Repository navigation
repsel: promotion census — make "zero values promoted" a visible, gated state instead of an invisible one #7106
Description
--opt-report (#6952) can tell you, for one build, which values got an unboxed
representation and which did not. What it cannot do is tell you that the answer
changed — or that it has been zero all along.
#7034 established the state of things by hand-instrumenting the compiler:
Ptr<Shape>promotes nothing at all onbenchmarks/app-patterns/kernels/batch.ts,
the object/property-heavy program that representation exists for.
PERRY_PTR_SHAPE_LOCALS=0versus the default produced a byte-identical
binary (4,294,904 bytes,__text3,804,424).- Across
benchmarks/suite, only 3 files promote a shape local, and each
promotes exactly one.
Nobody knew any of that until somebody went looking. That is the failure to
make impossible to repeat.
Ask
A per-workload, per-representation census recorded as a ratcheted baseline,
so a drop turns a build red.
Per representation, not aggregated: the interesting signal is Ptr<Shape> = 0
while canonical Str ≠ 0. One total hides exactly that.
The hard part
CLAUDE.md, "★ Four ways a gate can be unable to fail", case 4 — the gate runs
but its subject never did. A census that faithfully reports Ptr<Shape>: 0
and exits green is worth nothing: that is the state the project is in
today. So the census needs, beyond floors:
- floors that can actually be violated (the honest floor for
Ptr<Shape>on
real code is zero, and a zero floor can never fail); - a liveness assertion on the instrument itself — a fixture that definitely
promotes, which fails the census if its count reads zero, so that "zero
because nothing promoted" is distinguishable from "zero because nobody
increments the counter"; - verification by sabotage in both directions.
Related
- repsel row 1: scoping Ptr<Shape> beyond locals — measured, promotion is 0 on the motivating workload; build --opt-report (#6952) first #7034 — the scoping analysis with the original numbers.
- feat(cli): --opt-report — surface which values could not be statically typed, why, and whether the developer can fix it #6952 —
--opt-report, the data source. - opt-report v2: LocalId → Span side-table so --opt-report can print file:line and a source snippet #7036 — source spans for
--opt-report. Separate, additive.
Activity
docs: point the census at its own tracking issue (#7106, not #7035)
test(repsel): promotion census with a ratcheted, falsifiable floor (#…
test(repsel): make "no analysis ran" a visible, gated state
test(repsel): make "no analysis ran" a visible, gated state (#7113)
test(repsel): count CONSUMED promotions, not selected ones
test(repsel): count CONSUMED promotions, not selected ones (#7117)
docs(repsel): CI site-coverage assertion and changelog
test(repsel): gate per-site consumption coverage, and review follow-u…
proggeramlug commentedon Jul 31, 2026
The census's first verdict is measured and written up in #7128.
Headline for this issue specifically: the census works, and its
selected/consumedsplit (#7117/#7119) is what made the result honest.Corpus-wide
canonical-i32went 13 → 64 and from promoting in 2 of 18 realworkloads to 17 of 18 — and it changes the emitted object on 2 of them.
Without the consumed column the answer would have read "64 promotions, looks
great".
Also from that run, four instrumentation defects worth tracking here:
PERRY_CANONICAL_I32_LOCALS=0is not representation-isolating — it alsodisables all
Ptr<Shape>consumption (repsel_context_allows_ptr_shape: repsel_allows). Under that knob the census readsptr-shape7 selected /0 consumed with all six sites
NEVER FIRES.PERRY_CANONICAL_STR_LOCALS=0is not scoped toStrlocals either (it gatesnanbox_string_ref_boxed), so 24 of 26 workloads emit differently under it.__pshapeclone has zero call sites and is dead-stripped before theobject, so the 8
consumed_receiverentries count consumption insidefunctions that never ship.
Ptr<NumArray>: 6 selected, and turning it off leaves every real workloadbyte-identical.
And one that affects the census's own method: object emission is
nondeterministic on Linux (the module name embeds pid + nanotime), so
object-hash A/B is macOS-only today.
proggeramlug commentedon Aug 5, 2026
Delivered:
benchmarks/repsel_census/baseline.json,scripts/compiler_output_harness/repsel_census.py,tests/test_repsel_census.py, and arepsel-censusjob intest.ymlwith floors and sabotage checks.One follow-up remains and is tracked separately: the job is not yet in branch protection's required contexts, so it cannot turn a merge red.