[SYCL] Fix include loop and use-before-def

sycl/include/sycl/bit_cast.hpp

@@ -8,6 +8,8 @@
 
 #pragma once
 
+#include <sycl/detail/memcpy.hpp>
+
 #include <type_traits>
 
 #if __cpp_lib_bit_cast
@@ -17,11 +19,6 @@
 namespace sycl {
 __SYCL_INLINE_VER_NAMESPACE(_V1) {
 
-// forward decl
-namespace detail {
-inline void memcpy(void *Dst, const void *Src, std::size_t Size);
-}
-
 template <typename To, typename From>
 #if __cpp_lib_bit_cast || __has_builtin(__builtin_bit_cast)
 constexpr

sycl/include/sycl/detail/generic_type_lists.hpp

@@ -11,6 +11,7 @@
 #include <sycl/access/access.hpp>
 #include <sycl/detail/stl_type_traits.hpp>
 #include <sycl/detail/type_list.hpp>
+#include <sycl/half_type.hpp>
 
 #include <cstddef>
 
@@ -22,10 +23,6 @@ namespace sycl {
 __SYCL_INLINE_VER_NAMESPACE(_V1) {
 template <typename T, int N> class vec;
 template <typename Type, std::size_t NumElements> class marray;
-namespace detail::half_impl {
-class half;
-} // namespace detail::half_impl
-using half = detail::half_impl::half;
 } // __SYCL_INLINE_VER_NAMESPACE(_V1)
 } // namespace sycl
 

sycl/include/sycl/detail/helpers.hpp

@@ -13,6 +13,7 @@
 #include <sycl/access/access.hpp>
 #include <sycl/detail/common.hpp>
 #include <sycl/detail/export.hpp>
+#include <sycl/detail/memcpy.hpp>
 #include <sycl/detail/pi.hpp>
 #include <sycl/detail/type_traits.hpp>
 
@@ -35,13 +36,6 @@ template <typename Type, std::size_t NumElements> class marray;
 enum class memory_order;
 
 namespace detail {
-inline void memcpy(void *Dst, const void *Src, size_t Size) {
-  char *Destination = reinterpret_cast<char *>(Dst);
-  const char *Source = reinterpret_cast<const char *>(Src);
-  for (size_t I = 0; I < Size; ++I) {
-    Destination[I] = Source[I];
-  }
-}
 
 class context_impl;
 // The function returns list of events that can be passed to OpenCL API as

sycl/include/sycl/detail/memcpy.hpp

@@ -0,0 +1,25 @@
+//==---------------- memcpy.hpp - SYCL memcpy --------------------------==//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#pragma once
+
+#include <cstddef>
+
+namespace sycl {
+__SYCL_INLINE_VER_NAMESPACE(_V1) {
+namespace detail {
+inline void memcpy(void *Dst, const void *Src, size_t Size) {
+  char *Destination = reinterpret_cast<char *>(Dst);
+  const char *Source = reinterpret_cast<const char *>(Src);
+  for (size_t I = 0; I < Size; ++I) {
+    Destination[I] = Source[I];
+  }
+}
+} // namespace detail
+} // __SYCL_INLINE_VER_NAMESPACE(_V1)
+} // namespace sycl

sycl/include/sycl/detail/type_traits.hpp

@@ -12,6 +12,7 @@
 #include <sycl/detail/generic_type_lists.hpp>
 #include <sycl/detail/stl_type_traits.hpp>
 #include <sycl/detail/type_list.hpp>
+#include <sycl/detail/vector_traits.hpp>
 
 #include <array>
 #include <tuple>
@@ -85,7 +86,6 @@ template <typename T, typename R>
 using copy_cv_qualifiers_t = typename copy_cv_qualifiers<T, R>::type;
 
 template <int V> using int_constant = std::integral_constant<int, V>;
-
 // vector_size
 // scalars are interpreted as a vector of 1 length.
 template <typename T> struct vector_size_impl : int_constant<1> {};
@@ -94,16 +94,6 @@ struct vector_size_impl<vec<T, N>> : int_constant<N> {};
 template <typename T>
 struct vector_size : vector_size_impl<remove_cv_t<remove_reference_t<T>>> {};
 
-// 4.10.2.6 Memory layout and alignment
-template <typename T, int N>
-struct vector_alignment_impl
-    : conditional_t<N == 3, int_constant<sizeof(T) * 4>,
-                    int_constant<sizeof(T) * N>> {};
-
-template <typename T, int N>
-struct vector_alignment
-    : vector_alignment_impl<remove_cv_t<remove_reference_t<T>>, N> {};
-
 // vector_element
 template <typename T> struct vector_element_impl;
 template <typename T>

sycl/include/sycl/detail/vector_traits.hpp

@@ -0,0 +1,30 @@
+//==----------- vector_traits.hpp - SYCL vector size queries --------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#pragma once
+
+#include <sycl/detail/stl_type_traits.hpp>
+
+#include <type_traits>
+
+namespace sycl {
+__SYCL_INLINE_VER_NAMESPACE(_V1) {
+namespace detail {
+
+// 4.10.2.6 Memory layout and alignment
+template <typename T, int N>
+struct vector_alignment_impl
+    : conditional_t<N == 3, std::integral_constant<int, sizeof(T) * 4>,
+                    std::integral_constant<int, sizeof(T) * N>> {};
+
+template <typename T, int N>
+struct vector_alignment
+    : vector_alignment_impl<remove_cv_t<remove_reference_t<T>>, N> {};
+} // namespace detail
+} // __SYCL_INLINE_VER_NAMESPACE(_V1)
+} // namespace sycl

sycl/include/sycl/ext/intel/esimd.hpp

@@ -79,6 +79,7 @@
 #pragma clang diagnostic ignored "-Wpsabi"
 #endif // !defined(__SYCL_DEVICE_ONLY__) && defined(__clang__)
 
+#include <sycl/detail/type_traits.hpp>
 #include <sycl/ext/intel/esimd/alt_ui.hpp>
 #include <sycl/ext/intel/esimd/common.hpp>
 #include <sycl/ext/intel/esimd/detail/bfloat16_type_traits.hpp>

sycl/include/sycl/half_type.hpp

@@ -9,10 +9,11 @@
 #pragma once
 
 #include <sycl/aspects.hpp>
+#include <sycl/bit_cast.hpp>
 #include <sycl/detail/defines.hpp>
 #include <sycl/detail/export.hpp>
 #include <sycl/detail/iostream_proxy.hpp>
-#include <sycl/detail/type_traits.hpp>
+#include <sycl/detail/vector_traits.hpp>
 
 #include <functional>
 #include <limits>
@@ -33,6 +34,10 @@
 
 namespace sycl {
 __SYCL_INLINE_VER_NAMESPACE(_V1) {
+namespace detail::half_impl {
+class half;
+}
+using half = detail::half_impl::half;
 
 namespace ext::intel::esimd::detail {
 class WrapperElementTypeProxy;
@@ -248,8 +253,8 @@ using BIsRepresentationT = half;
 // hood. As a result half values will be converted to the integer and passed
 // as a kernel argument which is expected to be floating point number.
 template <int NumElements> struct half_vec {
-  alignas(detail::vector_alignment<StorageT, NumElements>::value)
-      StorageT s[NumElements];
+  alignas(
+      vector_alignment<StorageT, NumElements>::value) StorageT s[NumElements];
 
   __SYCL_CONSTEXPR_HALF half_vec() : s{0.0f} { initialize_data(); }
   constexpr void initialize_data() {