Re: Floating point differences at compile-time

[Don]

bearophile wrote:
I don't understand where the result differences come from in this code:
import std.math: sqrt, PI;
import std.stdio: writefln;

void main() {
  const double x1 = 3.0 * PI / 16.0;
  writefln("%.17f", sqrt(1.0 / x1));

  double x2 = 3.0 * PI / 16.0;
  writefln("%.17f", sqrt(1.0 / x2));

  real x3 = 3.0 * PI / 16.0;
  writefln("%.17f", sqrt(1.0 / x3));

  real x4 = 3.0L * PI / 16.0L;
  writefln("%.17f", sqrt(1.0L / x4));
}

This is caused by compile-time constant folding being done at higher 
precision than at run time; and compile-time doubles and floats are 
actually stored at real precision.

If you change 'const double' into 'enum double' for D2, the values for 
D1 and D2 will match.

But I agree, we ought to do better than this.




Output with various D compilers:

DMD1:
1.30294003174111994
1.30294003174111972
1.30294003174111979
1.30294003174111979

DMD2:
1.30294003174111972
1.30294003174111972
1.30294003174111979
1.30294003174111979

LDC:
1.30294003174111994
1.30294003174111994
1.30294003174111972
1.30294003174111972

I'd like the compiler(s) to give more deterministic results here.

Bye,
bearophile