i386: push <memory> ?

[Jeffrey A Law]

  In message < qanyb563z7w.fsf@frodo.colorado.edu >you write:
  > Note that the K6 optimization guide recommends using 'push mem'
  > instead of 'load reg; push reg' to reduce register pressure, since the
  > 'push mem' will be translated to a 'load reg, store reg' sequence
  > internally anyway. Seems like the Pentium Pro would be similiar.
Interesting.

This is _exactly_ the kind of thing that I would want to do with a 
more aggressive instruction splitter after reload.

The basic idea is after reload it is possible to know exactly what
registers are free at every insn.  Thus if you know an appropriate
register is free, then you go ahead and use the longer, but faster
sequence which requires an extra register.

There's numerous examples on various chips where this would be
profitable.


Jeff