[clang-cl] arm_neon.h: cannot mangle built-in type __mfp8

[tycho]

On the release/20.x and main branches, when targeting ARM64 on clang-cl (--target=aarch64-pc-windows-msvc --driver-mode=cl):

https://gcc.godbolt.org/z/Mx1Y8hze4

In file included from <source>:1:
/opt/compiler-explorer/clang-trunk-20250129/lib/clang/20/include/arm_neon.h(41124,1): error: cannot mangle this built-in type:  __mfp8 yet
 41124 | __ai __attribute__((target("fp8,neon"))) bfloat16x8_t vcvt1_bf16_mf8_fpm(mfloat8x8_t __p0, fpm_t __p1) {
       | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 41125 |   bfloat16x8_t __ret;
       |   ~~~~~~~~~~~~~~~~~~~
 41126 |   __ret = (bfloat16x8_t) __builtin_neon_vcvt1_bf16_mf8_fpm(__p0, __p1);
       |   ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 41127 |   return __ret;
       |   ~~~~~~~~~~~~~
 41128 | }
       | ~
/opt/compiler-explorer/clang-trunk-20250129/lib/clang/20/include/arm_neon.h(124,14): note: expanded from macro '__ai'
  124 | #define __ai static __inline__ __attribute__((__always_inline__, __nodebug__))
      |              ^
1 error generated.

Just including <arm_neon.h> causes a compile error, due to the __mfp8 type. This prevents compiling any Windows ARM64 code that includes the intrinsics header.

Can someone get this fixed before LLVM 20 release please?

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: Steven Noonan (tycho)

On the `release/20.x` and `main` branches, when targeting ARM64 on clang-cl (`--target=aarch64-pc-windows-msvc --driver-mode=cl`):

https://gcc.godbolt.org/z/Mx1Y8hze4

In file included from <source>:1:
/opt/compiler-explorer/clang-trunk-20250129/lib/clang/20/include/arm_neon.h(41124,1): error: cannot mangle this built-in type:  __mfp8 yet
 41124 | __ai __attribute__((target("fp8,neon"))) bfloat16x8_t vcvt1_bf16_mf8_fpm(mfloat8x8_t __p0, fpm_t __p1) {
       | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 41125 |   bfloat16x8_t __ret;
       |   ~~~~~~~~~~~~~~~~~~~
 41126 |   __ret = (bfloat16x8_t) __builtin_neon_vcvt1_bf16_mf8_fpm(__p0, __p1);
       |   ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 41127 |   return __ret;
       |   ~~~~~~~~~~~~~
 41128 | }
       | ~
/opt/compiler-explorer/clang-trunk-20250129/lib/clang/20/include/arm_neon.h(124,14): note: expanded from macro '__ai'
  124 | #define __ai static __inline__ __attribute__((__always_inline__, __nodebug__))
      |              ^
1 error generated.

Just including <arm_neon.h> causes a compile error, due to the __mfp8 type. This prevents compiling any Windows ARM64 code that includes the intrinsics header.

Can someone get this fixed before LLVM 20 release please?

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Steven Noonan (tycho)

On the `release/20.x` and `main` branches, when targeting ARM64 on clang-cl (`--target=aarch64-pc-windows-msvc --driver-mode=cl`):

https://gcc.godbolt.org/z/Mx1Y8hze4

In file included from <source>:1:
/opt/compiler-explorer/clang-trunk-20250129/lib/clang/20/include/arm_neon.h(41124,1): error: cannot mangle this built-in type:  __mfp8 yet
 41124 | __ai __attribute__((target("fp8,neon"))) bfloat16x8_t vcvt1_bf16_mf8_fpm(mfloat8x8_t __p0, fpm_t __p1) {
       | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 41125 |   bfloat16x8_t __ret;
       |   ~~~~~~~~~~~~~~~~~~~
 41126 |   __ret = (bfloat16x8_t) __builtin_neon_vcvt1_bf16_mf8_fpm(__p0, __p1);
       |   ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 41127 |   return __ret;
       |   ~~~~~~~~~~~~~
 41128 | }
       | ~
/opt/compiler-explorer/clang-trunk-20250129/lib/clang/20/include/arm_neon.h(124,14): note: expanded from macro '__ai'
  124 | #define __ai static __inline__ __attribute__((__always_inline__, __nodebug__))
      |              ^
1 error generated.

Just including <arm_neon.h> causes a compile error, due to the __mfp8 type. This prevents compiling any Windows ARM64 code that includes the intrinsics header.

Can someone get this fixed before LLVM 20 release please?