I've been digging in 'sort' package of standard Go library, studying concrete implementations of used sorting algorithms and one thing seems really strange to me - definition of Float64 Less() function:
func (x Float64Slice) Less(i, j int) bool { return x[i] < x[j] || (isNaN(x[i]) && !isNaN(x[j])) }
I quite confident, that floating-point values (in IEEE 754 semantics, at least) MUST NOT be compared directly. Why the comparison implemented this way is in standard Go lib?
I think you misinterpreted that floating-point values must not be compared directly.
This is true in this context: if you take two floating-point numbers, run different calculations on both which mathematically give the same results, then there are lot of chances that the resulting numbers are not equal, so relying on floating-point computation to check mathematical equality is a very, very bad idea.
However, it is perfectly sensed to sort a slice of floating-point numbers: they are comparable, they have an order. This can be either a good or a bad idea, depending on what is your usecase.
If your usecase is just to sort input floating-point data, then there is no problem.
If your usecase is related to a mathematical problem in a dense interval, this is probably a bad idea.
Anyways, just remember that it is generally a bad idea to check for floating-point equality, but comparing two floating-point numbers may have sense, otherwise you could never take any decisions by comparing them, and you often have to do that.