cp-struct-by-value — measured, still hint
Status
hint. Wall-clock JSONL: cp-struct-by-value.jsonl. Not corpus-triaged. The 2026-10-22 corpus verdict does not cover this slug.
Measurement
clang 21.0.0, -std=c11 -O2. The callee is in another translation unit. The record is eight int fields and unsigned char pad[480] (sizeof 512). Both sides call the function and add two fields plus the first and last pad bytes into a volatile accumulator. Median of 11.
| role | case | median_ms | ops | ratio |
|---|---|---|---|---|
| after | pass_pointer | 0.152 | 128000 | 1 |
| before | pass_by_value | 1.864 | 128000 | 12.263 |
A static noinline pair in one translation unit, same record and the same 128000 calls, was 0.153 ms versus 0.158 ms (ratio 0.968). The compiler removed the copy. This is not a general speedup.
The scanner fires when a parameter is struct Name x with no *, the struct is defined earlier in the same file, a lower bound of the members it understands is at least 512 bytes, and there are at least four such fields. That lower bound uses LP64 widths (int 4, long and pointer 8) and skips nested types, bit-fields, and other members it does not understand, so a large record can be missed. On a repeat, a 320-byte record was 5.848× and a 448-byte record was 6.781×, so those sizes are not matched.
Unblock
Do not set the default to error from this file alone. A same-file static function, a cold call, or a record the size heuristic underestimates can erase the difference. Corpus triage is still required before any default change.