Disabling COMPILER_RT_CRT_USE_EH_FRAME_REGISTRY by default breaks LLDB tests on AArch64

[mysterymath]

PR #83201 produced a breakage in the LLDB test suite for AArch64 on a Fuchsia builder:
https://luci-milo.appspot.com/ui/p/fuchsia/builders/toolchain.ci/clang-linux-arm64/b8754880571862677073/overview

The precise relationship between the two is unclear; it may be that the artifacts built during the LLDB test suite are built with a configuration that bucks the expectations of that PR.

[llvmbot]

@llvm/issue-subscribers-lldb

Author: Daniel Thornburgh (mysterymath)

PR #83201 produced a breakage in the LLDB test suite for AArch64 on a Fuchsia builder: https://luci-milo.appspot.com/ui/p/fuchsia/builders/toolchain.ci/clang-linux-arm64/b8754880571862677073/overview

The precise relationship between the two is unclear; it may be that the artifacts built during the LLDB test suite are built with a configuration that bucks the expectations of that PR.

[mysterymath]

@MaskRay

[MaskRay]

Just saw your Fuchsia report. Since LLVM libunwind implements __register_frame_info as an empty stub, and the test doesn't look like a statically linked executable, it strongly suggests to me that it might be an configuration problem on Fuchsia side, which I know nothing about.

Can you help investigate?

[MaskRay]

Ping @mysterymath : This looks like a Fuchsia configuration problem that should not block relanding #83201

[mysterymath]

Just saw your Fuchsia report. Since LLVM libunwind implements __register_frame_info as an empty stub, and the test doesn't look like a statically linked executable, it strongly suggests to me that it might be an configuration problem on Fuchsia side, which I know nothing about.

Sorry for dropping this one; does something specific suggest that the test isn't statically linked? IIRC the tests should build binaries using the clang built with the toolchain, and that clang is configured to statically link against compiler-rt.

[MaskRay]

Just saw your Fuchsia report. Since LLVM libunwind implements __register_frame_info as an empty stub, and the test doesn't look like a statically linked executable, it strongly suggests to me that it might be an configuration problem on Fuchsia side, which I know nothing about.

Sorry for dropping this one; does something specific suggest that the test isn't statically linked? IIRC the tests should build binaries using the clang built with the toolchain, and that clang is configured to statically link against compiler-rt.

Sorry for not replying to your comment in time. This feels very odd and I don't know what happened on the Fuchsia bot. Could you debug this?

In addition, LLVM libunwind implements __register_frame_info as an empty stub.