Test status
Test cases
test-many-klibcs implements:
- Build for each architecture (with a cross-compiler where needed)
- Run various programs (using qemu-user where needed) in static and shared configurations: a. Many self-test programs (through system shell) b. “sh -c exit” c. “sh -c ‘.../bin/true; exit’”
Test failures
2.0.10
- arm, i386, m68k, parisc, ppc, sh:
- Fail to build (with Linux 5.19 headers)
- arm64:
- 2a:malloctest2, 2a:setjmptest, 2a:sigint: Fail in shared build with GNU toolchain (gcc 12.1.0, binutils 2.38.90.20220713)
- 2c: Fails with SIGSEGV in static build with LLVM, but only on QEMU
- riscv64:
- 2a:sigint: Fails when built with Clang, but only on QEMU
- s390x:
- 2a:environ, 2a:opentest, 2a:setjmptest: Produce wrong output when built with Clang, but only on QEMU
- 2a:malloctest, 2a:malloctest2, 2c: Appear to loop forever when built with Clang, but only on QEMU
- 2a:sscanf: Exits with error when built with Clang, but only on QEMU
2.0.10+git20220806
- arm64:
- 2c: Fails with SIGSEGV in static build with LLVM, maybe only on QEMU
- riscv64:
- 2a:sigint: Fails when built with Clang, maybe only on QEMU
- s390x:
- 2a:environ, 2a:opentest, 2a:setjmptest: Produce wrong output when built with Clang, maybe only on QEMU
- 2a:malloctest, 2a:malloctest2, 2c: Appear to loop forever when built with Clang, maybe only on QEMU
- 2a:sscanf: Exits with error when built with Clang, maybe only on QEMU
Known bugs elsewhere
QEMU 3.1 had a bug in handling of struct timveval on sparc64, and a bug in the ELF loader. Additionally, it has a lot of bugs in signal handling on alpha and sparc. All of those are fixed in my local build, qemu-user-static_3.1+dfsg-2.1_amd64.deb.
QEMU 6.1 fixes those plus the s390x failures, but also has regressions for armhf, hppa, and riscv64.
Clang 13 has a bug that affects i386, -mregparm, and optimisation of printf() to puts(). The call to puts() passes arguments using the usual stack-based calling convention. This is now worked-around by disabling CONFIG_REGPARM.
LLD 13 has a bug affecting mips and mips64 that puts trap instructions (sigrie) instead of nops in branch delay slots. Patching these back to nops made the tests pass. I have only seen this in klibc.so, not in any executables.
LLD 13 seems to have a bug in PowerPC global register initialisation affecting most tests on ppc64.
LLD 13 is missing required features for riscv64 and sparc64.