quicopt IR — a model as plain data

The form a model takes between the interface that wrote it and the service that solves it: variables, expressions and constraints, with no trace of how they were authored. model() and friends build one of these on the way out; encode() turns it into the bytes the service reads. The shape is the service's published contract — these constructors track it, they never fork it.

Usage

ir_const(value)

ir_param(name, index = list())

ir_var(name, index = list())

ir_apply(op, args)

ir_reduce(op, idx, over, body, cond = NULL)

ir_source_ref(name)

ir_set_ref(name, args = list())

Arguments

value

A numeric constant.

name

The referenced name.

index

An index tuple (a list of integers and strings; list() for a scalar).

op

A catalog operator key, e.g. "+".

args

Enclosing bound indices the set is applied to (list() for a flat set).

idx

The bound dummy index name.

over

An ir_set_ref() the fold ranges across.

body

The folded expression.

cond

Keep a term only where cond is non-zero; NULL keeps every term.

Details

Nodes are plain lists tagged by a kind field. An index tuple is a plain list whose entries are integers (concrete coordinates) or strings (bound index names).

The constructors carry an ir_ prefix rather than mirroring the Python client's bare names: Reduce is a base R function, and this package extends base names, it does not mask them.