X-Git-Url: https://git.cryptolib.org/?p=avr-crypto-lib.git;a=blobdiff_plain;f=bmw%2Fbmw_small.c;h=f2bc22821ae3e219536a0df33494e66b5caf285e;hp=8a4b1652734605710eddee45eae90683b3346045;hb=4b5da1dc27a791b5c448274a3db09cd035b33493;hpb=302f07c97444e2f5f5560dc2fd9af58d4b591d45

diff --git a/bmw/bmw_small.c b/bmw/bmw_small.c
index 8a4b165..f2bc228 100644
--- a/bmw/bmw_small.c
+++ b/bmw/bmw_small.c
@@ -1,7 +1,7 @@
 /* bmw_small.c */
 /*
     This file is part of the AVR-Crypto-Lib.
-    Copyright (C) 2009  Daniel Otte (daniel.otte@rub.de)
+    Copyright (C) 2006-2015 Daniel Otte (bg@nerilex.org)
 
     This program is free software: you can redistribute it and/or modify
     it under the terms of the GNU General Public License as published by
@@ -19,7 +19,7 @@
 /*
  * \file    bmw_small.c
  * \author  Daniel Otte
- * \email   daniel.otte@rub.de
+ * \email   bg@nerilex.org
  * \date    2009-04-27
  * \license GPLv3 or later
  *
@@ -28,6 +28,7 @@
 #include <stdint.h>
 #include <string.h>
 #include <avr/pgmspace.h>
+#include "memxor.h"
 #include "bmw_small.h"
 
 
@@ -44,14 +45,18 @@
 #  define BUG24   1
 #endif
 
-#define F0_HACK 1
+#define F0_HACK 2
 
 #define DEBUG 0
 
+#ifndef F0_HACK
+#  define F0_HACK 0
+#endif
+
 #if DEBUG
  #include "cli.h"
 
- void ctx_dump(const bmw_small_ctx_t* ctx){
+ void ctx_dump(const bmw_small_ctx_t *ctx){
  	uint8_t i;
 	cli_putstr_P(PSTR("\r\n==== ctx dump ===="));
 	for(i=0; i<16;++i){
@@ -64,7 +69,7 @@
 	cli_hexdump(&(ctx->counter), 4);
  }
 
- void dump_x(const uint32_t* q, uint8_t elements, char x){
+ void dump_x(const uint32_t *q, uint8_t elements, char x){
 	uint8_t i;
  	cli_putstr_P(PSTR("\r\n==== "));
 	cli_putc(x);
@@ -83,6 +88,7 @@
  #define dump_x(a,b,c)
 #endif
 
+static
 uint32_t bmw_small_s0(uint32_t x){
 	uint32_t r;
 	r =   SHR32(x, 1)
@@ -92,6 +98,7 @@ uint32_t bmw_small_s0(uint32_t x){
 	return r;
 }
 
+static
 uint32_t bmw_small_s1(uint32_t x){
 	uint32_t r;
 	r =   SHR32(x, 1)
@@ -101,6 +108,7 @@ uint32_t bmw_small_s1(uint32_t x){
 	return r;
 }
 
+static
 uint32_t bmw_small_s2(uint32_t x){
 	uint32_t r;
 	r =   SHR32(x, 2)
@@ -110,6 +118,7 @@ uint32_t bmw_small_s2(uint32_t x){
 	return r;
 }
 
+static
 uint32_t bmw_small_s3(uint32_t x){
 	uint32_t r;
 	r =   SHR32(x, 2)
@@ -119,6 +128,7 @@ uint32_t bmw_small_s3(uint32_t x){
 	return r;
 }
 
+static
 uint32_t bmw_small_s4(uint32_t x){
 	uint32_t r;
 	r =   SHR32(x, 1)
@@ -126,6 +136,7 @@ uint32_t bmw_small_s4(uint32_t x){
 	return r;
 }
 
+static
 uint32_t bmw_small_s5(uint32_t x){
 	uint32_t r;
 	r =   SHR32(x, 2)
@@ -133,42 +144,49 @@ uint32_t bmw_small_s5(uint32_t x){
 	return r;
 }
 
+static
 uint32_t bmw_small_r1(uint32_t x){
 	uint32_t r;
 	r =   ROTL32(x, 3);
 	return r;
 }
 
+static
 uint32_t bmw_small_r2(uint32_t x){
 	uint32_t r;
 	r =   ROTL32(x, 7);
 	return r;
 }
 
+static
 uint32_t bmw_small_r3(uint32_t x){
 	uint32_t r;
 	r =   ROTL32(x, 13);
 	return r;
 }
 
+static
 uint32_t bmw_small_r4(uint32_t x){
 	uint32_t r;
 	r =   ROTL32(x, 16);
 	return r;
 }
 
+static
 uint32_t bmw_small_r5(uint32_t x){
 	uint32_t r;
 	r =   ROTR32(x, 13);
 	return r;
 }
 
+static
 uint32_t bmw_small_r6(uint32_t x){
 	uint32_t r;
 	r =   ROTR32(x, 9);
 	return r;
 }
 
+static
 uint32_t bmw_small_r7(uint32_t x){
 	uint32_t r;
 	r =   ROTR32(x, 5);
@@ -177,7 +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
 };
@@ -185,7 +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,
@@ -193,31 +211,33 @@ uint32_t k_lut[] PROGMEM = {
 	0x9555554cL, 0x9aaaaaa1L, 0x9ffffff6L,
 	0xa555554bL };
 
-
-uint32_t bmw_small_expand1(uint8_t j, const uint32_t* q, const void* m, const void* h){
+static
+uint32_t bmw_small_expand1(uint8_t j, const uint32_t *q, const void *m, const void *h){
 	uint32_t(*s[])(uint32_t) = {bmw_small_s1, bmw_small_s2, bmw_small_s3, bmw_small_s0};
-	uint32_t r=0;
+	uint32_t r;
 	uint8_t i;
 	/* r = 0x05555555*(j+16); */
-	for(i=0; i<16; ++i){
-		r += s[i%4](q[j+i]);
-	}
+
 #if TWEAK
-	r += (   ROTL32(((uint32_t*)m)[j&0xf],      ((j+0)&0xf)+1  )
+	r = (   ROTL32(((uint32_t*)m)[j&0xf],      ((j+0)&0xf)+1  )
 	       + ROTL32(((uint32_t*)m)[(j+3)&0xf],  ((j+3)&0xf)+1  )
 	       - ROTL32(((uint32_t*)m)[(j+10)&0xf], ((j+10)&0xf)+1 )
 	       + pgm_read_dword(k_lut+j)
 	     ) ^ ((uint32_t*)h)[(j+7)&0xf];
 #else
-	r += pgm_read_dword(k_lut+j);
+	r = pgm_read_dword(k_lut+j);
 	r += ((uint32_t*)m)[j&0xf];
 	r += ((uint32_t*)m)[(j+3)&0xf];
 	r -= ((uint32_t*)m)[(j+10)&0xf];
 #endif
+	for(i=0; i<16; ++i){
+		r += s[i%4](q[j+i]);
+	}
 	return r;
 }
 
-uint32_t bmw_small_expand2(uint8_t j, const uint32_t* q, const void* m, const void* h){
+static
+uint32_t bmw_small_expand2(uint8_t j, const uint32_t *q, const void *m, const void *h){
 	uint32_t(*rf[])(uint32_t) = {bmw_small_r1, bmw_small_r2, bmw_small_r3,
 	                             bmw_small_r4, bmw_small_r5, bmw_small_r6,
 							     bmw_small_r7};
@@ -251,9 +271,55 @@ uint32_t bmw_small_expand2(uint8_t j, const uint32_t* q, const void* m, const vo
 	return r;
 }
 
-#if F0_HACK
+#if F0_HACK==2
+/* to understand this implementation take a look at f0-opt-table.txt */
+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){
+	uint16_t hack_reg;
+	uint8_t c,i,j;
+	uint32_t(*s[])(uint32_t)={ bmw_small_s0, bmw_small_s1, bmw_small_s2,
+	                           bmw_small_s3, bmw_small_s4 };
+	for(i=0; i<16; ++i){
+		((uint32_t*)h)[i] ^= ((uint32_t*)m)[i];
+	}
+	dump_x(h, 16, 'T');
+	memset(q, 0, 4*16);
+	c=4;
+	do{
+		i=15;
+		j=pgm_read_byte(offset_table+c);
+		hack_reg=pgm_read_word(&(hack_table[c]));
+		do{
+			if(hack_reg&1){
+				q[i]-= h[j&15];
+			}else{
+				q[i]+= h[j&15];
+			}
+			--j;
+			hack_reg>>= 1;
+		}while(i--!=0);
+	}while(c--!=0);
+	dump_x(q, 16, 'W');
+	for(i=0; i<16; ++i){
+		q[i] = s[i%5](q[i]);
+	}
+#if TWEAK
+	for(i=0; i<16; ++i){
+		((uint32_t*)h)[i] ^= ((uint32_t*)m)[i];
+	}
+	for(i=0; i<16; ++i){
+		q[i] += h[(i+1)&0xf];
+	}
+#endif
+}
+#endif /* F0_HACK==2*/
+
+#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,
@@ -272,7 +338,8 @@ uint8_t f0_lut[] PROGMEM = {
 	12<<1, ( 4<<1)+1, ( 6<<1)+1, ( 9<<1)+1, (13<<1)+0
 };
 
-void bmw_small_f0(uint32_t* q, uint32_t* h, const void* m){
+static
+void bmw_small_f0(uint32_t *q, uint32_t *h, const void *m){
 	uint8_t i,j=-1,v,sign,l=0;
 	uint32_t(*s[])(uint32_t)={ bmw_small_s0, bmw_small_s1, bmw_small_s2,
 	                           bmw_small_s3, bmw_small_s4 };
@@ -308,11 +375,13 @@ void bmw_small_f0(uint32_t* q, uint32_t* h, const void* m){
 	for(i=0; i<16; ++i){
 		q[i] += h[(i+1)&0xf];
 	}
-#endif
+#endif /* TWEAK */
 }
+#endif /* F0_HACK==1 */
 
-#else
-void bmw_small_f0(uint32_t* q, uint32_t* h, const void* m){
+#if F0_HACK==0
+static
+void bmw_small_f0(uint32_t *q, uint32_t *h, const void *m){
 	uint8_t i;
 	uint32_t(*s[])(uint32_t)={ bmw_small_s0, bmw_small_s1, bmw_small_s2,
 	                           bmw_small_s3, bmw_small_s4 };
@@ -347,11 +416,12 @@ void bmw_small_f0(uint32_t* q, uint32_t* h, const void* m){
 	for(i=0; i<16; ++i){
 		q[i] += h[(i+1)&0xf];
 	}
-#endif
+#endif /* TWEAK */
 }
-#endif
+#endif /* F0_HACK==0 */
 
-void bmw_small_f1(uint32_t* q, const void* m, const void* h){
+static
+void bmw_small_f1(uint32_t *q, const void *m, const void *h){
 	uint8_t i;
 	q[16] = bmw_small_expand1(0, q, m, h);
 	q[17] = bmw_small_expand1(1, q, m, h);
@@ -360,7 +430,8 @@ void bmw_small_f1(uint32_t* q, const void* m, const void* h){
 	}
 }
 
-void bmw_small_f2(uint32_t* h, const uint32_t* q, const void* m){
+static
+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){
@@ -392,6 +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];
@@ -400,9 +472,21 @@ 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){
+void bmw_small_nextBlock(bmw_small_ctx_t *ctx, const void *block){
 	uint32_t q[32];
 	dump_x(block, 16, 'M');
 	bmw_small_f0(q, ctx->h, block);
@@ -414,43 +498,47 @@ void bmw_small_nextBlock(bmw_small_ctx_t* ctx, const void* block){
 	ctx_dump(ctx);
 }
 
-void bmw_small_lastBlock(bmw_small_ctx_t* ctx, const void* block, uint16_t length_b){
-	uint8_t buffer[64];
+void bmw_small_lastBlock(bmw_small_ctx_t *ctx, const void *block, uint16_t length_b){
+	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
 }
 
-void bmw224_init(bmw224_ctx_t* ctx){
+void bmw224_init(bmw224_ctx_t *ctx){
 	uint8_t i;
 	ctx->h[0] = 0x00010203;
 	for(i=1; i<16; ++i){
@@ -463,7 +551,7 @@ void bmw224_init(bmw224_ctx_t* ctx){
 	ctx_dump(ctx);
 }
 
-void bmw256_init(bmw256_ctx_t* ctx){
+void bmw256_init(bmw256_ctx_t *ctx){
 	uint8_t i;
 	ctx->h[0] = 0x40414243;
 	for(i=1; i<16; ++i){
@@ -473,31 +561,31 @@ void bmw256_init(bmw256_ctx_t* ctx){
 	ctx_dump(ctx);
 }
 
-void bmw224_nextBlock(bmw224_ctx_t* ctx, const void* block){
+void bmw224_nextBlock(bmw224_ctx_t *ctx, const void *block){
 	bmw_small_nextBlock(ctx, block);
 }
 
-void bmw256_nextBlock(bmw256_ctx_t* ctx, const void* block){
+void bmw256_nextBlock(bmw256_ctx_t *ctx, const void *block){
 	bmw_small_nextBlock(ctx, block);
 }
 
-void bmw224_lastBlock(bmw224_ctx_t* ctx, const void* block, uint16_t length_b){
+void bmw224_lastBlock(bmw224_ctx_t *ctx, const void *block, uint16_t length_b){
 	bmw_small_lastBlock(ctx, block, length_b);
 }
 
-void bmw256_lastBlock(bmw256_ctx_t* ctx, const void* block, uint16_t length_b){
+void bmw256_lastBlock(bmw256_ctx_t *ctx, const void *block, uint16_t length_b){
 	bmw_small_lastBlock(ctx, block, length_b);
 }
 
-void bmw224_ctx2hash(void* dest, const bmw224_ctx_t* ctx){
+void bmw224_ctx2hash(void *dest, const bmw224_ctx_t *ctx){
 	memcpy(dest, &(ctx->h[9]), 224/8);
 }
 
-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);
 }
 
-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){
@@ -509,7 +597,7 @@ void bmw224(void* dest, const void* msg, uint32_t length_b){
 	bmw224_ctx2hash(dest, &ctx);
 }
 
-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){