ldmd2 Segmentation fault: 11 for macOS 15.4 arm64

[calvin2021y]

try v1.40.1 and v1.40.1-beta1, crash immediate.

ldc2 work.

./ldmd2 -v
Segmentation fault: 11

[remy-moueza]

I have the same issue. A compilation with ldc2 works. However the compiled executable still fails with a segmentation fault.

I have downloaded a nightly build to check if that issue was fixed. I still encounter segmentation faults.
I also have XCode installed and freshly udpated.

I tried to run it through lldb to see if I could get a lead on what could be wrong with my set up. I did not understand much, but here is what I get, it might be helpful :

hello.d:

module hello;

void main() {
}

> ~/dev/conf/ldc2-06b2e42a-osx-arm64/bin/ldc2 hello.d
> ./hello
fish: Job 1, './hello' terminated by signal SIGSEGV (Address boundary error)
> lldb ./hello
(lldb) target create "./hello"
Current executable set to '/tmp/hello' (arm64).
(lldb) run
Process 30568 launched: '/tmp/hello' (arm64)
Process 30568 stopped
* thread #1, queue = 'com.apple.main-thread', stop reason = EXC_BAD_ACCESS (code=1, address=0x3c1f99f5530)
    frame #0: 0x000000018ab20bf4 libsystem_pthread.dylib`pthread_getspecific + 4
libsystem_pthread.dylib`pthread_getspecific:
->  0x18ab20bf4 <+4>: ldr    x0, [x8, x0, lsl #3]
    0x18ab20bf8 <+8>: ret

libsystem_pthread.dylib`pthread_setspecific:
    0x18ab20bfc <+0>: sub    x8, x0, #0xa
    0x18ab20c00 <+4>: cmp    x8, #0x2f5
Target 0: (hello) stopped.
(lldb)

Maybe the Mac OS 15.4 pthread dynamic library is incompatible with the one used to compile ldc2, or XCode libraries are causing trouble. I haven't tried to remove XCode yet.

> lldb -v
lldb-1700.0.9.46
Apple Swift version 6.1 (swiftlang-6.1.0.110.21 clang-1700.0.13.3)

> clang -v
Apple clang version 17.0.0 (clang-1700.0.13.3)
Target: arm64-apple-darwin24.4.0
Thread model: posix
InstalledDir: /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin

[Geod24]

dlang/dmd#21126

[yanok]

FWIW, I think the reason why ldc2 works but ldmd doesn't is the former is linked against mimalloc. @calvin2021y Maybe try linking your binary with mimalloc as well? Would that be a workaround?

[remy-moueza]

I have had success with Sönke Ludwig's approach : using ldc-1.28, last compiler known to work on 15.4, compiling llvm then ldc-1.40 with a fix described in that comment : dlang/dmd#21126 (comment) .
The new ldc2 compiler and its ldmd2 front end work as well as their compiled programs, though most of my sources only use the standard library. That's good enough for me right now, however the discussion is not over and the final fix may differ.

I followed the compilation instructions from this page : https://wiki.dlang.org/Building_LDC_from_source. The sole difficulty I found was providing a path to the zstd library, which was installed through homebrew. I launched the ldc ninja build with the LIBRARY_PATH set to my zstd location: LIBRARY_PATH=$LIBRARY_PATH:/opt/homebrew/Cellar/zstd/1.5.7/lib ninja

From what I understand of the dlang issue thread discussion, Apple may have altered their usage of the higher 16 or more bits of the mentionned header.firstTLVKey variable, that supposedly weren't used. Recompiling the whole compiler might not even been strictly necessary, since only the druntime library is affected as commented there : dlang/dmd#21126 (comment) .

[kinke]

The (preliminary) upstream fix has just landed in master.

[calvin2021y]

hi @kinke , is this 1.41.0-beta1 (DMD v2.111.0, LLVM 19.1.7) include this fixed patch?

macOS 15.4.1 still crash for arm64&x86_64 ldmd2

[Rhys-T]

Hmm… yeah, it's doing that on my machine too. I can see that the fix (60079c3) made it into 1.41.0-beta1, but it looks like maybe the release binaries of that version were themselves built with an earlier LDC that didn't have the fix yet. Don't know how to tell for sure - is there any documentation on how the official binaries are produced?

[yanok]

@Rhys-T I believe release binaries are built by Github Actions. Usually they build a bootstrap compiler from the released version first, such that the final binary is built by itself. But specifically for OSX/arm64 combination, compiling bootstrap LDC is disabled (see

ldc/.github/workflows/main.yml

Line 170 in 20d22b1

# FIXME: on macOS, mixing newer LLVM from bootstrap LDC with older LLVM from Xcode clang is problematic,

), so the final binary is built using the host LDC, so if it is affected by this bug, the final release binary will be affected as well.

BTW, I believe the bug with non-deterministic crashes that was the reason to disable using the bootstrap compiler, also goes away without mimalloc.

EDIT: we disable the bootstrap compiler for both Intel and ARM OSX