Subj : Re: Serious Problem with precision or if statement
To   : borland.public.cpp.borlandcpp
From : " Bruce Salzman" <bruce@nospam.org>
Date : Wed Aug 13 2003 03:07 pm

> #include <stdio.h>
>
> int main() {
>
>   int Phase = 3;
>
>   if( Phase/3.0 == 1.0 )
>     printf("equal");
>   else
>     printf("not equal");
>
>   return 1;
> }
> -----------------------
>
> Now we would expect "equal" to be printed, but to my
> (everyone's) surprise, "not equal" is printed.
>

IEEE floating point cannot represent certain numbers exactly. 1/3 is one of
them. Testing floating point numbers for equality is not recommended.
Usually one uses an "epsilon test" to determine if the numbers are "close
enough" to be declared equal.

> An even more extraordinary thing occurs when tracing the program.
> The expression ( Phase/3.0 == 1.0 ) evaluates to TRUE. However,
> the else(FALSE) branch is taken printing "not equal"!

How did you determine that the expression evaluates to true? When I step
through the assembly code in the debugger, (after the SAHF instruction) I
see the zero flag is clear (Z = 0) and the carry flag is set (C = 1). The
following JNE is (correcty) taken.

>
> I have tried several modifications to the if statement
> to no avail.
>
> if( (double)Phase/3 == 1.0 )
> if( Phase/3.0 == (double) 1 )
>
    Again, testing for equality is not going to do what you might expect.

> It is worthwhile to note that
> if( Phase/3 == 1 ) does return the correct answer.
>
    Integer division this time. 3 / 3 = 1

> Assigning the value (Phase/3.0) to an intermediate double x
> and then checking if( x == 1.0 ) also works.
>
> If anyone has time or wants some amusement, run my program.
> I am using Borland C++ 5.02 on an Intel Pentium III with Windows
> XP. I get the same results with a AMD Duron with Windows 98.
>
   Not surprised.

> Under LINUX compiled for i586, compiling and running my program
> with gcc returns "equal", the correct answer.

Don't know gcc, so I can't say what it's floating point math is up to.

HTH,
Bruce

.