On i386, growing crtl breaks main's exit code in programs that never call the new code

• Type: bug (C backend / 32-bit codegen or link — Track C)
• Opened: 2026-07-31 by Track B, after Track T's cross matrix caught a regression my native gate could not see.

Symptom

library_candidates/c-testsuite/tests/single-exec/00211.c:

extern int printf(const char *format, ...);
#define n 0xe
int main() { printf("n+1 = %d\n", n+1); }     /* implicit return 0 */

Compiled --target=i386 and run under tools/run_target.sh:

n+1 = 15      <- output CORRECT
rc=60         <- should be 0

The program prints the right thing and then exits with garbage. 00206.c and 00212.c do the same.

What changes it — and this is the alarming part

Nothing in the test. Adding unrelated functions to lib/crtl. Bisected across four Track B commits, with lib/crtl checked out at each and only the exit code observed:

00211.c calls none of those functions. And reverting stdlib.c alone did not fix it — the stdio.c change breaks it independently. Two unrelated additions each trigger it, and the failing value varies with what was added, which is the signature of a size- or layout-sensitive fault rather than a bug in any one function.

Plausible directions, none confirmed: the entry stub reading main's return from the wrong place once crtl crosses some size; a relocation or offset that overflows on 32-bit; lldiv's 64-bit division helper or its 16-byte struct-by-value return perturbing the module. Worth checking the i386 _start -> main -> exit path first, since output is fine and only the return value is wrong.

Why it went unnoticed until Track T

tools/gate.sh lib is x86-64 only. The whole point of "confirm native, offload the matrix" is this case, and it worked — the cross matrix caught it, my native gate could not have. Worth remembering when a Track B change touches lib/crtl: that code is compiled for every target.

Current state

Master is restored — lib/crtl and test/crtl_libc_oracle.c are back at aa3f561a7, keeping the i386-clean errno.h and strnlen. 00206/00211/00212/ 00213 exit 0 again.

Reverted and wanted back once this is fixed (both were verified correct against gcc on x86-64, and are in git at the SHAs above):

• div / ldiv / lldiv / llabs + the div_t family — C89/C99, absent entirely.
• The sscanf field-width fix. That one matters: %15s — the safe spelling — silently abandoned the whole scan and left the destination untouched. Also %*d suppression, %hd/%hhd writing the right width, %#o, and %.0d.

Gate

C tests green + self-host byte-identical, plus test-c-conformance-i386 and -riscv32 green with the reverted crtl work restored — the restoration is the regression test.

Log

• 2026-07-31 — resolved, commit a4dd22b3b.