From 038ee0e9129f69e49ec97c8b685a0e0515fff21c Mon Sep 17 00:00:00 2001
From: Corbin Crutchley <git@crutchcorn.dev>
Date: Sat, 25 Jul 2026 23:14:21 -0700
Subject: [PATCH] fix: clear upper lanes after SIMD reductions

---
 crates/aarch64/src/lib.rs | 27 ++++++++++++++++++++++-----
 1 file changed, 22 insertions(+), 5 deletions(-)

diff --git a/crates/aarch64/src/lib.rs b/crates/aarch64/src/lib.rs
index 86a2127..bbef3be 100644
--- a/crates/aarch64/src/lib.rs
+++ b/crates/aarch64/src/lib.rs
@@ -2299,6 +2299,7 @@ fn execute_neon_value_operation(
             let maximum = matches!(operation, Operation::NeonUnsignedMaximumAcross { .. });
             let result = unsigned_lane_extreme(read(source)?, source.size, element_size, maximum)
                 .ok_or_else(unsupported)?;
+            let destination = scalar_vector_destination(destination).ok_or_else(unsupported)?;
             Ok((destination, result))
         }
         Operation::NeonUnsignedMaximumPairwise {
@@ -2765,9 +2766,27 @@ fn execute_neon_byte_reduction(
             result = result.wrapping_add(byte);
         }
     }
+    let destination = if count_bits {
+        destination
+    } else {
+        scalar_vector_destination(destination).ok_or_else(unsupported)?
+    };
     Ok((destination, result))
 }
 
+fn scalar_vector_destination(destination: Place) -> Option<Place> {
+    match destination {
+        Place {
+            storage: Storage::Vector { offset: 0, .. },
+            size: 1..=16,
+        } => Some(Place {
+            storage: destination.storage,
+            size: 16,
+        }),
+        _ => None,
+    }
+}
+
 fn execute_neon_unsigned_multiply_long(
     state: &Aarch64State,
     scratch: &ScratchValues,
@@ -4352,6 +4371,7 @@ mod tests {
         let upper = 0xfeed_face_cafe_beef_u128 << 64;
         let mut state = Aarch64State::new(CODE_ADDRESS, GuestAddress::new(0x8000));
         state.set_vector(0, upper | 0x0103_070f_1f3f_7fff).unwrap();
+        state.set_vector(22, u128::MAX).unwrap();
         for _ in 0..RUST_VECTOR_COUNT_CODE.len() / 4 {
             interpreter.step(&mut state, &mut memory).unwrap();
         }
@@ -4401,10 +4421,7 @@ mod tests {
             state.vector(1),
             Some(u128::from(u32::MAX) << 96 | u128::from(u32::MAX) << 32)
         );
-        assert_eq!(
-            state.vector(2),
-            Some(vector_two_upper | u128::from(u32::MAX))
-        );
+        assert_eq!(state.vector(2), Some(u128::from(u32::MAX)));
 
         let mut memory = executable_memory(SELECT_AND_MINIMUM);
         let mut state = Aarch64State::new(CODE_ADDRESS, GuestAddress::new(0x8000));
@@ -4415,7 +4432,7 @@ mod tests {
         for _ in 0..2 {
             interpreter.step(&mut state, &mut memory).unwrap();
         }
-        assert_eq!(state.vector(0), Some(upper | 0xaa11_aa11_aa11_aa11));
+        assert_eq!(state.vector(0), Some(0xaa11));
 
         let mut memory = executable_memory(ZERO_COMPARE);
         let mut state = Aarch64State::new(CODE_ADDRESS, GuestAddress::new(0x8000));
-- 
2.51.2