[maarten-boot]

led keeps each input module in one memory piece: header, section
table, name table, strings, then the offsets of the section contents.
With USEMALLOC each piece is its own malloc block, but the offsets
inside it were only 4-aligned (the archive magic took sizeof(int),
int_align() rounded to sizeof(int), the name table followed the
20-byte header and section entries). On a 64-bit host struct outname,
struct ar_hdr and ind_t need 8, so linking with an archive used
them at misaligned addresses: undefined behaviour, reported by
-fsanitize=alignment on every such link, and a fault on a
strict-alignment host.

int_align() now rounds to _Alignof(max_align_t), each table starts
aligned, and modulsize() is computed from that layout.

Tested: make with PLATS = all, all tests pass; every linked program
is byte-identical to before (the hilo_bas examples differ between any
two builds because the BASIC compiler emits a heap address, a separate
bug). With -fsanitize=alignment led has no reports left (the
remaining ones are in util/ass).

🤖 Generated with Claude Code

[davidgiven]

Nice, thanks; we can probably dispense with the !USEMALLOC case entirely and remove the #ifdef. It would simplify the code.

[maarten-boot]

Done: USEMALLOC was always defined, so the brk() allocator was never compiled. 38656bd removes it with the #ifdefs (memory.c goes from 810 to 469 lines). On cpm, linux68k, linuxppc and pc86 every linked program is byte-identical to before and all tests pass; with
▎ -fsanitize=alignment, led has no reports.

[davidgiven]

Excellent, thanks. Merging.