X-Git-Url: https://git.cryptolib.org/?a=blobdiff_plain;f=bmw%2Fbmw_256-tinyasm.S;h=fbef4694c5c80af94d755bab121c225715e01714;hb=3a9f9d918fb95474996af535c225a7e187d88372;hp=6327bc1e4c9a568c36c9fd2180df3068b9fc64c1;hpb=52ef0158a4d47da53cf4595a7f9ee867947b4390;p=avr-crypto-lib.git

diff --git a/bmw/bmw_256-tinyasm.S b/bmw/bmw_256-tinyasm.S
index 6327bc1..fbef469 100644
--- a/bmw/bmw_256-tinyasm.S
+++ b/bmw/bmw_256-tinyasm.S
@@ -41,6 +41,11 @@ acc1 = 15
   param s: r20
 */
 shiftleft32:
+	tst r20
+	brpl 10f
+	neg r20
+	rjmp shiftright32
+10:
 	clr r0
 	cpi r20, 8
 	brlo bitrotateleft_1
@@ -49,7 +54,7 @@ shiftleft32:
 	mov r23, r22
 	clr r22
 	subi r20, 8
-	rjmp shiftleft32
+	rjmp 10b
 
 /******************************************************************************/
 /*
@@ -257,10 +262,11 @@ mov32_to_acc:
 
 f2_1_shift_table:
 ;	.byte 0x2B, 0x64, 0x66, 0x03, 0x51, 0x55, 0x87, 0x55
-	.byte 0x55, 0x87, 0x55, 0x51, 0x03, 0x66, 0x64, 0x2B
+;	.byte 0x55, 0x87, 0x55, 0x51, 0x03, 0x66, 0x64, 0x2B
+	.byte 5, -5, -7, 8, -5, 5, -1, 5, -3, 0, 6, -6, -4, 6, -11, 2
 f2_2_shift_table:
 ;	.byte (2<<1), (7<<1), (4<<1), (3<<1), (4<<1)+1, (6<<1)+1, (6<<1)
-	.byte (8<<1)+1, (6<<1), (6<<1)+1, (4<<1)+1, (3<<1), (4<<1), (7<<1), (2<<1)
+	.byte 8, -6, 6, 4, -3, -4, -7, -2
 expand2_rot_table:
 	.byte 3,7,13,16,19,23,27
 
@@ -273,7 +279,7 @@ f0_hacktable:
 
 
 /*******************************************************************************
-* uint32_t addelment(uint8_t j, const uint32_t* m, const uint32_t* h){
+* uint32_t addelment(uint8_t j, const uint32_t *m, const uint32_t *h){
 * 	uint32_t r;
 *	r  = pgm_read_dword(k_lut+j);
 *	r += rotl_addel(((uint32_t*)m)[j&0xf], j+0);
@@ -738,40 +744,21 @@ f2:
 	movw r26, h0
 	ldi r30, lo8(f2_1_shift_table)
 	ldi r31, hi8(f2_1_shift_table)
-	ldi r17, 16
+	ldi r17, 15
 10:
 ;---
 	movw r22, xh0
 	movw r24, xh2
-	cpi r17, 9
-	brge 15f
-	clr r1
-	rjmp 26f
-15:	lpm r20, Z+
-	mov r1, r20
-	andi r20, 0x0f
-	clt
-	cpi r17, 16
-	breq 20f
-	cpi r17, 11
-	brne 21f
-20:	set
-21:	brts 25f
-	rcall shiftright32
-	rjmp 26f
-25:	rcall shiftleft32
-26: rcall mov32_to_acc
+  	lpm r20, Z+
+	sbrc r17, 3
+ 	rcall shiftleft32
+    rcall mov32_to_acc
 ;---
 	rcall load32_from_Y
-	mov r20, r1
-	clr r1
-	swap r20
-	andi r20, 0x0f
-	brts 27f
+	lpm r20, Z+
+	sbrc r17, 3
 	rcall shiftleft32
-	rjmp 28f
-27:	rcall shiftright32
-28:	rcall eor32_to_acc
+	rcall eor32_to_acc
 ;---
 	rcall load32_from_X
 	rcall eor32_to_acc
@@ -779,7 +766,7 @@ f2:
 	adiw r26, 4
 ;---
 	dec r17
-	brne 10b
+	brpl 10b
 ;-----
 	sbiw r28, 4*8 /* Y points to q[24] */
 	movw r30, r28
@@ -800,20 +787,13 @@ f2:
 
 	movw r26, h0
 	ldi r17, 15
-	ldi r30, lo8(f2_2_shift_table)
-	ldi r31, hi8(f2_2_shift_table)
+	ldi r30, lo8(f2_2_shift_table-8)
+	ldi r31, hi8(f2_2_shift_table-8)
 10:	movw r22, xl0
 	movw r24, xl2
-	sbrc r17, 3
-	rjmp 20f
 	lpm r20, Z+
-	lsr r20
-	brcs 15f
-	rcall shiftright32
-	rjmp 20f
-15:
+	sbrs r17, 3
 	rcall shiftleft32
-20:
 	rcall mov32_to_acc
 	rcall load32_from_Y
 	rcall eor32_to_acc
@@ -1039,7 +1019,7 @@ bmw256_lastBlock:
 
 
 /*******************************************************************************
-* void bmw256_ctx2hash(void* dest, const bmw256_ctx_t* ctx){
+* void bmw256_ctx2hash(void *dest, const bmw256_ctx_t *ctx){
 *	memcpy(dest, &(ctx->h[8]), 256/8);
 * }
 *
@@ -1059,7 +1039,7 @@ bmw256_ctx2hash:
 	ret
 
 /*******************************************************************************
-* void bmw256(void* dest, const void* msg, uint32_t length_b){
+* void bmw256(void *dest, const void *msg, uint32_t length_b){
 *	bmw_small_ctx_t ctx;
 *	bmw256_init(&ctx);
 *	while(length_b>=BMW_SMALL_BLOCKSIZE){
@@ -1088,7 +1068,7 @@ dst1 =  11
 
 
 /*******************************************************************************
-* void bmw224(void* dest, const void* msg, uint32_t length_b){
+* void bmw224(void *dest, const void *msg, uint32_t length_b){
 *	bmw_small_ctx_t ctx;
 *	bmw224_init(&ctx);
 *	while(length_b>=BMW_SMALL_BLOCKSIZE){
@@ -1122,6 +1102,7 @@ bmw256:
 	push_range 28, 29
 	push_range 8, 9
 	stack_alloc_large 64+4
+	adiw r30, 1
 10:	movw ctx0, r30
 	movw dst0, r24
 	movw msg0, r22