I strongly agree with RalfJung and Nikita’s earlier response about this option: Making it “implementation defined” (?) within a single compiler seems to be the worst of all options.
Mostly I don’t see, how does this help at all catering to the people and frontends that want deterministic behaviour. For example, how does this help implement Clang’s __builtin_elementwise_minnum?
I think in fact this will just delay this inevitable discussion into the future. I don’t see any reason to wait. What is going to change significantly? I don’t see us switching away from compiling to SSE2 by default for decades to come so this will remain as expensive as it is now. Much 32 bit software and 32 bit GCC is even still compiling to x87 by default today right as we speak.
In particular in the world of maybe, I would also be opposed to this x86 backend change. Why pay performance for no guaranteed return?