progress in asm version of keccak (theta now in asm)

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

diff --git a/blake/blake_large.c b/blake/blake_large.c
index b0b6cebf1c4df679d6d24168066a70d9fa3488ef..9dc118bd1512cff454e24a3b581d31c64d2d12d6 100644 (file)
--- a/blake/blake_large.c
+++ b/blake/blake_large.c
@@ -33,7 +33,7 @@
 #include "blake_common.h"
 
 static
-uint64_t pgm_read_qword(void* p){
+uint64_t pgm_read_qword(const void* p){
        union{
                uint64_t v64;
                uint32_t v32[2];
@@ -44,7 +44,7 @@ uint64_t pgm_read_qword(void* p){
 }
 
 static
-uint64_t blake_c[] PROGMEM = {
+const uint64_t blake_c[] PROGMEM = {
    0x243F6A8885A308D3LL, 0x13198A2E03707344LL,
    0xA4093822299F31D0LL, 0x082EFA98EC4E6C89LL,
    0x452821E638D01377LL, 0xBE5466CF34E90C6CLL,
@@ -146,21 +146,24 @@ void blake_large_lastBlock(blake_large_ctx_t* ctx, const void* msg, uint16_t len
                msg = (uint8_t*)msg + BLAKE_LARGE_BLOCKSIZE_B;
                length_b -= BLAKE_LARGE_BLOCKSIZE;
        }
-       uint8_t buffer[128];
+       union {
+               uint8_t   v8[128];
+               uint64_t v64[ 16];
+       } buffer;
        uint64_t v[16];
        uint64_t ctr;
        ctr = ctx->counter*1024+length_b;
-       memset(buffer, 0, 128);
-       memcpy(buffer, msg, (length_b+7)/8);
-       buffer[length_b/8] |= 0x80 >> (length_b&0x7);
-       blake_large_changeendian(buffer, buffer);
+       memset(buffer.v8, 0, 128);
+       memcpy(buffer.v8, msg, (length_b+7)/8);
+       buffer.v8[length_b/8] |= 0x80 >> (length_b&0x7);
+       blake_large_changeendian(buffer.v8, buffer.v8);
        blake_large_expand(v, ctx);
        if(length_b>1024-128-2){
                v[12] ^= ctr;
                v[13] ^= ctr;
-               blake_large_compress(v, buffer);
+               blake_large_compress(v, buffer.v8);
                blake_large_collapse(ctx, v);
-               memset(buffer, 0, 128-8);
+               memset(buffer.v8, 0, 128-8);
                blake_large_expand(v, ctx);
        } else {
                if(length_b){
@@ -169,14 +172,14 @@ void blake_large_lastBlock(blake_large_ctx_t* ctx, const void* msg, uint16_t len
                }
        }
        if(ctx->appendone)
-               buffer[128-16-8] |= 0x01;
-       *((uint64_t*)(&(buffer[128-8]))) = ctr;
-       blake_large_compress(v, buffer);
+               buffer.v8[128-16-8] |= 0x01;
+       buffer.v64[15] = ctr;
+       blake_large_compress(v, buffer.v8);
        blake_large_collapse(ctx, v);
 
 }
 
-uint64_t blake512_iv[] PROGMEM = {
+const uint64_t blake512_iv[] PROGMEM = {
     0x6A09E667F3BCC908LL, 0xBB67AE8584CAA73BLL,
     0x3C6EF372FE94F82BLL, 0xA54FF53A5F1D36F1LL,
     0x510E527FADE682D1LL, 0x9B05688C2B3E6C1FLL,
@@ -193,7 +196,7 @@ void blake512_init(blake512_ctx_t* ctx){
        ctx->appendone = 1;
 }
 
-uint64_t blake384_iv[] PROGMEM = {
+const uint64_t blake384_iv[] PROGMEM = {
     0xCBBB9D5DC1059ED8LL, 0x629A292A367CD507LL,
     0x9159015A3070DD17LL, 0x152FECD8F70E5939LL,
     0x67332667FFC00B31LL, 0x8EB44A8768581511LL,