Here's an article on the topic: https://cryptoservices.github.io/fde/2018/11/30/undefined-be...
Most compilers will do something like "treat the bits like they represent a float, even though they mean something else when they're treated as an int."
But the language spec says the compiler is allowed to send an email to Bjarne Stroustrup with your home address so he can come over and personally set your computer on fire.
More seriously, the typical response to undefined behavior is for the compiler to optimize out whatever code may trigger it. It is great for performance and one of the reasons C and C++ often top the charts, but it may lead to weird bugs that depend on the optimization level.
For the union case, the compiler could completely remove the code that reads the float: because it is UB, it is something you shouldn't do, so it considers that you will never do it, so it can just remove the code. In the end your conversion function may do nothing at all (very optimized!). In practice, because it is so common, even though it is UB, it will usually do what you want it to do.
- There’s at least an ugly library to do it in C++
- There might be support baked directly into the language
- Or you could do it in Lisp, but that would be too easy