Why do I even bother

src/runtime_intrinsics.c

@@ -249,23 +249,25 @@ float julia_half_to_float(uint16_t param) {
 
 // starting with GCC 12 and Clang 15, we have _Float16 on most platforms
 // (but not on Windows; this may be a bug in the MSYS2 GCC compilers)
+// (but not on x86-darwin, because they reverted the ABI break only for libcalls https://github.com/llvm/llvm-project/commit/2bcf51c7f82ca7752d1bba390a2e0cb5fdd05ca9)
 #if ((defined(__GNUC__) && __GNUC__ > 11) || \
      (defined(__clang__) && __clang_major__ > 14)) && \
     !defined(_CPU_PPC64_) && !defined(_CPU_PPC_) && \
-    !defined(_OS_WINDOWS_)
+    !defined(_OS_WINDOWS_) && !(defined(_CPU_X86_64_) && defined(_OS_DARWIN_))
     #define FLOAT16_TYPE _Float16
     #define FLOAT16_TO_UINT16(x) (*(uint16_t*)&(x))
     #define FLOAT16_FROM_UINT16(x) (*(_Float16*)&(x))
 // on older compilers, we need to emulate the platform-specific ABI
-#elif defined(_CPU_X86_) || (defined(_CPU_X86_64_) && !defined(_OS_WINDOWS_))
+#elif defined(_CPU_X86_) || (defined(_CPU_X86_64_) && !(defined(_OS_WINDOWS_) || defined(_OS_DARWIN_)))
     // on x86, we can use __m128; except on Windows where x64 calling
     // conventions expect to pass __m128 by reference.
     #define FLOAT16_TYPE __m128
     #define FLOAT16_TO_UINT16(x) take_from_xmm(x)
     #define FLOAT16_FROM_UINT16(x) return_in_xmm(x)
-#elif defined(_CPU_PPC64_) || defined(_CPU_PPC_)
+#elif defined(_CPU_PPC64_) || defined(_CPU_PPC_) || defined(_OS_DARWIN_)
     // on PPC, pass Float16 as if it were an integer, similar to the old x86 ABI
     // before _Float16
+    // on Darwin we need to use uint16_t because libcalls use it
     #define FLOAT16_TYPE uint16_t
     #define FLOAT16_TO_UINT16(x) (x)
     #define FLOAT16_FROM_UINT16(x) (x)