Module control_flow 

Structural-mode candidate generation from the control-flow skeleton.

The other two candidate stages both describe a unit by the pieces it is made of: crate::candidate indexes statement windows and subtrees and pairs units that share one exactly, and crate::near_match treats those same pieces as a set and pairs units whose sets overlap. Both lose the same edit, and lose it hardest where code is smallest.

Inserting a statement rewrites every piece that encloses it. In a short function almost every piece encloses almost everything: the body block, the loop, the whole unit. A copy with two statements added can therefore share no window and no subtree with its original, so the exact stage proposes nothing and the set stage sees two disjoint sets — while a reader would call them the same function. The loss is not a threshold that could be relaxed; there is no overlap left to find.

What such an edit does leave untouched is the shape of the control flow. A statement that is not a loop, a branch, a match or a jump does not appear in crate::features::CfgFeature::skeleton_hash at all, so a unit and its gapped copy hash to the same skeleton. This stage indexes that hash and pairs the units that share one. It is an exact-match index like the seed layer, not an approximation: units either have the same skeleton or they do not.

A skeleton says much less than a subtree does — it is why this stage proposes rather than concludes, and every pair it emits is judged by crate::verify like any other. Three controls keep what it proposes bounded (AGENTS.md invariant 10), and each counts what it drops:

• a minimum skeleton size — a unit with fewer than ControlFlowConfig::min_ops control operations has a skeleton so common that sharing it is no evidence at all, and is not indexed;
• high-frequency suppression — a skeleton shared by more than ControlFlowConfig::posting_cap units is common structure rather than a family of copies, and its whole posting list is dropped;
• a length-ratio gate and a global pair budget — as in the near-match stage, a pair spanning too great a size difference is not a gapped copy, and posting lists are paired rarest-first so exhaustion sacrifices the lowest-signal candidates.

As in crate::candidate, the budget stops between posting lists and never inside one: a list compared only in part reaches grouping as a family whose members disagree, and comes back out as the stray pairs the ceiling allowed rather than as the one group it is. What a list costs is counted after the length-ratio gate, so a list of widely differing sizes is charged for the few pairs it really contributes.

Output is a pure function of the input: the index is ordered, pairing is deterministic, and the emitted pairs are sorted.

Structs

ControlFlowConfig

Tuning for control-flow candidate generation.

ControlFlowPair

A control-flow candidate: two units with the same control-flow skeleton. Canonical: a < b.

ControlFlowSet

The control-flow stage’s output: candidate unit pairs plus funnel counters.

ControlFlowStats

Counters describing what control-flow generation saw and dropped.

Constants

DEFAULT_MAX_LENGTH_RATIO

Default largest unit-size ratio a pair may span.

DEFAULT_MIN_OPS

Default smallest skeleton size, in control operations, for a unit to be indexed.

DEFAULT_PAIR_BUDGET

Default global candidate-pair upper bound.

DEFAULT_POSTING_CAP

Default posting-list cap; longer lists are common structure and dropped.

Functions

generate

Generate control-flow candidate unit pairs across files.