IROpName names 68 of 75 IR ops, so a missing backend arm for the other seven reports unknown

error: target xtensa: unsupported node in IR codegen: unknown

That is the whole diagnostic. The op was IR_CLASSREF.

Counted by parsing defs.inc's IR_* = n; constants against ir.inc's IROpName case arms, not by reading them:

IR_PROCADDR(38)  IR_CLASSREF(39)  IR_VMTADDR(58)  IR_IMTADDR(59)
IR_SET_SIGNAL(65)  IR_IO_LOCK(66)  IR_IO_UNLOCK(67)

Why it matters more than a cosmetic string

IROpName has exactly one load-bearing caller: the unsupported node in IR codegen error every backend raises for an op it does not implement. So for these seven ops the error names nothing, on every target, and the only way to learn which op is missing is to edit the backend and self-compile.

That is not hypothetical — it is how this was found, and it cost a build. The same shape as target xtensa: unsupported binary operator (div/mod/shifts pending), fixed on 2026-08-30 for the same reason: a diagnostic that cannot name its own subject makes the reader guess, and the guess is usually the message's own stale hint.

Fix

Seven lines in ir.inc's IROpName. The names are already the constant identifiers; use the same lowercase-without-prefix convention the other 68 use (procaddr, classref, vmtaddr, imtaddr, set_signal, io_lock, io_unlock).

Worth adding with them: a compile-time or startup assertion that IROpName covers every declared op, so the eighth gap cannot open silently. The count above was produced by a parser in ten seconds; nothing in the tree runs it.

Why filed and not fixed

ir.inc is shared Track A ground and a Track S stop-line.

Bound

Static count at f19e16b67bad, by parsing both files. The claim that IROpName has one load-bearing caller is from grepping its call sites. Not measured: how many of the seven have a missing arm on any given backend today — only IR_CLASSREF was observed to actually surface, on xtensa.

RESOLVED 2026-08-31 — frank-rust

Seven arms added; an independent parse of defs.inc against ir.inc now gives 75 declared, 75 named, 0 unnamed. The count was re-derived rather than taken from the ticket, and matched it exactly.

The assertion the ticket asked for exists and is wired

"Worth adding with them: a compile-time or startup assertion that IROpName covers every declared op... The count above was produced by a parser in ten seconds; nothing in the tree runs it."

tools/iropname_lint.py, wired into gate.sh. Sub-second, parses two files, builds nothing. "Nothing in the tree runs it" was the whole defect — the parser already existed as a one-off; what was missing was a caller.

End-to-end, against pinned

PXXDBG=a.ir:* on --threadsafe hello.pas, which emits IR_IO_LOCK / IR_IO_UNLOCK:

40 + 40 = 80 exactly, which is what makes it an identity rather than two numbers that happen to differ.

My first linter had the exact bug it was written to catch

The first version scanned the case body for IR_* tokens including comments — and the comment I had just written above the seven new arms names IR_CLASSREF. So deleting the IR_CLASSREF arm left the linter reporting clean.

That is abi.inc's dead review grep, reproduced from scratch, hours after I replaced it: a checker satisfied by prose. And it arrived by the same route frankwasm identified — the better the surrounding comment, the more convincing the dead check becomes. Writing a careful comment is what broke it.

The unit selftest did not catch it. Its synthetic fixtures had no comments, so it passed while the tool was blind on the real file. It was caught only by a real-tree control — delete a real arm, assert the file actually changed, then require exit 1. Comment stripping and a comment_only control are now both in.

The transferable bit: a selftest over synthetic fixtures tests the logic, not the input. The fixture is written by the same person with the same blind spot, so it inherits the assumption instead of challenging it. Run the control against the real file.

Also fixed

ir_codegen_xtensa.inc:3391 said IROpName has no entry for this op — true when written, false as of this commit, and it sits directly above the arm that exists because of it. Rewritten to past tense with the resolution rather than deleted: the original diagnostic is why that arm exists.

Log

• 2026-08-31 — resolved, commit 60f0f1982.