- Flag idempotent instructions.

- Exploit idempotent insns in scheduling, allowing them to be executed
  inside "too many" loops.

- Be more exact in what inames to duplicate in CSE pre-computes.

- check_double_use_of_hw_dimensions
- barrier insertion
- make scheduling of 16^3 mxm work

Implement user control over slab decomposition.
Use CSEs instead of register prefetch.
Remove CodeGenerationState.
Carry CSE replacement info in CodeGenerationMapper.
Carry CCM in ExecutionDomain.
Allow user to specify (unchecked) assumptions on parameters,
  to avoid unnecessary conditionals.
Parse desired operation from strings, extract CSEs.
Support for equality constraints in for loops.