This is my new favorite exhibit when I want to convince people that floating point is hard.

Table 7 shows that no widespread library covered all edge cases even for a trivial algorithm as (a+b)/2.