This is essentially what RISC-V does with its "Compressed" instruction set, except without the dictionary switching. They pulled a bunch of statistics over real-world machine code, ran it through compression, then reverse-engineered that compression to make something a bit more sensible to a compiler writer. I think this will work out vastly better than the haphazard patching of e.g. Thumb on ARM.
http://www.icsi.berkeley.edu/pubs/arch/EECS-2011-63.pdf