]> granicus.if.org Git - postgresql/commitdiff
Fix loss of fractional digits for large values in cash_numeric().
authorTom Lane <tgl@sss.pgh.pa.us>
Fri, 26 Jul 2019 15:59:00 +0000 (11:59 -0400)
committerTom Lane <tgl@sss.pgh.pa.us>
Fri, 26 Jul 2019 15:59:00 +0000 (11:59 -0400)
Money values exceeding about 18 digits (depending on lc_monetary)
could be inaccurately converted to numeric, due to select_div_scale()
deciding it didn't need to compute any fractional digits.  Force
its hand by setting the dscale of one division input to equal the
number of fractional digits we need.

In passing, rearrange the logic to not do useless work in locales
where money values are considered integral.

Per bug #15925 from Slawomir Chodnicki.  Back-patch to all supported
branches.

Discussion: https://postgr.es/m/15925-da9953e2674bb5c8@postgresql.org

src/backend/utils/adt/cash.c
src/test/regress/expected/money.out
src/test/regress/sql/money.sql

index c92e9d5046acba7296178f4c30d7dec2cba95d98..6515fc8ec6955dc8a9f001c2928ef9231b3c347c 100644 (file)
@@ -1032,13 +1032,8 @@ Datum
 cash_numeric(PG_FUNCTION_ARGS)
 {
        Cash            money = PG_GETARG_CASH(0);
-       Numeric         result;
+       Datum           result;
        int                     fpoint;
-       int64           scale;
-       int                     i;
-       Datum           amount;
-       Datum           numeric_scale;
-       Datum           quotient;
        struct lconv *lconvert = PGLC_localeconv();
 
        /* see comments about frac_digits in cash_in() */
@@ -1046,22 +1041,45 @@ cash_numeric(PG_FUNCTION_ARGS)
        if (fpoint < 0 || fpoint > 10)
                fpoint = 2;
 
-       /* compute required scale factor */
-       scale = 1;
-       for (i = 0; i < fpoint; i++)
-               scale *= 10;
+       /* convert the integral money value to numeric */
+       result = DirectFunctionCall1(int8_numeric, Int64GetDatum(money));
 
-       /* form the result as money / scale */
-       amount = DirectFunctionCall1(int8_numeric, Int64GetDatum(money));
-       numeric_scale = DirectFunctionCall1(int8_numeric, Int64GetDatum(scale));
-       quotient = DirectFunctionCall2(numeric_div, amount, numeric_scale);
+       /* scale appropriately, if needed */
+       if (fpoint > 0)
+       {
+               int64           scale;
+               int                     i;
+               Datum           numeric_scale;
+               Datum           quotient;
+
+               /* compute required scale factor */
+               scale = 1;
+               for (i = 0; i < fpoint; i++)
+                       scale *= 10;
+               numeric_scale = DirectFunctionCall1(int8_numeric,
+                                                                                       Int64GetDatum(scale));
 
-       /* forcibly round to exactly the intended number of digits */
-       result = DatumGetNumeric(DirectFunctionCall2(numeric_round,
-                                                                                                quotient,
-                                                                                                Int32GetDatum(fpoint)));
+               /*
+                * Given integral inputs approaching INT64_MAX, select_div_scale()
+                * might choose a result scale of zero, causing loss of fractional
+                * digits in the quotient.  We can ensure an exact result by setting
+                * the dscale of either input to be at least as large as the desired
+                * result scale.  numeric_round() will do that for us.
+                */
+               numeric_scale = DirectFunctionCall2(numeric_round,
+                                                                                       numeric_scale,
+                                                                                       Int32GetDatum(fpoint));
+
+               /* Now we can safely divide ... */
+               quotient = DirectFunctionCall2(numeric_div, result, numeric_scale);
+
+               /* ... and forcibly round to exactly the intended number of digits */
+               result = DirectFunctionCall2(numeric_round,
+                                                                        quotient,
+                                                                        Int32GetDatum(fpoint));
+       }
 
-       PG_RETURN_NUMERIC(result);
+       PG_RETURN_DATUM(result);
 }
 
 /* numeric_cash()
index ab86595fc0219343da3374c11303e7d2d435d7a8..fc71a72fed319b76848db757efb4092a54cb3a8d 100644 (file)
@@ -1,6 +1,8 @@
 --
 -- MONEY
 --
+-- Note that we assume lc_monetary has been set to C.
+--
 CREATE TABLE money_data (m money);
 INSERT INTO money_data VALUES ('123');
 SELECT * FROM money_data;
@@ -476,7 +478,7 @@ SELECT (-12345678901234567)::numeric::money;
  -$12,345,678,901,234,567.00
 (1 row)
 
--- Cast from money
+-- Cast from money to numeric
 SELECT '12345678901234567'::money::numeric;
        numeric        
 ----------------------
@@ -489,3 +491,15 @@ SELECT '-12345678901234567'::money::numeric;
  -12345678901234567.00
 (1 row)
 
+SELECT '92233720368547758.07'::money::numeric;
+       numeric        
+----------------------
+ 92233720368547758.07
+(1 row)
+
+SELECT '-92233720368547758.08'::money::numeric;
+        numeric        
+-----------------------
+ -92233720368547758.08
+(1 row)
+
index 37b9ecce1fc4fea57bfc79d41ef99e7d80892d9a..5e746286c9079e0444c509d81e02e0359ce3e83a 100644 (file)
@@ -1,6 +1,8 @@
 --
 -- MONEY
 --
+-- Note that we assume lc_monetary has been set to C.
+--
 
 CREATE TABLE money_data (m money);
 
@@ -122,6 +124,8 @@ SELECT (-1234567890)::int4::money;
 SELECT (-12345678901234567)::int8::money;
 SELECT (-12345678901234567)::numeric::money;
 
--- Cast from money
+-- Cast from money to numeric
 SELECT '12345678901234567'::money::numeric;
 SELECT '-12345678901234567'::money::numeric;
+SELECT '92233720368547758.07'::money::numeric;
+SELECT '-92233720368547758.08'::money::numeric;