Add shift by multiple constant

include/xsimd/arch/common/xsimd_common_arithmetic.hpp

@@ -17,6 +17,7 @@
 #include <type_traits>
 
 #include "../../types/xsimd_batch_constant.hpp"
+#include "../xsimd_common_fwd.hpp"
 #include "./xsimd_common_details.hpp"
 
 namespace xsimd
@@ -27,14 +28,23 @@ namespace xsimd
 
         using namespace types;
 
-        // bitwise_lshift
+        // bitwise_lshift multiple (dynamic)
         template <class A, class T, class /*=typename std::enable_if<std::is_integral<T>::value>::type*/>
         XSIMD_INLINE batch<T, A> bitwise_lshift(batch<T, A> const& self, batch<T, A> const& other, requires_arch<common>) noexcept
         {
             return detail::apply([](T x, T y) noexcept
                                  { return x << y; },
                                  self, other);
         }
+
+        // bitwise_lshift multiple (constant)
+        template <class A, class T, T... Vals, detail::enable_integral_t<T>>
+        XSIMD_INLINE batch<T, A> bitwise_lshift(batch<T, A> const& lhs, batch_constant<T, A, Vals...> const& rhs, requires_arch<common> req) noexcept
+        {
+            return bitwise_lshift(lhs, rhs.as_batch(), req);
+        }
+
+        // bitwise_lshift single (constant)
         template <size_t shift, class A, class T, class /*=typename std::enable_if<std::is_integral<T>::value>::type*/>
         XSIMD_INLINE batch<T, A> bitwise_lshift(batch<T, A> const& self, requires_arch<common>) noexcept
         {

include/xsimd/arch/utils/shifts.hpp

@@ -0,0 +1,71 @@
+/***************************************************************************
+ * Copyright (c) Johan Mabille, Sylvain Corlay, Wolf Vollprecht and         *
+ * Martin Renou                                                             *
+ * Copyright (c) QuantStack                                                 *
+ * Copyright (c) Serge Guelton                                              *
+ * Copyright (c) Marco Barbone                                              *
+ *                                                                          *
+ * Distributed under the terms of the BSD 3-Clause License.                 *
+ *                                                                          *
+ * The full license is in the file LICENSE, distributed with this software. *
+ ****************************************************************************/
+
+#ifndef XSIMD_UTILS_SHIFTS_HPP
+#define XSIMD_UTILS_SHIFTS_HPP
+
+#include "xsimd/config/xsimd_inline.hpp"
+#include "xsimd/types/xsimd_batch.hpp"
+
+namespace xsimd
+{
+    namespace kernel
+    {
+        namespace utils
+        {
+            template <typename I, I offset, I length, I... Vs>
+            struct select_stride
+            {
+                static constexpr I values_array[] = { Vs... };
+
+                template <typename K>
+                static constexpr K get(K i, K)
+                {
+                    return static_cast<K>(values_array[length * i + offset]);
+                }
+            };
+
+            template <typename I>
+            constexpr I lsb_mask(I bit_index)
+            {
+                return static_cast<I>((I { 1 } << bit_index) - I { 1 });
+            }
+
+            template <class T, class T2, class A, T... Vs>
+            XSIMD_INLINE batch<T, A> bitwise_lshift_as_twice_larger(
+                batch<T, A> const& self, batch_constant<T, A, Vs...>) noexcept
+            {
+                static_assert(sizeof(T2) == 2 * sizeof(T), "One size must be twice the other");
+
+                const auto self2 = bitwise_cast<T2>(self);
+
+                // Lower byte: shift as twice the size and mask bits flowing to higher byte.
+                constexpr auto shifts_lo = make_batch_constant<T2, select_stride<T, 0, 2, Vs...>, A>();
+                constexpr auto mask_lo = lsb_mask<T2>(8 * sizeof(T));
+                const auto shifted_lo = bitwise_lshift(self2, shifts_lo);
+                const batch<T2, A> batch_mask_lo { mask_lo };
+                const auto masked_lo = bitwise_and(shifted_lo, batch_mask_lo);
+
+                // Higher byte: mask bits that would flow from lower byte and shift as twice the size.
+                constexpr auto shifts_hi = make_batch_constant<T2, select_stride<T, 1, 2, Vs...>, A>();
+                constexpr auto mask_hi = mask_lo << (8 * sizeof(T));
+                const batch<T2, A> batch_mask_hi { mask_hi };
+                const auto masked_hi = bitwise_and(self2, batch_mask_hi);
+                const auto shifted_hi = bitwise_lshift(masked_hi, shifts_hi);
+
+                return bitwise_cast<T>(bitwise_or(masked_lo, shifted_hi));
+            }
+        }
+    }
+}
+
+#endif

include/xsimd/arch/xsimd_avx.hpp

@@ -250,7 +250,7 @@ namespace xsimd
                                       self, other);
         }
 
-        // bitwise_lshift
+        // bitwise_lshift single
         template <class A, class T, class = typename std::enable_if<std::is_integral<T>::value>::type>
         XSIMD_INLINE batch<T, A> bitwise_lshift(batch<T, A> const& self, int32_t other, requires_arch<avx>) noexcept
         {
@@ -259,6 +259,13 @@ namespace xsimd
                                       self, other);
         }
 
+        // bitwise_lshift multiple
+        template <class A, class T, T... Vals, detail::enable_integral_t<T> = 0>
+        XSIMD_INLINE batch<T, A> bitwise_lshift(batch<T, A> const& lhs, batch_constant<T, A, Vals...> const& rhs, requires_arch<avx> req) noexcept
+        {
+            return bitwise_lshift(lhs, rhs.as_batch(), req);
+        }
+
         // bitwise_not
         template <class A, class T, class = typename std::enable_if<std::is_integral<T>::value>::type>
         XSIMD_INLINE batch<T, A> bitwise_not(batch<T, A> const& self, requires_arch<avx>) noexcept

include/xsimd/arch/xsimd_avx2.hpp

@@ -13,12 +13,12 @@
 #define XSIMD_AVX2_HPP
 
 #include <complex>
+#include <limits>
 #include <type_traits>
 
 #include "../types/xsimd_avx2_register.hpp"
 #include "../types/xsimd_batch_constant.hpp"
-
-#include <limits>
+#include "./utils/shifts.hpp"
 
 namespace xsimd
 {
@@ -265,7 +265,19 @@ namespace xsimd
             return _mm256_xor_si256(self, _mm256_set1_epi32(-1));
         }
 
-        // bitwise_lshift
+        // bitwise_or
+        template <class A, class T, class = typename std::enable_if<std::is_integral<T>::value>::type>
+        XSIMD_INLINE batch<T, A> bitwise_or(batch<T, A> const& self, batch<T, A> const& other, requires_arch<avx2>) noexcept
+        {
+            return _mm256_or_si256(self, other);
+        }
+        template <class A, class T, class = typename std::enable_if<std::is_integral<T>::value>::type>
+        XSIMD_INLINE batch_bool<T, A> bitwise_or(batch_bool<T, A> const& self, batch_bool<T, A> const& other, requires_arch<avx2>) noexcept
+        {
+            return _mm256_or_si256(self, other);
+        }
+
+        // bitwise_lshift single (dynamic)
         template <class A, class T, class = typename std::enable_if<std::is_integral<T>::value>::type>
         XSIMD_INLINE batch<T, A> bitwise_lshift(batch<T, A> const& self, int32_t other, requires_arch<avx2>) noexcept
         {
@@ -287,6 +299,7 @@ namespace xsimd
             }
         }
 
+        // bitwise_lshift single (constant)
         template <size_t shift, class A, class T, class = typename std::enable_if<std::is_integral<T>::value>::type>
         XSIMD_INLINE batch<T, A> bitwise_lshift(batch<T, A> const& self, requires_arch<avx2>) noexcept
         {
@@ -315,6 +328,7 @@ namespace xsimd
             }
         }
 
+        // bitwise_lshift multiple (dynamic)
         template <class A, class T, class = typename std::enable_if<std::is_integral<T>::value>::type>
         XSIMD_INLINE batch<T, A> bitwise_lshift(batch<T, A> const& self, batch<T, A> const& other, requires_arch<avx2>) noexcept
         {
@@ -332,16 +346,41 @@ namespace xsimd
             }
         }
 
-        // bitwise_or
-        template <class A, class T, class = typename std::enable_if<std::is_integral<T>::value>::type>
-        XSIMD_INLINE batch<T, A> bitwise_or(batch<T, A> const& self, batch<T, A> const& other, requires_arch<avx2>) noexcept
+        // bitwise_lshift multiple (constant)
+        template <class T, class A, T... Vs, detail::enable_sized_integral_t<T, 8> = 0>
+        XSIMD_INLINE batch<T, A> bitwise_lshift(
+            batch<T, A> const& self, batch_constant<T, A, Vs...> shifts, requires_arch<avx2> req) noexcept
         {
-            return _mm256_or_si256(self, other);
+            return bitwise_lshift(self, shifts.as_batch(), req);
         }
-        template <class A, class T, class = typename std::enable_if<std::is_integral<T>::value>::type>
-        XSIMD_INLINE batch_bool<T, A> bitwise_or(batch_bool<T, A> const& self, batch_bool<T, A> const& other, requires_arch<avx2>) noexcept
+
+        template <class T, class A, T... Vs, detail::enable_sized_integral_t<T, 4> = 0>
+        XSIMD_INLINE batch<T, A> bitwise_lshift(
+            batch<T, A> const& self, batch_constant<T, A, Vs...> shifts, requires_arch<avx2> req) noexcept
         {
-            return _mm256_or_si256(self, other);
+            return bitwise_lshift(self, shifts.as_batch(), req);
+        }
+
+        template <class T, class A, T... Vs, detail::enable_sized_integral_t<T, 2> = 0>
+        XSIMD_INLINE batch<T, A> bitwise_lshift(
+            batch<T, A> const& self, batch_constant<T, A, Vs...> shifts, requires_arch<avx2>) noexcept
+        {
+            using uint_t = typename std::make_unsigned<T>::type;
+            return bitwise_cast<T>(
+                utils::bitwise_lshift_as_twice_larger<uint_t, uint32_t>(
+                    bitwise_cast<uint_t>(self),
+                    bitwise_cast<uint_t>(shifts)));
+        }
+
+        template <class T, class A, T... Vs, detail::enable_sized_integral_t<T, 1> = 0>
+        XSIMD_INLINE batch<T, A> bitwise_lshift(
+            batch<T, A> const& self, batch_constant<T, A, Vs...> shifts, requires_arch<avx2>) noexcept
+        {
+            using uint_t = typename std::make_unsigned<T>::type;
+            return bitwise_cast<T>(
+                utils::bitwise_lshift_as_twice_larger<uint_t, uint16_t>(
+                    bitwise_cast<uint_t>(self),
+                    bitwise_cast<uint_t>(shifts)));
         }
 
         // bitwise_rshift

include/xsimd/arch/xsimd_common_fwd.hpp

@@ -47,6 +47,8 @@ namespace xsimd
         XSIMD_INLINE batch<T, A> abs(batch<T, A> const& self, requires_arch<common>) noexcept;
         template <class A, class T, class = typename std::enable_if<std::is_integral<T>::value>::type>
         XSIMD_INLINE batch<T, A> bitwise_lshift(batch<T, A> const& self, batch<T, A> const& other, requires_arch<common>) noexcept;
+        template <class A, class T, T... Vals, detail::enable_integral_t<T> = 0>
+        XSIMD_INLINE batch<T, A> bitwise_lshift(batch<T, A> const& lhs, batch_constant<T, A, Vals...> const& rhs, requires_arch<common> req) noexcept;
         template <size_t shift, class A, class T, class = typename std::enable_if<std::is_integral<T>::value>::type>
         XSIMD_INLINE batch<T, A> bitwise_lshift(batch<T, A> const& self, requires_arch<common>) noexcept;
         template <class A, class T, class = typename std::enable_if<std::is_integral<T>::value>::type>

include/xsimd/arch/xsimd_sse2.hpp

@@ -18,6 +18,7 @@
 
 #include "../types/xsimd_batch_constant.hpp"
 #include "../types/xsimd_sse2_register.hpp"
+#include "./utils/shifts.hpp"
 
 namespace xsimd
 {
@@ -267,65 +268,6 @@ namespace xsimd
             return _mm_andnot_pd(other, self);
         }
 
-        // bitwise_lshift
-        template <class A, class T, class = typename std::enable_if<std::is_integral<T>::value>::type>
-        XSIMD_INLINE batch<T, A> bitwise_lshift(batch<T, A> const& self, int32_t other, requires_arch<sse2>) noexcept
-        {
-            XSIMD_IF_CONSTEXPR(sizeof(T) == 1)
-            {
-                return _mm_and_si128(_mm_set1_epi8(0xFF << other), _mm_slli_epi32(self, other));
-            }
-            else XSIMD_IF_CONSTEXPR(sizeof(T) == 2)
-            {
-                return _mm_slli_epi16(self, other);
-            }
-            else XSIMD_IF_CONSTEXPR(sizeof(T) == 4)
-            {
-                return _mm_slli_epi32(self, other);
-            }
-            else XSIMD_IF_CONSTEXPR(sizeof(T) == 8)
-            {
-                return _mm_slli_epi64(self, other);
-            }
-            else
-            {
-                assert(false && "unsupported arch/op combination");
-                return {};
-            }
-        }
-        template <size_t shift, class A, class T, class = typename std::enable_if<std::is_integral<T>::value>::type>
-        XSIMD_INLINE batch<T, A> bitwise_lshift(batch<T, A> const& self, requires_arch<sse2>) noexcept
-        {
-            constexpr auto bits = std::numeric_limits<T>::digits + std::numeric_limits<T>::is_signed;
-            static_assert(shift < bits, "Count must be less than the number of bits in T");
-            XSIMD_IF_CONSTEXPR(shift == 0)
-            {
-                return self;
-            }
-            else XSIMD_IF_CONSTEXPR(sizeof(T) == 1)
-            {
-                // 8-bit left shift via 16-bit shift + mask
-                __m128i shifted = _mm_slli_epi16(self, static_cast<int>(shift));
-                // TODO(C++17): without `if constexpr` we must ensure the compile-time shift does not overflow
-                constexpr uint8_t mask8 = static_cast<uint8_t>(sizeof(T) == 1 ? (~0u << shift) : 0);
-                const __m128i mask = _mm_set1_epi8(mask8);
-                return _mm_and_si128(shifted, mask);
-            }
-            else XSIMD_IF_CONSTEXPR(sizeof(T) == 2)
-            {
-                return _mm_slli_epi16(self, static_cast<int>(shift));
-            }
-            else XSIMD_IF_CONSTEXPR(sizeof(T) == 4)
-            {
-                return _mm_slli_epi32(self, static_cast<int>(shift));
-            }
-            else XSIMD_IF_CONSTEXPR(sizeof(T) == 8)
-            {
-                return _mm_slli_epi64(self, static_cast<int>(shift));
-            }
-            return bitwise_lshift<shift>(self, common {});
-        }
-
         // bitwise_not
         template <class A>
         XSIMD_INLINE batch<float, A> bitwise_not(batch<float, A> const& self, requires_arch<sse2>) noexcept
@@ -582,6 +524,83 @@ namespace xsimd
             return _mm_castpd_si128(self);
         }
 
+        // bitwise_lshift single (dynamic)
+        template <class A, class T, class = typename std::enable_if<std::is_integral<T>::value>::type>
+        XSIMD_INLINE batch<T, A> bitwise_lshift(batch<T, A> const& self, int32_t other, requires_arch<sse2>) noexcept
+        {
+            XSIMD_IF_CONSTEXPR(sizeof(T) == 1)
+            {
+                return _mm_and_si128(_mm_set1_epi8(0xFF << other), _mm_slli_epi32(self, other));
+            }
+            else XSIMD_IF_CONSTEXPR(sizeof(T) == 2)
+            {
+                return _mm_slli_epi16(self, other);
+            }
+            else XSIMD_IF_CONSTEXPR(sizeof(T) == 4)
+            {
+                return _mm_slli_epi32(self, other);
+            }
+            else XSIMD_IF_CONSTEXPR(sizeof(T) == 8)
+            {
+                return _mm_slli_epi64(self, other);
+            }
+            else
+            {
+                assert(false && "unsupported arch/op combination");
+                return {};
+            }
+        }
+
+        // bitwise_lshift single (constant)
+        template <size_t shift, class A, class T, class = typename std::enable_if<std::is_integral<T>::value>::type>
+        XSIMD_INLINE batch<T, A> bitwise_lshift(batch<T, A> const& self, requires_arch<sse2>) noexcept
+        {
+            constexpr auto bits = std::numeric_limits<T>::digits + std::numeric_limits<T>::is_signed;
+            static_assert(shift < bits, "Count must be less than the number of bits in T");
+            XSIMD_IF_CONSTEXPR(shift == 0)
+            {
+                return self;
+            }
+            else XSIMD_IF_CONSTEXPR(sizeof(T) == 1)
+            {
+                // 8-bit left shift via 16-bit shift + mask
+                __m128i shifted = _mm_slli_epi16(self, static_cast<int>(shift));
+                // TODO(C++17): without `if constexpr` we must ensure the compile-time shift does not overflow
+                constexpr uint8_t mask8 = static_cast<uint8_t>(sizeof(T) == 1 ? (~0u << shift) : 0);
+                const __m128i mask = _mm_set1_epi8(mask8);
+                return _mm_and_si128(shifted, mask);
+            }
+            else XSIMD_IF_CONSTEXPR(sizeof(T) == 2)
+            {
+                return _mm_slli_epi16(self, static_cast<int>(shift));
+            }
+            else XSIMD_IF_CONSTEXPR(sizeof(T) == 4)
+            {
+                return _mm_slli_epi32(self, static_cast<int>(shift));
+            }
+            else XSIMD_IF_CONSTEXPR(sizeof(T) == 8)
+            {
+                return _mm_slli_epi64(self, static_cast<int>(shift));
+            }
+            return bitwise_lshift<shift>(self, common {});
+        }
+
+        // bitwise_lshift multiple (constant)
+        template <class A, uint16_t... Vs>
+        XSIMD_INLINE batch<uint16_t, A> bitwise_lshift(
+            batch<uint16_t, A> const& self, batch_constant<uint16_t, A, Vs...>, requires_arch<sse2>) noexcept
+        {
+            constexpr auto mults = batch_constant<uint16_t, A, static_cast<uint16_t>(1u << Vs)...>();
+            return _mm_mullo_epi16(self, mults.as_batch());
+        }
+
+        template <class A, uint8_t... Vs>
+        XSIMD_INLINE batch<uint8_t, A> bitwise_lshift(
+            batch<uint8_t, A> const& self, batch_constant<uint8_t, A, Vs...> shifts, requires_arch<sse2>) noexcept
+        {
+            return utils::bitwise_lshift_as_twice_larger<uint8_t, uint16_t>(self, shifts);
+        }
+
         // broadcast
         template <class A>
         batch<float, A> XSIMD_INLINE broadcast(float val, requires_arch<sse2>) noexcept