Copying SI regs as HI subregs

[Michael Hayes]

D. Towner writes:

 > Is it more efficient to use RTL to generate the two subreg moves
 > explicitly?  Does this present opportunities for optimisation if gcc knows
 > that two such instructions are used, rather than a single atomic movsi
 > instruction? How can I go about actually generating the RTL if this is the
 > case?

At the moment you are probably better off by not exposing this too
early.  You should use a splitter pattern predicated on
reload_completed.  However, you will miss some low level insn
combination opportunities which hopefully the proposed RTL lowering
scheme should avoid.

Michael.