Please give a better example, they have exact representation in IEEE

They do, but long chain of computations can lead to weird results like 4.00000000561 instead of 4.
Working with floating-point numbers u shouldn't do things like
if (a == b) ...
The code should be replaced with something like
double epsilon = 1e-9;
if (a > b - epsilon && a < b + epsilon) ...
If u validate transactions/blocks the code
if (inputs == outputs) ...
can fail, u r supposed to write it as
if (inputs > outputs - epsilon) ...