Module `Tolk.Linearizer`

Kernel-to-program linearization.

Provides rewrite passes for splitting multi-range Ends and adding control-flow edges, plus the final priority-based toposort and Uop → Program emission.

The input kernel must be in linearize-ready form: exactly one Tolk_uop.Ops.t.Sink, no unlowered nodes (Tolk_uop.Ops.t.Reduce, Tolk_uop.Ops.t.Stage), gated loads must already carry an alternate value, and graph-stage Tolk_uop.Ops.t.If/Tolk_uop.Ops.t.Endif nodes are rejected. Gated stores are lowered after ordering into explicit If; Store; Endif program lines.

See also Codegen_lower.

val do_split_ends : Tolk_uop.Uop.t -> Tolk_uop.Uop.t option

do_split_ends node splits a multi-range End into nested single-range Ends. Returns None for non-End nodes or single-range Ends.

val pm_split_ends : Tolk_uop.Uop.t -> Tolk_uop.Uop.t

pm_split_ends root splits multi-range Tolk_uop.Ops.t.End nodes into nested single-range Ends, sorted by full range argument (descending).

val pm_add_control_flow : Tolk_uop.Uop.t -> Tolk_uop.Uop.t

pm_add_control_flow sink computes control-flow dependencies between sibling loops and adds ordering edges to Tolk_uop.Ops.t.Range nodes.

val linearize : Tolk_uop.Uop.t -> Program_spec.program

linearize sink performs priority-based topological ordering and emits a flat Program_spec.program. Gated stores are rewritten to explicit Tolk_uop.Ops.t.If, ungated Tolk_uop.Ops.t.Store, and Tolk_uop.Ops.t.Endif lines. Expects pm_split_ends and pm_add_control_flow to have already been applied.

Raises Failure if sink contains unlowered nodes, gated loads without an alternate value, empty groups, or graph-stage Tolk_uop.Ops.t.If/Tolk_uop.Ops.t.Endif nodes. Full program validation is performed by Tolk_uop.Spec.verify_list with Tolk_uop.Spec.program_spec.