.NET crashes when compiled with clang 21.1 on Fedora 43

[tmds]

Our Fedora 43 CI shows some issues when trying to compile runtime repo with clang 21.1

  /home/tester/workdir/runtime/src/native/libs/System.Net.Security.Native/pal_gssapi.c:247:11: error: implicit conversion when initializing 'char *' with an expression of type 'void *' is not permitted in C++ [-Werror,-Wimplicit-void-ptr-cast]
    247 |     char* ptrSlash = memchr(inputName, '/', inputNameLen);
        |           ^          ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

  /home/tester/workdir/runtime/src/native/libs/System.Native/pal_networking.c:396:17: error: jump from this goto statement to its label is incompatible with C++ [-Werror,-Wjump-misses-init]
    396 |                 goto cleanup;
        |                 ^
  /home/tester/workdir/runtime/src/native/libs/System.Native/pal_networking.c:411:10: note: jump bypasses variable initialization
    411 |     bool includeIPv6Loopback = true;
        |          ^
  /home/tester/workdir/runtime/src/native/libs/System.Native/pal_networking.c:410:10: note: jump bypasses variable initialization
    410 |     bool includeIPv4Loopback = true;
        |          ^

  /home/tester/workdir/runtime/src/native/libs/System.Native/pal_networking.c:1518:34: error: implicit conversion when assigning to 'uint8_t *' (aka 'unsigned char *') from type 'void *' is not permitted in C++ [-Werror,-Wimplicit-void-ptr-cast]
   1518 |     messageHeader->ControlBuffer = (void*)buffer;
        |                                  ^ ~~~~~~~~~~~~~

cc @jkotas @omajid

[dotnet-policy-service]

Tagging subscribers to this area: @dotnet/area-infrastructure-libraries
See info in area-owners.md if you want to be subscribed.

[tmds]

From https://releases.llvm.org/21.1.0/tools/clang/docs/ReleaseNotes.html, the C compiler is now emitting some warnings for C++ incompatibilities (-Wc++-compat), like -Wimplicit-void-ptr-cast and -Wjump-misses-init.

When I ignore those with -Wno-c++-compat, then it issues one more error for:

runtime/src/coreclr/debug/di/rspriv.h

Line 6380 in c1b153a

static const CorDebugUserState kInvalidUserState = CorDebugUserState(-1);

  /root/runtime/src/coreclr/debug/di/rspriv.h:6380:36: error: in-class initializer for static data member is not a constant expression
   6380 |     static const CorDebugUserState kInvalidUserState = CorDebugUserState(-1);
        |                                    ^                   ~~~~~~~~~~~~~~~~~~~~~
  /root/runtime/src/coreclr/debug/di/rspriv.h:6380:56: note: integer value -1 is outside the valid range of values [0, 511] for the enumeration type 'CorDebugUserState'
   6380 |     static const CorDebugUserState kInvalidUserState = CorDebugUserState(-1);
        |                                                        

[omajid]

Duplicate of #116997 ?

[tmds]

Yes. It might have been reported for an older clang. I didn't run into the CorDebugUserState error until after addressing the 21.1 c++-compat warnings.

These are the changes I made to deal with the compiler errors:

diff --git a/eng/native/configurecompiler.cmake b/eng/native/configurecompiler.cmake
index d54d9b6803b..9438cb5af7e 100644
--- a/eng/native/configurecompiler.cmake
+++ b/eng/native/configurecompiler.cmake
@@ -574,6 +574,7 @@ if (CLR_CMAKE_HOST_UNIX OR CLR_CMAKE_HOST_WASI)
   add_compile_options(-Wno-unused-function)
   add_compile_options(-Wno-tautological-compare)
   add_compile_options(-Wno-unknown-pragmas)
+  add_compile_options(-Wno-c++-compat)
 
   # Explicitly enabled warnings
   check_c_compiler_flag(-Wimplicit-fallthrough COMPILER_SUPPORTS_W_IMPLICIT_FALLTHROUGH)
diff --git a/src/coreclr/pal/prebuilt/inc/cordebug.h b/src/coreclr/pal/prebuilt/inc/cordebug.h
index 78b86c69ee9..1672d602d99 100644
--- a/src/coreclr/pal/prebuilt/inc/cordebug.h
+++ b/src/coreclr/pal/prebuilt/inc/cordebug.h
@@ -10042,6 +10042,7 @@ EXTERN_C const IID IID_ICorDebugRegisterSet2;
 typedef
 enum CorDebugUserState
     {
+        CORDEBUG_INVALID = -1,
         USER_STOP_REQUESTED = 0x1,
         USER_SUSPEND_REQUESTED  = 0x2,
         USER_BACKGROUND = 0x4,

[hamarb123]

I ran into this when trying to compile #117405 locally the other day btw - only just found this issue with the solution - would be nice if it was in normal dotnet/runtime so it can be compiled on macOS when the llvm homebrew package is installed :)

[tmds]

I put it here as a patch tot help people unblock themselves. I think the actual fix might be a bit more refined.

[jkotas]

I think the actual fix might be a bit more refined.

Yeah, I think we want to add the casts to make compilers happy instead of warning disable.

Are you interested in submitting PR with the fix?

[corngood]

We hit this in nixpkgs for the darwin runtime/sdk, presumably after a clang update (NixOS/nixpkgs#449614).

Applying the CORDEBUG_INVALID change above fixes the build, but for whatever reason I didn't have to add -Wno-c++-compat.

I'm getting intermittent segfaults from the runtime, for example when running dotnet new console. I'm not sure if this is related. Has anyone been able to test a runtime built with clang-21? I'll try to verify that it's clang version related, but I don't have root access on the darwin machine I'm testing, so my debugging options are a little limited.