Since it keeps coming up, I'll say again, never depend on floating point
equality checks (or transitivity). As an example, I just ran into a
problem with domino caused by C++ code which looked like:
float min=big_number;
loop over things:
if (thing.get_value() < min) min= thing.get_value()
loop over things again:
if (thing.get_value() == min) return thing;
Code like this can fail if:
- the type of thing.get_value() is not exactly the same as the type of min
- or thing.get_value() performs some computations and the compiler
decides to inline them
- or for whatever other reason, thing.get_value() is never written to
main memory
The reason is that on intel chips, floating point values are stored as
32 bit, 64 bit or 80 bit values depending on what the compiler feels
like doing. They are only truncated into the specified 32 bit (float) or
64 bit (double) values when stored as the appropriate value type in
cache or memory or when copied to a different type of register. If one
copy of a value is truncated to a 32 bits or 64 bits by being moved,
then it is no longer equal to its old value.
In this case the fix is trivial: just combine the two loops and keep
track of the min element as you go along. Other times you will have to
jump through more hoops.