The build
is a graph.

No-op includes graph loading and dependency checks. A zero action count is verified for every sample. These are warm prepared fixtures, not cold-machine boots.

A byte change in one leaf source recompiles one object, updates its static archive and relinks the executable. Setup is excluded; graph traversal, these three actions and bookkeeping are included.

The private header reaches one object; a module header reaches 10% of translation units; the global header reaches all native leaf objects. Generated-input changes run both generator actions and rebuild their complete consumers.

Clean full removes outputs before timing. Clean + rebuild includes a common removal/recreation command inside timing. Native graph size is translation units + 15 actions: generators, two additional objects, ten archives and a link.

Large full-build cells have three measured samples. Their min–max spreads remain visible in the distributions below; no high-percentile or universal ranking claim follows from them.

A separate graph contains independent deterministic FNV hash actions. It exposes loading, checking and dispatch costs; it is not C++ throughput and does not cover every DAG shape.

Every command uses the same allowed four logical CPUs. Worker limits are 1, 2 or 4. The heavier profile instantiates sixteen independent depth-128 templates per translation unit; its final binary behavior is unchanged. Compare workers within a fixed size/profile.

All 9,348 retained official observations, including warmups, pass the expected action-set and output-hash checks. A common wrapper records each generator, compiler, archive and link invocation. Missing, extra or duplicate work aborts the campaign.

Nift’s “depends” controls order, not invalidation. File dependencies include prerequisite outputs. Named Nift targets are forced; the targeted comparison starts with a dirty module for all three systems. An untimed subsequent full build verifies the final executable.

CPU time and maximum-child RSS are retained for every timed command. Linux wait4 RSS is a child high-water measure, not summed simultaneous memory across parallel compiler processes. Wall time remains the primary metric.

At 10,000 light translation units, Ninja has the lowest observed no-op median and Ninja has the lowest one-file median. At 1,000 heavy translation units / j4, GNU Make has the lowest full-build median. These are descriptive cell comparisons, not significance tests or a universal ranking.

Nift’s higher-level tracked-action model is directly exercised here, including generation, ordering and targeted builds. Any dependency-checking or dispatch cost stays in the measurements. The syscall diagnostic provides a bounded lead for future core work; no optimization was made for this campaign.

One fresh Linode, four logical CPUs and 8 GB RAM. Participants rotate within every round; scenario order rotates across graphs. Sources and known build state are restored, contents mutated and timestamp preconditions checked before the C timer starts. No cache flushing or cold-machine claim. Stdout/stderr are captured; default CLI status-output costs remain inside timing, while terminal painting is excluded. Every warmup and sample is preserved.

Make and Ninja edges contain the actual native inputs. Nift additionally retains its required tracked-item content and explicit file sidecars. The first trial’s extra Make/Ninja recipe-marker edges were rejected, corrected and recertified; a second trial’s one-rule-per-action Ninja encoding was also rejected in favour of shared compile/action rules. Neither provisional series is pooled here.

The matrix is bounded: every native scenario at 100/1,000 light; larger light graphs for scaling; heavy 100/1,000 for compiler and worker comparisons; independent tiny actions at 100/1,000/10,000. It is not every combination of size, profile and workers.

Nift source

c14d06835bca515a5281e41da0e40219c2d60ef0 · rebuilt final candidate

Suite source

e95bfe88ac9e1eaf67cfeb8f67020fc3373ab81e

Toolchain

GNU Make 4.4.1 / 1.13.2 / Nift v4.12.0 / g++ (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0 / GNU ld (GNU Binutils for Ubuntu) 2.42

Canonical evidence

Raw observations, definitions, action oracles, hashes and lifecycle

Reproduction

Fairness contract · Exact sample plan · Frozen source reconstruction

Diagnostics

Bounded no-op syscall handover; tracing counts are diagnostic, not elapsed-time results. Nift core was not modified.

Sample policy: 100 no-op; 20 modest increments; 5–10 multi-second work; three expensive large builds. All samples remain, with linear percentile interpolation at (n−1)q. Shared-host variability and this synthetic graph limit generalization.