← board

Busybox the way busybox builds

[[feature-c-corpus-busybox-multi-applet]] is done. It resolved at twelve applets; the unity has since been pushed to twenty-six applets, 82 translation units, 154 cases, byte-identical to gcc on x86-64 and aarch64. It got there on a UNITY build, and the unity is still the wrong long-run model.

Why the unity cannot be the answer

Not capacity — twenty-six is a lot more than the twelve this section was written about, and the argument did not depend on the number. Three busybox files assume they own their namespace, and gcc rejects the unity too, so none of this is a pxx defect:

file what it claims who it breaks
include/common_bufsiz.h enum { COMMON_BUFSIZE = 1024 }, no include guard ls, tail
shell/ash.c 40 #defines of ordinary names (optlist, eflag) whatever follows it
coreutils/uname.c:112 #define options "snrvmpioa" whatever follows it
coreutils/test.c globals macros over ordinary identifiers ash, in BOTH orders

The harness orders around the second, refuses the fourth, and can do nothing about the first or third. Every applet added from here meets more of them.

Measured at scale 2026-09-01, and this is the number that settles it. A 34-applet unity (the twelve plus grep sed sort touch chmod ln du df sync env dd tr cut basename dirname readlink true false seq tee md5sum stat) does not compile AT ALL, and every error is gcc's:

./coreutils/du.c:104:1     redefinition of struct or union 'struct globals'
./findutils/grep.c:196:1   redefinition of struct or union 'struct globals'
./editors/sed.c:172:1      redefinition of struct or union 'struct globals'
include/common_bufsiz.h:1  redeclaration of enumerator 'COMMON_BUFSIZE'   (x8)
./coreutils/dd.c:399       #define skip (Z.skip) -- expanded inside shell/ash.c
./coreutils/cut.c:57       two or more data types; conflicts with /usr/include/regex.h
./libbb/ptr_to_globals.c   conflicting types for 'ptr_to_globals'

struct globals is not an edge case — it is busybox's STANDARD pattern, one per applet, and a unity can hold exactly one. dd.c defining skip and having it expand inside ash.c is the ash-macro problem in the other direction, and no include order fixes a pair that each claim the other's identifiers.

So the unity's limit is not twelve and not any number: it is one applet per namespace-claiming pattern, and busybox's whole design is that pattern. This is why the rung is separate compilation and not "more applets".

Where this already is

tools/busybox_diff.sh --separate exists and works. Measured 2026-09-01, compiler 825c28a30c31:

Getting that far cost exactly one compiler fix: x & 0 never folded to a literal, so a constant-false branch kept its dead arm and make_human_readable_str became a real external reference from coreutils/ls.c (3056e214c). That was the only undefined symbol across all 41 units.

The link error is loud and is the blocker's business. The interesting failure is quiet: the binary links, runs, and diverges from the gcc oracle.

< cat: can't open '.../missing.txt': No such file or directory
> cat: can't open '.../missing.txt'
> cat: can't open '-u'

crtl state is object-local with no linkage at all, so a program that touches it by name reads its own untouched copy. The lost reason is split errno; -u becoming a filename is split optind. Both are measured as two-object repros in the blocker.

Read that failure before adding applets here. The smoke test — links, runs, --list/echo/ash all correct — passed the whole time. Only the 62-case differential caught it.

What finishing this looks like

  1. The blocker lands and --separate no longer needs -z muldefs.
  2. --separate goes GREEN against the oracle on the twelve applets the unity already does. That is the gate: match the unity before exceeding it.
  3. Then grow the applet set, which is what rung 3's image actually needs, and which the unity cannot do.
  4. aarch64 is out until --emit-obj has an object writer for it ([[feature-a-object-output-for-arm32-and-aarch64]]). x86-64 only for now, and say so rather than letting --separate look like the stronger claim on both axes at once.

MET 2026-09-01 — x86-64

busybox-diff: applets=cat echo ash mkdir rm cp mv pwd wc head sleep printf sort
              chmod ln sync env tr cut basename dirname readlink true false seq
              md5sum   translation units=82
  ORACLE  gcc unity build (154 cases)
  ORACLE  busybox agrees with the gcc unity
  note    x86_64   82 objects linked separately (27765544 bytes)
  PASS    x86_64   byte-identical to the gcc oracle over 154 cases
  note    aarch64  skipped: --emit-obj has no object writer for this target
busybox-diff: GREEN

compiler 816f18f7784d. -Wl,-z,muldefs is gone from the harness rather than merely unneeded — the flag was the thing hiding the second bug.

Against the four completion bullets above:

  1. The blocker landed (frankA) and -z muldefs came out. That exposed the SECOND link failure, which the flag had been suppressing all along: multiple definition of bb_ascii_isalnum / bb_strtoi32 / is_tty_secure / new_tls_state — static functions in include/libbb.h, emitted GLOBAL. Filed and fixed as [[bug-a-static-c-functions-are-emitted-as-global-symbols]].
  2. --separate is GREEN against the oracle on the twenty-six applets the unity does, not the twelve this ticket was written against. Gate met.
  3. Growing the applet set is rung 3's.
  4. aarch64 unchanged: out until [[feature-a-object-output-for-arm32-and-aarch64]].

Worth keeping: the errno/optind divergence this ticket was mostly about is gone, and the smoke test would never have told anyone. --list, echo and ash arithmetic were all correct through both bugs. Only the 154-case differential moved.

Log