One of the things I've always wanted is autovectorisation by the CPU - imagine if there was a REP ADDSB/W/D/Q instruction (and naturally, repeated variants of the other ALU operations.) It could make use of the full memory bandwidth of any processor by reading and summing entire cache lines the fastest way the current microarchitecture can, and it'd also be future-proof in that future models may make this faster if they e.g. introduce a wider memory bus. Before the various versions of SSE there was MMX, and now AVX, so the fastest way to do something like sum bytes in memory changes with each processor model; but with autovectorisation in hardware, programs wouldn't need to be recompiled to take advantage of things like wider buses.

Of course, the reason why "string ALU instructions" haven't been present may just be because most programs wouldn't need them and only some would receive a huge performance boost, but then again, the same could be said for the AES extensions and various other special-purpose instructions like CRC32...