GCC and Clang have builtins for byteswapping that we can use to avoid having to write assembly for each architecture. GCC 8.3 does recognize the C implementation of SDL_Swap32/SDL_Swap64 as a byteswap and generate optimized code, but it doesn't quite generate optimal code for SDL_Swap16.

On ARMv7 with GCC 8.3, the SDL_Swap16 C implementation generates:
lsr	r0, r3, #8
orr	r0, r0, r3, lsl #8

but using __builtin_bswap16, we get the optimal implementation using rev16:
rev16	r0, r0

Rather than depending on the optimizer to figure out the 32 and 64 bit byteswaps, we can just use the builtins for all of them to guarantee it.