MDEV-37788 converted the uintNkorr() and intNstore() macros to use
memcpy() and byte swap intrinsics, but the floating point and short/long
conversion macros in big_endian.h and myisampack.h still accessed data
one byte at a time, and those in little_endian.h were a mix of both.
Compilers may emit slow byte-at-a-time loads and stores for such code.
Reimplement the macros with memcpy(), and with MY_BSWAP32/MY_BSWAP64
where the byte order differs from the host. big_endian.h and
little_endian.h no longer differ, so move the macros to my_byteorder.h
and remove the two headers, which are no longer installed. Implement
mi_float4store() etc. in myisampack.h with the same helper macros.
Remove the unused ulongget() macro, and the code for the mixed-endian
floating point layout (a little-endian CPU with big-endian floating
point word order) from the macros, change_double_for_sort() and dtoa.c.
It only applied to the obsolete ARM FPA format.
Reimplement mach_double_read(), mach_double_write(), mach_float_read()
and mach_float_write() in InnoDB with float8get(), float8store(),
float4get() and float4store(), instead of copying bytes in a loop.
The stored formats do not change. The unit test byte_order-t now checks
the byte layout of the floating point macros and the sign extension of
the native byte order macros, so no MTR test is added.
d1034cb490