"extended" real type precision

[Pi]

According to this page http://wiki.lazarus.freepascal.org/Variables_and_Data_Types
the "extended" real type has 19-20 significant digits, but in my function it seems to be less:

eg.
procedure TForm1.Button3Click(Sender: TObject);
var the_nr: extended;
begin
  the_nr:= (1000/1001);
  showmessage(floattostr(the_nr));
end;   

running the above function shows "0.99900099900099"
if "extended" has 19-20 significant digits, shouldn't there be at least another 9 on the end?
eg. 0.999000999000999 or 0.999000999000999001 ?

Is it something to do with using floattostr - does that only work up to 15-16 significant digits because it can only use up to "double" instead of "extended"?  Should I be using something other than floattostr to convert an extended to a string keeping all 19-20 significant digits (is there a function that does or would I need to write one?)?

edit: just found it it is the floattostr function that is causing the reduced precision as I've now half-written a (not very good) function to convert an extended to string.  So I wondered if there is a better/standard way to convert an extended to string without losing precision?

[photor]

The Extended type is suppressed on Win64 platforms.

[LV]

Free Pascal suppresses the Extended floating-point type on the Win64 platform because Microsoft’s 64-bit ABI (Application Binary Interface) for Windows does not support the 80-bit floating-point type (which Extended represents on 32-bit x86 platforms).

Code: Pascal  [Select][+][-]

1. procedure TForm1.Button1Click(Sender: TObject);
2. var
3.   a, b: Extended;
4.   s: string;
5. begin
6.   a := (1000/1001);
7.   b := 0.123456789012345678;
8.   s := FloatToStrF(a, ffGeneral, 18, 0) +
9.        #13#10 +
10.        FloatToStrF(b, ffGeneral, 18, 0);
11.   memo1.Clear;
12.   memo1.Lines.Add(s);
13. end;
14.  

output win32:

Code: Text  [Select][+][-]

1. 0.999000999000999
2. 0.12345678901234568
3.  

output win64:

Code: Text  [Select][+][-]

1. 0.999000999000999
2. 0.123456789012346
3.  

Workarounds: If you need more precision, you can use libraries (maybe GNU MPFR Library, AlgLib, Lmath, ...).