simd constexpr vec: more improvement

g-truc · Sep 15, 2024 · 5716d8e · 5716d8e
1 parent d5f8676
commit 5716d8e
Show file tree

Hide file tree

Showing 3 changed files with 19 additions and 17 deletions.
diff --git a/glm/detail/simd_constexpr/element.hpp b/glm/detail/simd_constexpr/element.hpp
@@ -3,19 +3,19 @@ namespace glm::detail
 {
 	consteval bool NotEmpty(length_t I, length_t L) { return I <= L; }
 	struct Empty {};
-	struct RowTwo {
+	struct  GLM_TRIVIAL RowTwo {
 		[[no_unique_address]] Empty y; [[no_unique_address]] Empty g; [[no_unique_address]] Empty t;
 	};
-	struct RowThree {
+	struct GLM_TRIVIAL RowThree {
 		[[no_unique_address]] Empty z; [[no_unique_address]] Empty b; [[no_unique_address]] Empty p;
 	};
-	struct RowFour {
+	struct GLM_TRIVIAL RowFour {
 		[[no_unique_address]] Empty w; [[no_unique_address]] Empty a; [[no_unique_address]] Empty q;
 	};
 	template <qualifier Q, typename T, length_t L>
 	struct ElementCollection;
 	template <qualifier Q, typename T>
-	struct ElementCollection<Q, T, 4> {
+	struct  GLM_TRIVIAL ElementCollection<Q, T, 4> {
 		using data_t = typename detail::storage<4, T, detail::is_aligned<Q>::value>::type;
 		union
 		{
@@ -31,7 +31,7 @@ namespace glm::detail
 
 
 	template <qualifier Q, typename T>
-	struct ElementCollection<Q, T, 3> : RowFour {
+	struct GLM_TRIVIAL ElementCollection<Q, T, 3> : RowFour {
 		using data_t = typename detail::storage<3, T, detail::is_aligned<Q>::value>::type;
 		using RowFour::w;
 		using RowFour::a;
@@ -47,7 +47,7 @@ namespace glm::detail
 		};
 	};
 	template <qualifier Q, typename T>
-	struct ElementCollection<Q, T, 2> : RowThree, RowFour {
+	struct GLM_TRIVIAL ElementCollection<Q, T, 2> : RowThree, RowFour {
 		using data_t = typename detail::storage<2, T, detail::is_aligned<Q>::value>::type;
 		using RowThree::z;
 		using RowThree::b;
@@ -65,7 +65,7 @@ namespace glm::detail
 		};
 	};
 	template <qualifier Q, typename T>
-	struct ElementCollection<Q, T, 1> : RowTwo, RowThree, RowFour {
+	struct GLM_TRIVIAL ElementCollection<Q, T, 1> : RowTwo, RowThree, RowFour {
 		using data_t = typename detail::storage<1, T, detail::is_aligned<Q>::value>::type;
 		using RowTwo::y;
 		using RowTwo::g;

diff --git a/glm/detail/simd_constexpr/simd_helpers.inl b/glm/detail/simd_constexpr/simd_helpers.inl
@@ -69,7 +69,7 @@ namespace glm::detail
 		{
 			//assuming that number of scalars is always the same as the length of the to-be-constructed vector
 			gcc_vec_t v;
-			std::array<T, sizeof...(scalars)> pack{scalars...};
+			std::array<T, sizeof...(scalars)> pack{static_cast<T>(scalars)...};
 			for (int i = 0; i != sizeof...(scalars); i++ ) {
 				v[i] = pack[i];
 				pack[i].T::~T();

diff --git a/glm/detail/simd_constexpr/vec.hpp b/glm/detail/simd_constexpr/vec.hpp
@@ -16,6 +16,11 @@
 #include <ranges>
 namespace glm
 {
+#ifdef __clang__
+#define GLM_TRIVIAL __attribute__((trivial_abi))
+#else
+#define GLM_TRIVIAL
+#endif
 	template <length_t L>
 	concept NotVec1 = !std::is_same_v<std::integral_constant<length_t, L>, std::integral_constant<length_t, 1>>;
 	template <qualifier Q>
@@ -119,7 +124,7 @@ namespace glm
 	template <length_t L, typename T, qualifier Q>
 	using EC = detail::ElementCollection<Q, T, L>;
 	template<length_t L, typename T, qualifier Q>
-	struct vec : detail::ElementCollection<Q, T, L>
+	struct vec GLM_TRIVIAL : detail::ElementCollection<Q, T, L>
 	{
 		// -- Data --
 		using detail::ElementCollection<Q, T, L>::x;
@@ -282,7 +287,7 @@ namespace glm
 			}
 		}
 		template <length_t len>
-		using RetArr = T[len];
+		using RetArr = std::array<T, len>;
 
 		template <typename Vs0>
 		static constexpr length_t ctor_mixed_constexpr_single_get_length()
@@ -353,19 +358,16 @@ namespace glm
 					const auto params = std::tuple{vecOrScalar...};
 
 					const auto arr = ctor_mixed_constexpr_single(std::get<0>(params));
-					if (arr) [[likely]]
-						std::memcpy(aa.a.p.begin()+i, arr, sizeof(T)*lengths[0]);
+					std::ranges::copy(arr.cbegin(), aa.a.p.begin()+i, aa.a.p.begin()+i + sizeof(T)*lengths[0]);
 					constexpr auto i2 = i + lengths[0];
 
 					if constexpr (sizeof...(VecOrScalar) > 1) {
 						const auto arr2 = ctor_mixed_constexpr_single(std::get<1>(params));
-						if (arr2) [[likely]]
-							std::memcpy(aa.a.p.begin()+i2, arr2, sizeof(T)*lengths[1]);
+						std::ranges::copy(arr2.cbegin(), aa.a.p.begin()+i2, aa.a.p.begin()+i2 + sizeof(T)*lengths[1]);
 						constexpr auto i3 = i2 + lengths[1];
 						if constexpr (sizeof...(VecOrScalar) > 2) {
 							const auto arr3 = ctor_mixed_constexpr_single(std::get<2>(params));
-							if (arr3) [[likely]]
-								std::memcpy(aa.a.p.begin()+i3, arr3, sizeof(T)*lengths[2]);
+							std::ranges::copy(arr3.cbegin(), aa.a.p.begin()+i3, aa.a.p.begin()+i3 + sizeof(T)*lengths[2]);
 						}
 					}
 
@@ -460,7 +462,7 @@ namespace glm
 		}
 
 		template<typename Tx>
-		inline GLM_CONSTEXPR vec<L, T, Q> & operator*=(vec<L, Tx, Q> const& __restrict__ v) __restrict__
+		inline GLM_CONSTEXPR vec<L, T, Q> & operator*=(vec<L, Tx, Q> const& __restrict__ v)
 		{
 			if constexpr (L < 3) {
 				this->data *= v.data;