X-Git-Url: https://git.cryptolib.org/?a=blobdiff_plain;f=bmw%2Fbmw_small.c;h=b41e923ac2f8d0776951c485667e0fdeee80c571;hb=872dff3138536a68b6dd96b182a386d242bee6cb;hp=9f9393562e42fd5a3ef789f866b70ed4e4c123d6;hpb=aa060d30f8600acb9cc5bcf9fa3d97a2fbe5ccda;p=avr-crypto-lib.git

diff --git a/bmw/bmw_small.c b/bmw/bmw_small.c
index 9f93935..b41e923 100644
--- a/bmw/bmw_small.c
+++ b/bmw/bmw_small.c
@@ -28,6 +28,7 @@
 #include <stdint.h>
 #include <string.h>
 #include <avr/pgmspace.h>
+#include "memxor.h"
 #include "bmw_small.h"
 
 
@@ -194,7 +195,7 @@ uint32_t bmw_small_r7(uint32_t x){
 /*
 #define K 0x05555555L
 static
-uint32_t k_lut[] PROGMEM = {
+const uint32_t k_lut[] PROGMEM = {
 	16L*K, 17L*K, 18L*K, 19L*K, 20L*K, 21L*K, 22L*K, 23L*K,
 	24L*K, 25L*K, 26L*K, 27L*K, 28L*K, 29L*K, 30L*K, 31L*K
 };
@@ -202,7 +203,7 @@ uint32_t k_lut[] PROGMEM = {
 /* same as above but precomputed to avoid compiler warnings */
 
 static
-uint32_t k_lut[] PROGMEM = {
+const uint32_t k_lut[] PROGMEM = {
 	0x55555550L, 0x5aaaaaa5L, 0x5ffffffaL,
 	0x6555554fL, 0x6aaaaaa4L, 0x6ffffff9L,
 	0x7555554eL, 0x7aaaaaa3L, 0x7ffffff8L,
@@ -272,8 +273,8 @@ uint32_t bmw_small_expand2(uint8_t j, const uint32_t* q, const void* m, const vo
 
 #if F0_HACK==2
 /* to understand this implementation take a look at f0-opt-table.txt */
-static uint16_t hack_table[5] PROGMEM = { 0x0311, 0xDDB3, 0x2A79, 0x07AA, 0x51C2 };
-static uint8_t  offset_table[5] PROGMEM = { 4+16, 6+16, 9+16, 12+16, 13+16 };
+static const uint16_t hack_table[5] PROGMEM = { 0x0311, 0xDDB3, 0x2A79, 0x07AA, 0x51C2 };
+static const uint8_t  offset_table[5] PROGMEM = { 4+16, 6+16, 9+16, 12+16, 13+16 };
 
 static
 void bmw_small_f0(uint32_t* q, uint32_t* h, const void* m){
@@ -318,7 +319,7 @@ void bmw_small_f0(uint32_t* q, uint32_t* h, const void* m){
 
 #if F0_HACK==1
 static
-uint8_t f0_lut[] PROGMEM = {
+const uint8_t f0_lut[] PROGMEM = {
 	 5<<1, ( 7<<1)+1, (10<<1)+0, (13<<1)+0, (14<<1)+0,
 	 6<<1, ( 8<<1)+1, (11<<1)+0, (14<<1)+0, (15<<1)+1,
 	 0<<1, ( 7<<1)+0, ( 9<<1)+0, (12<<1)+1, (15<<1)+0,
@@ -430,7 +431,7 @@ void bmw_small_f1(uint32_t* q, const void* m, const void* h){
 }
 
 static
-void bmw_small_f2(uint32_t* h, const uint32_t* q, const void* m){
+void bmw_small_f2(uint32_t* h, uint32_t* q, const void* m){
 	uint32_t xl=0, xh;
 	uint8_t i;
 	for(i=16;i<24;++i){
@@ -462,6 +463,7 @@ void bmw_small_f2(uint32_t* h, const uint32_t* q, const void* m){
 		h[8+i] ^= xh ^ q[24+i];
 		h[8+i] += ROTL32(h[(4+i)%8],i+9);
 	}
+/*
 	h[ 8] += SHL32(xl, 8) ^ q[23] ^ q[ 8];
 	h[ 9] += SHR32(xl, 6) ^ q[16] ^ q[ 9];
 	h[10] += SHL32(xl, 6) ^ q[17] ^ q[10];
@@ -470,6 +472,18 @@ void bmw_small_f2(uint32_t* h, const uint32_t* q, const void* m){
 	h[13] += SHR32(xl, 4) ^ q[20] ^ q[13];
 	h[14] += SHR32(xl, 7) ^ q[21] ^ q[14];
 	h[15] += SHR32(xl, 2) ^ q[22] ^ q[15];
+*/
+	memxor(q+9, q+16, 7*4);
+	q[8] ^= q[23];
+	h[ 8] += SHL32(xl, 8) ^ q[ 8];
+	h[ 9] += SHR32(xl, 6) ^ q[ 9];
+	h[10] += SHL32(xl, 6) ^ q[10];
+	h[11] += SHL32(xl, 4) ^ q[11];
+	h[12] += SHR32(xl, 3) ^ q[12];
+	h[13] += SHR32(xl, 4) ^ q[13];
+	h[14] += SHR32(xl, 7) ^ q[14];
+	h[15] += SHR32(xl, 2) ^ q[15];
+
 }
 
 void bmw_small_nextBlock(bmw_small_ctx_t* ctx, const void* block){
@@ -485,38 +499,42 @@ void bmw_small_nextBlock(bmw_small_ctx_t* ctx, const void* block){
 }
 
 void bmw_small_lastBlock(bmw_small_ctx_t* ctx, const void* block, uint16_t length_b){
-	uint8_t buffer[64];
+	union {
+		uint8_t   v8[64];
+		uint32_t v32[16];
+		uint64_t v64[ 8];
+	} buffer;
 	while(length_b >= BMW_SMALL_BLOCKSIZE){
 		bmw_small_nextBlock(ctx, block);
 		length_b -= BMW_SMALL_BLOCKSIZE;
 		block = (uint8_t*)block + BMW_SMALL_BLOCKSIZE_B;
 	}
-	memset(buffer, 0, 64);
-	memcpy(buffer, block, (length_b+7)/8);
-	buffer[length_b>>3] |= 0x80 >> (length_b&0x07);
+	memset(buffer.v8, 0, 64);
+	memcpy(buffer.v8, block, (length_b+7)/8);
+	buffer.v8[length_b>>3] |= 0x80 >> (length_b&0x07);
 	if(length_b+1>64*8-64){
-		bmw_small_nextBlock(ctx, buffer);
-		memset(buffer, 0, 64-8);
+		bmw_small_nextBlock(ctx, buffer.v8);
+		memset(buffer.v8, 0, 64-8);
 		ctx->counter -= 1;
 	}
-	*((uint64_t*)&(buffer[64-8])) = (uint64_t)(ctx->counter*512LL)+(uint64_t)length_b;
-	bmw_small_nextBlock(ctx, buffer);
+	buffer.v64[7] = (uint64_t)(ctx->counter*512LL)+(uint64_t)length_b;
+	bmw_small_nextBlock(ctx, buffer.v8);
 #if TWEAK
 	uint8_t i;
 	uint32_t q[32];
-	memset(buffer, 0xaa, 64);
+	memset(buffer.v8, 0xaa, 64);
 	for(i=0; i<16;++i){
-		buffer[i*4] = i+0xa0;
+		buffer.v8[i*4] = i+0xa0;
 	}
-//	dump_x(buffer, 16, 'A');
+//	dump_x(buffer.v8, 16, 'A');
 	dump_x(ctx->h, 16, 'M');
-	bmw_small_f0(q, (uint32_t*)buffer, ctx->h);
-	dump_x(buffer, 16, 'a');
+	bmw_small_f0(q, buffer.v32, ctx->h);
+	dump_x(buffer.v8, 16, 'a');
 	dump_x(q, 16, 'Q');
-	bmw_small_f1(q, ctx->h, (uint32_t*)buffer);
+	bmw_small_f1(q, ctx->h, buffer.v32);
 	dump_x(q, 32, 'Q');
-	bmw_small_f2((uint32_t*)buffer, q, ctx->h);
-	memcpy(ctx->h, buffer, 64);
+	bmw_small_f2(buffer.v32, q, ctx->h);
+	memcpy(ctx->h, buffer.v8, 64);
 #endif
 }