Fixed bug #22129: A small double precision number becomes zero

Better checks for underflow/overflow

mysql-test/r/type_float.result

@@ -272,3 +272,10 @@ desc t3;
 Field	Type	Null	Key	Default	Extra
 a	double			0	
 drop table t1,t2,t3;
+select 1e-308, 1.00000001e-300, 100000000e-300;
+1e-308	1.00000001e-300	100000000e-300
+0	1.00000001e-300	1e-292
+select 10e307;
+10e307
+1e+308
+End of 4.1 tests

mysql-test/t/type_float.test

@@ -179,4 +179,13 @@ show warnings;
 desc t3;
 drop table t1,t2,t3;
 
-# End of 4.1 tests
+#
+# Bug #22129: A small double precision number becomes zero
+#
+# check if underflows are detected correctly
+select 1e-308, 1.00000001e-300, 100000000e-300;
+
+# check if overflows are detected correctly
+select 10e307;
+
+--echo End of 4.1 tests

strings/strtod.c

@@ -30,7 +30,8 @@
 #include <m_ctype.h>
 
 #define MAX_DBL_EXP	308
-#define MAX_RESULT_FOR_MAX_EXP 1.79769313486232
+#define MAX_RESULT_FOR_MAX_EXP 1.7976931348623157
+#define MIN_RESULT_FOR_MIN_EXP 2.225073858507202
 static double scaler10[] = {
   1.0, 1e10, 1e20, 1e30, 1e40, 1e50, 1e60, 1e70, 1e80, 1e90
 };
@@ -57,10 +58,11 @@ double my_strtod(const char *str, char **end_ptr, int *error)
 {
   double result= 0.0;
   uint negative= 0, ndigits, dec_digits= 0, neg_exp= 0;
-  int exp= 0, digits_after_dec_point= 0;
+  int exp= 0, digits_after_dec_point= 0, tmp_exp;
   const char *old_str, *end= *end_ptr, *start_of_number;
   char next_char;
   my_bool overflow=0;
+  double scaler= 1.0;
 
   *error= 0;
   if (str >= end)
@@ -91,6 +93,7 @@ double my_strtod(const char *str, char **end_ptr, int *error)
   while ((next_char= *str) >= '0' && next_char <= '9')
   {
     result= result*10.0 + (next_char - '0');
+    scaler= scaler*10.0;
     if (++str == end)
     {
       next_char= 0;                             /* Found end of string */
@@ -114,6 +117,7 @@ double my_strtod(const char *str, char **end_ptr, int *error)
     {
       result= result*10.0 + (next_char - '0');
       digits_after_dec_point++;
+      scaler= scaler*10.0;
       if (++str == end)
       {
         next_char= 0;
@@ -144,39 +148,54 @@ double my_strtod(const char *str, char **end_ptr, int *error)
       } while (str < end && my_isdigit(&my_charset_latin1, *str));
     }
   }
-  if ((exp= (neg_exp ? exp + digits_after_dec_point :
-             exp - digits_after_dec_point)))
+  tmp_exp= neg_exp ? exp + digits_after_dec_point : exp - digits_after_dec_point;
+  if (tmp_exp)
   {
-    double scaler;
-    if (exp < 0)
-    {
-      exp= -exp;
-      neg_exp= 1;                               /* neg_exp was 0 before */
-    }
-    if (exp + ndigits >= MAX_DBL_EXP + 1 && result)
+    int order;
+    /*
+      Check for underflow/overflow.
+      order is such an integer number that f = C * 10 ^ order,
+      where f is the resulting floating point number and 1 <= C < 10.
+      Here we compute the modulus
+    */
+    order= exp + (neg_exp ? -1 : 1) * (ndigits - 1);
+    if (order < 0)
+      order= -order;
+    if (order >= MAX_DBL_EXP && result)
     {
-      /*
-        This is not 100 % as we actually will give an owerflow for
-        17E307 but not for 1.7E308 but lets cut some corners to make life
-        simpler
-      */
-      if (exp + ndigits > MAX_DBL_EXP + 1 ||
-          result >= MAX_RESULT_FOR_MAX_EXP)
+      double c;
+      /* Compute modulus of C (see comment above) */
+      c= result / scaler * 10.0;
+      if (neg_exp)
       {
-        if (neg_exp)
+        if (order > MAX_DBL_EXP || c < MIN_RESULT_FOR_MIN_EXP)
+        {
           result= 0.0;
-        else
+          goto done;
+        }
+      }
+      else
+      {
+        if (order > MAX_DBL_EXP || c > MAX_RESULT_FOR_MAX_EXP)
+        {
           overflow= 1;
-        goto done;
+          goto done;
+        }
       }
     }
-    scaler= 1.0;
+
+    exp= tmp_exp;
+    if (exp < 0)
+    {
+      exp= -exp;
+      neg_exp= 1;                               /* neg_exp was 0 before */
+    }
     while (exp >= 100)
     {
-      scaler*= 1.0e100;
+      result= neg_exp ? result/1.0e100 : result*1.0e100;
       exp-= 100;
     }
-    scaler*= scaler10[exp/10]*scaler1[exp%10];
+    scaler= scaler10[exp/10]*scaler1[exp%10];
     if (neg_exp)
       result/= scaler;
     else