today I looked into this issue , but it's not reproducible from top of llvm/clang trunk. I can find both __cxa_end_cleanup and __cxa_end_cleanup_impl function bodies from the generated assembly code, and I can build libcxxabi with '-fintegrated-as'. I use scripts from https://github.com/loganchien/libcxx-scripts to cross-compile it for ARM target from a X86 host. If the build step is correct, probably it has been already fixed.

I just checked and it's still affecting Android (reverting https://android-review.googlesource.com/#/c/110170/ and trying to build libc++ for the generic ARM target is enough to repro).

Will have to do some more digging to figure out what we do differently that is causing this though.

Testcase? This isn't a bug report without one.

This is caused by the fact that we use -ffunction-sections -fdata-sections and -Wl,--gc-sections in the compiler flags.

__cxa_end_cleanup ends up being thrown out by --gc-sections in the linker.

If I remove -Wl,--gc-sections, the warning goes away.

Well, sure, because you've turned off the thing that it depends on.

Do you have a better testcase here?

I can reproduce this issue, but the description in the thread is not accurate.

Short Answer

This section flag for .text.__cxa_end_cleanup() is missing.

PATCH: http://reviews.llvm.org/D15613

Long Answer

This problem will only occur when:

1. Assemble cxa_exception.cpp with integrated assembler (-fintegrated-as)

2. Use ld.gold as the linker (-fuse-ld=gold)

3. Enable GC section (-Wl,--gc-sections)

4. Treat warning as fatal (-Wl,--fatal-warning)

5. Generate one section per function or data (-ffunction-sections -fdata-sections)

When I was writing cxa_exception.cpp several months ago, I expected that .text.__cxa_end_cleanup_impl section won't be discarded because there is a relocation in .text.__cxa_end_cleanup section which refers to .text.__cxa_end_cleanup_impl section. However, this doesn't work due to two reasons:

1. In contrast to GNU assembler, the integrated assembler will not set default section flags to .text sections. As a result, SHF_ALLOC section flag is missing from the output of the integrated assembler.

2. ld.gold will ignore the sections without SHF_ALLOC. Consequently, .text.__cxa_end_cleanup section is not considered as a root set. ld.gold will discard .text.__cxa_end_cleanup_impl section and emit a warning for the relocations referring to the discarded section (which will be treated as an error.)

This is the reason why it only happens when we are compiling with integrated assembler.

A temporary workaround is to specify the section flags explicitly:
http://reviews.llvm.org/D15613

I am working on a long-term solution for lib/MC/MCParser so that integrated assembler can be more compatible with GNU assembler.

This commit should fix the problem:
http://reviews.llvm.org/rL256241

Change status to RESOLVED.

Feel free to re-open and let me know if the commit didn't fix the problem.

mentioned in issue #21794