+/*
+ for(i = 0; i < 5; ++i){
+ for(j = 0; j < 5; ++j){
+ a[j][i] ^= b[(4 + i) % 5][0];
+ }
+ }
+ for(i = 0; i < 5; ++i){
+ for(j = 0; j < 5; ++j){
+ a[j][i] ^= rotate64_1bit_left(b[(i + 1) % 5][0]);
+ }
+ }
+
+*/
+ sbiw XL, 5 * 8
+
+ sbiw r24, 40
+/* a[0..4][0]{0..7} ^= b[4][0]{0..7} */
+ rcall theta_2a
+/* a[0..4][1]{0..7} ^= b[0][0]{0..7} */
+ subi XL, lo8(4 * 5 * 8 + 8)
+ sbci XH, hi8(4 * 5 * 8 + 8)
+ rcall theta_2a
+/* a[0..4][2]{0..7} ^= b[1][0]{0..7} */
+ adiw XL, 4 * 8
+ rcall theta_2a
+/* a[0..4][3]{0..7} ^= b[2][0]{0..7} */
+ adiw XL, 4 * 8
+ rcall theta_2a
+/* a[0..4][4]{0..7} ^= b[3][0]{0..7} */
+ adiw XL, 4 * 8
+ rcall theta_2a
+
+ ret
+
+ ldi r20, 8
+
+10:
+ movw ZL, r24
+ ld r21, X+
+ .irp r, 0, 1, 2, 3, 4
+ ld r22, Z
+ eor r22, r21
+ st Z, r22
+ .if \r != 4
+ adiw ZL, 40
+ .endif
+ .endr
+ adiw r24, 1
+ dec r20
+ brne 10b
+