Re: [PATCH] Add some documentation on how to call internal functions - Mailing list pgsql-hackers

On 9 Oct 2024, at 2:57 PM, Pavel Stehule <pavel.stehule@gmail.com> wrote:

Hi

I am checking this patch

čt 12. 9. 2024 v 11:42 odesílatel Florents Tselai <florents.tselai@gmail.com> napsal:

Hi, 

The documentation on extending using C functions, 

leaves a blank on how to call other internal Postgres functions.

This can leave first-timers hanging.

I think it’d be helpful to spend some paragraphs on discussing the DirectFunctionCall API.

Here’s an attempt (also a PR[0]).

Here’s the relevant HTML snippet for convenience:

To call another version-1 function, you can use DirectFunctionCalln(func, arg1,...,argn). This is particularly useful when you want to call functions defined in the standard internal library, by using an interface similar to their SQL signature.

Different flavors of similar macros can be found in fmgr.h. The main point though is that they expect a C function name to call as their first argument (or its Oid in some cases), and actual arguments should be supplied as Datums. They always return Datum.

For example, to call the starts_with(text, text) from C, you can search through the catalog and find out that its C implementation is based on the Datum text_starts_with(PG_FUNCTION_ARGS) function.

In fmgr.h there are also available macros the facilitate conversions between C types and Datum. For example to turn text* into Datum, you can use DatumGetTextPP(X). If your extension defines additional types, it is usually convenient to define similar macros for these types too.

I’ve also added the below example function:

PG_FUNCTION_INFO_V1(t_starts_with);

Datum
t_starts_with(PG_FUNCTION_ARGS)
{    Datum t1 = PG_GETARG_DATUM(0);    Datum t2 = PG_GETARG_DATUM(1);    bool  bool_res;
    Datum datum_res = DirectFunctionCall2(text_starts_with, t1, t2);    bool_res = DatumGetBool(datum_res);
    PG_RETURN_BOOL(bool_res);
}

PS1: I was not sure if src/tutorial is still relevant with this part of the documentation.

If so, it needs updating too.

I have few points

1. I am missing warning so NULL is not supported by DirectFunctionCall - on input and on output. Any argument should not be null, and the result should not be null.

You’re right.

2. The example looks little bit not consistent

I propose

    Text *t1 = PG_GETARG_TEXT_PP(0);    Text *t2 = PG_GETARG_TEXT_PP(1);    bool  result;
    result = DatumGetBool(
                  DirectFunctionCall2(text_starts_with,
                                      PointerGetDatum(t1),
                                      PointerGetDatum(t2));
    PG_RETURN_BOOL(result);

or

    Datum t1 = PG_GETARG_DATUM(0);    Datum t2 = PG_GETARG_DATUM(1);    Datum result;
    result = DirectFunctionCall2(text_starts_with, t1, t2);
    PG_RETURN_DATUM(result);

Both examples can show some interesting patterns (maybe can be used both)

The example can be extended a bit more.

I like your first proposal.

My primary purpose with this patch is to show explicitly how 

one goes from SQL types, to Datum and/or char*/text if necessary,

use some internal C function (even if it doesn’t have an SQL equivalent).

and then return back an SQL type.

So, simply GET_ARG_DATUM() and then PG_RETURN_DATUM() defeats the purpose.