x86-64 div-by-zero abort stub is emitted unconditionally, even for division-free programs

• Track: A (codegen). Tag: O (size). x86-64 only.
• Found: 2026-07-16, alongside [[feature-opt-rtti-emit-on-use]] while dissecting the frozen hello demo's size.

Symptom

EmitDiv0Stub (parser.inc ~24210) runs for every x86-64 build:

if TargetArch = TARGET_X86_64 then EmitDiv0Stub;

It emits the shared div-by-zero abort stub plus the 37-byte string "Runtime error 200 (division by zero)" into .data — even when the program contains no div/mod//-on-integer site that could ever branch to it. ~40 B of dead code+data in a division-free binary. Only visible as noise on a tiny frozen build, but it is pure waste there.

Fix

Gate the stub on "an integer div/mod site was actually emitted" — set a flag when a div/mod IR op is lowered on x86-64, and emit the stub (and intern its string) only if that flag is set. Division-free programs then carry neither the stub nor the string. The div/mod codegen already references Div0StubAddr; that reference is what the flag tracks.

Notes

• Correctness is unaffected: a program with no division never branches to the stub, so removing it when unused changes nothing observable.
• Low priority — it is a small, x86-64-only size trim. The bigger baseline item is the RTTI one ([[feature-opt-rtti-emit-on-use]]), which also hits ESP.
• Gate = self-host byte-identical (the self-host compiler HAS division, so its own build is unchanged) + a division-free program shrinks + quick.

Log

• 2026-07-18 (fable-O) LANDED, with a twist. Prescan (DetectPascalRuntimeNeeds gains needsDivStub: tkDiv/tkMod tokens) gates the stub — but the TICKET PREMISE was incomplete: the stub is NOT dead in division-free programs when builtinheap is pulled, because builtinheap's OWN early div sites (compiled before PXXDivZero registers; FindProc misses mid-unit) fall back to the stub. And the default-RTL uses injection precedes the scan, so needsHeap (kept in the condition) retains the stub for every default build → Pascal output byte-identical to pinned across the board. The trim only fires for genuinely heap-free division-free sources. The real find: the C and NilPy drivers NEVER emitted the stub — their builtinheap early div sites emitted call 0 (into the entry stub) on the div-by-zero path. Latent silent-corruption bug, exposed by the new safety net (Div0StubAddr = 0 at a stub-needing site → hard compiler error) and fixed: both drivers now emit the stub unconditionally, BEFORE their RTL pulls. Gates: self-host byte-identical, test-opt green, C 220/220, test-nilpy, hasdiv/nodiv byte-identical to pinned, C div-by-zero exits RE200.

• 2026-07-18 — resolved, commit 938c0154.

• 2026-07-18 follow-up (fable-O): borg's zig reds at f5c8fbec were REAL and mine to finish — the Zig, Rust, Erlang (eparser), BASIC (bparser) and Ada (aparser) drivers ALL had the same never-emitted-stub hole as C/NilPy; the safety net turned their latent call 0 into hard compile errors. All six non-Pascal drivers now emit the stub after their entry stubs (Ada gets a jump-over — its driver had no entry jmp). Verified: zig skeleton+perft, rust chess engine, erlang skeleton, .bas goto/gosub, ada skeleton all green; self-host fixedpoint, C 220/220, nilpy unchanged.