I don't think there's going to be any performance difference at all compared to the basic case, which actually is another reason why this is so good. Ultimately, all these struct just hold a double, which means that in memory they're exactly a double and nothing more, there's no type tag or anything like that. And all the code is static so there is no vtable, just the double value. The compiler does the rest, not the runtime.
It can make a difference due to weird ABIs. For example, in the ARM procedure call standard, 64-bit values can be returned in r0 and r1, but 64-bit structures can't (no matter whether they contain two 32-bit values or just one 64-bit value). However, I'd call that fairly niche - it's hard to imagine an application that would notice a significant performance difference.
Scientific computing will hit this sort of problem just because the kernel is so large it actually needs to be broken into sub-procedures for icache reasons. But also, the penalty is large, because there is a full store/load cycle involved, which will clear caches and increase memory bandwidth requirements.
(Still, I'd love to see newtypes in C++)