Part of #9. Phase 3 — the capstone. The chapter the whole book was building to: TRPL
Fearless Concurrency, applied to your own compute-heavy analyzer.
Learning goals
rayon data parallelism, Send/Sync as design constraints (your arena/NodeId choice from
issue 11 pays off HERE — index-based graphs are trivially Send), a Kahn-style ready-queue
scheduler, and the hard discipline: determinism as a tested property, not a hope.
Task
Parallelize by the dependency structure (references/dataflow-construction.md § Parallel execution model):
- Level-1 per-file build + stages 1–4 per callable: embarrassingly parallel — rayon fan-out
(par_iter). The whole-program alias solve (such as it is) runs CONCURRENTLY with stages 1–4
and joins before stage 6.
- Stages 6–7: ready-queue wavefront over the SCC condensation DAG — per-SCC dependency
counters; a completed summary decrements dependents, enqueues the ready. The SCC is the atomic
unit (its internal fixpoint stays on one worker). Strictly better than level barriers.
-j/--jobs N: output must be BYTE-IDENTICAL to --jobs 1. The monotone framework makes
the fixpoint schedule-independent; discovery ORDER is what varies — so never assign ids or
emit during parallel execution: collect, then sort by (signature, node_id). --jobs 1 stays
the debug mode forever.- Sequential-first: every gate green at
--jobs 1 BEFORE parallelizing; the sequential output is
then the differential oracle.
- Watch memory, not just CPU: release per-callable CFG/DFG structures once the PDG is emitted.
Gate (determinism gate)
--jobs 8 output byte-identical to --jobs 1 on the fixture — for the per-callable fan-out
AND the SCC wavefront (diff the files in the test).- A wall-clock improvement is demonstrated on a real crate (tokio or similar) and recorded in
the PR description.
- When this closes, the epic closes: symbol table → call graph → CFG → PDG → SDG → slicing →
taint → parallel and deterministic. That is a mature analyzer, and you wrote every line.
Part of #9. Phase 3 — the capstone. The chapter the whole book was building to: TRPL
Fearless Concurrency, applied to your own compute-heavy analyzer.
Learning goals
rayondata parallelism,Send/Syncas design constraints (your arena/NodeId choice fromissue 11 pays off HERE — index-based graphs are trivially Send), a Kahn-style ready-queue
scheduler, and the hard discipline: determinism as a tested property, not a hope.
Task
Parallelize by the dependency structure (
references/dataflow-construction.md § Parallel execution model):(
par_iter). The whole-program alias solve (such as it is) runs CONCURRENTLY with stages 1–4and joins before stage 6.
counters; a completed summary decrements dependents, enqueues the ready. The SCC is the atomic
unit (its internal fixpoint stays on one worker). Strictly better than level barriers.
-j/--jobs N: output must be BYTE-IDENTICAL to--jobs 1. The monotone framework makesthe fixpoint schedule-independent; discovery ORDER is what varies — so never assign ids or
emit during parallel execution: collect, then sort by (signature, node_id).
--jobs 1staysthe debug mode forever.
--jobs 1BEFORE parallelizing; the sequential output isthen the differential oracle.
Gate (determinism gate)
--jobs 8output byte-identical to--jobs 1on the fixture — for the per-callable fan-outAND the SCC wavefront (diff the files in the test).
the PR description.
taint → parallel and deterministic. That is a mature analyzer, and you wrote every line.