Обсуждение: PG_FREE_IF_COPY extraneous in numeric_cmp?

Hi hackers,

I have a question on the code below:

Datum
numeric_cmp(PG_FUNCTION_ARGS)
{
  Numeric num1 = PG_GETARG_NUMERIC(0);
  Numeric num2 = PG_GETARG_NUMERIC(1);
  int result;

  result = cmp_numerics(num1, num2);

  PG_FREE_IF_COPY(num1, 0);
  PG_FREE_IF_COPY(num2, 1);

  PG_RETURN_INT32(result);
}

It seems to me that num1 is a copy of fcinfo->arg[0]. It is passed to
the function cmp_numerics(), It's value remains the same after the
call. Also, cmp_numerics() does not have a handle to fcinfo, so it
can't modify fcinfo->arg[0].

Isn't it true that pfree() will never be called by PG_FREE_IF_COPY?

Cheers,
-cktan

CK Tan <cktan@vitessedata.com> writes:
> Isn't it true that pfree() will never be called by PG_FREE_IF_COPY?

No.  You're forgetting the possibility that PG_GETARG_NUMERIC will
have to de-toast a toasted input.  Granted, numerics are seldom
going to be long enough to get compressed or pushed out-of-line;
but that's possible, and what's very possible is that they'll have
a short header.

            regards, tom lane

Thanks!

On Fri, Feb 24, 2023 at 2:16 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
>
> CK Tan <cktan@vitessedata.com> writes:
> > Isn't it true that pfree() will never be called by PG_FREE_IF_COPY?
>
> No.  You're forgetting the possibility that PG_GETARG_NUMERIC will
> have to de-toast a toasted input.  Granted, numerics are seldom
> going to be long enough to get compressed or pushed out-of-line;
> but that's possible, and what's very possible is that they'll have
> a short header.
>
>                         regards, tom lane