I saw an article describing something similar while working on the automatic generation of multimedia instructions (MMX, SSE and the like). While most works tend to focus on generating optimized binary (or source code using special constructs to direct the use of multimedia opcodes), one group was studying instead the possibility of applying the same kind of optimization directly on the binaries, thus using an approach similar to one you picture.

Unfortunately, i can't find the references of the article i read... I'll post it if i find it later... In any case, it should be referenced on citeseer (hxxp://www.citeseer.com), and referenced through keywords like 'SIMD', 'vectorization' or 'multimedia instruction set', so you may want to try to look there if you are interested.