← board

The inliner accepts only functions — procedures are a third admission axis

The measurement

Identical bodies, one a function and one a procedure, -O3:

function  AddF(a, b: Integer): Integer;         begin AddF := a + b; end;
procedure SetG(v: Integer);                     begin g := v; end;

call AddF is gone; call SetG remains. PXXDBG=a.inline shows only AddF retained — the procedure is rejected at retention, so no shape validator ever sees it. TryRetainInlineBody:

if not Procs[procIdx].IsFunc then Exit;

Why this is its own axis and not part of an existing ticket

The inliner's admission has three independent gates:

axis governed by ticket
return TYPE InlineScalarTk feature-opt-inline-float-and-record-returning-leaves (float half landed ec95c2beb; record half open)
body SHAPE InlineExprSimple / TryRetainInlineStmtBody feature-inline-nonleaf-and-branch-locals
is-FUNCTION the line above none — this ticket

feature-inline-nonleaf-and-branch-locals extends shape (branch bodies, while/for, deeper non-leaf) but every shape it discusses is inside a function. Nothing widens the first gate.

What it is worth, measured — and what it is NOT worth

A slot-op-shaped microbench (small branchy procedure writing through a pointer param, six calls per iteration), -O3, min-of-5 interleaved, contended box:
out-of-line calls 0.200 s 1.00x
hand-inlined 0.140 s 1.43x
no dispatch at all 0.050 s 4.00x

1.43x in this shape, ~3.3 ns per call. Read the third row before ranking this: in the workload it was found under, removing the calls recovers 1.43x while removing the classification recovers 2.80x more. Do not take this ticket expecting to close the NilPy subscript gap — that is option 3 there, and it is a frontend type-inference question, not an inliner one.

Its value is instead general: procedures are everywhere, and none of them has ever been inlinable. That generality is unmeasured, and measuring it is the first step of taking this ticket, not an assumption to start from (the ~37%-advertised / ~5-7%-delivered campaign is the standing warning).

Scope sketch, smallest useful slice first

A procedure has no Result, so the Result placeholder machinery does not apply and the rest of the splice does. The natural first slice is by-value scalar params, straight-line assignment bodies — which SetG already is.

Note what that slice does NOT reach, and why the slot ops need more: a var param is IsRef and rejected outright, and PyVarSlotClear(dst: PPyVarRec) writes through dst^, which InlineExprSimple rejects as a deref. So the slot-op shape needs by-ref params or deref-target assignment on top. Slice the ticket accordingly rather than promising the slot ops from the start.

Gate

Track O's two: PROMISE measured as delivered value on a real workload (not a census of how many procedure calls exist), PROOF from Track T's full tier. Behind -O3 first. tools/optfuzz.sh is the designated net for splice machinery — it covers this one, unlike the float axis, since pasmith generates integer procedures freely ([[bug-t-pasmith-returns-only-integer-kinds-so-optfuzz-is-blind-to-the-return-type-axis]]).