Some really good reasons: https://stackoverflow.com/a/22135559/740553

It basically assumes all maths is finite and defined, then ignores how floating point arithmetic actually works, optimizing based purely on "what the operations suggest should work if we wrote them on paper" (alongside using approximations of certain functions that are super fast, while also being guaranteed inaccurate)

It reorders instructions in ways that are mathematically but not computationally equivalent (as is the nature of FP). This also breaks IEEE compliance.