctionCallInfo, it should just do it.  I doubt that providing
syntactic-sugar macros for these cases is useful.


Call-Site Coding Conventions
----------------------------

There are many places in the system that call either a specific function
(for example, the parser invokes "textin" by name in places) or a
particular group of functions that have a common argument list (for
example, the optimizer invokes selectivity estimation functions with
a fixed argument list).  These places will need to change, but we should
try to avoid making them significantly uglier than before.

Places that invoke an arbitrary function with an arbitrary argument list
can simply be changed to fill a FunctionCallInfoData structure directly;
that'll be no worse and possibly cleaner than what they do now.

When invoking a specific built-in function by name, we have generally
just written something like
	result = textin ( ... args ... )
which will not work after textin() is converted to the new call style.
I suggest that code like this be converted to use "helper" functions
that will create and fill in a FunctionCallInfoData struct.  For
example, if textin is being called with one argument, it'd look
something like
	result = DirectFunctionCall1(textin, PointerGetDatum(argument));
These helper routines will have declarations like
	Datum DirectFunctionCall2(PGFunction func, Datum arg1, Datum arg2);
Note it will be the caller's responsibility to convert to and from
Datum; appropriate conversion macros should be used.

The DirectFunctionCallN routines will not bother to fill in
fcinfo->flinfo (indeed cannot, since they have no idea about an OID for
the target function); they will just set it NULL.  This is unlikely to
bother any built-in function that could be called this way.  Note also
that this style of coding cannot pass a NULL input value nor cope with
a NULL result (it couldn't before, either!).  We can make the helper
routines ereport an error if they see that the function returns a NULL.

When invoking a function that has a known argument signature, we have
usually written either
	result = fmgr(targetfuncOid, ... args ... );
or
	result = fmgr_ptr(FmgrInfo *finfo, ... args ... );
depending on whether an FmgrInfo lookup has been done yet or not.
This kind of code can be recast using helper routines, in the same
style as above:
	result = OidFunctionCall1(funcOid, PointerGetDatum(argument));
	result = FunctionCall2(funcCallInfo,
	                       PointerGetDatum(argument),
	                       Int32GetDatum(argument));
Again, this style of coding does not allow for expressing NULL inputs
or receiving a NULL result.

As with the callee-side situation, I propose adding argument conversion
macros that hide whether int8, float4, and float8 are pass-by-value or
pass-by-reference.

The existing helper functions fmgr(), fmgr_c(), etc will be left in
place until all uses of them are gone.  Of course their internals will
have to change in the first step of implementation, but they can
continue to support the same external appearance.


Support for TOAST-Able Data Types
---------------------------------

For TOAST-able data types, the PG_GETARG macro will deliver a de-TOASTed
data value.  There might be a few cases where the still-toasted value is
wanted, but the vast majority of cases want the de-toasted result, so
that will be the default.  To get the argument value without causing
de-toasting, use PG_GETARG_RAW_VARLENA_P(n).

Some functions require a modifiable copy of their input values.  In these
cases, it's silly to do an extra copy step if we copied the data anyway
to de-TOAST it.  Therefore, each toastable datatype has an additional
fetch macro, for example PG_GETARG_TEXT_P_COPY(n), which delivers a
guaranteed-fresh copy, combining this with the detoasting step if possible.

There is also a PG_FREE_IF_COPY(ptr,n) macro, which pfree's the given
pointer if and only if it is different from the original value of the n'th
argument.  This can be used to free the de-toasted value of the n'th
argument, if it was actually de-toasted.  Currently, doing this is not
necessary for the majority of functions because the core backend code
releases temporary space periodically, so that memory leaked in function
execution isn't a big problem.  However, as of 7.1 memory leaks in
functions that are called by index searches will not be cleaned up until
end of transaction.  Therefore, functions that are listed in pg_amop or
pg_amproc should be careful not to leak detoasted copies, and so these
functions do need to use PG_FREE_IF_COPY() for toastable inputs.

A function should never try to re-TOAST its result value; it should just
deliver an untoasted result that's been palloc'd in the current memory
context.  When and if the value is actually stored into a tuple, the
tuple toaster will decide whether toasting is needed.


Functions Accepting or Returning Sets
-------------------------------------

[ this section revised 29-Aug-2002 for 7.3 ]

If a function is marked in pg_proc as returning a set, then it is called
with fcinfo->resultinfo pointing to a node of type ReturnSetInfo.  A
function that desires to return a set should raise an error "called in
context that does not accept a set result" if resultinfo is NULL or does
not point to a ReturnSetInfo node.

There are currently two modes in which a function can return a set result:
value-per-call, or materialize.  In value-per-call mode, the function returns
one value each time it is called, and finally reports "done" when it has no
more values to return.  In materialize mode, the function's output set is
instantiated in a Tuplestore object; all the values are returned in one call.
Additional modes might be added in future.

ReturnSetInfo contains a field "allowedModes" which is set (by the caller)
to a bitmask that's the OR of the modes the caller can support.  The actual
mode used by the function is returned in another field "returnMode".  For
backwards-compatibility reasons, returnMode is initialized to value-per-call
and need only be changed if the function wants to use a different mode.
The function should ereport() if it cannot use any of the modes the caller is
willing to support.

Value-per-call mode works like this: ReturnSetInfo contains a field
"isDone", which should be set to one of these values:

    ExprSingleResult             /* expression does not return a set */
    ExprMultipleResult           /* this result is an element of a set */
    ExprEndResult                /* there are no more elements in the set */

(the caller will initialize it to ExprSingleResult).  If the function simply
returns a Datum without touching ReturnSetInfo, then the call is over and a
single-item set has been returned.  To return a set, the function must set
isDone to ExprMultipleResult for each set element.  After all elements have
been returned, the next call should set isDone to ExprEndResult and return a
null result.  (Note it is possible to return an empty set by doing this on
the first call.)

The ReturnSetInfo node also contains a link to the ExprContext within which
the function is being evaluated.  This is useful for value-per-call functions
that need to close down internal state when they are not run to completion:
they can register a shutdown callback function in the ExprContext.

Materialize mode works like this: the function creates a Tuplestore holding
the (possibly empty) result set, and returns it.  There are no multiple calls.
The function must also return a TupleDesc that indicates the tuple structure.
The Tuplestore and TupleDesc should be created in the context
econtext->ecxt_per_query_memory (note this will *not* be the context the
function is called in).  The function stores pointers to the Tuplestore and
TupleDesc into ReturnSetInfo, sets returnMode to indicate materialize mode,
and returns null.  isDone is not used and should be left at ExprSingleResult.

The Tuplestore must be created with randomAccess = true if
SFRM_Materialize_Random is set in allowedModes, but it can (and preferably
should) be created with randomAccess = false if not.  Callers that can support
both ValuePerCall and Materialize mode will set SFRM_Materialize_Preferred,
or not, depending on which mode they prefer.

If available, the expected tuple descriptor is passed in ReturnSetInfo;
in other contexts the expectedDesc field will be NULL.  The function need
not pay attention to expectedDesc, but it may be useful in special cases.

There is no support for functions accepting sets; instead, the function will
be called multiple times, once for each element of the input set.


Notes About Function Handlers
-----------------------------

Handlers for classes of functions should find life much easier and
cleaner in this design.  The OID of the called function is directly
reachable from the passed parameters; we don't need the global variable
fmgr_pl_finfo anymore.  Also, by modifying fcinfo->flinfo->fn_extra,
the handler can cache lookup info to avoid repeat lookups when the same
function is invoked many times.  (fn_extra can only be used as a hint,
since callers are not required to re-use an FmgrInfo struct.
But in performance-critical paths they normally will do so.)

If the handler wants to allocate memory to hold fn_extra data, it should
NOT do so in CurrentMemoryContext, since the current context may well be
much shorter-lived than the context where the FmgrInfo is.  Instead,
allocate the memory in context flinfo->fn_mcxt, or in a long-lived cache
context.  fn_mcxt normally points at the context that was
CurrentMemoryContext at the time the FmgrInfo structure was created;
in any case it is required to be a context at least as long-lived as the
FmgrInfo itself.


Telling the Difference Between Old- and New-Style Functions
-----------------------------------------------------------

During the conversion process, we carried two different pg_language
entries, "internal" and "newinternal", for internal functions.  The
function manager used the language code to distinguish which calling
convention to use.  (Old-style internal functions were supported via
a function handler.)  As of Nov. 2000, no old-style internal functions
remain, so we can drop support for them.  We will remove the old "internal"
pg_language entry and rename "newinternal" to "internal".

The interim solution for dynamically-loaded compiled functions has been
similar: two pg_language entries "C" and "newC".  This naming convention
is not desirable for the long run, and yet we cannot stop supporting
old-style user functions.  Instead, it seems better to use just one
pg_language entry "C", and require the dynamically-loaded library to
provide additional information that identifies new-style functions.
This avoids compatibility problems --- for example, existing dump
scripts will identify PL language handlers as being in language "C",
which would be wrong under the "newC" convention.  Also, this approach
should generalize more conveniently for future extensions to the function
interface specification.

Given a dynamically loaded function named "foo" (note that the name being
considered here is the link-symbol name, not the SQL-level function name),
the function manager will look for another function in the same dynamically
loaded library named "pg_finfo_foo".  If this second function does not
exist, then foo is assumed to be called old-style, thus ensuring backwards
compatibility with existing libraries.  If the info function does exist,
it is expected to have the signature

	Pg_finfo_record * pg_finfo_foo (void);

The info function will be called by the fmgr, and must return a pointer
to a Pg_finfo_record struct.  (The returned struct will typically be a
statically allocated constant in the dynamic-link library.)  The current
definition of the struct is just

	typedef struct {
		int	api_version;
	} Pg_finfo_record;

where api_version is 0 to indicate old-style or 1 to indicate new-style
calling convention.  In future releases, additional fields may be defined
after api_version, but these additional fields will only be used if
api_version is greater than 1.

These details will be hidden from the author of a dynamically loaded
function by using a macro.  To define a new-style dynamically loaded
function named foo, write

	PG_FUNCTION_INFO_V1(foo);

	Datum
	foo(PG_FUNCTION_ARGS)
	{
		...
	}

The function itself is written using the same conventions as for new-style
internal functions; you just need to add the PG_FUNCTION_INFO_V1() macro.
Note that old-style and new-style functions can be intermixed in the same
library, depending on whether or not you write a PG_FUNCTION_INFO_V1() for
each one.

The SQL declaration for a dynamically-loaded function is CREATE FUNCTION
foo ... LANGUAGE C regardless of whether it is old- or new-style.

New-style dynamic functions will be invoked directly by fmgr, and will
therefore have the same performance as internal functions after the initial
pg_proc lookup overhead.  Old-style dynamic functions will be invoked via
a handler, and will therefore have a small performance penalty.

To allow old-style dynamic functions to work safely on toastable datatypes,
the handler for old-style functions will automatically detoast toastable
arguments before passing them to the old-style function.  A new-style
function is expected to take care of toasted arguments by using the
standard argument access macros defined above.