The DWARF target set is written down three times, and the only correct statement is dead code

Found by tools/docaudit.py targets (frankD's derive-don't-match audit) over compiler/**, from a elfwriter.inc hit whose comment sat one line above a condition contradicting it.

What is true, measured rather than read

-g emits 4 .debug sections on x86-64, aarch64, i386 and arm32; xtensa and riscv32 are excluded. Verified by compiling the same program for each target and counting sections with readelf -S.

The three places it is written, and why reading any one of them misleads

Three comments additionally asserted "x86-64 only" (elfwriter.inc x2, compiler.pas's -g arm). Those are fixed — comment-only, self-host byte-identical, 7c4f7ce26297.

The trap is worth stating because it caught me. Reading the condition directly under the corrected comment gives the answer "two targets", and that is a reasoned answer from real code that is still wrong — the second writer is 190 lines away. Only compiling for all four gave "four". The near-miss is the finding: this is the shape where measuring beats reading even when you are reading the implementation itself.

What is NOT fixed, and why it needs care rather than a quick edit

The obvious fix is to make the dead authority live: doDebug := DebugInfo and DbgArchSupported at all three sites, collapsing three duplicated target lists into one. It is probably behaviour-preserving, because each writer only ever runs for its own target family — but that binding was not proved, and the writers also serve exec / shared / --emit-obj modes that this ticket did not exercise. A change that is "probably equivalent" to a gate on debug-info emission is exactly the kind this repo asks you to prove rather than assume.

So: three lists, one dead authority, and a fix that is one line per site once someone proves the writer/target binding. Do that with a measurement across every target x every output mode, not by reading the dispatch.

Why it is a bug ticket and not an audit note

A function whose only content is a correct statement, that nothing calls, is worse than no function: it reads as the authority. Anyone grepping DbgArchSupported finds the right answer and reasonably concludes it is enforced. That is a false statement about a gate, told at the point of maximum authority — the same class as bug-a-threadsafe-is-x86-64-only-is-asserted-in-five-places-and-has-been-false-since-july, and this one has an extra twist: the correct sentence is the dead one.

Gate

The comment half is done and byte-identical. The code half: make compiler/pascal26 plus -g emitting the same section count on all four supported targets and none on xtensa/riscv32, across exec and shared output.

Log

• 2026-09-01 — resolved, commit ae16da1d2.