The specializer accepts a representation of the program targeted to the abstract Mill operation set, typically produced by a compiler. The representation has no knowledge of which Mill family member will be the execution target. The specializer transforms the abstract representation into a concrete representation which reflects the capabilities of a particular target member.
In operation, the specializer first substitutes emulation code for any abstract operations that are not present in native hardware on the target. It then places the program operations into instructions, ordered such that produced results are available when needed by consuming operations. The placing is constrained because each instruction slot is only able to execute a subset of the operation set. Lastly, the specializer determines the belt position numbers that let consuming operations reference their arguments. If necessary (due to the limited size of the belt), the specializer will introduce spill and fill operations to temporarily save operands to the scratchpad.
The specialized code is cached in the ELF code files or load modules, where it can be found by subsequent linking or loading steps.