It is also a good choice to use Intel 80251 that has also a second source from Temic.

251 has two modes of operation: - binary mode is compatible to standard 51 but running much faster (at least 4 times at the same clock), and source mode - has extended features like 16-bit and also 32-bit instructions. 1 kByte of SRAM is available in the CPU. 256 Kbyte linear address space, etc. An application written in C can be after recompilation with C251 up to 16 times faster than the one compiled with C51 and running on standard 51. If you like the idea to group four registers like R0..R3 to one long register DR0 and increment it in one instruction INC DR0,#1, it is the rigth choice. By using the paged bus mode, you can make it run even faster at the same clock.

But 251 is like XA (I'm not sure about the XA) an 8 bit machine seeing from the outside (has an 8 bit data bus). If you need a real 16 bit machine, take a look at Infineon 166 or 167 beasts, very-very powerful, but not 51 compatible.