bcachefs: Use memcpy_u64s_small() for copying keys

Small performance optimization; an open coded loop is better than rep ;
movsq for small copies.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

fs/bcachefs/bkey.c

@@ -360,7 +360,7 @@ bool bch2_bkey_pack(struct bkey_packed *out, const struct bkey_i *in,
 	memmove_u64s((u64 *) out + format->key_u64s,
 		     &in->v,
 		     bkey_val_u64s(&in->k));
-	memcpy_u64s(out, &tmp, format->key_u64s);
+	memcpy_u64s_small(out, &tmp, format->key_u64s);
 
 	return true;
 }

fs/bcachefs/bkey_sort.c

@@ -188,7 +188,7 @@ unsigned bch2_sort_keys(struct bkey_packed *dst,
 		}
 
 		if (bkey_deleted(in)) {
-			memcpy_u64s(out, in, bkeyp_key_u64s(f, in));
+			memcpy_u64s_small(out, in, bkeyp_key_u64s(f, in));
 			set_bkeyp_val_u64s(f, out, 0);
 		} else {
 			bkey_copy(out, in);

fs/bcachefs/bset.c

@@ -1022,7 +1022,7 @@ void bch2_bset_insert(struct btree *b,
 		set_btree_bset_end(b, t);
 	}
 
-	memcpy_u64s(where, src,
+	memcpy_u64s_small(where, src,
 		    bkeyp_key_u64s(f, src));
 	memcpy_u64s(bkeyp_val(f, where), &insert->v,
 		    bkeyp_val_u64s(f, src));