T2 = 22
T3 = 23
-.global serpent_lt
serpent_lt:
push_range 4, 17
movw r26, r24
pop_range 4, 17
ret
-
#endif
+
T0 = 22
T1 = 23
T2 = 24
brne 1b
ret
-.global serpent_lt
serpent_lt:
/* X0 := X0 <<< 13 */
movw r26, r24
rcall memrotr32
ret
-.global serpent_inv_lt
serpent_inv_lt:
/* X0 := X0 >>> 5 */
movw r26, r24
/* trashes:
* r20-r25, r30-r31
*/
-.global serpent_gen_w
serpent_gen_w:
movw r30, r24
/* ^i^b[0]*/
stack_free 32
ret
+/*
+ * void serpent_enc(void* buffer, const serpent_ctx_t* ctx){
+ */
+/*
+ * param buffer is passed in r24:r25
+ * param ctx is passed in r22:r23
+ */
+.global serpent_enc
+serpent_enc:
+
+ push_ r12, r13, r14, r15, r16
+ clr r16
+ movw r14, r24
+ movw r12, r22
+1:
+ movw r24, r14
+ movw r22, r12
+ ldi r20, 16
+ add r12, r20
+ adc r13, r1
+ clr r21
+ rcall memxor
+ movw r24, r14
+ mov r22, r16
+ rcall sbox128
+ movw r24, r14
+ rcall serpent_lt
+
+ inc r16
+ cpi r16, 31
+ brne 1b
+
+ movw r24, r14
+ movw r22, r12
+ ldi r20, 16
+ add r12, r20
+ adc r13, r1
+ clr r21
+ rcall memxor
+ movw r24, r14
+ mov r22, r16
+ rcall sbox128
+
+ inc r16
+ movw r24, r14
+ movw r22, r12
+ ldi r20, 16
+ clr r21
+ pop_ r16, r15, r14, r13, r12
+ rjmp memxor
+/*
+ * void serpent_dec(void* buffer, const serpent_ctx_t* ctx){
+ */
+/*
+ * param buffer is passed in r24:r25
+ * param ctx is passed in r22:r23
+ */
+.global serpent_dec
+serpent_dec:
+ push_ r12, r13, r14, r15, r16
+ movw r14, r24
+// ldi r16, lo8(32*16)
+// add r22, r16
+ ldi r16, hi8(32*16)
+ add r23, r16
+ movw r12, r22
+ ldi r20, 16
+ clr r21
+ rcall memxor
+
+ movw r24, r14
+ ldi r22, 31
+ rcall inv_sbox128
+
+ movw r24, r14
+ ldi r20, 16
+ sub r12, r20
+ sbc r13, r1
+ movw r22, r12
+ clr r21
+ rcall memxor
+ ldi r16, 31
+1:
+ dec r16
+ movw r24, r14
+ rcall serpent_inv_lt
+ movw r24, r14
+ mov r22, r16
+ rcall inv_sbox128
+ movw r24, r14
+ ldi r20, 16
+ sub r12, r20
+ sbc r13, r1
+ movw r22, r12
+ clr r21
+ rcall memxor
+
+ tst r16
+ brne 1b
+ pop_ r16, r15, r14, r13, r12
+ ret
+
+
+