RFA: pervasive SSE codegen inefficiency

[Paolo Bonzini]

> So basically, pick a problem you see, and fix it.

The RTL infrastructure is exceptionally good at doing some things, and 
exceptionally bad at doing some others.  Sometimes, take into account 
the coding style and it is good and bad at the same time.  :-(  CSE, 
flow, etc. come to mind.

All you can do is try to improve on the bad items and be careful not to 
cripple the good ones.

FWIW, it's the same thing we have been doing with fwprop.  After 
removing code from pre-existing RTL passes (which you don't need because 
the allocator is already crippled...), benchmark, fix regressions, and 
iterate from step 2.

Paolo