Specifically, I think the original code will always do what you expect as long as sizeof(long) == sizeof(float), and the alignment is the same. In reality no compiler is gonna do weird stuff unless your hardware target is weird.
If you use a union, your code won't read or write out of bounds, since it will save space for the larger type. Ideally you would also specify _Float32_t and int32_t for some pseudoportability (at least before endianness gets involved), although I'm aware this was not available in 1999.
compiler engineers have been disappointing our 'surely no compiler would ever be so perverse as to do x' expectations for 25 years now, and it doesn't seem that they're likely to stop soon
