#pragma pack is IGNORED — struct layout silently differs from gcc

• Type: bug (C frontend — record layout; silent ABI mismatch)
• Track: C — C frontend (record layout) — file-lane A for symtab.inc layout code
• Status: done bug-c-bitfield-packing-sizeof-vs-gcc and blamed the bitfield storage-unit algorithm. That diagnosis was wrong, and anyone following it would have rewritten code that is already correct. See "What the old ticket got wrong" below.

The defect

#pragma pack(N) is parsed away and has no effect on layout.

#pragma pack(1)
struct A { char c; int i; };     /* gcc: 5   pxx: 8 */
#pragma pack()
struct B { char c; int i; };     /* gcc: 8   pxx: 8  (agree) */

Struct A contains no bitfields at all. The alignment cap that pack imposes is simply not applied: every field keeps its natural alignment, so i is padded to offset 4 instead of sitting at offset 1.

This is an ABI mismatch, and a silent one — the struct compiles, runs, and has the wrong size and the wrong field offsets. Any packed on-disk/on-wire structure (file headers, network frames, hardware register maps, anything a real C corpus does with #pragma pack) is laid out wrong with no diagnostic.

What the old ticket got wrong (verified 2026-07-14)

It claimed the bitfield storage-unit algorithm was at fault and pointed at RecFieldBitShift. Measured against gcc on six varied bitfield structs — :3/:5, :31/:2, char + int:5, a :0 separator, long long:40 + :20, char:4/:4/short:9 — pxx and gcc agree on every one. The bitfield layout is gcc-identical.

The original repro (packed=8 vs gcc's packed=8 -> a bitfield struct under pack) mixed the two things together and blamed the wrong one: it is pack that is missing, and the bitfields in that struct were incidental.

Wanted

• Honour #pragma pack(N) / #pragma pack(push, N) / #pragma pack(pop): cap each member's alignment at N, and cap the record's own alignment at N.
• __attribute__((packed)) is the same knob and should land with it.
• Apply to bitfield storage units too (they take the cap like any other member) — but note the bitfield algorithm itself needs no change.

Repro / gate

The four-struct program above must match gcc's sizeof for all four, plus a field-offset check (size alone can coincide). Then the C corpus: tools/run_c_conformance.sh green, and the real corpora (sqlite/zlib/lua/tcc) unchanged.

Log

• 2026-07-14 — resolved, commit 56365087.