Optimize hot loops with direct access, LUTs, and closed-form formulas

1 files changed, 8 insertions, 3 deletions

diff --git a/src/kernels.h b/src/kernels.h
index 45d25b7..57066c0 100644
--- a/src/kernels.h
+++ b/src/kernels.h
@@ -208,9 +208,14 @@ static inline TERM_ALWAYS_INLINE void bytes_copy_small(uint8_t *restrict dst,
         memcpy(&b, src + n - 4, 4);
         memcpy(dst, &a, 4);
         memcpy(dst + n - 4, &b, 4);
-    } else {
-        for (int k = 0; k < n; k++)
-            dst[k] = (uint8_t)src[k];
+    } else if (n >= 2) {
+        uint16_t a, b;
+        memcpy(&a, src, 2);
+        memcpy(&b, src + n - 2, 2);
+        memcpy(dst, &a, 2);
+        memcpy(dst + n - 2, &b, 2);
+    } else if (n == 1) {
+        dst[0] = (uint8_t)src[0];
     }
 }