FEAT: Optimize shift_remove_index using MaybeUninit

src/impl_methods.rs

@@ -6,7 +6,7 @@
 // option. This file may not be copied, modified, or distributed
 // except according to those terms.
 
-use std::mem::{size_of, ManuallyDrop};
+use std::mem::{size_of, ManuallyDrop, MaybeUninit};
 use alloc::slice;
 use alloc::vec;
 use alloc::vec::Vec;
@@ -2462,10 +2462,24 @@ where
             Zip::from(tail.lanes_mut(axis)).for_each(|lane| {
                 // TODO give ArrayViewMut1 a split_first method?
                 let (mut fst, mut rest) = lane.split_at(Axis(0), 1);
+
+                // Logically we do a circular swap here, all elements in a chain
+                // Using MaybeUninit to avoid unecessary writes in the safe swap solution
+                //
+                //  for elt in rest.iter_mut() {
+                //      std::mem::swap(dst, elt);
+                //      dst = elt;
+                //  }
+                //
+                let mut slot = MaybeUninit::<A>::uninit();
                 let mut dst = &mut fst[0];
-                for elt in rest.iter_mut() {
-                    std::mem::swap(dst, elt);
-                    dst = elt;
+                unsafe {
+                    slot.as_mut_ptr().copy_from_nonoverlapping(dst, 1);
+                    for elt in rest.iter_mut() {
+                        (dst as *mut A).copy_from_nonoverlapping(elt, 1);
+                        dst = elt;
+                    }
+                    (dst as *mut A).copy_from_nonoverlapping(slot.as_ptr(), 1);
                 }
             });
         }