[HIP] Fix __clang_hip_cmath.hip for ambiguity (#101341)

If there is a type T which can be converted to both float and double etc
but itself is not specialized for __numeric_type, and it is called for
math functions eg. fma, it will cause ambiguity with test function of
__numeric_type.

Since test is not template, this error is not bypassed by SFINAE. This
is a design flaw of __numeric_type. This patch fixes clang wrapper
header to use SFINAE to avoid such ambiguity.

Fixes: SWDEV-461604

Fixes: #101239

Author: yxsamliu

clang/lib/Headers/__clang_hip_cmath.h

@@ -395,7 +395,12 @@ template <class _Tp> struct __numeric_type {
   // No support for long double, use double instead.
   static double __test(long double);
 
-  typedef decltype(__test(declval<_Tp>())) type;
+  template <typename _U>
+  static auto __test_impl(int) -> decltype(__test(declval<_U>()));
+
+  template <typename _U> static void __test_impl(...);
+
+  typedef decltype(__test_impl<_Tp>(0)) type;
   static const bool value = !is_same<type, void>::value;
 };
 

clang/test/Headers/__clang_hip_cmath.hip

@@ -87,3 +87,22 @@ extern "C" __device__ float test_sin_f32(float x) {
 extern "C" __device__ float test_cos_f32(float x) {
   return cos(x);
 }
+
+// Check user defined type which can be converted to float and double but not
+// specializes __numeric_type will not cause ambiguity diagnostics.
+struct user_bfloat16 {
+  __host__ __device__ user_bfloat16(float);
+  operator float();
+  operator double();
+};
+
+namespace user_namespace {
+  __device__ user_bfloat16 fma(const user_bfloat16 a, const user_bfloat16 b, const user_bfloat16 c) {
+    return a;
+  }
+
+  __global__ void test_fma() {
+    user_bfloat16 a = 1.0f, b = 2.0f;
+    fma(a, b, b);
+  }
+}