integer moves of floating point values

[Matthew Fluet]

> Isn't  it more efficient, in terms of number of registers used, to use only a
> single register?  I.e., instead of loading the high and low parts of the real
> into  different  32-bit  registers and then storing them, just load the high,
> store the high and then load the low and store the  low.   

That should be what the code is doing.  Did you see an example that I sent
where this isn't the case?

> I would think that making all of this invisible except
> to the actual code generator would be the way to go, and  then  the  aliasing
> wouldn't  be  a  problem, but I guess that then certain possible (very local)
> optimizations wouldn't be possible.

This is done at the very last stage of register allocation; everything up
to that point is doing operations in terms of abstract memory locations.
But, the aliasing problem doesn't go away.  When moving a float from SD(8)
to SD(12) I need to do the high word first, whereas moving the float in
the opposite direction requires doing the low word first.

Ah... if you thought that I first decided to implement the mem-mem f.p.
move using a gpr, did optimizations, and then did register allocation,
that is not the case.  During register allocation, I see a f.p. move and
make some decision how it's going to be implemented and immediately output
the allocated assembly.