- Only iterating over the graph once should be slightly faster
- We don't need to insert the middle of register allocation logic but
  can perform it once before.
- We can gather statistics on the prepared graph before spilling/regalloc has
  happened.

(printf now works with floating point)

When remove_bads replaces a block, it destroys the additional information
attached to the block (e.g. dominance). Commit 27215754 asserted these
properties in error.

This is a common pattern for x % c.

Also added Mux(a >= 0, 0, 0xFFFFFFFF) => ~a >>s 31.