[SYCL] Workaround for seg fault in vec::convert<> for OpenCL CPU at…

… O0 (#14498)

There seems to be a bug in CodeGen for OpenCL CPU which causes a
seg-fault in `vec::convert<>`, when converting to `vec<long, 8>` at O0.
To unblock #14317, this PR proposes a
non-functional change as a workaround for the CodeGen bug.

sycl/include/sycl/vector_preview.hpp

@@ -489,9 +489,9 @@ class vec : public detail::vec_arith<DataT, NumElements> {
           !std::is_same_v<convertT, bool>;
 
       if constexpr (canUseNativeVectorConvert) {
-        Result.m_Data = sycl::bit_cast<decltype(Result.m_Data)>(
-            detail::convertImpl<T, R, roundingMode, NumElements, OpenCLVecT,
-                                OpenCLVecR>(NativeVector));
+        auto val = detail::convertImpl<T, R, roundingMode, NumElements, OpenCLVecT,
+                                OpenCLVecR>(NativeVector);
+        Result.m_Data = sycl::bit_cast<decltype(Result.m_Data)>(val);
       } else
 #endif // __SYCL_DEVICE_ONLY__
       {