[Issue 13474] New: 32 bit DMD optimizer FP arithmetic bug

[via Digitalmars-d-bugs]

https://issues.dlang.org/show_bug.cgi?id=13474

          Issue ID: 13474
           Summary: 32 bit DMD optimizer FP arithmetic bug
           Product: D
           Version: D2
          Hardware: x86
                OS: All
            Status: NEW
          Severity: critical
          Priority: P1
         Component: DMD
          Assignee: nobody at puremagic.com
          Reporter: ilyayaroshenko at gmail.com

KBN summation return wrong result when compiled with 
dmd -m32 -O

With -m64 dmd works correctly
LDC works correctly both m32 and m64

+++++++++++++++++++++++++++++++++

/**
Kahan-Babuška-Neumaier summation algorithm
+++++++++
s := x[1]
c := 0
FOR i := 2 TO n DO
    t := s + x[i]
    IF ABS(s) >= ABS(x[i]) THEN
        c := c + ((s-t)+x[i])
    ELSE
        c := c + ((x[i]-t)+s)
    END IF
    s := t
END DO
s := s + c
+++++++++
*/
F sumKBN(Range, F = Unqual!(ForeachType!Range))(Range r, F s = 0.0) 
{
    F c = 0.0;
        foreach(F x; r)
        {
            F t = s + x;
            if(s.fabs >= x.fabs)
            {
                F y = s-t;
                c += y+x;
            }
            else
            {
                F y = x-t;
                c += y+s;
            }
            s = t;
        }
    return s + c;
}

unittest 
{
    import std.algorithm : map;
    auto ar = [1, 1e100, 1, -1e100].map!(a => a*10000);
    double r = 20000;
    assert(r == ar.sumKBN);    // fails with dmd -m32 -O
}

+++++++++++++++++++++++++++++++++

Git Reference: https://github.com/D-Programming-Language/phobos/pull/2513

--