From 8aadb13160ec0599291b0add9260bde721375640 Mon Sep 17 00:00:00 2001
From: Vladimir Saraikin <saraikin.v@phystech.edu>
Date: Tue, 7 Jul 2026 23:31:44 +0200
Subject: [PATCH 1/2] fix(neon): under-aligned store in generic swizzle kernel

---
 include/xsimd/arch/xsimd_neon.hpp |  4 +++-
 test/test_batch_manip.cpp         | 18 ++++++++++++++++++
 2 files changed, 21 insertions(+), 1 deletion(-)

diff --git a/include/xsimd/arch/xsimd_neon.hpp b/include/xsimd/arch/xsimd_neon.hpp
index 8ea2756f7..48be9b88e 100644
--- a/include/xsimd/arch/xsimd_neon.hpp
+++ b/include/xsimd/arch/xsimd_neon.hpp
@@ -3308,7 +3308,9 @@ namespace xsimd
                                          requires_arch<neon>) noexcept
         {
             static_assert(batch<T, A>::size == sizeof...(idx), "valid swizzle indices");
-            std::array<T, batch<T, A>::size> data;
+            // std::array is only aligned to alignof(T), while store_aligned requires
+            // the full batch alignment
+            alignas(A::alignment()) std::array<T, batch<T, A>::size> data;
             self.store_aligned(data.data());
             return set(batch<T, A>(), A(), data[idx]...);
         }
diff --git a/test/test_batch_manip.cpp b/test/test_batch_manip.cpp
index 93d0e2a46..c864ae5f7 100644
--- a/test/test_batch_manip.cpp
+++ b/test/test_batch_manip.cpp
@@ -330,6 +330,24 @@ TEST_CASE_TEMPLATE("[swizzle]", B, BATCH_SWIZZLE_TYPES)
     SUBCASE("rotate_right") { swizzle_test<B>().rotate_right(); }
 }
 
+#if XSIMD_WITH_NEON
+// Regression test for https://github.com/xtensor-stack/xsimd/issues/1260: the generic
+// neon swizzle kernel stored to an std::array aligned only to alignof(T), tripping the
+// alignment assertion in store_aligned for the 8- and 16-bit batches that use it
+TEST_CASE_TEMPLATE("[swizzle generic neon kernel]", B, xsimd::batch<uint8_t, xsimd::neon>, xsimd::batch<int8_t, xsimd::neon>, xsimd::batch<uint16_t, xsimd::neon>, xsimd::batch<int16_t, xsimd::neon>)
+{
+    XSIMD_SWIZZLE_PATTERN_CASE(Reversor);
+    XSIMD_SWIZZLE_PATTERN_CASE(EvenThenOdd);
+    XSIMD_SWIZZLE_PATTERN_CASE(RotateRight1);
+    SUBCASE("reduce_max")
+    {
+        auto lhs = swizzle_test<B>::make_lhs();
+        auto b = B::load_unaligned(lhs.data());
+        CHECK_EQ(xsimd::reduce_max(b), lhs[B::size - 1]);
+    }
+}
+#endif
+
 #undef XSIMD_SWIZZLE_PATTERN_CASE
 
 #endif /* XSIMD_NO_SUPPORTED_ARCHITECTURE */

From 933a762cee7359c072fe7dd7cd289e377828c152 Mon Sep 17 00:00:00 2001
From: Vladimir Saraikin <saraikin.v@phystech.edu>
Date: Wed, 8 Jul 2026 12:02:43 +0200
Subject: [PATCH 2/2] test: run narrow swizzle/reduce_max on every supported
 architecture

---
 test/test_batch_manip.cpp | 38 +++++++++++++++++++++++++++++---------
 1 file changed, 29 insertions(+), 9 deletions(-)

diff --git a/test/test_batch_manip.cpp b/test/test_batch_manip.cpp
index c864ae5f7..8fdc3cfd2 100644
--- a/test/test_batch_manip.cpp
+++ b/test/test_batch_manip.cpp
@@ -330,23 +330,43 @@ TEST_CASE_TEMPLATE("[swizzle]", B, BATCH_SWIZZLE_TYPES)
     SUBCASE("rotate_right") { swizzle_test<B>().rotate_right(); }
 }
 
-#if XSIMD_WITH_NEON
 // Regression test for https://github.com/xtensor-stack/xsimd/issues/1260: the generic
-// neon swizzle kernel stored to an std::array aligned only to alignof(T), tripping the
-// alignment assertion in store_aligned for the 8- and 16-bit batches that use it
-TEST_CASE_TEMPLATE("[swizzle generic neon kernel]", B, xsimd::batch<uint8_t, xsimd::neon>, xsimd::batch<int8_t, xsimd::neon>, xsimd::batch<uint16_t, xsimd::neon>, xsimd::batch<int16_t, xsimd::neon>)
+// swizzle kernel stored to an std::array aligned only to alignof(T), tripping the
+// alignment assertion in store_aligned for the 8- and 16-bit batches that use it.
+// Run on every supported architecture, not just the default one, so the generic
+// kernels of the narrower archs are exercised too
+struct check_narrow_swizzle_and_reduce
 {
-    XSIMD_SWIZZLE_PATTERN_CASE(Reversor);
-    XSIMD_SWIZZLE_PATTERN_CASE(EvenThenOdd);
-    XSIMD_SWIZZLE_PATTERN_CASE(RotateRight1);
-    SUBCASE("reduce_max")
+    template <class Arch>
+    void operator()(Arch) const
     {
+        if (!Arch::available())
+        {
+            return;
+        }
+        check_for<uint8_t, Arch>();
+        check_for<int8_t, Arch>();
+        check_for<uint16_t, Arch>();
+        check_for<int16_t, Arch>();
+    }
+
+    template <class T, class Arch>
+    void check_for() const
+    {
+        using B = xsimd::batch<T, Arch>;
+        swizzle_test<B>().template run<xsimd::Reversor>();
+        swizzle_test<B>().template run<xsimd::EvenThenOdd>();
+        swizzle_test<B>().template run<xsimd::RotateRight1>();
         auto lhs = swizzle_test<B>::make_lhs();
         auto b = B::load_unaligned(lhs.data());
         CHECK_EQ(xsimd::reduce_max(b), lhs[B::size - 1]);
     }
+};
+
+TEST_CASE("[narrow swizzle and reduce_max on every supported architecture]")
+{
+    xsimd::supported_architectures::for_each(check_narrow_swizzle_and_reduce {});
 }
-#endif
 
 #undef XSIMD_SWIZZLE_PATTERN_CASE