Back end (summary)

The back end is probably the easiest part to understand,
since we've already developed a keen understanding of 
quasiquotation.

Produces a C function for each rule
Trades space for speed
Not that much space either: ~ 30 insns / rule at -O2
Pre-calculates a lot
Only needs stat(), system(), and write()
Many other tradeoffs possible

Can perform domain-specific optimizations
Topological sort done at compile time
Missing build rules directly fail when called
Dead build rules not compiled into executable
Common dependency DAG nodes refer to same function
All messages pre-formatted
All string lengths known, for write()