C frontend — shared-IR touch points that belong to Track A

• Type: coordination / heads-up
• Track: A (compiler core / shared IR) — raised by Track D (C frontend, worktree feat/cfront).
• Opened: 2026-06-25
• Status: done
• Closed 2026-06-29 (board cleanup): all flagged shared-IR nodes landed (AN_TERNARY/AN_COMMA/AN_INCDEC/AN_COMPOUND_ASSIGN/AN_SWITCH in defs.inc+ir.inc); ticket already carried a RESOLVED section but stayed in backlog.
• Why: the C body frontend lowers to the SAME shared AST/IR as the Pascal frontend. Where building it required touching shared lowering (not just clexer/cparser), Track A should own/review the semantics so a future Pascal refactor doesn't silently regress C, and vice-versa.

1. Value-bearing top-level AN_EXIT now routes through the Halt terminate path

ir.inc (AN_EXIT lowering): a return <expr> from a top-level C main arrives as an AN_EXIT with ASTLeft <> -1 while CurProc < 0. Pascal's program-level Exit never carries a value, so the original lowering discarded it and codegen emitted EmitExit(0). Added branch: when CurProc < 0 and the node has a value, terminate with that value (the AN_HALT path) so the C process exit code is the returned int.

• Guarded by CurProc < 0 and a present value → never fires for Pascal → self-host stays byte-identical (verified).
• Landed in commit 202ae37 (Track D, branch feat/cfront).
• Ask of Track A: confirm this is acceptable shared-IR behavior, or tell us to move it behind a C-frontend-only hook. Slice D later made main a real proc (return via RetSymIdx + entry exit_group(eax)), so this branch is now only a fallback for the inline-main model — it can likely be retired once the C frontend fully commits to the proc-based entry; flagging so A decides.

2. Conditional expression ?: — possible missing AST node

C's ternary a ? b : c is an expression that yields a value; pxx has no conditional-expression AST node (Pascal has no ternary). Track D can desugar it in the C frontend via a hidden temp + AN_IF statement, but a first-class AN_TERNARY (or a generic conditional-expression lowering) would be cleaner and reusable. Ask of Track A: do you want to own a shared conditional-expression node, or should Track D desugar locally? Currently deferred (no ?: in the fixtures yet); not blocking, but lua/sqlite use it heavily.

3. Already-tracked shared bug the C path also hits

bug-esp-not-always-boolean — not <non-boolean-typed expr> (e.g. a function call returning Int64) is typed boolean, so ~0 in C const-eval yields 1 not -1. Root cause is the shared front-end not-typing, owned by Track A. Track D filed bug-c-const-eval-bitwise-not with a C-local XOR-with-(-1) workaround option; the general fix stays with A.

Log

• 2026-06-25 — opened by Track D after slices A–D. Going forward, Track D files a Track A ticket whenever the C frontend needs a shared-IR/AST change (e.g. a new node) rather than absorbing it into the C-only path.

RESOLVED 2026-06-26 — value-bearing nodes reconciled (read this before adding an AST node)

What went wrong: BOTH tracks implemented the value-bearing nodes in parallel (Track C on feat/cfront, Track A on master) with different node numbers and different node shapes. Merging stacked two definitions → duplicate identifiers + mismatched lowering. Avoidable: the rule below already existed; it was crossed.

Canonical (master, after merge 556ad7dd) — these are DONE, do not re-add:

Next free AN_ number = 72. Highest in use = AN_SWITCH=71.

Rules to avoid a repeat (reinforced):

1. C never invents a shared AST node number unilaterally. Need one → file a Track A ticket (as section 2 did for AN_TERNARY) and use the number A assigns.
2. If C must prototype a node before A lands it, claim from a RESERVED HIGH range (AN_C_PROTO_BASE = 100+) and record it HERE the same commit, so it can never collide with A's next sequential number, and A sees the reservation.
3. The C frontend's cparser builds nodes in master's documented shape (table above). Lowering for these lives in ir.inc and is Track A's; C only emits.
4. AN_CASE/AN_DEFAULT are shared Pascal+C node kinds (17/48) — C switch reuses them, does not redefine.