big bug fixed, still some problems with flow control

[avr-crypto-lib.git] / blake_small.c

diff --git a/blake_small.c b/blake_small.c
index edf5cda4e4facdd89fcf1f37e4aad9c1a72b0bfa..75a28374c4772d06105d1e60c325daffba8c9be8 100644 (file)
--- a/blake_small.c
+++ b/blake_small.c
@@ -32,32 +32,9 @@
 #include "blake_small.h"
 #include "blake_common.h"
 
-#define DEBUG 0
-#define DEBUG_2 0
+#define BUG_1 1 /* bug compatibility for zero length message */
+#define BUG_2 1 /* bug compatibility for messages of length%512=505...511 */
 
-#define BUG_1 0 /* bug compatibility for zero length message */
-#define BUG_2 0 /* bug compatibility for messages of length%512=505...511 */
-
-
-#if DEBUG_2
- #include "cli.h"
-#endif
-
-#if DEBUG
- #include "cli.h"
- void dump_v(uint32_t* v){
-       uint8_t i;
-       cli_putstr_P(PSTR("\r\n=== v dump ==="));
-       for(i=0; i<16; ++i){
-               if(i%8==0)
-                       cli_putstr_P(PSTR("\r\n\t"));
-               cli_hexdump_rev(v+i, 4);
-               cli_putc(' ');
-       }
- }
-#else
- #define dump_v(v)
-#endif
 
 uint32_t blake_c[] PROGMEM = {
    0x243F6A88, 0x85A308D3,
@@ -77,37 +54,6 @@ uint32_t blake_c[] PROGMEM = {
                                                    ((0x00ff0000&(a))>>8)| \
                                                    (a)>>24 )
 
-void blake_small_g(uint8_t r, uint8_t i, uint32_t* v, const uint32_t* m){
-       uint8_t a,b,c,d, s0, s1;
-       a = pgm_read_byte(blake_index_lut+4*i+0);
-       b = pgm_read_byte(blake_index_lut+4*i+1);
-       c = pgm_read_byte(blake_index_lut+4*i+2);
-       d = pgm_read_byte(blake_index_lut+4*i+3);
-       s0 = pgm_read_byte(blake_sigma+16*r+2*i+0);
-       s1 = pgm_read_byte(blake_sigma+16*r+2*i+1);
-#if DEBUG
-       if(i==0){
-               cli_putstr_P(PSTR("\r\n s0 = "));
-               cli_hexdump(&s0, 1);
-               cli_putstr_P(PSTR("   s1 = "));
-               cli_hexdump(&s1, 1);
-               cli_putstr_P(PSTR("\r\n m[s0] = "));
-               cli_hexdump_rev(m+s0, 4);
-               cli_putstr_P(PSTR("\r\n m[s1] = "));
-               cli_hexdump_rev(m+s1, 4);
-       }
-#endif         
-       v[a] += v[b] + (m[s0] ^ pgm_read_dword(&(blake_c[s1])));
-       v[d]  = ROTR32(v[d]^v[a], 16);
-       v[c] += v[d];
-       v[b]  = ROTR32(v[b]^v[c], 12);  
-       v[a] += v[b] + (m[s1] ^ pgm_read_dword(&(blake_c[s0])));
-       v[d]  = ROTR32(v[d]^v[a], 8);
-       v[c] += v[d];
-       v[b]  = ROTR32(v[b]^v[c], 7);
-
-}
-
 void blake_small_expand(uint32_t* v, const blake_small_ctx_t* ctx){
        uint8_t i;
        memcpy(v, ctx->h, 8*4);
@@ -127,26 +73,36 @@ void blake_small_changeendian(void* dest, const void* src){
 
 void blake_small_compress(uint32_t* v,const void* m){
        uint8_t r,i;
-#if DEBUG      
-       cli_putstr_P(PSTR("\r\n== compress 32 =="));
-       dump_v(v);
-#endif
-#if DEBUG_2
-       cli_putstr_P(PSTR("\r\n=== message block ===\r\n m ="));
-       cli_hexdump_block(m, 512/8, 4, 16);
-#endif
+       uint8_t a,b,c,d, s0, s1;
+       uint32_t lv[4];
        for(r=0; r<10; ++r){
                for(i=0; i<8; ++i){
-                       blake_small_g(r, i, v, (uint32_t*)m);
-#if DEBUG
-                       if(1){
-                               cli_putstr_P(PSTR("\r\n ROUND: "));
-                               cli_hexdump(&r,1);
-                               cli_putstr_P(PSTR("    I: "));
-                               cli_hexdump(&i,1);
-                               dump_v(v);
-                       }
-#endif
+       //              blake_small_g(r, i, v, (uint32_t*)m);
+                       a = pgm_read_byte(blake_index_lut+4*i+0);
+                       b = pgm_read_byte(blake_index_lut+4*i+1);
+                       c = pgm_read_byte(blake_index_lut+4*i+2);
+                       d = pgm_read_byte(blake_index_lut+4*i+3);
+                       s0 = pgm_read_byte(blake_sigma+16*r+2*i+0);
+                       s1 = pgm_read_byte(blake_sigma+16*r+2*i+1);
+                       lv[0] = v[a];
+                       lv[1] = v[b];
+                       lv[2] = v[c];
+                       lv[3] = v[d];
+                       
+                       lv[0] += lv[1] + (((uint32_t*)m)[s0] ^ pgm_read_dword(&(blake_c[s1])));
+                       lv[3]  = ROTR32(lv[3]^lv[0], 16);
+                       lv[2] += lv[3];
+                       lv[1]  = ROTR32(lv[1]^lv[2], 12);       
+                       lv[0] += lv[1] + (((uint32_t*)m)[s1] ^ pgm_read_dword(&(blake_c[s0])));
+                       lv[3]  = ROTR32(lv[3]^lv[0], 8);
+                       lv[2] += lv[3];
+                       lv[1]  = ROTR32(lv[1]^lv[2], 7);
+
+                       v[a] = lv[0];
+                       v[b] = lv[1];
+                       v[c] = lv[2];
+                       v[d] = lv[3];
+
                }
        }
 }