Merge pull request #302 from konrad-kruczynski/fast_double_to_from_int64

Double <-> Int64 conversion is much faster now.

Instead of using GetBytes and converting bytes to long, which involves allocation and deallocation of small temporary array, the simple cpu register load is in fact done now. This results in a much faster operation.

With this sample program:

var stopwatch = Stopwatch.StartNew();
var sum = 0L;
for(var i = 0; i < 100000000; i++)
{
    sum += BitConverter.DoubleToInt64Bits(99.9);
}
stopwatch.Stop();
Console.WriteLine(stopwatch.Elapsed);
my results are:

00:00:00.6614282 with my patch;
00:00:19.8135516 without my patch and Boehm GC;
00:00:10.2296766 without my patch and SGEN GC.
I'm doing a lot of this operations in my serialization framework, so it is quite essential to be fast ;) The performance improvement is due to no-allocation nature of the patch vs the old code, which allocated a temporary array (difference between Boehm and SGEN is explained the same way).

mcs/class/corlib/System/BitConverter.cs

@@ -60,14 +60,14 @@ namespace System
 			return b [2] == 0xf0;
 		}
 
-		public static long DoubleToInt64Bits (double value)
+		public unsafe static long DoubleToInt64Bits (double value)
 		{
-			return ToInt64 (GetBytes (value), 0);
+			return *(long *) &value;
 		}
 
-		public static double Int64BitsToDouble (long value)
+		public unsafe static double Int64BitsToDouble (long value)
 		{
-			return ToDouble (GetBytes (value), 0);
+			return *(double *) &value;
 		}
 
 		internal static double InternalInt64BitsToDouble (long value)