clearing up inclusion structure for easier integration in supercop

[avr-crypto-lib.git] / blake / blake_large.c

diff --git a/blake/blake_large.c b/blake/blake_large.c
index 43b0fa78bdc49401898120b0b900e0e08da033b1..af967e2555eee61a45c9b8ba9261e2a595fab4de 100644 (file)
--- a/blake/blake_large.c
+++ b/blake/blake_large.c
@@ -22,7 +22,7 @@
  * \email   daniel.otte@rub.de
  * \date    2009-05-08
  * \license GPLv3 or later
- * 
+ *
  */
 
 #include <stdint.h>
@@ -32,9 +32,7 @@
 #include "blake_large.h"
 #include "blake_common.h"
 
-#define BUG_3 1 /* bug compatibility with reference code */
-#define BUG_4 1 /* bug compatibility with reference code */
-
+static
 uint64_t pgm_read_qword(void* p){
        union{
                uint64_t v64;
@@ -46,7 +44,7 @@ uint64_t pgm_read_qword(void* p){
 }
 
 static
-uint64_t blake_c[] PROGMEM = {   
+uint64_t blake_c[] PROGMEM = {
    0x243F6A8885A308D3LL, 0x13198A2E03707344LL,
    0xA4093822299F31D0LL, 0x082EFA98EC4E6C89LL,
    0x452821E638D01377LL, 0xBE5466CF34E90C6CLL,
@@ -55,18 +53,18 @@ uint64_t blake_c[] PROGMEM = {
    0x2FFD72DBD01ADFB7LL, 0xB8E1AFED6A267E96LL,
    0xBA7C9045F12C7F99LL, 0x24A19947B3916CF7LL,
    0x0801F2E2858EFC16LL, 0x636920D871574E69LL
-
 };
 
 
 
-#define ROTL64(a, n) (((a)<<(n))|((a)>>(64-(n)))) 
-#define ROTR64(a, n) (((a)>>(n))|((a)<<(64-(n)))) 
+#define ROTL64(a, n) (((a)<<(n))|((a)>>(64-(n))))
+#define ROTR64(a, n) (((a)>>(n))|((a)<<(64-(n))))
 #define CHANGE_ENDIAN32(a) (((a)<<24)| \
                             ((0x0000ff00&(a))<<8)| \
                                                    ((0x00ff0000&(a))>>8)| \
                                                    (a)>>24 )
 
+static
 void blake_large_expand(uint64_t* v, const blake_large_ctx_t* ctx){
        uint8_t i;
        memcpy(v, ctx->h, 8*8);
@@ -74,9 +72,10 @@ void blake_large_expand(uint64_t* v, const blake_large_ctx_t* ctx){
                v[8+i] = pgm_read_qword(&(blake_c[i]));
        }
        memxor((uint8_t*)v+8, ctx->s, 4*8);
-       
+
 }
 
+static
 void blake_large_changeendian(void* dest, const void* src){
        uint8_t i;
        uint32_t tmp;
@@ -87,12 +86,12 @@ void blake_large_changeendian(void* dest, const void* src){
        }
 }
 
+static
 void blake_large_compress(uint64_t* v,const void* m){
        uint8_t r,i;
        uint8_t a,b,c,d, s0, s1;
        for(r=0; r<14; ++r){
                for(i=0; i<8; ++i){
-       //              blake_large_g(r%10, i, v, (uint64_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);
@@ -102,7 +101,7 @@ void blake_large_compress(uint64_t* v,const void* m){
                        v[a] += v[b] + (((uint64_t*)m)[s0] ^ pgm_read_qword(&(blake_c[s1])));
                        v[d]  = ROTR64(v[d]^v[a], 32);
                        v[c] += v[d];
-                       v[b]  = ROTR64(v[b]^v[c], 25);  
+                       v[b]  = ROTR64(v[b]^v[c], 25);
                        v[a] += v[b] + (((uint64_t*)m)[s1] ^ pgm_read_qword(&(blake_c[s0])));
                        v[d]  = ROTR64(v[d]^v[a], 16);
                        v[c] += v[d];
@@ -111,12 +110,13 @@ void blake_large_compress(uint64_t* v,const void* m){
        }
 }
 
+static
 void blake_large_collapse(blake_large_ctx_t* ctx, uint64_t* v){
        uint8_t i;
        for(i=0; i<8; ++i){
                ctx->h[i] ^= ctx->s[i%4] ^ v[i] ^ v[8+i];
        }
-}      
+}
 
 void blake_large_nextBlock(blake_large_ctx_t* ctx, const void* msg){
        uint64_t v[16];
@@ -147,39 +147,28 @@ void blake_large_lastBlock(blake_large_ctx_t* ctx, const void* msg, uint16_t len
        ctr = ctx->counter*1024+length_b;
        memset(buffer, 0, 128);
        memcpy(buffer, msg, (length_b+7)/8);
-       buffer[length_b/8] |= 0x80 >> (length_b%8);
+       buffer[length_b/8] |= 0x80 >> (length_b&0x7);
        blake_large_changeendian(buffer, buffer);
        blake_large_expand(v, ctx);
-#if BUG_3
-       uint8_t x=0;
-       if(length_b%1024<895 && length_b%8)
-               x=0x40;
-       v[12] ^= ctr + x;
-       v[13] ^= ctr + x;
-       
-#else  
-       v[12] ^= ctr;
-       v[13] ^= ctr;
-#endif
        if(length_b>1024-128-2){
-#if BUG_4
-               if(length_b<1017){
-                       blake_large_compress(v, buffer);
-                       blake_large_collapse(ctx, v);
-               }       
-#else  
+               v[12] ^= ctr;
+               v[13] ^= ctr;
                blake_large_compress(v, buffer);
                blake_large_collapse(ctx, v);
-#endif
                memset(buffer, 0, 128-8);
                blake_large_expand(v, ctx);
+       } else {
+               if(length_b){
+                       v[12] ^= ctr;
+                       v[13] ^= ctr;
+               }
        }
        if(ctx->appendone)
-               buffer[128-16-8] |= 0x01;       
+               buffer[128-16-8] |= 0x01;
        *((uint64_t*)(&(buffer[128-8]))) = ctr;
        blake_large_compress(v, buffer);
        blake_large_collapse(ctx, v);
-       
+
 }
 
 uint64_t blake64_iv[] PROGMEM = {