even more serpent in asm

serpent-asm.S

@@ -538,6 +538,9 @@ static uint32_t serpent_gen_w(uint32_t * b, uint8_t i){
  * param i is passed in r22
  * return value is returned in r22.r23.r24.r25
  */
+ /* trashes:
+  *  r20-r25, r30-r31
+  */
 .global serpent_gen_w
 serpent_gen_w:
 	movw r30, r24
@@ -589,3 +592,120 @@ serpent_gen_w:
 	brne 1b
 	ret
 
+/*
+ * void serpent_init(const void* key, uint16_t keysize_b, serpent_ctx_t* ctx)
+ */
+/*
+ * param key     is passed in r24:r25
+ * param keysize is passed in r22:r23
+ * param ctx     is passed in r20:r21
+ */
+.global serpent_init
+serpent_init:
+    stack_alloc 32
+    adiw r30, 1
+	push_ r30, r31
+    movw r26, r22
+    adiw r26, 7
+    tst r27
+    breq 1f
+	ldi r26, 32
+	rjmp 2f
+1:
+	lsr r26
+	lsr r26
+	lsr r26
+2:	
+	mov r22, r26
+	bst r22, 5 /* store in T if we have to do the "append 1 thing"*/
+	ldi r27, 32
+3:	/* set buffer to zero */
+	st Z+, r1
+	dec r27
+	brne 3b
+	
+	movw r26, r24 /* X points to the key */
+	sbiw r30, 32
+	tst r22
+	breq 5f /* if keylength_b==0 */
+4:	/* copy keybytes to buffer */
+	ld r19, X+
+	st Z+, r19
+	dec r22
+	brne 4b
+5:
+	brts 7f /* if keylength_b == 256 */
+	ldi r18, 0x01
+	andi r22, 0x07
+	brne 6f
+	st Z, r18
+	rjmp 7f
+6:	/* shift the one to the right position */
+	lsl r18
+	dec r22
+	brne 6b
+	or r18, r19
+	st -Z, r18
+7: /* post "appending 1 thing" buffer is ready for subkey generation */
+	movw r26, r20  /* X points to the context */
+	
+	pop_ r19, r18 /* r18:r19 points to the buffer */
+	push r16
+	clr r16
+8:
+	movw r24, r18
+	mov  r22, r16
+	rcall serpent_gen_w
+	movw r30, r18
+	ldi r20, 7*4
+1: /* the memmove */
+	ldd r0, Z+4
+	st Z+, r0
+	dec r20
+	brne 1b
+  /* store new word in buffer and context */	
+	st Z+, r22
+	st Z+, r23
+	st Z+, r24
+	st Z+, r25
+	st X+, r22
+	st X+, r23
+	st X+, r24
+	st X+, r25
+	
+	inc r16
+	cpi r16, 132
+	brne 8b	
+	
+	push_ r28, r29
+	movw r28, r26
+	subi r28, lo8(132*4)
+	sbci r29, hi8(132*4)
+	ldi r16, 33
+2:
+	movw r24, r28
+	adiw r28, 16
+	ldi r22, 2
+	add r22, r16
+	rcall sbox128
+	dec r16
+	brne 2b
+	pop_ r29, r28, r16
+	stack_free 32
+	ret
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+

serpent.c

@@ -91,14 +91,15 @@ static uint32_t serpent_gen_w(uint32_t * b, uint8_t i){
 	return ret;
 } 
 */
-void serpent_init(const void* key, uint16_t keysize, serpent_ctx_t* ctx){
+#if 0
+void serpent_init(const void* key, uint16_t keysize_b, serpent_ctx_t* ctx){
 	uint32_t buffer[8];
 	uint8_t i,j;
-	if(keysize<256){
+	if(keysize_b<256){
 		/* keysize is less than 256 bit, padding needed */
 		memset(buffer, 0, 32);
-		memcpy(buffer, key, (keysize+7)/8);
-		((uint8_t*)buffer)[keysize/8] |= 1<<(keysize%8);
+		memcpy(buffer, key, (keysize_b+7)/8);
+		((uint8_t*)buffer)[keysize_b/8] |= 1<<(keysize_b%8);
 	} else {
 		/* keysize is 256 bit */
 		memcpy(buffer, key, 32); 
@@ -114,7 +115,7 @@ void serpent_init(const void* key, uint16_t keysize, serpent_ctx_t* ctx){
 		sbox128(ctx->k[i],3-i);
 	}
 }
-
+#endif
 
 void serpent_enc(void* buffer, const serpent_ctx_t* ctx){
 	uint8_t i;

serpent.h

@@ -39,7 +39,7 @@ typedef struct serpent_ctx_st {
 
 
 /* key must be 256bit (32 byte) large! */
-void serpent_init(const void* key, uint16_t keysize, serpent_ctx_t* ctx);
+void serpent_init(const void* key, uint16_t keysize_b, serpent_ctx_t* ctx);
 void serpent_enc(void* buffer, const serpent_ctx_t* ctx);
 void serpent_dec(void* buffer, const serpent_ctx_t* ctx);
 

test_src/main-serpent-test.c

@@ -62,6 +62,19 @@ void testrun_nessie_serpent(void){
 	nessie_bc_run();
 }
 
+void testrun_test_serpent(void){
+	uint8_t key[32];
+	serpent_ctx_t ctx;
+	uint8_t i;
+	memset(key, 0, 16);
+	serpent_init(key, 128, &ctx);
+	for(i=0; i<33; ++i){
+		uart_putstr_P(PSTR("\r\n subkekey "));	
+		uart_hexdump(&i, 1);
+		uart_putstr_P(PSTR(" : "));	
+		uart_hexdump(ctx.k[i], 16);
+	}
+}
 
 void testrun_performance_serpent(void){
 	uint64_t t;
@@ -114,7 +127,7 @@ int main (void){
 	uart_putstr_P(PSTR(")\r\nloaded and running\r\n"));
 
 	PGM_P    u   = PSTR("nessie\0test\0performance\0");
-	void_fpt v[] = {testrun_nessie_serpent, testrun_nessie_serpent, testrun_performance_serpent};
+	void_fpt v[] = {testrun_nessie_serpent, testrun_test_serpent, testrun_performance_serpent};
 
 	while(1){ 
 		if (!getnextwordn(str,20)){DEBUG_S("DBG: W1\r\n"); goto error;}