Well, you’ve correctly typified the situation: C/C++ have no way of doing a full signed int/unsigned int comparison with a single compare.
I would be surprised if promotion to int64 was faster than doing two comparisons. In my experience, compilers are quite good at realizing that a subexpression like that is pure (has no side effects) and thus that there’s no need for a second branch. (You can also explicitly opt-out of short circuiting using bitwise-or: (x < 0) | (x < y).) In contrast, my experience is that compilers tend NOT to do much special-case optimization on integers greater than the native word size, so (int64)x < (int64)y is quite likely to actually do a full int comparison.
Bottom line, there’s no incantation which is guaranteed to produce the best possible machine code on any processor, but for the most common compilers on the most common processors, I would guess that the two-comparison form would be no slower than the promotion-to-int64 form.
EDIT: Some mucking about on Godbolt confirms that on ARM32, GCC puts way too much machinery into the int64 approach. VC does the same on x86. With x64, though, the int64 approach is actually one instruction shorter (since the promotion and 64-bit comparison are trivial). The pipelining might make the actual performance go either way, though. https://godbolt.org/g/wyG4yC