• Saaa (13/13) Jul 25 2009 I thought my previous precision question was so retarded it didn't need ...
• downs (25/30) Jul 25 2009 Here's an incredibly simple hack.
• Saaa (19/19) Jul 25 2009 Ok it finally hit me :(

"Saaa" <empty needmail.com> writes:

I thought my previous precision question was so retarded it didn't need any 
explanation; anybody here could easily tell me my fault in reasoning.

Here is my question again but this time with examples.

Is there some formatting which lets me print out a floating point in full 
precision?

Observation: 6 digits is not full precision:

float f=3.4028234; // 0x7f7f ffff
writefln("%.8g",f); // prints 3.4028234
writefln(f.dig); // prints 6
writefln(3.4028234f.dig); // prints 6

Shouldn't f.dig print out 8?
.dig =  number of decimal digits of precision

I still think I don't get something, somewhere. 

downs <default_357-line yahoo.de> writes:

Saaa wrote:
 I thought my previous precision question was so retarded it didn't need any
 explanation; anybody here could easily tell me my fault in reasoning.

 Here is my question again but this time with examples.

 Is there some formatting which lets me print out a floating point in full
 precision?

Here's an incredibly simple hack.

import std.stdio, std.string;

string ftoaFull(float f) {
  if (f < 0) return "-" ~ ftoaFull(-f);
  auto start = f;
  auto res = toString(cast(int) f) ~ ".";
  while (true) {
    f -= cast(int) f;
    f *= 10;
    res ~= "0123456789"[cast(int) f];
    // The critical step
    if (cast(float) res.atof() == start) return res;
  }
}

void main() {
  writefln(ftoaFull(0f));
  writefln(ftoaFull(-5.234));
  writefln(ftoaFull(1f / 3f)); // problematic
}

Output:

0.0
-5.234
0.33333334

And there you have it.

"Saaa" <empty needmail.com> writes:

Ok it finally hit me :(
The 24 fraction bits in a float aren't used like an integer (2^24 giving max 
7 decimal digits precision)
they are used by halving the value of the previous bit: 1, 0.5, 0.25 ... 
like fractions !!!
this of course has a much wider decimal range (something like 20 or so) 
because not all numbers are
represented.

Thus, the formatting %.100g and downs code both just give the correct answer 
.

Now, there are 2^24 different floats being represented (ignoring exponent) 
so the question for me now is:
What is the minimal (decimal) representation of a float for which the 
following holds:

float f; //for all possible floats
string s = format(f);
float f2 = to!(float)(s);
assert(f == f2);

Does your second code hold, downs?