[Mono] AOT compiler segfaults when building tests

[steveisok]

Discovered in #56316

When we go to AOT our functional tests, the aot compiler crashes with a segmentation fault (11). When peeling back what's going on, the backtrace shows us:

* thread #1, name = 'tid_103', queue = 'com.apple.main-thread', stop reason = EXC_BAD_ACCESS (code=1, address=0x44)
    frame #0: 0x00000001001487e1 mono-aot-cross`create_jit_info(cfg=<unavailable>, method_to_compile=<unavailable>) at mini.c:2595:18 [opt]
   2592					int end_offset;
   2593					if (ec->handler_offset + ec->handler_len < header->code_size) {
   2594						tblock = cfg->cil_offset_to_bb [ec->handler_offset + ec->handler_len];
-> 2595						if (tblock->native_offset) {
   2596							end_offset = tblock->native_offset;
   2597						} else {
   2598							int j, end;

* thread #1, name = 'tid_103', queue = 'com.apple.main-thread', stop reason = EXC_BAD_ACCESS (code=1, address=0x44)
  * frame #0: 0x00000001001487e1 mono-aot-cross`create_jit_info(cfg=<unavailable>, method_to_compile=<unavailable>) at mini.c:2595:18 [opt]
    frame #1: 0x000000010014792c mono-aot-cross`mini_method_compile(method=<unavailable>, opts=<unavailable>, flags=JIT_FLAG_AOT | JIT_FLAG_FULL_AOT | JIT_FLAG_LLVM, parts=0, aot_method_index=179) at mini.c:3924:2 [opt]
    frame #2: 0x00000001001d9f1d mono-aot-cross`compile_method(acfg=0x0000000101008c00, method=0x0000000101a0fca0) at aot-compiler.c:8884:8 [opt]
    frame #3: 0x00000001001ca8cb mono-aot-cross`mono_compile_assembly [inlined] compile_methods(acfg=0x0000000101008c00) at aot-compiler.c:12356:3 [opt]
    frame #4: 0x00000001001ca74c mono-aot-cross`mono_compile_assembly(ass=0x0000000100f09970, opts=<unavailable>, aot_options=<unavailable>, global_aot_state=0x00007ffeefbff068) at aot-compiler.c:14125:2 [opt]
    frame #5: 0x00000001001b962c mono-aot-cross`mono_main at driver.c:1434:10 [opt]
    frame #6: 0x00000001001b9598 mono-aot-cross`mono_main(argc=<unavailable>, argv=<unavailable>) at driver.c:2682:3 [opt]
    frame #7: 0x0000000100003a73 mono-aot-cross`main [inlined] mono_main_with_options(argc=<unavailable>, argv=<unavailable>) at main.c:54:9 [opt]
    frame #8: 0x0000000100003a5f mono-aot-cross`main(argc=5, argv=0x00007ffeefbff2a0) at main.c:397:9 [opt]
    frame #9: 0x00007fff203fbf5d libdyld.dylib`start + 1
  thread #2, name = 'SGen worker'
    frame #0: 0x00007fff203adcde libsystem_kernel.dylib`__psynch_cvwait + 10
    frame #1: 0x00007fff203e0e49 libsystem_pthread.dylib`_pthread_cond_wait + 1298
    frame #2: 0x000000010013e5e3 mono-aot-cross`thread_func [inlined] mono_os_cond_wait(cond=<unavailable>, mutex=<unavailable>) at mono-os-mutex.h:219:8 [opt]
    frame #3: 0x000000010013e5c8 mono-aot-cross`thread_func at sgen-thread-pool.c:167:3 [opt]
    frame #4: 0x000000010013e4cb mono-aot-cross`thread_func(data=<unavailable>) at sgen-thread-pool.c:198:3 [opt]
    frame #5: 0x00007fff203e08fc libsystem_pthread.dylib`_pthread_start + 224
    frame #6: 0x00007fff203dc443 libsystem_pthread.dylib`thread_start + 15

If we go to the 2nd frame and print the important method details, we get:

method->name = "Trim"
method->klass->name_space = "System.Buffers"
method->klass->name = "TlsOverPerCoreLockedStacksArrayPool`1"

Which leads to this block that was changed in the PR #56316 https://github.com/dotnet/runtime/blob/96673df0d1ba24cf2be4ec9529a6ba54f7d97902/src/libraries/System.Private.CoreLib/src/System/Buffers/TlsOverPerCoreLockedStacksArrayPool.cs#L187-L280

@vargaz it looks like foreach may be causing problems under certain conditions.

[vargaz]

Which test suite is this ?
That PR has AOT failures on a lot of assemblies, not just corlib.

[steveisok]

The iOS and Android AOT functional tests.

https://github.com/dotnet/runtime/tree/main/src/tests/FunctionalTests/Android/Device_Emulator

[vargaz]

This was caused by a linker problem.

[stephentoub]

This was caused by a linker problem.

Isn't it still a problem? The other code generators are accepting this apparently; shouldn't the AOT compiler not seg fault in this case?

My PR is also still blocked until something is fixed somewhere.