As we know, the practice "not fully IEEE 754 conformant, but faster" exists. This is what compilers optionally do (options related to floating-point model
floating-point model