multiplies by powers of two

[Matthew Fluet]

> I did notice that there is a lot of funny arrangement of code which makes
> the number of false positives not so surprising.  It was as if the generated
> code was intermingling 2 unrelated tasks.

After register allocation, this is more true.  Trying to hold values in
register, delaying writes to memory, where aliasing kicks in can sometimes
produce strange orderings.  But I'm doing a lot of peephole optimizations
before register allocation, just treating operands as abstract memory
locations.  In the initial translation, this should look a lot like the
machine IL, so that was the source of my surprise.