Thread overview
Floating-Point arithmetic in dlang - Difference to other languages
6 days ago
Jan Hönig
6 days ago
mipri
6 days ago
Basile B.
5 days ago
H. S. Teoh
6 days ago
Today i have stumbled on Hacker News into: https://0.30000000000000004.com/

I am learning D, that's why i have to ask.

Why does

    writefln("%.17f", .1+.2);

not evaluate into: 0.30000000000000004, like C++
but rather to: 0.29999999999999999

Many other languages evaluate to 0.30000000000000004 as well.
6 days ago
On Tuesday, 3 December 2019 at 09:22:49 UTC, Jan Hönig wrote:
> Today i have stumbled on Hacker News into: https://0.30000000000000004.com/
>
> I am learning D, that's why i have to ask.
>
> Why does
>
>     writefln("%.17f", .1+.2);
>
> not evaluate into: 0.30000000000000004, like C++
> but rather to: 0.29999999999999999
>
> Many other languages evaluate to 0.30000000000000004 as well.

You can get this result in D as well:

  $ rdmd --eval 'double a = .1, b = .2; writefln("%.17f", a+b)'
  0.30000000000000004

https://dlang.org/spec/float.html explains it: real (80bit,
'long double' in C) floats are used in your first calculation,
and doubles are used in my revised example.

Most other languages give you the double result for very
reasonable historical reasons, described here:
https://retrocomputing.stackexchange.com/questions/9751/did-any-compiler-fully-use-intel-x87-80-bit-floating-point/9760#9760

D's behavior is a minor 'miss': https://nwcpp.org/Oct-2019.html

... I wanted to include a C example that gives the D result,
but it seems to be trickier to force 80 bit intermediates.
6 days ago
On Tuesday, 3 December 2019 at 09:22:49 UTC, Jan Hönig wrote:
> not evaluate into: 0.30000000000000004, like C++
> but rather to: 0.29999999999999999

You get the same in C++ with:

#include <cstdio>
int main()
{
    printf("%.17f",double(0.1L + 0.2L));
}
6 days ago
On Tuesday, 3 December 2019 at 09:52:18 UTC, mipri wrote:
> Most other languages give you the double result for very
> reasonable historical reasons

Not only historical, it is also for numerical reasons. You can get very unpredictable results if you do compares and compiletime evalution is different from runtime evaluation.


6 days ago
On Tuesday, 3 December 2019 at 09:22:49 UTC, Jan Hönig wrote:
> Today i have stumbled on Hacker News into: https://0.30000000000000004.com/
>
> I am learning D, that's why i have to ask.
>
> Why does
>
>     writefln("%.17f", .1+.2);
>
> not evaluate into: 0.30000000000000004, like C++
> but rather to: 0.29999999999999999
>
> Many other languages evaluate to 0.30000000000000004 as well.

D lang could have a very good sprintf replacement.
5 days ago
On Tue, Dec 03, 2019 at 09:22:49AM +0000, Jan Hönig via Digitalmars-d-learn wrote:
> Today i have stumbled on Hacker News into: https://0.30000000000000004.com/
> 
> I am learning D, that's why i have to ask.
> 
> Why does
> 
>     writefln("%.17f", .1+.2);
> 
> not evaluate into: 0.30000000000000004, like C++
> but rather to: 0.29999999999999999
> 
> Many other languages evaluate to 0.30000000000000004 as well.

In short, because they use 64-bit floats for intermediate values
(`double`), whereas D defaults to 80-bit intermediates (`real`).

The use of a more precise intermediate value is also indicated by the
fact that 0.29999999999999999 is closer to the real answer (error of
1e-17) than 0.30000000000000004 (error of 4e-17).

If you explicitly specify to use `double`, then you should get the same answer as C++.

	double a=.1, b=.2;
	writefln("%.17f", a+b);


T

-- 
MSDOS = MicroSoft's Denial Of Service