numeric.c

PostgreSQL7.4.6 for Linux

	init_var(&result);

	set_var_from_str(buf, &result);
	res = make_result(&result);

	free_var(&result);

	PG_RETURN_NUMERIC(res);
}


Datum
numeric_float4(PG_FUNCTION_ARGS)
{
	Numeric		num = PG_GETARG_NUMERIC(0);
	char	   *tmp;
	Datum		result;

	if (NUMERIC_IS_NAN(num))
		PG_RETURN_FLOAT4((float4) NAN);

	tmp = DatumGetCString(DirectFunctionCall1(numeric_out,
											  NumericGetDatum(num)));

	result = DirectFunctionCall1(float4in, CStringGetDatum(tmp));

	pfree(tmp);

	PG_RETURN_DATUM(result);
}


Datum
text_numeric(PG_FUNCTION_ARGS)
{
	text	   *str = PG_GETARG_TEXT_P(0);
	int			len;
	char	   *s;
	Datum		result;

	len = (VARSIZE(str) - VARHDRSZ);
	s = palloc(len + 1);
	memcpy(s, VARDATA(str), len);
	*(s + len) = '\0';

	result = DirectFunctionCall3(numeric_in, CStringGetDatum(s),
								 ObjectIdGetDatum(0), Int32GetDatum(-1));

	pfree(s);

	return result;
}

Datum
numeric_text(PG_FUNCTION_ARGS)
{
	/* val is numeric, but easier to leave it as Datum */
	Datum		val = PG_GETARG_DATUM(0);
	char	   *s;
	int			len;
	text	   *result;

	s = DatumGetCString(DirectFunctionCall1(numeric_out, val));
	len = strlen(s);

	result = (text *) palloc(VARHDRSZ + len);

	VARATT_SIZEP(result) = len + VARHDRSZ;
	memcpy(VARDATA(result), s, len);

	pfree(s);

	PG_RETURN_TEXT_P(result);
}


/* ----------------------------------------------------------------------
 *
 * Aggregate functions
 *
 * The transition datatype for all these aggregates is a 3-element array
 * of Numeric, holding the values N, sum(X), sum(X*X) in that order.
 *
 * We represent N as a numeric mainly to avoid having to build a special
 * datatype; it's unlikely it'd overflow an int4, but ...
 *
 * ----------------------------------------------------------------------
 */

static ArrayType *
do_numeric_accum(ArrayType *transarray, Numeric newval)
{
	Datum	   *transdatums;
	int			ndatums;
	Datum		N,
				sumX,
				sumX2;
	ArrayType  *result;

	/* We assume the input is array of numeric */
	deconstruct_array(transarray,
					  NUMERICOID, -1, false, 'i',
					  &transdatums, &ndatums);
	if (ndatums != 3)
		elog(ERROR, "expected 3-element numeric array");
	N = transdatums[0];
	sumX = transdatums[1];
	sumX2 = transdatums[2];

	N = DirectFunctionCall1(numeric_inc, N);
	sumX = DirectFunctionCall2(numeric_add, sumX,
							   NumericGetDatum(newval));
	sumX2 = DirectFunctionCall2(numeric_add, sumX2,
								DirectFunctionCall2(numeric_mul,
												 NumericGetDatum(newval),
											   NumericGetDatum(newval)));

	transdatums[0] = N;
	transdatums[1] = sumX;
	transdatums[2] = sumX2;

	result = construct_array(transdatums, 3,
							 NUMERICOID, -1, false, 'i');

	return result;
}

Datum
numeric_accum(PG_FUNCTION_ARGS)
{
	ArrayType  *transarray = PG_GETARG_ARRAYTYPE_P(0);
	Numeric		newval = PG_GETARG_NUMERIC(1);

	PG_RETURN_ARRAYTYPE_P(do_numeric_accum(transarray, newval));
}

/*
 * Integer data types all use Numeric accumulators to share code and
 * avoid risk of overflow.	For int2 and int4 inputs, Numeric accumulation
 * is overkill for the N and sum(X) values, but definitely not overkill
 * for the sum(X*X) value.	Hence, we use int2_accum and int4_accum only
 * for stddev/variance --- there are faster special-purpose accumulator
 * routines for SUM and AVG of these datatypes.
 */

Datum
int2_accum(PG_FUNCTION_ARGS)
{
	ArrayType  *transarray = PG_GETARG_ARRAYTYPE_P(0);
	Datum		newval2 = PG_GETARG_DATUM(1);
	Numeric		newval;

	newval = DatumGetNumeric(DirectFunctionCall1(int2_numeric, newval2));

	PG_RETURN_ARRAYTYPE_P(do_numeric_accum(transarray, newval));
}

Datum
int4_accum(PG_FUNCTION_ARGS)
{
	ArrayType  *transarray = PG_GETARG_ARRAYTYPE_P(0);
	Datum		newval4 = PG_GETARG_DATUM(1);
	Numeric		newval;

	newval = DatumGetNumeric(DirectFunctionCall1(int4_numeric, newval4));

	PG_RETURN_ARRAYTYPE_P(do_numeric_accum(transarray, newval));
}

Datum
int8_accum(PG_FUNCTION_ARGS)
{
	ArrayType  *transarray = PG_GETARG_ARRAYTYPE_P(0);
	Datum		newval8 = PG_GETARG_DATUM(1);
	Numeric		newval;

	newval = DatumGetNumeric(DirectFunctionCall1(int8_numeric, newval8));

	PG_RETURN_ARRAYTYPE_P(do_numeric_accum(transarray, newval));
}

Datum
numeric_avg(PG_FUNCTION_ARGS)
{
	ArrayType  *transarray = PG_GETARG_ARRAYTYPE_P(0);
	Datum	   *transdatums;
	int			ndatums;
	Numeric		N,
				sumX;

	/* We assume the input is array of numeric */
	deconstruct_array(transarray,
					  NUMERICOID, -1, false, 'i',
					  &transdatums, &ndatums);
	if (ndatums != 3)
		elog(ERROR, "expected 3-element numeric array");
	N = DatumGetNumeric(transdatums[0]);
	sumX = DatumGetNumeric(transdatums[1]);
	/* ignore sumX2 */

	/* SQL92 defines AVG of no values to be NULL */
	/* N is zero iff no digits (cf. numeric_uminus) */
	if (N->varlen == NUMERIC_HDRSZ)
		PG_RETURN_NULL();

	PG_RETURN_DATUM(DirectFunctionCall2(numeric_div,
										NumericGetDatum(sumX),
										NumericGetDatum(N)));
}

Datum
numeric_variance(PG_FUNCTION_ARGS)
{
	ArrayType  *transarray = PG_GETARG_ARRAYTYPE_P(0);
	Datum	   *transdatums;
	int			ndatums;
	Numeric		N,
				sumX,
				sumX2,
				res;
	NumericVar	vN,
				vsumX,
				vsumX2,
				vNminus1;
	int			rscale;

	/* We assume the input is array of numeric */
	deconstruct_array(transarray,
					  NUMERICOID, -1, false, 'i',
					  &transdatums, &ndatums);
	if (ndatums != 3)
		elog(ERROR, "expected 3-element numeric array");
	N = DatumGetNumeric(transdatums[0]);
	sumX = DatumGetNumeric(transdatums[1]);
	sumX2 = DatumGetNumeric(transdatums[2]);

	if (NUMERIC_IS_NAN(N) || NUMERIC_IS_NAN(sumX) || NUMERIC_IS_NAN(sumX2))
		PG_RETURN_NUMERIC(make_result(&const_nan));

	/* Sample variance is undefined when N is 0 or 1, so return NULL */
	init_var(&vN);
	set_var_from_num(N, &vN);

	if (cmp_var(&vN, &const_one) <= 0)
	{
		free_var(&vN);
		PG_RETURN_NULL();
	}

	init_var(&vNminus1);
	sub_var(&vN, &const_one, &vNminus1);

	init_var(&vsumX);
	set_var_from_num(sumX, &vsumX);
	init_var(&vsumX2);
	set_var_from_num(sumX2, &vsumX2);

	/* compute rscale for mul_var calls */
	rscale = vsumX.dscale * 2;

	mul_var(&vsumX, &vsumX, &vsumX, rscale);	/* vsumX = sumX * sumX */
	mul_var(&vN, &vsumX2, &vsumX2, rscale);		/* vsumX2 = N * sumX2 */
	sub_var(&vsumX2, &vsumX, &vsumX2);	/* N * sumX2 - sumX * sumX */

	if (cmp_var(&vsumX2, &const_zero) <= 0)
	{
		/* Watch out for roundoff error producing a negative numerator */
		res = make_result(&const_zero);
	}
	else
	{
		mul_var(&vN, &vNminus1, &vNminus1, 0);	/* N * (N - 1) */
		rscale = select_div_scale(&vsumX2, &vNminus1);
		div_var(&vsumX2, &vNminus1, &vsumX, rscale);	/* variance */

		res = make_result(&vsumX);
	}

	free_var(&vN);
	free_var(&vNminus1);
	free_var(&vsumX);
	free_var(&vsumX2);

	PG_RETURN_NUMERIC(res);
}

Datum
numeric_stddev(PG_FUNCTION_ARGS)
{
	ArrayType  *transarray = PG_GETARG_ARRAYTYPE_P(0);
	Datum	   *transdatums;
	int			ndatums;
	Numeric		N,
				sumX,
				sumX2,
				res;
	NumericVar	vN,
				vsumX,
				vsumX2,
				vNminus1;
	int			rscale;

	/* We assume the input is array of numeric */
	deconstruct_array(transarray,
					  NUMERICOID, -1, false, 'i',
					  &transdatums, &ndatums);
	if (ndatums != 3)
		elog(ERROR, "expected 3-element numeric array");
	N = DatumGetNumeric(transdatums[0]);
	sumX = DatumGetNumeric(transdatums[1]);
	sumX2 = DatumGetNumeric(transdatums[2]);

	if (NUMERIC_IS_NAN(N) || NUMERIC_IS_NAN(sumX) || NUMERIC_IS_NAN(sumX2))
		PG_RETURN_NUMERIC(make_result(&const_nan));

	/* Sample stddev is undefined when N is 0 or 1, so return NULL */
	init_var(&vN);
	set_var_from_num(N, &vN);

	if (cmp_var(&vN, &const_one) <= 0)
	{
		free_var(&vN);
		PG_RETURN_NULL();
	}

	init_var(&vNminus1);
	sub_var(&vN, &const_one, &vNminus1);

	init_var(&vsumX);
	set_var_from_num(sumX, &vsumX);
	init_var(&vsumX2);
	set_var_from_num(sumX2, &vsumX2);

	/* compute rscale for mul_var calls */
	rscale = vsumX.dscale * 2;

	mul_var(&vsumX, &vsumX, &vsumX, rscale);	/* vsumX = sumX * sumX */
	mul_var(&vN, &vsumX2, &vsumX2, rscale);		/* vsumX2 = N * sumX2 */
	sub_var(&vsumX2, &vsumX, &vsumX2);	/* N * sumX2 - sumX * sumX */

	if (cmp_var(&vsumX2, &const_zero) <= 0)
	{
		/* Watch out for roundoff error producing a negative numerator */
		res = make_result(&const_zero);
	}
	else
	{
		mul_var(&vN, &vNminus1, &vNminus1, 0);	/* N * (N - 1) */
		rscale = select_div_scale(&vsumX2, &vNminus1);
		div_var(&vsumX2, &vNminus1, &vsumX, rscale);	/* variance */
		sqrt_var(&vsumX, &vsumX, rscale);		/* stddev */

		res = make_result(&vsumX);
	}

	free_var(&vN);
	free_var(&vNminus1);
	free_var(&vsumX);
	free_var(&vsumX2);

	PG_RETURN_NUMERIC(res);
}


/*
 * SUM transition functions for integer datatypes.
 *
 * To avoid overflow, we use accumulators wider than the input datatype.
 * A Numeric accumulator is needed for int8 input; for int4 and int2
 * inputs, we use int8 accumulators which should be sufficient for practical
 * purposes.  (The latter two therefore don't really belong in this file,
 * but we keep them here anyway.)
 *
 * Because SQL92 defines the SUM() of no values to be NULL, not zero,
 * the initial condition of the transition data value needs to be NULL. This
 * means we can't rely on ExecAgg to automatically insert the first non-null
 * data value into the transition data: it doesn't know how to do the type
 * conversion.	The upshot is that these routines have to be marked non-strict
 * and handle substitution of the first non-null input themselves.
 */

Datum
int2_sum(PG_FUNCTION_ARGS)
{
	int64		oldsum;
	int64		newval;

	if (PG_ARGISNULL(0))
	{
		/* No non-null input seen so far... */
		if (PG_ARGISNULL(1))
			PG_RETURN_NULL();	/* still no non-null */
		/* This is the first non-null input. */
		newval = (int64) PG_GETARG_INT16(1);
		PG_RETURN_INT64(newval);
	}

	oldsum = PG_GETARG_INT64(0);

	/* Leave sum unchanged if new input is null. */
	if (PG_ARGISNULL(1))
		PG_RETURN_INT64(oldsum);

	/* OK to do the addition. */
	newval = oldsum + (int64) PG_GETARG_INT16(1);

	PG_RETURN_INT64(newval);
}

Datum
int4_sum(PG_FUNCTION_ARGS)
{
	int64		oldsum;
	int64		newval;

	if (PG_ARGISNULL(0))
	{
		/* No non-null input seen so far... */
		if (PG_ARGISNULL(1))
			PG_RETURN_NULL();	/* still no non-null */
		/* This is the first non-null input. */
		newval = (int64) PG_GETARG_INT32(1);
		PG_RETURN_INT64(newval);
	}

	oldsum = PG_GETARG_INT64(0);

	/* Leave sum unchanged if new input is null. */
	if (PG_ARGISNULL(1))
		PG_RETURN_INT64(oldsum);

	/* OK to do the addition. */
	newval = oldsum + (int64) PG_GETARG_INT32(1);

	PG_RETURN_INT64(newval);
}

Datum
int8_sum(PG_FUNCTION_ARGS)
{
	Numeric		oldsum;
	Datum		newval;

	if (PG_ARGISNULL(0))
	{
		/* No non-null input seen so far... */
		if (PG_ARGISNULL(1))
			PG_RETURN_NULL();	/* still no non-null */
		/* This is the first non-null input. */
		newval = DirectFunctionCall1(int8_numeric, PG_GETARG_DATUM(1));
		PG_RETURN_DATUM(newval);
	}

	oldsum = PG_GETARG_NUMERIC(0);

	/* Leave sum unchanged if new input is null. */
	if (PG_ARGISNULL(1))
		PG_RETURN_NUMERIC(oldsum);

	/* OK to do the addition. */
	newval = DirectFunctionCall1(int8_numeric, PG_GETARG_DATUM(1));

	PG_RETURN_DATUM(DirectFunctionCall2(numeric_add,
										NumericGetDatum(oldsum), newval));
}


/*
 * Routines for avg(int2) and avg(int4).  The transition datatype
 * is a two-element int8 array, holding count and sum.
 */

typedef struct Int8TransTypeData
{
#ifndef INT64_IS_BUSTED
	int64		count;
	int64		sum;
#else
	/* "int64" isn't really 64 bits, so fake up properly-aligned fields */
	int32		count;
	int32		pad1;
	int32		sum;
	int32		pad2;
#endif
} Int8TransTypeData;

Datum
int2_avg_accum(PG_FUNCTION_ARGS)
{
	ArrayType  *transarray = PG_GETARG_ARRAYTYPE_P_COPY(0);
	int16		newval = PG_GETARG_INT16(1);
	Int8TransTypeData *transdata;

	/*
	 * We copied the input array, so it's okay to scribble on it directly.
	 */
	if (ARR_SIZE(transarray) != ARR_OVERHEAD(1) + sizeof(Int8TransTypeData))
		elog(ERROR, "expected 2-element int8 array");
	transdata = (Int8TransTypeData *) ARR_DATA_PTR(transarray);

	transdata->count++;
	transdata->sum += newval;

	PG_RETURN_ARRAYTYPE_P(transarray);
}

Datum
int4_avg_accum(PG_FUNCTION_ARGS)
{
	ArrayType  *transarray = PG_GETARG_ARRAYTYPE_P_COPY(0);
	int32		newval = PG_GETARG_INT32(1);
	Int8TransTypeData *transdata;

	/*
	 * We copied the input array, so it's okay to scribble on it directly.
	 */
	if (ARR_SIZE(transarray) != ARR_OVERHEAD(1) + sizeof(Int8TransTypeData))
		elog(ERROR, "expected 2-element int8 array");
	transdata = (Int8TransTypeData *) ARR_DATA_PTR(transarray);

	transdata->count++;
	transdata->sum += newval;

	PG_RETURN_ARRAYTYPE_P(transarray);
}

Datum
int8_avg(PG_FUNCTION_ARGS)
{
	ArrayType  *transarray = PG_GETARG_ARRAYTYPE_P(0);
	Int8TransTypeData *transdata;
	Datum		countd,
				sumd;

	if (ARR_SIZE(transarray) != ARR_OVERHEAD(1) + sizeof(Int8TransTypeData))
		elog(ERROR, "expected 2-element int8 array");
	transdata = (Int8TransTypeData *) ARR_DATA_PTR(transarray);

	/* SQL92 defines AVG of no values to be NULL */
	if (transdata->count == 0)
		PG_RETURN_NULL();

	countd = DirectFunctionCall1(int8_numeric,
								 Int64GetDatumFast(transdata->count));
	sumd = DirectFunctionCall1(int8_numeric,
							   Int64GetDatumFast(transdata->sum));

	PG_RETURN_DATUM(DirectFunctionCall2(numeric_div, sumd, countd));
}


/* ----------------------------------------------------------------------
 *
 * Debug support
 *
 * ----------------------------------------------------------------------
 */

#ifdef NUMERIC_DEBUG

/*
 * dump_numeric() - Dump a value in the db storage format for debugging
 */
static void
dump_numeric(const char *str, Numeric num)
{
	NumericDigit *digits = (NumericDigit *) num->n_data;
	int			ndigits;
	int			i;

	ndigits = (num->varlen - NUMERIC_HDRSZ) / sizeof(NumericDigit);

	printf("%s: NUMERIC w=%d d=%d ", str, num->n_weight, NUMERIC_DSCALE(num));
	switch (NUMERIC_SIGN(num))
	{
		case NUMERIC_POS:
			printf("POS");
			break;
		case NUMERIC_NEG:
			printf("NEG");
			break;
		case NUMERIC_NAN:
			printf("NaN");
			break;
		default:
			printf("SIGN=0x%x", NUMERIC_SIGN(num));
			break;
	}

	for (i = 0; i < ndigits; i++)
		printf(" %0*d", DEC_DIGITS, digits[i]);
	printf("\n");
}


/*
 * dump_var() - Dump a value in the variable format for debugging
 */