fixing shuffle

Author: DiamonDinoia

include/xsimd/arch/common/xsimd_common_swizzle.hpp

@@ -232,12 +232,8 @@ namespace xsimd
             template <uint32_t... Values>
             XSIMD_INLINE constexpr std::uint32_t shuffle() noexcept
             {
-                return shuffle_impl<0,
-                                    sizeof...(Values),
-                                    log2_c<sizeof...(Values)>::value,
-                                    Values...>::value;
+                return shuffle_impl<0, sizeof...(Values), log2_c<sizeof...(Values)>::value, Values...>::value;
             }
-
             template <uint32_t... Values>
             XSIMD_INLINE constexpr std::uint32_t mod_shuffle() noexcept
             {

include/xsimd/arch/xsimd_avx2.hpp

@@ -964,14 +964,15 @@ namespace xsimd
                 constexpr auto imm = ((V0 & 1) << 0) | ((V1 & 1) << 1) | ((V2 & 1) << 2) | ((V3 & 1) << 3);
                 return _mm256_permute_pd(self, imm);
             }
+            constexpr auto imm = detail::mod_shuffle<V0, V1, V2, V3>();
             // fallback to full 4-element permute
-            return _mm256_permute4x64_pd(self, detail::shuffle<V0, V1, V2, V3>());
+            return _mm256_permute4x64_pd(self, imm);
         }
 
         template <class A, uint64_t V0, uint64_t V1, uint64_t V2, uint64_t V3>
         XSIMD_INLINE batch<uint64_t, A> swizzle(batch<uint64_t, A> const& self, batch_constant<uint64_t, A, V0, V1, V2, V3>, requires_arch<avx2>) noexcept
         {
-            constexpr auto mask = detail::shuffle<V0, V1, V2, V3>();
+            constexpr auto mask = detail::mod_shuffle<V0, V1, V2, V3>();
             return _mm256_permute4x64_epi64(self, mask);
         }
         template <class A, uint64_t V0, uint64_t V1, uint64_t V2, uint64_t V3>

include/xsimd/arch/xsimd_sse2.hpp

@@ -1282,16 +1282,13 @@ namespace xsimd
         template <class A, class T, class _ = typename std::enable_if<(sizeof(T) <= 2), void>::type>
         XSIMD_INLINE T reduce_max(batch<T, A> const& self, requires_arch<sse2>) noexcept
         {
-            constexpr auto mask0 = detail::shuffle<2, 3, 0, 0>();
-            batch<T, A> step0 = _mm_shuffle_epi32(self, mask0);
+            batch<T, A> step0 = _mm_shuffle_epi32(self, detail::shuffle<2, 3, 0, 0>());
             batch<T, A> acc0 = max(self, step0);
 
-            constexpr auto mask1 = detail::shuffle<1, 0, 0, 0>();
-            batch<T, A> step1 = _mm_shuffle_epi32(acc0, mask1);
+            batch<T, A> step1 = _mm_shuffle_epi32(acc0, detail::shuffle<1, 0, 0, 0>());
             batch<T, A> acc1 = max(acc0, step1);
 
-            constexpr auto mask2 = detail::shuffle<1, 0, 0, 0>();
-            batch<T, A> step2 = _mm_shufflelo_epi16(acc1, mask2);
+            batch<T, A> step2 = _mm_shufflelo_epi16(acc1, detail::shuffle<1, 0, 0, 0>());
             batch<T, A> acc2 = max(acc1, step2);
             if (sizeof(T) == 2)
                 return first(acc2, A {});
@@ -1304,16 +1301,13 @@ namespace xsimd
         template <class A, class T, class _ = typename std::enable_if<(sizeof(T) <= 2), void>::type>
         XSIMD_INLINE T reduce_min(batch<T, A> const& self, requires_arch<sse2>) noexcept
         {
-            constexpr auto mask0 = detail::shuffle<2, 3, 0, 0>();
-            batch<T, A> step0 = _mm_shuffle_epi32(self, mask0);
+            batch<T, A> step0 = _mm_shuffle_epi32(self, detail::shuffle<2, 3, 0, 0>());
             batch<T, A> acc0 = min(self, step0);
 
-            constexpr auto mask1 = detail::shuffle<1, 0, 0, 0>();
-            batch<T, A> step1 = _mm_shuffle_epi32(acc0, mask1);
+            batch<T, A> step1 = _mm_shuffle_epi32(acc0, detail::shuffle<1, 0, 0, 0>());
             batch<T, A> acc1 = min(acc0, step1);
 
-            constexpr auto mask2 = detail::shuffle<1, 0, 0, 0>();
-            batch<T, A> step2 = _mm_shufflelo_epi16(acc1, mask2);
+            batch<T, A> step2 = _mm_shufflelo_epi16(acc1, detail::shuffle<1, 0, 0, 0>());
             batch<T, A> acc2 = min(acc1, step2);
             if (sizeof(T) == 2)
                 return first(acc2, A {});