Bootstrap: measure how splittable the spine is before splitting it (#22642)

[PurHur]

Measurement only — nothing in the build consumes it yet. It exists to decide whether split-compiling the spine is worth attempting, before anyone commits to a partitioning.

Why

The gen-0 rebuild compiles test/selfhost/compiler_lib_spine_smoke/main.php — 6,521 literal require_once lines — as one translation unit, single-threaded, for hours. The obvious remedy is the split-compilation shape already proven for helper units (#15889): per-chunk .o with a content fingerprint, emitted in parallel, merged at link.

The risk is not the chunking, it is the edges. JIT\Call\ExternalMethod lowers a method call on a class that is not in the current module to __value__writeNull, silently (#579). Split the spine naively and every cross-chunk call becomes a silent null — a miscompile far worse than a slow build.

What it measures

Static pass over the spine require list: which file declares which class, which classes each file references (new, static call, extends/implements, trait use, instanceof, catch, param/return/property types), and how many references cross a candidate chunk boundary.

Result — the spine is much less entangled than its size suggests

6,521 files, 8,068 declared classes:

strategy	chunks	largest chunk	cross-chunk refs	ratio
top	5	4,662	443	2.9%
ext	77	2,121	529	3.5%
dir	93	2,121	679	4.4%

Only 3–4% of intra-spine references cross a directory boundary, and they concentrate into a few core chunks — heaviest pairs are ext/intl > lib/VM (82), ext/standard > lib/VM (73), lib/JIT > lib/VM (45). A further 40,825 references leave the spine entirely (builtin/vendor) and are unaffected by chunking.

What this says about the next step

A split is worth attempting. ext looks like the sweet spot — 77 chunks for parallelism at 3.5% coupling, and the coupling is concentrated enough that binding the core chunks' symbols by declaration (as HelperRuntimeCache::tryProvide already does) would cover most edges.

The remaining obstacle is ext/standard at 2,121 files: it would dominate wall time regardless of how many other chunks run in parallel, so it needs sub-splitting before parallelism actually pays off.

When a real chunk build is attempted, count the edges that failed to bind with PHP_COMPILER_REPORT_EXTERNAL_STUBS=1 (#22967) — that turns the silent-null risk into a number.

Test plan

• php -l clean
• all three strategies run to completion on the current spine (1 unparsed file)
• --json output parses

🤖 Generated with Claude Code