]> granicus.if.org Git - postgresql/commitdiff
Re-implement division for numeric values using the traditional "schoolbook"
authorTom Lane <tgl@sss.pgh.pa.us>
Fri, 4 Apr 2008 18:45:36 +0000 (18:45 +0000)
committerTom Lane <tgl@sss.pgh.pa.us>
Fri, 4 Apr 2008 18:45:36 +0000 (18:45 +0000)
algorithm.  This is a good deal slower than our old roundoff-error-prone
code for long inputs, so we keep the old code for use in the transcendental
functions, where everything is approximate anyway.  Also create a
user-accessible function div(numeric, numeric) to provide access to the
exact result of trunc(x/y) --- since the regular numeric / operator will
round off its result, simply computing that expression in SQL doesn't
reliably give the desired answer.  This fixes bug #3387 and various related
corner cases, and improves the usefulness of PG for high-precision integer
arithmetic.

doc/src/sgml/func.sgml
src/backend/utils/adt/numeric.c
src/include/catalog/catversion.h
src/include/catalog/pg_proc.h
src/include/utils/builtins.h
src/test/regress/expected/numeric.out
src/test/regress/sql/numeric.sql

index 1fa0fc6c2a3fa3ad0f09d7222d59ebfbd649f1fe..0811a282db2ee0a8fe00bf60d584a86a5dada0dc 100644 (file)
@@ -1,4 +1,4 @@
-<!-- $PostgreSQL: pgsql/doc/src/sgml/func.sgml,v 1.426 2008/04/04 16:57:21 momjian Exp $ -->
+<!-- $PostgreSQL: pgsql/doc/src/sgml/func.sgml,v 1.427 2008/04/04 18:45:36 tgl Exp $ -->
 
  <chapter id="functions">
   <title>Functions and Operators</title>
    <indexterm>
     <primary>degrees</primary>
    </indexterm>
+   <indexterm>
+    <primary>div</primary>
+   </indexterm>
    <indexterm>
     <primary>exp</primary>
    </indexterm>
        <entry><literal>28.6478897565412</literal></entry>
       </row>
 
+      <row>
+       <entry><literal><function>div</function>(<parameter>y</parameter> <type>numeric</>,
+        <parameter>x</parameter> <type>numeric</>)</literal></entry>
+       <entry><type>numeric</></entry>
+       <entry>integer quotient of <parameter>y</parameter>/<parameter>x</parameter></entry>
+       <entry><literal>div(9,4)</literal></entry>
+       <entry><literal>2</literal></entry>
+      </row>
+
       <row>
        <entry><literal><function>exp</function>(<type>dp</type> or <type>numeric</type>)</literal></entry>
        <entry>(same as input)</entry>
index 2ebfee52e4559a758210d86e6327c29ddab519fd..86765d5d5325f686046f46d152625f181a1df303 100644 (file)
@@ -14,7 +14,7 @@
  * Copyright (c) 1998-2008, PostgreSQL Global Development Group
  *
  * IDENTIFICATION
- *       $PostgreSQL: pgsql/src/backend/utils/adt/numeric.c,v 1.108 2008/01/01 19:45:52 momjian Exp $
+ *       $PostgreSQL: pgsql/src/backend/utils/adt/numeric.c,v 1.109 2008/04/04 18:45:36 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -53,7 +53,7 @@
  * NBASE that's less than sqrt(INT_MAX), in practice we are only interested
  * in NBASE a power of ten, so that I/O conversions and decimal rounding
  * are easy.  Also, it's actually more efficient if NBASE is rather less than
- * sqrt(INT_MAX), so that there is "headroom" for mul_var and div_var to
+ * sqrt(INT_MAX), so that there is "headroom" for mul_var and div_var_fast to
  * postpone processing carries.
  * ----------
  */
@@ -90,6 +90,10 @@ typedef int16 NumericDigit;
 
 
 /* ----------
+ * NumericVar is the format we use for arithmetic.  The digit-array part
+ * is the same as the NumericData storage format, but the header is more
+ * complex.
+ *
  * The value represented by a NumericVar is determined by the sign, weight,
  * ndigits, and digits[] array.
  * Note: the first digit of a NumericVar's value is assumed to be multiplied
@@ -100,7 +104,7 @@ typedef int16 NumericDigit;
  * NumericVar. digits points at the first digit in actual use (the one
  * with the specified weight). We normally leave an unused digit or two
  * (preset to zeroes) between buf and digits, so that there is room to store
- * a carry out of the top digit without special pushups.  We just need to
+ * a carry out of the top digit without reallocating space.  We just need to
  * decrement digits (and increment weight) to make room for the carry digit.
  * (There is no such extra space in a numeric value stored in the database,
  * only in a NumericVar in memory.)
@@ -265,6 +269,8 @@ static void mul_var(NumericVar *var1, NumericVar *var2, NumericVar *result,
                int rscale);
 static void div_var(NumericVar *var1, NumericVar *var2, NumericVar *result,
                int rscale, bool round);
+static void div_var_fast(NumericVar *var1, NumericVar *var2, NumericVar *result,
+               int rscale, bool round);
 static int     select_div_scale(NumericVar *var1, NumericVar *var2);
 static void mod_var(NumericVar *var1, NumericVar *var2, NumericVar *result);
 static void ceil_var(NumericVar *var, NumericVar *result);
@@ -1419,6 +1425,52 @@ numeric_div(PG_FUNCTION_ARGS)
 }
 
 
+/*
+ * numeric_div_trunc() -
+ *
+ *     Divide one numeric into another, truncating the result to an integer
+ */
+Datum
+numeric_div_trunc(PG_FUNCTION_ARGS)
+{
+       Numeric         num1 = PG_GETARG_NUMERIC(0);
+       Numeric         num2 = PG_GETARG_NUMERIC(1);
+       NumericVar      arg1;
+       NumericVar      arg2;
+       NumericVar      result;
+       Numeric         res;
+
+       /*
+        * Handle NaN
+        */
+       if (NUMERIC_IS_NAN(num1) || NUMERIC_IS_NAN(num2))
+               PG_RETURN_NUMERIC(make_result(&const_nan));
+
+       /*
+        * Unpack the arguments
+        */
+       init_var(&arg1);
+       init_var(&arg2);
+       init_var(&result);
+
+       set_var_from_num(num1, &arg1);
+       set_var_from_num(num2, &arg2);
+
+       /*
+        * Do the divide and return the result
+        */
+       div_var(&arg1, &arg2, &result, 0, false);
+
+       res = make_result(&result);
+
+       free_var(&arg1);
+       free_var(&arg2);
+       free_var(&result);
+
+       PG_RETURN_NUMERIC(res);
+}
+
+
 /*
  * numeric_mod() -
  *
@@ -4036,12 +4088,291 @@ mul_var(NumericVar *var1, NumericVar *var2, NumericVar *result,
 /*
  * div_var() -
  *
- *     Division on variable level. Quotient of var1 / var2 is stored
- *     in result.      Result is rounded to no more than rscale fractional digits.
+ *     Division on variable level. Quotient of var1 / var2 is stored in result.
+ *     The quotient is figured to exactly rscale fractional digits.
+ *     If round is true, it is rounded at the rscale'th digit; if false, it
+ *     is truncated (towards zero) at that digit.
  */
 static void
 div_var(NumericVar *var1, NumericVar *var2, NumericVar *result,
                int rscale, bool round)
+{
+       int                     div_ndigits;
+       int                     res_ndigits;
+       int                     res_sign;
+       int                     res_weight;
+       int                     carry;
+       int                     borrow;
+       int                     divisor1;
+       int                     divisor2;
+       NumericDigit *dividend;
+       NumericDigit *divisor;
+       NumericDigit *res_digits;
+       int                     i;
+       int                     j;
+
+       /* copy these values into local vars for speed in inner loop */
+       int                     var1ndigits = var1->ndigits;
+       int                     var2ndigits = var2->ndigits;
+
+       /*
+        * First of all division by zero check; we must not be handed an
+        * unnormalized divisor.
+        */
+       if (var2ndigits == 0 || var2->digits[0] == 0)
+               ereport(ERROR,
+                               (errcode(ERRCODE_DIVISION_BY_ZERO),
+                                errmsg("division by zero")));
+
+       /*
+        * Now result zero check
+        */
+       if (var1ndigits == 0)
+       {
+               zero_var(result);
+               result->dscale = rscale;
+               return;
+       }
+
+       /*
+        * Determine the result sign, weight and number of digits to calculate.
+        * The weight figured here is correct if the emitted quotient has no
+        * leading zero digits; otherwise strip_var() will fix things up.
+        */
+       if (var1->sign == var2->sign)
+               res_sign = NUMERIC_POS;
+       else
+               res_sign = NUMERIC_NEG;
+       res_weight = var1->weight - var2->weight;
+       /* The number of accurate result digits we need to produce: */
+       res_ndigits = res_weight + 1 + (rscale + DEC_DIGITS - 1) / DEC_DIGITS;
+       /* ... but always at least 1 */
+       res_ndigits = Max(res_ndigits, 1);
+       /* If rounding needed, figure one more digit to ensure correct result */
+       if (round)
+               res_ndigits++;
+       /*
+        * The working dividend normally requires res_ndigits + var2ndigits
+        * digits, but make it at least var1ndigits so we can load all of var1
+        * into it.  (There will be an additional digit dividend[0] in the
+        * dividend space, but for consistency with Knuth's notation we don't
+        * count that in div_ndigits.)
+        */
+       div_ndigits = res_ndigits + var2ndigits;
+       div_ndigits = Max(div_ndigits, var1ndigits);
+
+       /*
+        * We need a workspace with room for the working dividend (div_ndigits+1
+        * digits) plus room for the possibly-normalized divisor (var2ndigits
+        * digits).  It is convenient also to have a zero at divisor[0] with
+        * the actual divisor data in divisor[1 .. var2ndigits].  Transferring the
+        * digits into the workspace also allows us to realloc the result (which
+        * might be the same as either input var) before we begin the main loop.
+        * Note that we use palloc0 to ensure that divisor[0], dividend[0], and
+        * any additional dividend positions beyond var1ndigits, start out 0.
+        */
+       dividend = (NumericDigit *)
+               palloc0((div_ndigits + var2ndigits + 2) * sizeof(NumericDigit));
+       divisor = dividend + (div_ndigits + 1);
+       memcpy(dividend + 1, var1->digits, var1ndigits * sizeof(NumericDigit));
+       memcpy(divisor + 1, var2->digits, var2ndigits * sizeof(NumericDigit));
+
+       /*
+        * Now we can realloc the result to hold the generated quotient digits.
+        */
+       alloc_var(result, res_ndigits);
+       res_digits = result->digits;
+
+       if (var2ndigits == 1)
+       {
+               /*
+                * If there's only a single divisor digit, we can use a fast path
+                * (cf. Knuth section 4.3.1 exercise 16).
+                */
+               divisor1 = divisor[1];
+               carry = 0;
+               for (i = 0; i < res_ndigits; i++)
+               {
+                       carry = carry * NBASE + dividend[i + 1];
+                       res_digits[i] = carry / divisor1;
+                       carry = carry % divisor1;
+               }
+       }
+       else
+       {
+               /*
+                * The full multiple-place algorithm is taken from Knuth volume 2,
+                * Algorithm 4.3.1D.
+                *
+                * We need the first divisor digit to be >= NBASE/2.  If it isn't,
+                * make it so by scaling up both the divisor and dividend by the
+                * factor "d".  (The reason for allocating dividend[0] above is to
+                * leave room for possible carry here.)
+                */
+               if (divisor[1] < HALF_NBASE)
+               {
+                       int             d = NBASE / (divisor[1] + 1);
+
+                       carry = 0;
+                       for (i = var2ndigits; i > 0; i--)
+                       {
+                               carry += divisor[i] * d;
+                               divisor[i] = carry % NBASE;
+                               carry = carry / NBASE;
+                       }
+                       Assert(carry == 0);
+                       carry = 0;
+                       /* at this point only var1ndigits of dividend can be nonzero */
+                       for (i = var1ndigits; i >= 0; i--)
+                       {
+                               carry += dividend[i] * d;
+                               dividend[i] = carry % NBASE;
+                               carry = carry / NBASE;
+                       }
+                       Assert(carry == 0);
+                       Assert(divisor[1] >= HALF_NBASE);
+               }
+               /* First 2 divisor digits are used repeatedly in main loop */
+               divisor1 = divisor[1];
+               divisor2 = divisor[2];
+
+               /*
+                * Begin the main loop.  Each iteration of this loop produces the
+                * j'th quotient digit by dividing dividend[j .. j + var2ndigits]
+                * by the divisor; this is essentially the same as the common manual
+                * procedure for long division.
+                */
+               for (j = 0; j < res_ndigits; j++)
+               {
+                       /* Estimate quotient digit from the first two dividend digits */
+                       int             next2digits = dividend[j] * NBASE + dividend[j+1];
+                       int             qhat;
+
+                       /*
+                        * If next2digits are 0, then quotient digit must be 0 and there's
+                        * no need to adjust the working dividend.  It's worth testing
+                        * here to fall out ASAP when processing trailing zeroes in
+                        * a dividend.
+                        */
+                       if (next2digits == 0)
+                       {
+                               res_digits[j] = 0;
+                               continue;
+                       }
+
+                       if (dividend[j] == divisor1)
+                               qhat = NBASE - 1;
+                       else
+                               qhat = next2digits / divisor1;
+                       /*
+                        * Adjust quotient digit if it's too large.  Knuth proves that
+                        * after this step, the quotient digit will be either correct
+                        * or just one too large.  (Note: it's OK to use dividend[j+2]
+                        * here because we know the divisor length is at least 2.)
+                        */
+                       while (divisor2 * qhat >
+                                  (next2digits - qhat * divisor1) * NBASE + dividend[j+2])
+                               qhat--;
+
+                       /* As above, need do nothing more when quotient digit is 0 */
+                       if (qhat > 0)
+                       {
+                               /*
+                                * Multiply the divisor by qhat, and subtract that from the
+                                * working dividend.  "carry" tracks the multiplication,
+                                * "borrow" the subtraction (could we fold these together?)
+                                */
+                               carry = 0;
+                               borrow = 0;
+                               for (i = var2ndigits; i >= 0; i--)
+                               {
+                                       carry += divisor[i] * qhat;
+                                       borrow -= carry % NBASE;
+                                       carry = carry / NBASE;
+                                       borrow += dividend[j + i];
+                                       if (borrow < 0)
+                                       {
+                                               dividend[j + i] = borrow + NBASE;
+                                               borrow = -1;
+                                       }
+                                       else
+                                       {
+                                               dividend[j + i] = borrow;
+                                               borrow = 0;
+                                       }
+                               }
+                               Assert(carry == 0);
+
+                               /*
+                                * If we got a borrow out of the top dividend digit, then
+                                * indeed qhat was one too large.  Fix it, and add back the
+                                * divisor to correct the working dividend.  (Knuth proves
+                                * that this will occur only about 3/NBASE of the time; hence,
+                                * it's a good idea to test this code with small NBASE to be
+                                * sure this section gets exercised.)
+                                */
+                               if (borrow)
+                               {
+                                       qhat--;
+                                       carry = 0;
+                                       for (i = var2ndigits; i >= 0; i--)
+                                       {
+                                               carry += dividend[j + i] + divisor[i];
+                                               if (carry >= NBASE)
+                                               {
+                                                       dividend[j + i] = carry - NBASE;
+                                                       carry = 1;
+                                               }
+                                               else
+                                               {
+                                                       dividend[j + i] = carry;
+                                                       carry = 0;
+                                               }
+                                       }
+                                       /* A carry should occur here to cancel the borrow above */
+                                       Assert(carry == 1);
+                               }
+                       }
+
+                       /* And we're done with this quotient digit */
+                       res_digits[j] = qhat;
+               }
+       }
+
+       pfree(dividend);
+
+       /*
+        * Finally, round or truncate the result to the requested precision.
+        */
+       result->weight = res_weight;
+       result->sign = res_sign;
+
+       /* Round or truncate to target rscale (and set result->dscale) */
+       if (round)
+               round_var(result, rscale);
+       else
+               trunc_var(result, rscale);
+
+       /* Strip leading and trailing zeroes */
+       strip_var(result);
+}
+
+
+/*
+ * div_var_fast() -
+ *
+ *     This has the same API as div_var, but is implemented using the division
+ *     algorithm from the "FM" library, rather than Knuth's schoolbook-division
+ *     approach.  This is significantly faster but can produce inaccurate
+ *     results, because it sometimes has to propagate rounding to the left,
+ *     and so we can never be entirely sure that we know the requested digits
+ *     exactly.  We compute DIV_GUARD_DIGITS extra digits, but there is
+ *     no certainty that that's enough.  We use this only in the transcendental
+ *     function calculation routines, where everything is approximate anyway.
+ */
+static void
+div_var_fast(NumericVar *var1, NumericVar *var2, NumericVar *result,
+                        int rscale, bool round)
 {
        int                     div_ndigits;
        int                     res_sign;
@@ -4367,30 +4698,21 @@ static void
 mod_var(NumericVar *var1, NumericVar *var2, NumericVar *result)
 {
        NumericVar      tmp;
-       int                     rscale;
 
        init_var(&tmp);
 
        /* ---------
         * We do this using the equation
         *              mod(x,y) = x - trunc(x/y)*y
-        * We set rscale the same way numeric_div and numeric_mul do
-        * to get the right answer from the equation.  The final result,
-        * however, need not be displayed to more precision than the inputs.
+        * div_var can be persuaded to give us trunc(x/y) directly.
         * ----------
         */
-       rscale = select_div_scale(var1, var2);
+       div_var(var1, var2, &tmp, 0, false);
 
-       div_var(var1, var2, &tmp, rscale, false);
-
-       trunc_var(&tmp, 0);
-
-       mul_var(var2, &tmp, &tmp, var2->dscale + tmp.dscale);
+       mul_var(var2, &tmp, &tmp, var2->dscale);
 
        sub_var(var1, &tmp, result);
 
-       round_var(result, Max(var1->dscale, var2->dscale));
-
        free_var(&tmp);
 }
 
@@ -4497,7 +4819,7 @@ sqrt_var(NumericVar *arg, NumericVar *result, int rscale)
 
        for (;;)
        {
-               div_var(&tmp_arg, result, &tmp_val, local_rscale, true);
+               div_var_fast(&tmp_arg, result, &tmp_val, local_rscale, true);
 
                add_var(result, &tmp_val, result);
                mul_var(result, &const_zero_point_five, result, local_rscale);
@@ -4587,7 +4909,7 @@ exp_var(NumericVar *arg, NumericVar *result, int rscale)
 
        /* Compensate for input sign, and round to requested rscale */
        if (xneg)
-               div_var(&const_one, result, result, rscale, true);
+               div_var_fast(&const_one, result, result, rscale, true);
        else
                round_var(result, rscale);
 
@@ -4652,7 +4974,7 @@ exp_var_internal(NumericVar *arg, NumericVar *result, int rscale)
                add_var(&ni, &const_one, &ni);
                mul_var(&xpow, &x, &xpow, local_rscale);
                mul_var(&ifac, &ni, &ifac, 0);
-               div_var(&xpow, &ifac, &elem, local_rscale, true);
+               div_var_fast(&xpow, &ifac, &elem, local_rscale, true);
 
                if (elem.ndigits == 0)
                        break;
@@ -4736,7 +5058,7 @@ ln_var(NumericVar *arg, NumericVar *result, int rscale)
         */
        sub_var(&x, &const_one, result);
        add_var(&x, &const_one, &elem);
-       div_var(result, &elem, result, local_rscale, true);
+       div_var_fast(result, &elem, result, local_rscale, true);
        set_var_from_var(result, &xx);
        mul_var(result, result, &x, local_rscale);
 
@@ -4746,7 +5068,7 @@ ln_var(NumericVar *arg, NumericVar *result, int rscale)
        {
                add_var(&ni, &const_two, &ni);
                mul_var(&xx, &x, &xx, local_rscale);
-               div_var(&xx, &ni, &elem, local_rscale, true);
+               div_var_fast(&xx, &ni, &elem, local_rscale, true);
 
                if (elem.ndigits == 0)
                        break;
@@ -4816,7 +5138,7 @@ log_var(NumericVar *base, NumericVar *num, NumericVar *result)
        /* Select scale for division result */
        rscale = select_div_scale(&ln_num, &ln_base);
 
-       div_var(&ln_num, &ln_base, result, rscale, true);
+       div_var_fast(&ln_num, &ln_base, result, rscale, true);
 
        free_var(&ln_num);
        free_var(&ln_base);
@@ -4990,7 +5312,7 @@ power_var_int(NumericVar *base, int exp, NumericVar *result, int rscale)
 
        /* Compensate for input sign, and round to requested rscale */
        if (neg)
-               div_var(&const_one, result, result, rscale, true);
+               div_var_fast(&const_one, result, result, rscale, true);
        else
                round_var(result, rscale);
 }
@@ -5361,8 +5683,8 @@ round_var(NumericVar *var, int rscale)
 /*
  * trunc_var
  *
- * Truncate the value of a variable at rscale decimal digits after the
- * decimal point.  NOTE: we allow rscale < 0 here, implying
+ * Truncate (towards zero) the value of a variable at rscale decimal digits
+ * after the decimal point.  NOTE: we allow rscale < 0 here, implying
  * truncation before the decimal point.
  */
 static void
index 009cf8abf664ff9d4de1ecbe25e1477520dd56df..f4821984e91d0d2d9466dd74235b1cc841aa86ac 100644 (file)
@@ -37,7 +37,7 @@
  * Portions Copyright (c) 1996-2008, PostgreSQL Global Development Group
  * Portions Copyright (c) 1994, Regents of the University of California
  *
- * $PostgreSQL: pgsql/src/include/catalog/catversion.h,v 1.444 2008/03/23 00:24:19 tgl Exp $
+ * $PostgreSQL: pgsql/src/include/catalog/catversion.h,v 1.445 2008/04/04 18:45:36 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -53,6 +53,6 @@
  */
 
 /*                                                     yyyymmddN */
-#define CATALOG_VERSION_NO     200803222
+#define CATALOG_VERSION_NO     200804041
 
 #endif
index fba36a018d312ec249ce40bff99f13b118924436..2abaeea211efcca1fd6a50cfcac2caa65e896977 100644 (file)
@@ -7,7 +7,7 @@
  * Portions Copyright (c) 1996-2008, PostgreSQL Global Development Group
  * Portions Copyright (c) 1994, Regents of the University of California
  *
- * $PostgreSQL: pgsql/src/include/catalog/pg_proc.h,v 1.486 2008/04/04 16:57:21 momjian Exp $
+ * $PostgreSQL: pgsql/src/include/catalog/pg_proc.h,v 1.487 2008/04/04 18:45:36 tgl Exp $
  *
  * NOTES
  *       The script catalog/genbki.sh reads this file and generates .bki
@@ -1115,7 +1115,7 @@ DESCR("does not match LIKE expression");
 DATA(insert OID =  860 (  bpchar                  PGNSP PGUID 12 1 0 f f t f i 1 1042 "18" _null_ _null_ _null_        char_bpchar - _null_ _null_ ));
 DESCR("convert char to char()");
 
-DATA(insert OID = 861 ( current_database          PGNSP PGUID 12 1 0 f f t f i 0 19 "" _null_ _null_ _null_ current_database - _null_ _null_ ));
+DATA(insert OID = 861 ( current_database          PGNSP PGUID 12 1 0 f f t f s 0 19 "" _null_ _null_ _null_ current_database - _null_ _null_ ));
 DESCR("returns the current database");
 DATA(insert OID = 817 (  current_query        PGNSP PGUID 12 1 0 f f f f v 0 25  "" _null_ _null_ _null_  current_query - _null_ _null_ ));
 DESCR("returns the currently executing query");
@@ -2573,6 +2573,10 @@ DATA(insert OID = 1745 ( float4                                  PGNSP PGUID 12 1 0 f f t f i 1 700 "1700" _n
 DESCR("(internal)");
 DATA(insert OID = 1746 ( float8                                        PGNSP PGUID 12 1 0 f f t f i 1 701 "1700" _null_ _null_ _null_  numeric_float8 - _null_ _null_ ));
 DESCR("(internal)");
+DATA(insert OID = 1973 ( div                                   PGNSP PGUID 12 1 0 f f t f i 2 1700 "1700 1700" _null_ _null_ _null_    numeric_div_trunc - _null_ _null_ ));
+DESCR("trunc(x/y)");
+DATA(insert OID = 1980 ( numeric_div_trunc             PGNSP PGUID 12 1 0 f f t f i 2 1700 "1700 1700" _null_ _null_ _null_    numeric_div_trunc - _null_ _null_ ));
+DESCR("trunc(x/y)");
 DATA(insert OID = 2170 ( width_bucket                  PGNSP PGUID 12 1 0 f f t f i 4 23 "1700 1700 1700 23" _null_ _null_ _null_      width_bucket_numeric - _null_ _null_ ));
 DESCR("bucket number of operand in equidepth histogram");
 
index f80802f9bffdf0875cdb171a7bbc678a201ee5e9..c14fcab72e854ef2242cd13673e0542bb0ffa343 100644 (file)
@@ -7,7 +7,7 @@
  * Portions Copyright (c) 1996-2008, PostgreSQL Global Development Group
  * Portions Copyright (c) 1994, Regents of the University of California
  *
- * $PostgreSQL: pgsql/src/include/utils/builtins.h,v 1.311 2008/04/04 16:57:21 momjian Exp $
+ * $PostgreSQL: pgsql/src/include/utils/builtins.h,v 1.312 2008/04/04 18:45:36 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -845,6 +845,7 @@ extern Datum numeric_add(PG_FUNCTION_ARGS);
 extern Datum numeric_sub(PG_FUNCTION_ARGS);
 extern Datum numeric_mul(PG_FUNCTION_ARGS);
 extern Datum numeric_div(PG_FUNCTION_ARGS);
+extern Datum numeric_div_trunc(PG_FUNCTION_ARGS);
 extern Datum numeric_mod(PG_FUNCTION_ARGS);
 extern Datum numeric_inc(PG_FUNCTION_ARGS);
 extern Datum numeric_smaller(PG_FUNCTION_ARGS);
index 7840990bf7dfc0e2dec02fb6e6f9334a7bad7c69..1c5047455e89b22508aa2af8145bdfac12988f2f 100644 (file)
@@ -1260,3 +1260,84 @@ SELECT * FROM num_input_test;
  -555.50
 (5 rows)
 
+--
+-- Test some corner cases for division
+--
+select 999999999999999999999::numeric/1000000000000000000000;
+        ?column?        
+------------------------
+ 1.00000000000000000000
+(1 row)
+
+select div(999999999999999999999::numeric,1000000000000000000000);
+ div 
+-----
+   0
+(1 row)
+
+select mod(999999999999999999999::numeric,1000000000000000000000);
+          mod          
+-----------------------
+ 999999999999999999999
+(1 row)
+
+select div(-9999999999999999999999::numeric,1000000000000000000000);
+ div 
+-----
+  -9
+(1 row)
+
+select mod(-9999999999999999999999::numeric,1000000000000000000000);
+          mod           
+------------------------
+ -999999999999999999999
+(1 row)
+
+select div(-9999999999999999999999::numeric,1000000000000000000000)*1000000000000000000000 + mod(-9999999999999999999999::numeric,1000000000000000000000);
+        ?column?         
+-------------------------
+ -9999999999999999999999
+(1 row)
+
+select mod (70.0,70) ;
+ mod 
+-----
+ 0.0
+(1 row)
+
+select div (70.0,70) ;
+ div 
+-----
+   1
+(1 row)
+
+select 70.0 / 70 ;
+        ?column?        
+------------------------
+ 1.00000000000000000000
+(1 row)
+
+select 12345678901234567890 % 123;
+ ?column? 
+----------
+       78
+(1 row)
+
+select 12345678901234567890 / 123;
+      ?column?      
+--------------------
+ 100371373180768845
+(1 row)
+
+select div(12345678901234567890, 123);
+        div         
+--------------------
+ 100371373180768844
+(1 row)
+
+select div(12345678901234567890, 123) * 123 + 12345678901234567890 % 123;
+       ?column?       
+----------------------
+ 12345678901234567890
+(1 row)
+
index dc1452f9ef0a7c08770945742a0d890c4c061500..9fd6bba31eeee8fdf7806f5fe5c8bf0bf65406f4 100644 (file)
@@ -805,3 +805,21 @@ INSERT INTO num_input_test(n1) VALUES ('');
 INSERT INTO num_input_test(n1) VALUES (' N aN ');
 
 SELECT * FROM num_input_test;
+
+--
+-- Test some corner cases for division
+--
+
+select 999999999999999999999::numeric/1000000000000000000000;
+select div(999999999999999999999::numeric,1000000000000000000000);
+select mod(999999999999999999999::numeric,1000000000000000000000);
+select div(-9999999999999999999999::numeric,1000000000000000000000);
+select mod(-9999999999999999999999::numeric,1000000000000000000000);
+select div(-9999999999999999999999::numeric,1000000000000000000000)*1000000000000000000000 + mod(-9999999999999999999999::numeric,1000000000000000000000);
+select mod (70.0,70) ;
+select div (70.0,70) ;
+select 70.0 / 70 ;
+select 12345678901234567890 % 123;
+select 12345678901234567890 / 123;
+select div(12345678901234567890, 123);
+select div(12345678901234567890, 123) * 123 + 12345678901234567890 % 123;