CC @andrurogerz

@llvm/pr-subscribers-llvm-support

Author: Martin Storsjö (mstorsjo)

diff --git a/llvm/include/llvm/Support/Atomic.h b/llvm/include/llvm/Support/Atomic.h
index a8445fddc1a85..4dcef19359ce5 100644
--- a/llvm/include/llvm/Support/Atomic.h
+++ b/llvm/include/llvm/Support/Atomic.h
@@ -17,6 +17,7 @@
 #ifndef LLVM_SUPPORT_ATOMIC_H
 #define LLVM_SUPPORT_ATOMIC_H
 
+#include "llvm/Support/Compiler.h"
 #include "llvm/Support/DataTypes.h"
 
 // Windows will at times define MemoryFence.
@@ -26,16 +27,16 @@
 
 namespace llvm {
   namespace sys {
-    void MemoryFence();
+    LLVM_ABI void MemoryFence();
 
 #ifdef _MSC_VER
     typedef long cas_flag;
 #else
     typedef uint32_t cas_flag;
 #endif
-    cas_flag CompareAndSwap(volatile cas_flag* ptr,
-                            cas_flag new_value,
-                            cas_flag old_value);
+    LLVM_ABI cas_flag CompareAndSwap(volatile cas_flag* ptr,
+                                     cas_flag new_value,
+                                     cas_flag old_value);
   }
 }
 

✅ With the latest revision this PR passed the C/C++ code formatter.

lgtm, thanks!

@andrurogerz As for why this had been missed; I had the understanding that the current annotations were made with some automated tool - shouldn't that have caught these declarations? They should be available and exported in all builds, even if most users (with LLVM_THREADING_USE_STD_CALL_ONCE=1) don't end up using them?

As for why this had been missed; I had the understanding that the current annotations were made with some automated tool - shouldn't that have caught these declarations?

Correct, the annotations are automatically added by the IDS tool, but it is run on each header individually and this is one was missed. There will likely be more that pop-up over the coming weeks.

FWIW, I may have spoken too soon, re

This fixes building LLD for mingw targets with LLVM_BUILD_LLVM_DYLIB_VIS=ON.

As in, it does fix it, if building with Clang, but it still mostly uses __attribute__((visibility("default"))) and hidden in that configuration (and doesn't work as is out of the box in builds with GCC) - it did seem a bit too good to be true. It will require a number of further tweaks to take these annotations fully in use in the form of dllexport/import for mingw - I'll try it out soon.

But this fix is one step on the way in any case.