Fix implementation of rotate_left on avx2 and improve test coverage for rotate_left

Author: serge-sans-paille

include/xsimd/arch/xsimd_avx2.hpp

@@ -659,7 +659,15 @@ namespace xsimd
         template <size_t N, class A>
         XSIMD_INLINE batch<uint8_t, A> rotate_left(batch<uint8_t, A> const& self, requires_arch<avx2>) noexcept
         {
-            return _mm256_alignr_epi8(self, self, N);
+            auto other = _mm256_permute2x128_si256(self, self, 0x1);
+            if (N < 16)
+            {
+                return _mm256_alignr_epi8(other, self, 2 * N);
+            }
+            else
+            {
+                return _mm256_alignr_epi8(self, other, 2 * (N - 16));
+            }
         }
         template <size_t N, class A>
         XSIMD_INLINE batch<int8_t, A> rotate_left(batch<int8_t, A> const& self, requires_arch<avx2>) noexcept
@@ -669,7 +677,15 @@ namespace xsimd
         template <size_t N, class A>
         XSIMD_INLINE batch<uint16_t, A> rotate_left(batch<uint16_t, A> const& self, requires_arch<avx2>) noexcept
         {
-            return _mm256_alignr_epi8(self, self, 2 * N);
+            auto other = _mm256_permute2x128_si256(self, self, 0x1);
+            if (N < 8)
+            {
+                return _mm256_alignr_epi8(other, self, 2 * N);
+            }
+            else
+            {
+                return _mm256_alignr_epi8(self, other, 2 * (N - 8));
+            }
         }
         template <size_t N, class A>
         XSIMD_INLINE batch<int16_t, A> rotate_left(batch<int16_t, A> const& self, requires_arch<avx2>) noexcept
@@ -886,6 +902,7 @@ namespace xsimd
         }
 
         // swizzle (dynamic mask)
+
         template <class A>
         XSIMD_INLINE batch<float, A> swizzle(batch<float, A> const& self, batch<uint32_t, A> mask, requires_arch<avx2>) noexcept
         {

test/test_batch_manip.cpp

@@ -20,7 +20,7 @@ namespace xsimd
     struct init_swizzle_base
     {
         using swizzle_vector_type = std::array<T, N>;
-        swizzle_vector_type lhs_in, exped_reverse, exped_fill, exped_dup, exped_ror, exped_rol;
+        swizzle_vector_type lhs_in, exped_reverse, exped_fill, exped_dup, exped_ror, exped_rol, exped_rol2;
 
         template <int... Indices>
         std::vector<swizzle_vector_type> create_swizzle_vectors()
@@ -42,12 +42,14 @@ namespace xsimd
                 exped_dup[i] = lhs_in[2 * (i / 2)];
                 exped_ror[i] = lhs_in[(i - 1) % N];
                 exped_rol[i] = lhs_in[(i + 1) % N];
+                exped_rol2[i] = lhs_in[(i + N - 1) % N];
             }
             vects.push_back(std::move(exped_reverse));
             vects.push_back(std::move(exped_fill));
             vects.push_back(std::move(exped_dup));
             vects.push_back(std::move(exped_ror));
             vects.push_back(std::move(exped_rol));
+            vects.push_back(std::move(exped_rol2));
 
             return vects;
         }
@@ -176,6 +178,20 @@ struct swizzle_test
         CHECK_BATCH_EQ(b_res, b_exped);
     }
 
+    void rotate_left_inv()
+    {
+        xsimd::init_swizzle_base<value_type, size> swizzle_base;
+        auto swizzle_vecs = swizzle_base.create_swizzle_vectors();
+        auto v_lhs = swizzle_vecs[0];
+        auto v_exped = swizzle_vecs[6];
+
+        B b_lhs = B::load_unaligned(v_lhs.data());
+        B b_exped = B::load_unaligned(v_exped.data());
+
+        B b_res = xsimd::rotate_left<size - 1>(b_lhs);
+        CHECK_BATCH_EQ(b_res, b_exped);
+    }
+
     void swizzle_reverse()
     {
         xsimd::init_swizzle_base<value_type, size> swizzle_base;
@@ -248,6 +264,7 @@ TEST_CASE_TEMPLATE("[swizzle]", B, BATCH_SWIZZLE_TYPES)
     SUBCASE("rotate")
     {
         Test.rotate_left();
+        Test.rotate_left_inv();
         Test.rotate_right();
     }