summaryrefslogtreecommitdiff
path: root/drivers/staging/skein
diff options
context:
space:
mode:
Diffstat (limited to 'drivers/staging/skein')
-rw-r--r--drivers/staging/skein/skein_api.c3
-rw-r--r--drivers/staging/skein/skein_base.c90
-rw-r--r--drivers/staging/skein/skein_base.h45
-rw-r--r--drivers/staging/skein/skein_block.c92
-rw-r--r--drivers/staging/skein/skein_generic.c6
-rw-r--r--drivers/staging/skein/threefish_api.h2
-rw-r--r--drivers/staging/skein/threefish_block.c2144
7 files changed, 1192 insertions, 1190 deletions
diff --git a/drivers/staging/skein/skein_api.c b/drivers/staging/skein/skein_api.c
index 36f849fbb..cab26e736 100644
--- a/drivers/staging/skein/skein_api.c
+++ b/drivers/staging/skein/skein_api.c
@@ -165,7 +165,6 @@ int skein_update(struct skein_ctx *ctx, const u8 *msg,
break;
}
return ret;
-
}
int skein_update_bits(struct skein_ctx *ctx, const u8 *msg,
@@ -210,7 +209,7 @@ int skein_update_bits(struct skein_ctx *ctx, const u8 *msg,
/* internal sanity check: there IS a partial byte in the buffer! */
skein_assert(length != 0);
/* partial byte bit mask */
- mask = (u8) (1u << (7 - (msg_bit_cnt & 7)));
+ mask = (u8)(1u << (7 - (msg_bit_cnt & 7)));
/* apply bit padding on final byte (in the buffer) */
up[length - 1] = (u8)((up[length - 1] & (0 - mask)) | mask);
diff --git a/drivers/staging/skein/skein_base.c b/drivers/staging/skein/skein_base.c
index 25a01ca76..c24a57396 100644
--- a/drivers/staging/skein/skein_base.c
+++ b/drivers/staging/skein/skein_base.c
@@ -58,7 +58,7 @@ int skein_256_init(struct skein_256_ctx *ctx, size_t hash_bit_len)
cfg.w[1] = skein_swap64(hash_bit_len);
cfg.w[2] = skein_swap64(SKEIN_CFG_TREE_INFO_SEQUENTIAL);
/* zero pad config block */
- memset(&cfg.w[3], 0, sizeof(cfg) - 3*sizeof(cfg.w[0]));
+ memset(&cfg.w[3], 0, sizeof(cfg) - 3 * sizeof(cfg.w[0]));
/* compute the initial chaining values from config block */
/* zero the chaining variables */
@@ -98,7 +98,7 @@ int skein_256_init_ext(struct skein_256_ctx *ctx, size_t hash_bit_len,
skein_assert(sizeof(cfg.b) >= sizeof(ctx->x));
/* do a mini-Init right here */
/* set output hash bit count = state size */
- ctx->h.hash_bit_len = 8*sizeof(ctx->x);
+ ctx->h.hash_bit_len = 8 * sizeof(ctx->x);
/* set tweaks: T0 = 0; T1 = KEY type */
skein_start_new_type(ctx, KEY);
/* zero the initial chaining variables */
@@ -171,7 +171,7 @@ int skein_256_update(struct skein_256_ctx *ctx, const u8 *msg,
*/
if (msg_byte_cnt > SKEIN_256_BLOCK_BYTES) {
/* number of full blocks to process */
- n = (msg_byte_cnt-1) / SKEIN_256_BLOCK_BYTES;
+ n = (msg_byte_cnt - 1) / SKEIN_256_BLOCK_BYTES;
skein_256_process_block(ctx, msg, n,
SKEIN_256_BLOCK_BYTES);
msg_byte_cnt -= n * SKEIN_256_BLOCK_BYTES;
@@ -205,7 +205,7 @@ int skein_256_final(struct skein_256_ctx *ctx, u8 *hash_val)
/* zero pad b[] if necessary */
if (ctx->h.b_cnt < SKEIN_256_BLOCK_BYTES)
memset(&ctx->b[ctx->h.b_cnt], 0,
- SKEIN_256_BLOCK_BYTES - ctx->h.b_cnt);
+ SKEIN_256_BLOCK_BYTES - ctx->h.b_cnt);
/* process the final block */
skein_256_process_block(ctx, ctx->b, 1, ctx->h.b_cnt);
@@ -219,19 +219,19 @@ int skein_256_final(struct skein_256_ctx *ctx, u8 *hash_val)
memset(ctx->b, 0, sizeof(ctx->b));
/* keep a local copy of counter mode "key" */
memcpy(x, ctx->x, sizeof(x));
- for (i = 0; i*SKEIN_256_BLOCK_BYTES < byte_cnt; i++) {
+ for (i = 0; i * SKEIN_256_BLOCK_BYTES < byte_cnt; i++) {
/* build the counter block */
- ((u64 *)ctx->b)[0] = skein_swap64((u64) i);
+ ((u64 *)ctx->b)[0] = skein_swap64((u64)i);
skein_start_new_type(ctx, OUT_FINAL);
/* run "counter mode" */
skein_256_process_block(ctx, ctx->b, 1, sizeof(u64));
/* number of output bytes left to go */
- n = byte_cnt - i*SKEIN_256_BLOCK_BYTES;
+ n = byte_cnt - i * SKEIN_256_BLOCK_BYTES;
if (n >= SKEIN_256_BLOCK_BYTES)
n = SKEIN_256_BLOCK_BYTES;
/* "output" the ctr mode bytes */
- skein_put64_lsb_first(hash_val+i*SKEIN_256_BLOCK_BYTES, ctx->x,
- n);
+ skein_put64_lsb_first(hash_val + (i * SKEIN_256_BLOCK_BYTES),
+ ctx->x, n);
/* restore the counter mode key for next time */
memcpy(ctx->x, x, sizeof(x));
}
@@ -282,7 +282,7 @@ int skein_512_init(struct skein_512_ctx *ctx, size_t hash_bit_len)
cfg.w[1] = skein_swap64(hash_bit_len);
cfg.w[2] = skein_swap64(SKEIN_CFG_TREE_INFO_SEQUENTIAL);
/* zero pad config block */
- memset(&cfg.w[3], 0, sizeof(cfg) - 3*sizeof(cfg.w[0]));
+ memset(&cfg.w[3], 0, sizeof(cfg) - 3 * sizeof(cfg.w[0]));
/* compute the initial chaining values from config block */
/* zero the chaining variables */
@@ -326,7 +326,7 @@ int skein_512_init_ext(struct skein_512_ctx *ctx, size_t hash_bit_len,
skein_assert(sizeof(cfg.b) >= sizeof(ctx->x));
/* do a mini-Init right here */
/* set output hash bit count = state size */
- ctx->h.hash_bit_len = 8*sizeof(ctx->x);
+ ctx->h.hash_bit_len = 8 * sizeof(ctx->x);
/* set tweaks: T0 = 0; T1 = KEY type */
skein_start_new_type(ctx, KEY);
/* zero the initial chaining variables */
@@ -398,7 +398,7 @@ int skein_512_update(struct skein_512_ctx *ctx, const u8 *msg,
*/
if (msg_byte_cnt > SKEIN_512_BLOCK_BYTES) {
/* number of full blocks to process */
- n = (msg_byte_cnt-1) / SKEIN_512_BLOCK_BYTES;
+ n = (msg_byte_cnt - 1) / SKEIN_512_BLOCK_BYTES;
skein_512_process_block(ctx, msg, n,
SKEIN_512_BLOCK_BYTES);
msg_byte_cnt -= n * SKEIN_512_BLOCK_BYTES;
@@ -432,7 +432,7 @@ int skein_512_final(struct skein_512_ctx *ctx, u8 *hash_val)
/* zero pad b[] if necessary */
if (ctx->h.b_cnt < SKEIN_512_BLOCK_BYTES)
memset(&ctx->b[ctx->h.b_cnt], 0,
- SKEIN_512_BLOCK_BYTES - ctx->h.b_cnt);
+ SKEIN_512_BLOCK_BYTES - ctx->h.b_cnt);
/* process the final block */
skein_512_process_block(ctx, ctx->b, 1, ctx->h.b_cnt);
@@ -446,19 +446,19 @@ int skein_512_final(struct skein_512_ctx *ctx, u8 *hash_val)
memset(ctx->b, 0, sizeof(ctx->b));
/* keep a local copy of counter mode "key" */
memcpy(x, ctx->x, sizeof(x));
- for (i = 0; i*SKEIN_512_BLOCK_BYTES < byte_cnt; i++) {
+ for (i = 0; i * SKEIN_512_BLOCK_BYTES < byte_cnt; i++) {
/* build the counter block */
- ((u64 *)ctx->b)[0] = skein_swap64((u64) i);
+ ((u64 *)ctx->b)[0] = skein_swap64((u64)i);
skein_start_new_type(ctx, OUT_FINAL);
/* run "counter mode" */
skein_512_process_block(ctx, ctx->b, 1, sizeof(u64));
/* number of output bytes left to go */
- n = byte_cnt - i*SKEIN_512_BLOCK_BYTES;
+ n = byte_cnt - i * SKEIN_512_BLOCK_BYTES;
if (n >= SKEIN_512_BLOCK_BYTES)
n = SKEIN_512_BLOCK_BYTES;
/* "output" the ctr mode bytes */
- skein_put64_lsb_first(hash_val+i*SKEIN_512_BLOCK_BYTES, ctx->x,
- n);
+ skein_put64_lsb_first(hash_val + (i * SKEIN_512_BLOCK_BYTES),
+ ctx->x, n);
/* restore the counter mode key for next time */
memcpy(ctx->x, x, sizeof(x));
}
@@ -506,7 +506,7 @@ int skein_1024_init(struct skein_1024_ctx *ctx, size_t hash_bit_len)
cfg.w[1] = skein_swap64(hash_bit_len);
cfg.w[2] = skein_swap64(SKEIN_CFG_TREE_INFO_SEQUENTIAL);
/* zero pad config block */
- memset(&cfg.w[3], 0, sizeof(cfg) - 3*sizeof(cfg.w[0]));
+ memset(&cfg.w[3], 0, sizeof(cfg) - 3 * sizeof(cfg.w[0]));
/* compute the initial chaining values from config block */
/* zero the chaining variables */
@@ -547,7 +547,7 @@ int skein_1024_init_ext(struct skein_1024_ctx *ctx, size_t hash_bit_len,
skein_assert(sizeof(cfg.b) >= sizeof(ctx->x));
/* do a mini-Init right here */
/* set output hash bit count = state size */
- ctx->h.hash_bit_len = 8*sizeof(ctx->x);
+ ctx->h.hash_bit_len = 8 * sizeof(ctx->x);
/* set tweaks: T0 = 0; T1 = KEY type */
skein_start_new_type(ctx, KEY);
/* zero the initial chaining variables */
@@ -620,7 +620,7 @@ int skein_1024_update(struct skein_1024_ctx *ctx, const u8 *msg,
*/
if (msg_byte_cnt > SKEIN_1024_BLOCK_BYTES) {
/* number of full blocks to process */
- n = (msg_byte_cnt-1) / SKEIN_1024_BLOCK_BYTES;
+ n = (msg_byte_cnt - 1) / SKEIN_1024_BLOCK_BYTES;
skein_1024_process_block(ctx, msg, n,
SKEIN_1024_BLOCK_BYTES);
msg_byte_cnt -= n * SKEIN_1024_BLOCK_BYTES;
@@ -654,7 +654,7 @@ int skein_1024_final(struct skein_1024_ctx *ctx, u8 *hash_val)
/* zero pad b[] if necessary */
if (ctx->h.b_cnt < SKEIN_1024_BLOCK_BYTES)
memset(&ctx->b[ctx->h.b_cnt], 0,
- SKEIN_1024_BLOCK_BYTES - ctx->h.b_cnt);
+ SKEIN_1024_BLOCK_BYTES - ctx->h.b_cnt);
/* process the final block */
skein_1024_process_block(ctx, ctx->b, 1, ctx->h.b_cnt);
@@ -668,19 +668,19 @@ int skein_1024_final(struct skein_1024_ctx *ctx, u8 *hash_val)
memset(ctx->b, 0, sizeof(ctx->b));
/* keep a local copy of counter mode "key" */
memcpy(x, ctx->x, sizeof(x));
- for (i = 0; i*SKEIN_1024_BLOCK_BYTES < byte_cnt; i++) {
+ for (i = 0; i * SKEIN_1024_BLOCK_BYTES < byte_cnt; i++) {
/* build the counter block */
- ((u64 *)ctx->b)[0] = skein_swap64((u64) i);
+ ((u64 *)ctx->b)[0] = skein_swap64((u64)i);
skein_start_new_type(ctx, OUT_FINAL);
/* run "counter mode" */
skein_1024_process_block(ctx, ctx->b, 1, sizeof(u64));
/* number of output bytes left to go */
- n = byte_cnt - i*SKEIN_1024_BLOCK_BYTES;
+ n = byte_cnt - i * SKEIN_1024_BLOCK_BYTES;
if (n >= SKEIN_1024_BLOCK_BYTES)
n = SKEIN_1024_BLOCK_BYTES;
/* "output" the ctr mode bytes */
- skein_put64_lsb_first(hash_val+i*SKEIN_1024_BLOCK_BYTES, ctx->x,
- n);
+ skein_put64_lsb_first(hash_val + (i * SKEIN_1024_BLOCK_BYTES),
+ ctx->x, n);
/* restore the counter mode key for next time */
memcpy(ctx->x, x, sizeof(x));
}
@@ -702,7 +702,7 @@ int skein_256_final_pad(struct skein_256_ctx *ctx, u8 *hash_val)
/* zero pad b[] if necessary */
if (ctx->h.b_cnt < SKEIN_256_BLOCK_BYTES)
memset(&ctx->b[ctx->h.b_cnt], 0,
- SKEIN_256_BLOCK_BYTES - ctx->h.b_cnt);
+ SKEIN_256_BLOCK_BYTES - ctx->h.b_cnt);
/* process the final block */
skein_256_process_block(ctx, ctx->b, 1, ctx->h.b_cnt);
@@ -724,7 +724,7 @@ int skein_512_final_pad(struct skein_512_ctx *ctx, u8 *hash_val)
/* zero pad b[] if necessary */
if (ctx->h.b_cnt < SKEIN_512_BLOCK_BYTES)
memset(&ctx->b[ctx->h.b_cnt], 0,
- SKEIN_512_BLOCK_BYTES - ctx->h.b_cnt);
+ SKEIN_512_BLOCK_BYTES - ctx->h.b_cnt);
/* process the final block */
skein_512_process_block(ctx, ctx->b, 1, ctx->h.b_cnt);
@@ -746,7 +746,7 @@ int skein_1024_final_pad(struct skein_1024_ctx *ctx, u8 *hash_val)
/* zero pad b[] if necessary */
if (ctx->h.b_cnt < SKEIN_1024_BLOCK_BYTES)
memset(&ctx->b[ctx->h.b_cnt], 0,
- SKEIN_1024_BLOCK_BYTES - ctx->h.b_cnt);
+ SKEIN_1024_BLOCK_BYTES - ctx->h.b_cnt);
/* process the final block */
skein_1024_process_block(ctx, ctx->b, 1, ctx->h.b_cnt);
@@ -775,19 +775,19 @@ int skein_256_output(struct skein_256_ctx *ctx, u8 *hash_val)
memset(ctx->b, 0, sizeof(ctx->b));
/* keep a local copy of counter mode "key" */
memcpy(x, ctx->x, sizeof(x));
- for (i = 0; i*SKEIN_256_BLOCK_BYTES < byte_cnt; i++) {
+ for (i = 0; i * SKEIN_256_BLOCK_BYTES < byte_cnt; i++) {
/* build the counter block */
- ((u64 *)ctx->b)[0] = skein_swap64((u64) i);
+ ((u64 *)ctx->b)[0] = skein_swap64((u64)i);
skein_start_new_type(ctx, OUT_FINAL);
/* run "counter mode" */
skein_256_process_block(ctx, ctx->b, 1, sizeof(u64));
/* number of output bytes left to go */
- n = byte_cnt - i*SKEIN_256_BLOCK_BYTES;
+ n = byte_cnt - i * SKEIN_256_BLOCK_BYTES;
if (n >= SKEIN_256_BLOCK_BYTES)
n = SKEIN_256_BLOCK_BYTES;
/* "output" the ctr mode bytes */
- skein_put64_lsb_first(hash_val+i*SKEIN_256_BLOCK_BYTES, ctx->x,
- n);
+ skein_put64_lsb_first(hash_val + (i * SKEIN_256_BLOCK_BYTES),
+ ctx->x, n);
/* restore the counter mode key for next time */
memcpy(ctx->x, x, sizeof(x));
}
@@ -812,19 +812,19 @@ int skein_512_output(struct skein_512_ctx *ctx, u8 *hash_val)
memset(ctx->b, 0, sizeof(ctx->b));
/* keep a local copy of counter mode "key" */
memcpy(x, ctx->x, sizeof(x));
- for (i = 0; i*SKEIN_512_BLOCK_BYTES < byte_cnt; i++) {
+ for (i = 0; i * SKEIN_512_BLOCK_BYTES < byte_cnt; i++) {
/* build the counter block */
- ((u64 *)ctx->b)[0] = skein_swap64((u64) i);
+ ((u64 *)ctx->b)[0] = skein_swap64((u64)i);
skein_start_new_type(ctx, OUT_FINAL);
/* run "counter mode" */
skein_512_process_block(ctx, ctx->b, 1, sizeof(u64));
/* number of output bytes left to go */
- n = byte_cnt - i*SKEIN_512_BLOCK_BYTES;
+ n = byte_cnt - i * SKEIN_512_BLOCK_BYTES;
if (n >= SKEIN_512_BLOCK_BYTES)
n = SKEIN_512_BLOCK_BYTES;
/* "output" the ctr mode bytes */
- skein_put64_lsb_first(hash_val+i*SKEIN_512_BLOCK_BYTES, ctx->x,
- n);
+ skein_put64_lsb_first(hash_val + (i * SKEIN_512_BLOCK_BYTES),
+ ctx->x, n);
/* restore the counter mode key for next time */
memcpy(ctx->x, x, sizeof(x));
}
@@ -849,19 +849,19 @@ int skein_1024_output(struct skein_1024_ctx *ctx, u8 *hash_val)
memset(ctx->b, 0, sizeof(ctx->b));
/* keep a local copy of counter mode "key" */
memcpy(x, ctx->x, sizeof(x));
- for (i = 0; i*SKEIN_1024_BLOCK_BYTES < byte_cnt; i++) {
+ for (i = 0; i * SKEIN_1024_BLOCK_BYTES < byte_cnt; i++) {
/* build the counter block */
- ((u64 *)ctx->b)[0] = skein_swap64((u64) i);
+ ((u64 *)ctx->b)[0] = skein_swap64((u64)i);
skein_start_new_type(ctx, OUT_FINAL);
/* run "counter mode" */
skein_1024_process_block(ctx, ctx->b, 1, sizeof(u64));
/* number of output bytes left to go */
- n = byte_cnt - i*SKEIN_1024_BLOCK_BYTES;
+ n = byte_cnt - i * SKEIN_1024_BLOCK_BYTES;
if (n >= SKEIN_1024_BLOCK_BYTES)
n = SKEIN_1024_BLOCK_BYTES;
/* "output" the ctr mode bytes */
- skein_put64_lsb_first(hash_val+i*SKEIN_1024_BLOCK_BYTES, ctx->x,
- n);
+ skein_put64_lsb_first(hash_val + (i * SKEIN_1024_BLOCK_BYTES),
+ ctx->x, n);
/* restore the counter mode key for next time */
memcpy(ctx->x, x, sizeof(x));
}
diff --git a/drivers/staging/skein/skein_base.h b/drivers/staging/skein/skein_base.h
index 3c7f8ad36..dc464f334 100644
--- a/drivers/staging/skein/skein_base.h
+++ b/drivers/staging/skein/skein_base.h
@@ -32,7 +32,7 @@
/* below two prototype assume we are handed aligned data */
#define skein_put64_lsb_first(dst08, src64, b_cnt) memcpy(dst08, src64, b_cnt)
#define skein_get64_lsb_first(dst64, src08, w_cnt) \
- memcpy(dst64, src08, 8*(w_cnt))
+ memcpy(dst64, src08, 8 * (w_cnt))
#define skein_swap64(w64) (w64)
enum {
@@ -48,17 +48,17 @@ enum {
#define SKEIN_1024_STATE_WORDS 16
#define SKEIN_MAX_STATE_WORDS 16
-#define SKEIN_256_STATE_BYTES (8*SKEIN_256_STATE_WORDS)
-#define SKEIN_512_STATE_BYTES (8*SKEIN_512_STATE_WORDS)
-#define SKEIN_1024_STATE_BYTES (8*SKEIN_1024_STATE_WORDS)
+#define SKEIN_256_STATE_BYTES (8 * SKEIN_256_STATE_WORDS)
+#define SKEIN_512_STATE_BYTES (8 * SKEIN_512_STATE_WORDS)
+#define SKEIN_1024_STATE_BYTES (8 * SKEIN_1024_STATE_WORDS)
-#define SKEIN_256_STATE_BITS (64*SKEIN_256_STATE_WORDS)
-#define SKEIN_512_STATE_BITS (64*SKEIN_512_STATE_WORDS)
-#define SKEIN_1024_STATE_BITS (64*SKEIN_1024_STATE_WORDS)
+#define SKEIN_256_STATE_BITS (64 * SKEIN_256_STATE_WORDS)
+#define SKEIN_512_STATE_BITS (64 * SKEIN_512_STATE_WORDS)
+#define SKEIN_1024_STATE_BITS (64 * SKEIN_1024_STATE_WORDS)
-#define SKEIN_256_BLOCK_BYTES (8*SKEIN_256_STATE_WORDS)
-#define SKEIN_512_BLOCK_BYTES (8*SKEIN_512_STATE_WORDS)
-#define SKEIN_1024_BLOCK_BYTES (8*SKEIN_1024_STATE_WORDS)
+#define SKEIN_256_BLOCK_BYTES (8 * SKEIN_256_STATE_WORDS)
+#define SKEIN_512_BLOCK_BYTES (8 * SKEIN_512_STATE_WORDS)
+#define SKEIN_1024_BLOCK_BYTES (8 * SKEIN_1024_STATE_WORDS)
struct skein_ctx_hdr {
size_t hash_bit_len; /* size of hash result, in bits */
@@ -84,11 +84,6 @@ struct skein_1024_ctx { /* 1024-bit Skein hash context structure */
u8 b[SKEIN_1024_BLOCK_BYTES]; /* partial block buf (8-byte aligned) */
};
-static inline u64 rotl_64(u64 x, u8 N)
-{
- return (x << N) | (x >> (64 - N));
-}
-
/* Skein APIs for (incremental) "straight hashing" */
int skein_256_init(struct skein_256_ctx *ctx, size_t hash_bit_len);
int skein_512_init(struct skein_512_ctx *ctx, size_t hash_bit_len);
@@ -162,13 +157,13 @@ int skein_1024_output(struct skein_1024_ctx *ctx, u8 *hash_val);
#define SKEIN_T1_POS_FINAL SKEIN_T1_BIT(127) /* 127 final blk flag */
/* tweak word tweak[1]: flag bit definition(s) */
-#define SKEIN_T1_FLAG_FIRST (((u64) 1) << SKEIN_T1_POS_FIRST)
-#define SKEIN_T1_FLAG_FINAL (((u64) 1) << SKEIN_T1_POS_FINAL)
-#define SKEIN_T1_FLAG_BIT_PAD (((u64) 1) << SKEIN_T1_POS_BIT_PAD)
+#define SKEIN_T1_FLAG_FIRST (((u64)1) << SKEIN_T1_POS_FIRST)
+#define SKEIN_T1_FLAG_FINAL (((u64)1) << SKEIN_T1_POS_FINAL)
+#define SKEIN_T1_FLAG_BIT_PAD (((u64)1) << SKEIN_T1_POS_BIT_PAD)
/* tweak word tweak[1]: tree level bit field mask */
#define SKEIN_T1_TREE_LVL_MASK (((u64)0x7F) << SKEIN_T1_POS_TREE_LVL)
-#define SKEIN_T1_TREE_LEVEL(n) (((u64) (n)) << SKEIN_T1_POS_TREE_LVL)
+#define SKEIN_T1_TREE_LEVEL(n) (((u64)(n)) << SKEIN_T1_POS_TREE_LVL)
/* tweak word tweak[1]: block type field */
#define SKEIN_BLK_TYPE_KEY (0) /* key, for MAC and KDF */
@@ -181,7 +176,7 @@ int skein_1024_output(struct skein_1024_ctx *ctx, u8 *hash_val);
#define SKEIN_BLK_TYPE_OUT (63) /* output stage */
#define SKEIN_BLK_TYPE_MASK (63) /* bit field mask */
-#define SKEIN_T1_BLK_TYPE(T) (((u64) (SKEIN_BLK_TYPE_##T)) << \
+#define SKEIN_T1_BLK_TYPE(T) (((u64)(SKEIN_BLK_TYPE_##T)) << \
SKEIN_T1_POS_BLK_TYPE)
#define SKEIN_T1_BLK_TYPE_KEY SKEIN_T1_BLK_TYPE(KEY) /* for MAC and KDF */
#define SKEIN_T1_BLK_TYPE_CFG SKEIN_T1_BLK_TYPE(CFG) /* config block */
@@ -204,11 +199,11 @@ int skein_1024_output(struct skein_1024_ctx *ctx, u8 *hash_val);
#define SKEIN_ID_STRING_LE (0x33414853) /* "SHA3" (little-endian)*/
#endif
-#define SKEIN_MK_64(hi32, lo32) ((lo32) + (((u64) (hi32)) << 32))
+#define SKEIN_MK_64(hi32, lo32) ((lo32) + (((u64)(hi32)) << 32))
#define SKEIN_SCHEMA_VER SKEIN_MK_64(SKEIN_VERSION, SKEIN_ID_STRING_LE)
#define SKEIN_KS_PARITY SKEIN_MK_64(0x1BD11BDA, 0xA9FC1A22)
-#define SKEIN_CFG_STR_LEN (4*8)
+#define SKEIN_CFG_STR_LEN (4 * 8)
/* bit field definitions in config block tree_info word */
#define SKEIN_CFG_TREE_LEAF_SIZE_POS (0)
@@ -327,9 +322,9 @@ enum {
#define SKEIN_512_ROUNDS_TOTAL (72)
#define SKEIN_1024_ROUNDS_TOTAL (80)
#else /* allow command-line define in range 8*(5..14) */
-#define SKEIN_256_ROUNDS_TOTAL (8*((((SKEIN_ROUNDS/100) + 5) % 10) + 5))
-#define SKEIN_512_ROUNDS_TOTAL (8*((((SKEIN_ROUNDS/10) + 5) % 10) + 5))
-#define SKEIN_1024_ROUNDS_TOTAL (8*((((SKEIN_ROUNDS) + 5) % 10) + 5))
+#define SKEIN_256_ROUNDS_TOTAL (8 * ((((SKEIN_ROUNDS / 100) + 5) % 10) + 5))
+#define SKEIN_512_ROUNDS_TOTAL (8 * ((((SKEIN_ROUNDS / 10) + 5) % 10) + 5))
+#define SKEIN_1024_ROUNDS_TOTAL (8 * ((((SKEIN_ROUNDS) + 5) % 10) + 5))
#endif
#endif /* ifndef _SKEIN_H_ */
diff --git a/drivers/staging/skein/skein_block.c b/drivers/staging/skein/skein_block.c
index 45b47327e..59a0a8a82 100644
--- a/drivers/staging/skein/skein_block.c
+++ b/drivers/staging/skein/skein_block.c
@@ -15,6 +15,7 @@
************************************************************************/
#include <linux/string.h>
+#include <linux/bitops.h>
#include "skein_base.h"
#include "skein_block.h"
@@ -59,10 +60,10 @@
#define ROUND256(p0, p1, p2, p3, ROT, r_num) \
do { \
X##p0 += X##p1; \
- X##p1 = rotl_64(X##p1, ROT##_0); \
+ X##p1 = rol64(X##p1, ROT##_0); \
X##p1 ^= X##p0; \
X##p2 += X##p3; \
- X##p3 = rotl_64(X##p3, ROT##_1); \
+ X##p3 = rol64(X##p3, ROT##_1); \
X##p3 ^= X##p2; \
} while (0)
@@ -120,10 +121,10 @@
#if !(SKEIN_USE_ASM & 512)
#undef RCNT
-#define RCNT (SKEIN_512_ROUNDS_TOTAL/8)
+#define RCNT (SKEIN_512_ROUNDS_TOTAL / 8)
#ifdef SKEIN_LOOP /* configure how much to unroll the loop */
-#define SKEIN_UNROLL_512 (((SKEIN_LOOP)/10)%10)
+#define SKEIN_UNROLL_512 (((SKEIN_LOOP) / 10) % 10)
#else
#define SKEIN_UNROLL_512 (0)
#endif
@@ -136,15 +137,16 @@
#define ROUND512(p0, p1, p2, p3, p4, p5, p6, p7, ROT, r_num) \
do { \
X##p0 += X##p1; \
- X##p1 = rotl_64(X##p1, ROT##_0); \
+ X##p1 = rol64(X##p1, ROT##_0); \
X##p1 ^= X##p0; \
X##p2 += X##p3; \
- X##p3 = rotl_64(X##p3, ROT##_1); \
+ X##p3 = rol64(X##p3, ROT##_1); \
X##p3 ^= X##p2; \
X##p4 += X##p5; \
- X##p5 = rotl_64(X##p5, ROT##_2); \
+ X##p5 = rol64(X##p5, ROT##_2); \
X##p5 ^= X##p4; \
- X##p6 += X##p7; X##p7 = rotl_64(X##p7, ROT##_3);\
+ X##p6 += X##p7; \
+ X##p7 = rol64(X##p7, ROT##_3); \
X##p7 ^= X##p6; \
} while (0)
@@ -200,7 +202,7 @@
} while (0)
#define R512_UNROLL_R(NN) \
((SKEIN_UNROLL_512 == 0 && \
- SKEIN_512_ROUNDS_TOTAL/8 > (NN)) || \
+ SKEIN_512_ROUNDS_TOTAL / 8 > (NN)) || \
(SKEIN_UNROLL_512 > (NN)))
#if (SKEIN_UNROLL_512 > 14)
@@ -210,7 +212,7 @@
#if !(SKEIN_USE_ASM & 1024)
#undef RCNT
-#define RCNT (SKEIN_1024_ROUNDS_TOTAL/8)
+#define RCNT (SKEIN_1024_ROUNDS_TOTAL / 8)
#ifdef SKEIN_LOOP /* configure how much to unroll the loop */
#define SKEIN_UNROLL_1024 ((SKEIN_LOOP) % 10)
#else
@@ -226,28 +228,28 @@
pF, ROT, r_num) \
do { \
X##p0 += X##p1; \
- X##p1 = rotl_64(X##p1, ROT##_0); \
+ X##p1 = rol64(X##p1, ROT##_0); \
X##p1 ^= X##p0; \
X##p2 += X##p3; \
- X##p3 = rotl_64(X##p3, ROT##_1); \
+ X##p3 = rol64(X##p3, ROT##_1); \
X##p3 ^= X##p2; \
X##p4 += X##p5; \
- X##p5 = rotl_64(X##p5, ROT##_2); \
+ X##p5 = rol64(X##p5, ROT##_2); \
X##p5 ^= X##p4; \
X##p6 += X##p7; \
- X##p7 = rotl_64(X##p7, ROT##_3); \
+ X##p7 = rol64(X##p7, ROT##_3); \
X##p7 ^= X##p6; \
X##p8 += X##p9; \
- X##p9 = rotl_64(X##p9, ROT##_4); \
+ X##p9 = rol64(X##p9, ROT##_4); \
X##p9 ^= X##p8; \
X##pA += X##pB; \
- X##pB = rotl_64(X##pB, ROT##_5); \
+ X##pB = rol64(X##pB, ROT##_5); \
X##pB ^= X##pA; \
X##pC += X##pD; \
- X##pD = rotl_64(X##pD, ROT##_6); \
+ X##pD = rol64(X##pD, ROT##_6); \
X##pD ^= X##pC; \
X##pE += X##pF; \
- X##pF = rotl_64(X##pF, ROT##_7); \
+ X##pF = rol64(X##pF, ROT##_7); \
X##pF ^= X##pE; \
} while (0)
@@ -311,28 +313,28 @@
#define R1024_8_ROUNDS(R) \
do { \
R1024(00, 01, 02, 03, 04, 05, 06, 07, 08, 09, 10, 11, 12, \
- 13, 14, 15, R1024_0, 8*(R) + 1); \
+ 13, 14, 15, R1024_0, 8 * (R) + 1); \
R1024(00, 09, 02, 13, 06, 11, 04, 15, 10, 07, 12, 03, 14, \
- 05, 08, 01, R1024_1, 8*(R) + 2); \
+ 05, 08, 01, R1024_1, 8 * (R) + 2); \
R1024(00, 07, 02, 05, 04, 03, 06, 01, 12, 15, 14, 13, 08, \
- 11, 10, 09, R1024_2, 8*(R) + 3); \
+ 11, 10, 09, R1024_2, 8 * (R) + 3); \
R1024(00, 15, 02, 11, 06, 13, 04, 09, 14, 01, 08, 05, 10, \
- 03, 12, 07, R1024_3, 8*(R) + 4); \
- I1024(2*(R)); \
+ 03, 12, 07, R1024_3, 8 * (R) + 4); \
+ I1024(2 * (R)); \
R1024(00, 01, 02, 03, 04, 05, 06, 07, 08, 09, 10, 11, 12, \
- 13, 14, 15, R1024_4, 8*(R) + 5); \
+ 13, 14, 15, R1024_4, 8 * (R) + 5); \
R1024(00, 09, 02, 13, 06, 11, 04, 15, 10, 07, 12, 03, 14, \
- 05, 08, 01, R1024_5, 8*(R) + 6); \
+ 05, 08, 01, R1024_5, 8 * (R) + 6); \
R1024(00, 07, 02, 05, 04, 03, 06, 01, 12, 15, 14, 13, 08, \
- 11, 10, 09, R1024_6, 8*(R) + 7); \
+ 11, 10, 09, R1024_6, 8 * (R) + 7); \
R1024(00, 15, 02, 11, 06, 13, 04, 09, 14, 01, 08, 05, 10, \
- 03, 12, 07, R1024_7, 8*(R) + 8); \
- I1024(2*(R)+1); \
+ 03, 12, 07, R1024_7, 8 * (R) + 8); \
+ I1024(2 * (R) + 1); \
} while (0)
#define R1024_UNROLL_R(NN) \
((SKEIN_UNROLL_1024 == 0 && \
- SKEIN_1024_ROUNDS_TOTAL/8 > (NN)) || \
+ SKEIN_1024_ROUNDS_TOTAL / 8 > (NN)) || \
(SKEIN_UNROLL_1024 > (NN)))
#if (SKEIN_UNROLL_1024 > 14)
@@ -351,10 +353,10 @@ void skein_256_process_block(struct skein_256_ctx *ctx, const u8 *blk_ptr,
size_t r;
#if SKEIN_UNROLL_256
/* key schedule: chaining vars + tweak + "rot"*/
- u64 kw[WCNT+4+RCNT*2];
+ u64 kw[WCNT + 4 + (RCNT * 2)];
#else
/* key schedule words : chaining vars + tweak */
- u64 kw[WCNT+4];
+ u64 kw[WCNT + 4];
#endif
u64 X0, X1, X2, X3; /* local copy of context vars, for speed */
u64 w[WCNT]; /* local copy of input block */
@@ -460,9 +462,10 @@ void skein_256_process_block(struct skein_256_ctx *ctx, const u8 *blk_ptr,
#if defined(SKEIN_CODE_SIZE) || defined(SKEIN_PERF)
size_t skein_256_process_block_code_size(void)
{
- return ((u8 *) skein_256_process_block_code_size) -
- ((u8 *) skein_256_process_block);
+ return ((u8 *)skein_256_process_block_code_size) -
+ ((u8 *)skein_256_process_block);
}
+
unsigned int skein_256_unroll_cnt(void)
{
return SKEIN_UNROLL_256;
@@ -480,9 +483,11 @@ void skein_512_process_block(struct skein_512_ctx *ctx, const u8 *blk_ptr,
};
size_t r;
#if SKEIN_UNROLL_512
- u64 kw[WCNT+4+RCNT*2]; /* key sched: chaining vars + tweak + "rot"*/
+ /* key sched: chaining vars + tweak + "rot"*/
+ u64 kw[WCNT + 4 + RCNT * 2];
#else
- u64 kw[WCNT+4]; /* key schedule words : chaining vars + tweak */
+ /* key schedule words : chaining vars + tweak */
+ u64 kw[WCNT + 4];
#endif
u64 X0, X1, X2, X3, X4, X5, X6, X7; /* local copies, for speed */
u64 w[WCNT]; /* local copy of input block */
@@ -543,7 +548,6 @@ void skein_512_process_block(struct skein_512_ctx *ctx, const u8 *blk_ptr,
for (r = 1;
r < (SKEIN_UNROLL_512 ? 2 * RCNT : 2);
r += (SKEIN_UNROLL_512 ? 2 * SKEIN_UNROLL_512 : 1)) {
-
R512_8_ROUNDS(0);
#if R512_UNROLL_R(1)
@@ -609,9 +613,10 @@ void skein_512_process_block(struct skein_512_ctx *ctx, const u8 *blk_ptr,
#if defined(SKEIN_CODE_SIZE) || defined(SKEIN_PERF)
size_t skein_512_process_block_code_size(void)
{
- return ((u8 *) skein_512_process_block_code_size) -
- ((u8 *) skein_512_process_block);
+ return ((u8 *)skein_512_process_block_code_size) -
+ ((u8 *)skein_512_process_block);
}
+
unsigned int skein_512_unroll_cnt(void)
{
return SKEIN_UNROLL_512;
@@ -629,9 +634,11 @@ void skein_1024_process_block(struct skein_1024_ctx *ctx, const u8 *blk_ptr,
};
size_t r;
#if (SKEIN_UNROLL_1024 != 0)
- u64 kw[WCNT+4+RCNT*2]; /* key sched: chaining vars + tweak + "rot" */
+ /* key sched: chaining vars + tweak + "rot" */
+ u64 kw[WCNT + 4 + (RCNT * 2)];
#else
- u64 kw[WCNT+4]; /* key schedule words : chaining vars + tweak */
+ /* key schedule words : chaining vars + tweak */
+ u64 kw[WCNT + 4];
#endif
/* local copy of vars, for speed */
@@ -771,9 +778,10 @@ void skein_1024_process_block(struct skein_1024_ctx *ctx, const u8 *blk_ptr,
#if defined(SKEIN_CODE_SIZE) || defined(SKEIN_PERF)
size_t skein_1024_process_block_code_size(void)
{
- return ((u8 *) skein_1024_process_block_code_size) -
- ((u8 *) skein_1024_process_block);
+ return ((u8 *)skein_1024_process_block_code_size) -
+ ((u8 *)skein_1024_process_block);
}
+
unsigned int skein_1024_unroll_cnt(void)
{
return SKEIN_UNROLL_1024;
diff --git a/drivers/staging/skein/skein_generic.c b/drivers/staging/skein/skein_generic.c
index e29b9abaa..11f5e530a 100644
--- a/drivers/staging/skein/skein_generic.c
+++ b/drivers/staging/skein/skein_generic.c
@@ -27,7 +27,7 @@ static int skein256_init(struct shash_desc *desc)
}
static int skein256_update(struct shash_desc *desc, const u8 *data,
- unsigned int len)
+ unsigned int len)
{
return skein_256_update((struct skein_256_ctx *)shash_desc_ctx(desc),
data, len);
@@ -62,7 +62,7 @@ static int skein512_init(struct shash_desc *desc)
}
static int skein512_update(struct shash_desc *desc, const u8 *data,
- unsigned int len)
+ unsigned int len)
{
return skein_512_update((struct skein_512_ctx *)shash_desc_ctx(desc),
data, len);
@@ -97,7 +97,7 @@ static int skein1024_init(struct shash_desc *desc)
}
static int skein1024_update(struct shash_desc *desc, const u8 *data,
- unsigned int len)
+ unsigned int len)
{
return skein_1024_update((struct skein_1024_ctx *)shash_desc_ctx(desc),
data, len);
diff --git a/drivers/staging/skein/threefish_api.h b/drivers/staging/skein/threefish_api.h
index 8e0a0b77e..615e46757 100644
--- a/drivers/staging/skein/threefish_api.h
+++ b/drivers/staging/skein/threefish_api.h
@@ -52,7 +52,7 @@ enum threefish_size {
*/
struct threefish_key {
u64 state_size;
- u64 key[SKEIN_MAX_STATE_WORDS+1]; /* max number of key words*/
+ u64 key[SKEIN_MAX_STATE_WORDS + 1]; /* max number of key words*/
u64 tweak[3];
};
diff --git a/drivers/staging/skein/threefish_block.c b/drivers/staging/skein/threefish_block.c
index e19ac4368..a95563fad 100644
--- a/drivers/staging/skein/threefish_block.c
+++ b/drivers/staging/skein/threefish_block.c
@@ -512,622 +512,622 @@ void threefish_decrypt_256(struct threefish_key *key_ctx, u64 *input,
b2 -= k0 + t1;
b3 -= k1 + 18;
tmp = b3 ^ b0;
- b3 = (tmp >> 32) | (tmp << (64 - 32));
+ b3 = ror64(tmp, 32);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 32) | (tmp << (64 - 32));
+ b1 = ror64(tmp, 32);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 58) | (tmp << (64 - 58));
+ b1 = ror64(tmp, 58);
b0 -= b1;
tmp = b3 ^ b2;
- b3 = (tmp >> 22) | (tmp << (64 - 22));
+ b3 = ror64(tmp, 22);
b2 -= b3;
tmp = b3 ^ b0;
- b3 = (tmp >> 46) | (tmp << (64 - 46));
+ b3 = ror64(tmp, 46);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 12) | (tmp << (64 - 12));
+ b1 = ror64(tmp, 12);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 25) | (tmp << (64 - 25));
+ b1 = ror64(tmp, 25);
b0 -= b1 + k2;
b1 -= k3 + t2;
tmp = b3 ^ b2;
- b3 = (tmp >> 33) | (tmp << (64 - 33));
+ b3 = ror64(tmp, 33);
b2 -= b3 + k4 + t0;
b3 -= k0 + 17;
tmp = b3 ^ b0;
- b3 = (tmp >> 5) | (tmp << (64 - 5));
+ b3 = ror64(tmp, 5);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 37) | (tmp << (64 - 37));
+ b1 = ror64(tmp, 37);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 23) | (tmp << (64 - 23));
+ b1 = ror64(tmp, 23);
b0 -= b1;
tmp = b3 ^ b2;
- b3 = (tmp >> 40) | (tmp << (64 - 40));
+ b3 = ror64(tmp, 40);
b2 -= b3;
tmp = b3 ^ b0;
- b3 = (tmp >> 52) | (tmp << (64 - 52));
+ b3 = ror64(tmp, 52);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 57) | (tmp << (64 - 57));
+ b1 = ror64(tmp, 57);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 14) | (tmp << (64 - 14));
+ b1 = ror64(tmp, 14);
b0 -= b1 + k1;
b1 -= k2 + t1;
tmp = b3 ^ b2;
- b3 = (tmp >> 16) | (tmp << (64 - 16));
+ b3 = ror64(tmp, 16);
b2 -= b3 + k3 + t2;
b3 -= k4 + 16;
tmp = b3 ^ b0;
- b3 = (tmp >> 32) | (tmp << (64 - 32));
+ b3 = ror64(tmp, 32);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 32) | (tmp << (64 - 32));
+ b1 = ror64(tmp, 32);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 58) | (tmp << (64 - 58));
+ b1 = ror64(tmp, 58);
b0 -= b1;
tmp = b3 ^ b2;
- b3 = (tmp >> 22) | (tmp << (64 - 22));
+ b3 = ror64(tmp, 22);
b2 -= b3;
tmp = b3 ^ b0;
- b3 = (tmp >> 46) | (tmp << (64 - 46));
+ b3 = ror64(tmp, 46);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 12) | (tmp << (64 - 12));
+ b1 = ror64(tmp, 12);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 25) | (tmp << (64 - 25));
+ b1 = ror64(tmp, 25);
b0 -= b1 + k0;
b1 -= k1 + t0;
tmp = b3 ^ b2;
- b3 = (tmp >> 33) | (tmp << (64 - 33));
+ b3 = ror64(tmp, 33);
b2 -= b3 + k2 + t1;
b3 -= k3 + 15;
tmp = b3 ^ b0;
- b3 = (tmp >> 5) | (tmp << (64 - 5));
+ b3 = ror64(tmp, 5);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 37) | (tmp << (64 - 37));
+ b1 = ror64(tmp, 37);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 23) | (tmp << (64 - 23));
+ b1 = ror64(tmp, 23);
b0 -= b1;
tmp = b3 ^ b2;
- b3 = (tmp >> 40) | (tmp << (64 - 40));
+ b3 = ror64(tmp, 40);
b2 -= b3;
tmp = b3 ^ b0;
- b3 = (tmp >> 52) | (tmp << (64 - 52));
+ b3 = ror64(tmp, 52);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 57) | (tmp << (64 - 57));
+ b1 = ror64(tmp, 57);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 14) | (tmp << (64 - 14));
+ b1 = ror64(tmp, 14);
b0 -= b1 + k4;
b1 -= k0 + t2;
tmp = b3 ^ b2;
- b3 = (tmp >> 16) | (tmp << (64 - 16));
+ b3 = ror64(tmp, 16);
b2 -= b3 + k1 + t0;
b3 -= k2 + 14;
tmp = b3 ^ b0;
- b3 = (tmp >> 32) | (tmp << (64 - 32));
+ b3 = ror64(tmp, 32);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 32) | (tmp << (64 - 32));
+ b1 = ror64(tmp, 32);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 58) | (tmp << (64 - 58));
+ b1 = ror64(tmp, 58);
b0 -= b1;
tmp = b3 ^ b2;
- b3 = (tmp >> 22) | (tmp << (64 - 22));
+ b3 = ror64(tmp, 22);
b2 -= b3;
tmp = b3 ^ b0;
- b3 = (tmp >> 46) | (tmp << (64 - 46));
+ b3 = ror64(tmp, 46);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 12) | (tmp << (64 - 12));
+ b1 = ror64(tmp, 12);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 25) | (tmp << (64 - 25));
+ b1 = ror64(tmp, 25);
b0 -= b1 + k3;
b1 -= k4 + t1;
tmp = b3 ^ b2;
- b3 = (tmp >> 33) | (tmp << (64 - 33));
+ b3 = ror64(tmp, 33);
b2 -= b3 + k0 + t2;
b3 -= k1 + 13;
tmp = b3 ^ b0;
- b3 = (tmp >> 5) | (tmp << (64 - 5));
+ b3 = ror64(tmp, 5);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 37) | (tmp << (64 - 37));
+ b1 = ror64(tmp, 37);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 23) | (tmp << (64 - 23));
+ b1 = ror64(tmp, 23);
b0 -= b1;
tmp = b3 ^ b2;
- b3 = (tmp >> 40) | (tmp << (64 - 40));
+ b3 = ror64(tmp, 40);
b2 -= b3;
tmp = b3 ^ b0;
- b3 = (tmp >> 52) | (tmp << (64 - 52));
+ b3 = ror64(tmp, 52);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 57) | (tmp << (64 - 57));
+ b1 = ror64(tmp, 57);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 14) | (tmp << (64 - 14));
+ b1 = ror64(tmp, 14);
b0 -= b1 + k2;
b1 -= k3 + t0;
tmp = b3 ^ b2;
- b3 = (tmp >> 16) | (tmp << (64 - 16));
+ b3 = ror64(tmp, 16);
b2 -= b3 + k4 + t1;
b3 -= k0 + 12;
tmp = b3 ^ b0;
- b3 = (tmp >> 32) | (tmp << (64 - 32));
+ b3 = ror64(tmp, 32);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 32) | (tmp << (64 - 32));
+ b1 = ror64(tmp, 32);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 58) | (tmp << (64 - 58));
+ b1 = ror64(tmp, 58);
b0 -= b1;
tmp = b3 ^ b2;
- b3 = (tmp >> 22) | (tmp << (64 - 22));
+ b3 = ror64(tmp, 22);
b2 -= b3;
tmp = b3 ^ b0;
- b3 = (tmp >> 46) | (tmp << (64 - 46));
+ b3 = ror64(tmp, 46);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 12) | (tmp << (64 - 12));
+ b1 = ror64(tmp, 12);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 25) | (tmp << (64 - 25));
+ b1 = ror64(tmp, 25);
b0 -= b1 + k1;
b1 -= k2 + t2;
tmp = b3 ^ b2;
- b3 = (tmp >> 33) | (tmp << (64 - 33));
+ b3 = ror64(tmp, 33);
b2 -= b3 + k3 + t0;
b3 -= k4 + 11;
tmp = b3 ^ b0;
- b3 = (tmp >> 5) | (tmp << (64 - 5));
+ b3 = ror64(tmp, 5);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 37) | (tmp << (64 - 37));
+ b1 = ror64(tmp, 37);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 23) | (tmp << (64 - 23));
+ b1 = ror64(tmp, 23);
b0 -= b1;
tmp = b3 ^ b2;
- b3 = (tmp >> 40) | (tmp << (64 - 40));
+ b3 = ror64(tmp, 40);
b2 -= b3;
tmp = b3 ^ b0;
- b3 = (tmp >> 52) | (tmp << (64 - 52));
+ b3 = ror64(tmp, 52);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 57) | (tmp << (64 - 57));
+ b1 = ror64(tmp, 57);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 14) | (tmp << (64 - 14));
+ b1 = ror64(tmp, 14);
b0 -= b1 + k0;
b1 -= k1 + t1;
tmp = b3 ^ b2;
- b3 = (tmp >> 16) | (tmp << (64 - 16));
+ b3 = ror64(tmp, 16);
b2 -= b3 + k2 + t2;
b3 -= k3 + 10;
tmp = b3 ^ b0;
- b3 = (tmp >> 32) | (tmp << (64 - 32));
+ b3 = ror64(tmp, 32);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 32) | (tmp << (64 - 32));
+ b1 = ror64(tmp, 32);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 58) | (tmp << (64 - 58));
+ b1 = ror64(tmp, 58);
b0 -= b1;
tmp = b3 ^ b2;
- b3 = (tmp >> 22) | (tmp << (64 - 22));
+ b3 = ror64(tmp, 22);
b2 -= b3;
tmp = b3 ^ b0;
- b3 = (tmp >> 46) | (tmp << (64 - 46));
+ b3 = ror64(tmp, 46);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 12) | (tmp << (64 - 12));
+ b1 = ror64(tmp, 12);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 25) | (tmp << (64 - 25));
+ b1 = ror64(tmp, 25);
b0 -= b1 + k4;
b1 -= k0 + t0;
tmp = b3 ^ b2;
- b3 = (tmp >> 33) | (tmp << (64 - 33));
+ b3 = ror64(tmp, 33);
b2 -= b3 + k1 + t1;
b3 -= k2 + 9;
tmp = b3 ^ b0;
- b3 = (tmp >> 5) | (tmp << (64 - 5));
+ b3 = ror64(tmp, 5);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 37) | (tmp << (64 - 37));
+ b1 = ror64(tmp, 37);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 23) | (tmp << (64 - 23));
+ b1 = ror64(tmp, 23);
b0 -= b1;
tmp = b3 ^ b2;
- b3 = (tmp >> 40) | (tmp << (64 - 40));
+ b3 = ror64(tmp, 40);
b2 -= b3;
tmp = b3 ^ b0;
- b3 = (tmp >> 52) | (tmp << (64 - 52));
+ b3 = ror64(tmp, 52);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 57) | (tmp << (64 - 57));
+ b1 = ror64(tmp, 57);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 14) | (tmp << (64 - 14));
+ b1 = ror64(tmp, 14);
b0 -= b1 + k3;
b1 -= k4 + t2;
tmp = b3 ^ b2;
- b3 = (tmp >> 16) | (tmp << (64 - 16));
+ b3 = ror64(tmp, 16);
b2 -= b3 + k0 + t0;
b3 -= k1 + 8;
tmp = b3 ^ b0;
- b3 = (tmp >> 32) | (tmp << (64 - 32));
+ b3 = ror64(tmp, 32);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 32) | (tmp << (64 - 32));
+ b1 = ror64(tmp, 32);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 58) | (tmp << (64 - 58));
+ b1 = ror64(tmp, 58);
b0 -= b1;
tmp = b3 ^ b2;
- b3 = (tmp >> 22) | (tmp << (64 - 22));
+ b3 = ror64(tmp, 22);
b2 -= b3;
tmp = b3 ^ b0;
- b3 = (tmp >> 46) | (tmp << (64 - 46));
+ b3 = ror64(tmp, 46);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 12) | (tmp << (64 - 12));
+ b1 = ror64(tmp, 12);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 25) | (tmp << (64 - 25));
+ b1 = ror64(tmp, 25);
b0 -= b1 + k2;
b1 -= k3 + t1;
tmp = b3 ^ b2;
- b3 = (tmp >> 33) | (tmp << (64 - 33));
+ b3 = ror64(tmp, 33);
b2 -= b3 + k4 + t2;
b3 -= k0 + 7;
tmp = b3 ^ b0;
- b3 = (tmp >> 5) | (tmp << (64 - 5));
+ b3 = ror64(tmp, 5);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 37) | (tmp << (64 - 37));
+ b1 = ror64(tmp, 37);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 23) | (tmp << (64 - 23));
+ b1 = ror64(tmp, 23);
b0 -= b1;
tmp = b3 ^ b2;
- b3 = (tmp >> 40) | (tmp << (64 - 40));
+ b3 = ror64(tmp, 40);
b2 -= b3;
tmp = b3 ^ b0;
- b3 = (tmp >> 52) | (tmp << (64 - 52));
+ b3 = ror64(tmp, 52);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 57) | (tmp << (64 - 57));
+ b1 = ror64(tmp, 57);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 14) | (tmp << (64 - 14));
+ b1 = ror64(tmp, 14);
b0 -= b1 + k1;
b1 -= k2 + t0;
tmp = b3 ^ b2;
- b3 = (tmp >> 16) | (tmp << (64 - 16));
+ b3 = ror64(tmp, 16);
b2 -= b3 + k3 + t1;
b3 -= k4 + 6;
tmp = b3 ^ b0;
- b3 = (tmp >> 32) | (tmp << (64 - 32));
+ b3 = ror64(tmp, 32);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 32) | (tmp << (64 - 32));
+ b1 = ror64(tmp, 32);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 58) | (tmp << (64 - 58));
+ b1 = ror64(tmp, 58);
b0 -= b1;
tmp = b3 ^ b2;
- b3 = (tmp >> 22) | (tmp << (64 - 22));
+ b3 = ror64(tmp, 22);
b2 -= b3;
tmp = b3 ^ b0;
- b3 = (tmp >> 46) | (tmp << (64 - 46));
+ b3 = ror64(tmp, 46);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 12) | (tmp << (64 - 12));
+ b1 = ror64(tmp, 12);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 25) | (tmp << (64 - 25));
+ b1 = ror64(tmp, 25);
b0 -= b1 + k0;
b1 -= k1 + t2;
tmp = b3 ^ b2;
- b3 = (tmp >> 33) | (tmp << (64 - 33));
+ b3 = ror64(tmp, 33);
b2 -= b3 + k2 + t0;
b3 -= k3 + 5;
tmp = b3 ^ b0;
- b3 = (tmp >> 5) | (tmp << (64 - 5));
+ b3 = ror64(tmp, 5);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 37) | (tmp << (64 - 37));
+ b1 = ror64(tmp, 37);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 23) | (tmp << (64 - 23));
+ b1 = ror64(tmp, 23);
b0 -= b1;
tmp = b3 ^ b2;
- b3 = (tmp >> 40) | (tmp << (64 - 40));
+ b3 = ror64(tmp, 40);
b2 -= b3;
tmp = b3 ^ b0;
- b3 = (tmp >> 52) | (tmp << (64 - 52));
+ b3 = ror64(tmp, 52);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 57) | (tmp << (64 - 57));
+ b1 = ror64(tmp, 57);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 14) | (tmp << (64 - 14));
+ b1 = ror64(tmp, 14);
b0 -= b1 + k4;
b1 -= k0 + t1;
tmp = b3 ^ b2;
- b3 = (tmp >> 16) | (tmp << (64 - 16));
+ b3 = ror64(tmp, 16);
b2 -= b3 + k1 + t2;
b3 -= k2 + 4;
tmp = b3 ^ b0;
- b3 = (tmp >> 32) | (tmp << (64 - 32));
+ b3 = ror64(tmp, 32);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 32) | (tmp << (64 - 32));
+ b1 = ror64(tmp, 32);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 58) | (tmp << (64 - 58));
+ b1 = ror64(tmp, 58);
b0 -= b1;
tmp = b3 ^ b2;
- b3 = (tmp >> 22) | (tmp << (64 - 22));
+ b3 = ror64(tmp, 22);
b2 -= b3;
tmp = b3 ^ b0;
- b3 = (tmp >> 46) | (tmp << (64 - 46));
+ b3 = ror64(tmp, 46);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 12) | (tmp << (64 - 12));
+ b1 = ror64(tmp, 12);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 25) | (tmp << (64 - 25));
+ b1 = ror64(tmp, 25);
b0 -= b1 + k3;
b1 -= k4 + t0;
tmp = b3 ^ b2;
- b3 = (tmp >> 33) | (tmp << (64 - 33));
+ b3 = ror64(tmp, 33);
b2 -= b3 + k0 + t1;
b3 -= k1 + 3;
tmp = b3 ^ b0;
- b3 = (tmp >> 5) | (tmp << (64 - 5));
+ b3 = ror64(tmp, 5);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 37) | (tmp << (64 - 37));
+ b1 = ror64(tmp, 37);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 23) | (tmp << (64 - 23));
+ b1 = ror64(tmp, 23);
b0 -= b1;
tmp = b3 ^ b2;
- b3 = (tmp >> 40) | (tmp << (64 - 40));
+ b3 = ror64(tmp, 40);
b2 -= b3;
tmp = b3 ^ b0;
- b3 = (tmp >> 52) | (tmp << (64 - 52));
+ b3 = ror64(tmp, 52);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 57) | (tmp << (64 - 57));
+ b1 = ror64(tmp, 57);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 14) | (tmp << (64 - 14));
+ b1 = ror64(tmp, 14);
b0 -= b1 + k2;
b1 -= k3 + t2;
tmp = b3 ^ b2;
- b3 = (tmp >> 16) | (tmp << (64 - 16));
+ b3 = ror64(tmp, 16);
b2 -= b3 + k4 + t0;
b3 -= k0 + 2;
tmp = b3 ^ b0;
- b3 = (tmp >> 32) | (tmp << (64 - 32));
+ b3 = ror64(tmp, 32);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 32) | (tmp << (64 - 32));
+ b1 = ror64(tmp, 32);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 58) | (tmp << (64 - 58));
+ b1 = ror64(tmp, 58);
b0 -= b1;
tmp = b3 ^ b2;
- b3 = (tmp >> 22) | (tmp << (64 - 22));
+ b3 = ror64(tmp, 22);
b2 -= b3;
tmp = b3 ^ b0;
- b3 = (tmp >> 46) | (tmp << (64 - 46));
+ b3 = ror64(tmp, 46);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 12) | (tmp << (64 - 12));
+ b1 = ror64(tmp, 12);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 25) | (tmp << (64 - 25));
+ b1 = ror64(tmp, 25);
b0 -= b1 + k1;
b1 -= k2 + t1;
tmp = b3 ^ b2;
- b3 = (tmp >> 33) | (tmp << (64 - 33));
+ b3 = ror64(tmp, 33);
b2 -= b3 + k3 + t2;
b3 -= k4 + 1;
tmp = b3 ^ b0;
- b3 = (tmp >> 5) | (tmp << (64 - 5));
+ b3 = ror64(tmp, 5);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 37) | (tmp << (64 - 37));
+ b1 = ror64(tmp, 37);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 23) | (tmp << (64 - 23));
+ b1 = ror64(tmp, 23);
b0 -= b1;
tmp = b3 ^ b2;
- b3 = (tmp >> 40) | (tmp << (64 - 40));
+ b3 = ror64(tmp, 40);
b2 -= b3;
tmp = b3 ^ b0;
- b3 = (tmp >> 52) | (tmp << (64 - 52));
+ b3 = ror64(tmp, 52);
b0 -= b3;
tmp = b1 ^ b2;
- b1 = (tmp >> 57) | (tmp << (64 - 57));
+ b1 = ror64(tmp, 57);
b2 -= b1;
tmp = b1 ^ b0;
- b1 = (tmp >> 14) | (tmp << (64 - 14));
+ b1 = ror64(tmp, 14);
b0 -= b1 + k0;
b1 -= k1 + t0;
tmp = b3 ^ b2;
- b3 = (tmp >> 16) | (tmp << (64 - 16));
+ b3 = ror64(tmp, 16);
b2 -= b3 + k2 + t1;
b3 -= k3;
@@ -2125,1226 +2125,1226 @@ void threefish_decrypt_512(struct threefish_key *key_ctx, u64 *input,
b7 -= k7 + 18;
tmp = b3 ^ b4;
- b3 = (tmp >> 22) | (tmp << (64 - 22));
+ b3 = ror64(tmp, 22);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 56) | (tmp << (64 - 56));
+ b5 = ror64(tmp, 56);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 35) | (tmp << (64 - 35));
+ b7 = ror64(tmp, 35);
b0 -= b7;
tmp = b1 ^ b6;
- b1 = (tmp >> 8) | (tmp << (64 - 8));
+ b1 = ror64(tmp, 8);
b6 -= b1;
tmp = b7 ^ b2;
- b7 = (tmp >> 43) | (tmp << (64 - 43));
+ b7 = ror64(tmp, 43);
b2 -= b7;
tmp = b5 ^ b0;
- b5 = (tmp >> 39) | (tmp << (64 - 39));
+ b5 = ror64(tmp, 39);
b0 -= b5;
tmp = b3 ^ b6;
- b3 = (tmp >> 29) | (tmp << (64 - 29));
+ b3 = ror64(tmp, 29);
b6 -= b3;
tmp = b1 ^ b4;
- b1 = (tmp >> 25) | (tmp << (64 - 25));
+ b1 = ror64(tmp, 25);
b4 -= b1;
tmp = b3 ^ b0;
- b3 = (tmp >> 17) | (tmp << (64 - 17));
+ b3 = ror64(tmp, 17);
b0 -= b3;
tmp = b5 ^ b6;
- b5 = (tmp >> 10) | (tmp << (64 - 10));
+ b5 = ror64(tmp, 10);
b6 -= b5;
tmp = b7 ^ b4;
- b7 = (tmp >> 50) | (tmp << (64 - 50));
+ b7 = ror64(tmp, 50);
b4 -= b7;
tmp = b1 ^ b2;
- b1 = (tmp >> 13) | (tmp << (64 - 13));
+ b1 = ror64(tmp, 13);
b2 -= b1;
tmp = b7 ^ b6;
- b7 = (tmp >> 24) | (tmp << (64 - 24));
+ b7 = ror64(tmp, 24);
b6 -= b7 + k5 + t0;
b7 -= k6 + 17;
tmp = b5 ^ b4;
- b5 = (tmp >> 34) | (tmp << (64 - 34));
+ b5 = ror64(tmp, 34);
b4 -= b5 + k3;
b5 -= k4 + t2;
tmp = b3 ^ b2;
- b3 = (tmp >> 30) | (tmp << (64 - 30));
+ b3 = ror64(tmp, 30);
b2 -= b3 + k1;
b3 -= k2;
tmp = b1 ^ b0;
- b1 = (tmp >> 39) | (tmp << (64 - 39));
+ b1 = ror64(tmp, 39);
b0 -= b1 + k8;
b1 -= k0;
tmp = b3 ^ b4;
- b3 = (tmp >> 56) | (tmp << (64 - 56));
+ b3 = ror64(tmp, 56);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 54) | (tmp << (64 - 54));
+ b5 = ror64(tmp, 54);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 9) | (tmp << (64 - 9));
+ b7 = ror64(tmp, 9);
b0 -= b7;
tmp = b1 ^ b6;
- b1 = (tmp >> 44) | (tmp << (64 - 44));
+ b1 = ror64(tmp, 44);
b6 -= b1;
tmp = b7 ^ b2;
- b7 = (tmp >> 39) | (tmp << (64 - 39));
+ b7 = ror64(tmp, 39);
b2 -= b7;
tmp = b5 ^ b0;
- b5 = (tmp >> 36) | (tmp << (64 - 36));
+ b5 = ror64(tmp, 36);
b0 -= b5;
tmp = b3 ^ b6;
- b3 = (tmp >> 49) | (tmp << (64 - 49));
+ b3 = ror64(tmp, 49);
b6 -= b3;
tmp = b1 ^ b4;
- b1 = (tmp >> 17) | (tmp << (64 - 17));
+ b1 = ror64(tmp, 17);
b4 -= b1;
tmp = b3 ^ b0;
- b3 = (tmp >> 42) | (tmp << (64 - 42));
+ b3 = ror64(tmp, 42);
b0 -= b3;
tmp = b5 ^ b6;
- b5 = (tmp >> 14) | (tmp << (64 - 14));
+ b5 = ror64(tmp, 14);
b6 -= b5;
tmp = b7 ^ b4;
- b7 = (tmp >> 27) | (tmp << (64 - 27));
+ b7 = ror64(tmp, 27);
b4 -= b7;
tmp = b1 ^ b2;
- b1 = (tmp >> 33) | (tmp << (64 - 33));
+ b1 = ror64(tmp, 33);
b2 -= b1;
tmp = b7 ^ b6;
- b7 = (tmp >> 37) | (tmp << (64 - 37));
+ b7 = ror64(tmp, 37);
b6 -= b7 + k4 + t2;
b7 -= k5 + 16;
tmp = b5 ^ b4;
- b5 = (tmp >> 19) | (tmp << (64 - 19));
+ b5 = ror64(tmp, 19);
b4 -= b5 + k2;
b5 -= k3 + t1;
tmp = b3 ^ b2;
- b3 = (tmp >> 36) | (tmp << (64 - 36));
+ b3 = ror64(tmp, 36);
b2 -= b3 + k0;
b3 -= k1;
tmp = b1 ^ b0;
- b1 = (tmp >> 46) | (tmp << (64 - 46));
+ b1 = ror64(tmp, 46);
b0 -= b1 + k7;
b1 -= k8;
tmp = b3 ^ b4;
- b3 = (tmp >> 22) | (tmp << (64 - 22));
+ b3 = ror64(tmp, 22);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 56) | (tmp << (64 - 56));
+ b5 = ror64(tmp, 56);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 35) | (tmp << (64 - 35));
+ b7 = ror64(tmp, 35);
b0 -= b7;
tmp = b1 ^ b6;
- b1 = (tmp >> 8) | (tmp << (64 - 8));
+ b1 = ror64(tmp, 8);
b6 -= b1;
tmp = b7 ^ b2;
- b7 = (tmp >> 43) | (tmp << (64 - 43));
+ b7 = ror64(tmp, 43);
b2 -= b7;
tmp = b5 ^ b0;
- b5 = (tmp >> 39) | (tmp << (64 - 39));
+ b5 = ror64(tmp, 39);
b0 -= b5;
tmp = b3 ^ b6;
- b3 = (tmp >> 29) | (tmp << (64 - 29));
+ b3 = ror64(tmp, 29);
b6 -= b3;
tmp = b1 ^ b4;
- b1 = (tmp >> 25) | (tmp << (64 - 25));
+ b1 = ror64(tmp, 25);
b4 -= b1;
tmp = b3 ^ b0;
- b3 = (tmp >> 17) | (tmp << (64 - 17));
+ b3 = ror64(tmp, 17);
b0 -= b3;
tmp = b5 ^ b6;
- b5 = (tmp >> 10) | (tmp << (64 - 10));
+ b5 = ror64(tmp, 10);
b6 -= b5;
tmp = b7 ^ b4;
- b7 = (tmp >> 50) | (tmp << (64 - 50));
+ b7 = ror64(tmp, 50);
b4 -= b7;
tmp = b1 ^ b2;
- b1 = (tmp >> 13) | (tmp << (64 - 13));
+ b1 = ror64(tmp, 13);
b2 -= b1;
tmp = b7 ^ b6;
- b7 = (tmp >> 24) | (tmp << (64 - 24));
+ b7 = ror64(tmp, 24);
b6 -= b7 + k3 + t1;
b7 -= k4 + 15;
tmp = b5 ^ b4;
- b5 = (tmp >> 34) | (tmp << (64 - 34));
+ b5 = ror64(tmp, 34);
b4 -= b5 + k1;
b5 -= k2 + t0;
tmp = b3 ^ b2;
- b3 = (tmp >> 30) | (tmp << (64 - 30));
+ b3 = ror64(tmp, 30);
b2 -= b3 + k8;
b3 -= k0;
tmp = b1 ^ b0;
- b1 = (tmp >> 39) | (tmp << (64 - 39));
+ b1 = ror64(tmp, 39);
b0 -= b1 + k6;
b1 -= k7;
tmp = b3 ^ b4;
- b3 = (tmp >> 56) | (tmp << (64 - 56));
+ b3 = ror64(tmp, 56);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 54) | (tmp << (64 - 54));
+ b5 = ror64(tmp, 54);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 9) | (tmp << (64 - 9));
+ b7 = ror64(tmp, 9);
b0 -= b7;
tmp = b1 ^ b6;
- b1 = (tmp >> 44) | (tmp << (64 - 44));
+ b1 = ror64(tmp, 44);
b6 -= b1;
tmp = b7 ^ b2;
- b7 = (tmp >> 39) | (tmp << (64 - 39));
+ b7 = ror64(tmp, 39);
b2 -= b7;
tmp = b5 ^ b0;
- b5 = (tmp >> 36) | (tmp << (64 - 36));
+ b5 = ror64(tmp, 36);
b0 -= b5;
tmp = b3 ^ b6;
- b3 = (tmp >> 49) | (tmp << (64 - 49));
+ b3 = ror64(tmp, 49);
b6 -= b3;
tmp = b1 ^ b4;
- b1 = (tmp >> 17) | (tmp << (64 - 17));
+ b1 = ror64(tmp, 17);
b4 -= b1;
tmp = b3 ^ b0;
- b3 = (tmp >> 42) | (tmp << (64 - 42));
+ b3 = ror64(tmp, 42);
b0 -= b3;
tmp = b5 ^ b6;
- b5 = (tmp >> 14) | (tmp << (64 - 14));
+ b5 = ror64(tmp, 14);
b6 -= b5;
tmp = b7 ^ b4;
- b7 = (tmp >> 27) | (tmp << (64 - 27));
+ b7 = ror64(tmp, 27);
b4 -= b7;
tmp = b1 ^ b2;
- b1 = (tmp >> 33) | (tmp << (64 - 33));
+ b1 = ror64(tmp, 33);
b2 -= b1;
tmp = b7 ^ b6;
- b7 = (tmp >> 37) | (tmp << (64 - 37));
+ b7 = ror64(tmp, 37);
b6 -= b7 + k2 + t0;
b7 -= k3 + 14;
tmp = b5 ^ b4;
- b5 = (tmp >> 19) | (tmp << (64 - 19));
+ b5 = ror64(tmp, 19);
b4 -= b5 + k0;
b5 -= k1 + t2;
tmp = b3 ^ b2;
- b3 = (tmp >> 36) | (tmp << (64 - 36));
+ b3 = ror64(tmp, 36);
b2 -= b3 + k7;
b3 -= k8;
tmp = b1 ^ b0;
- b1 = (tmp >> 46) | (tmp << (64 - 46));
+ b1 = ror64(tmp, 46);
b0 -= b1 + k5;
b1 -= k6;
tmp = b3 ^ b4;
- b3 = (tmp >> 22) | (tmp << (64 - 22));
+ b3 = ror64(tmp, 22);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 56) | (tmp << (64 - 56));
+ b5 = ror64(tmp, 56);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 35) | (tmp << (64 - 35));
+ b7 = ror64(tmp, 35);
b0 -= b7;
tmp = b1 ^ b6;
- b1 = (tmp >> 8) | (tmp << (64 - 8));
+ b1 = ror64(tmp, 8);
b6 -= b1;
tmp = b7 ^ b2;
- b7 = (tmp >> 43) | (tmp << (64 - 43));
+ b7 = ror64(tmp, 43);
b2 -= b7;
tmp = b5 ^ b0;
- b5 = (tmp >> 39) | (tmp << (64 - 39));
+ b5 = ror64(tmp, 39);
b0 -= b5;
tmp = b3 ^ b6;
- b3 = (tmp >> 29) | (tmp << (64 - 29));
+ b3 = ror64(tmp, 29);
b6 -= b3;
tmp = b1 ^ b4;
- b1 = (tmp >> 25) | (tmp << (64 - 25));
+ b1 = ror64(tmp, 25);
b4 -= b1;
tmp = b3 ^ b0;
- b3 = (tmp >> 17) | (tmp << (64 - 17));
+ b3 = ror64(tmp, 17);
b0 -= b3;
tmp = b5 ^ b6;
- b5 = (tmp >> 10) | (tmp << (64 - 10));
+ b5 = ror64(tmp, 10);
b6 -= b5;
tmp = b7 ^ b4;
- b7 = (tmp >> 50) | (tmp << (64 - 50));
+ b7 = ror64(tmp, 50);
b4 -= b7;
tmp = b1 ^ b2;
- b1 = (tmp >> 13) | (tmp << (64 - 13));
+ b1 = ror64(tmp, 13);
b2 -= b1;
tmp = b7 ^ b6;
- b7 = (tmp >> 24) | (tmp << (64 - 24));
+ b7 = ror64(tmp, 24);
b6 -= b7 + k1 + t2;
b7 -= k2 + 13;
tmp = b5 ^ b4;
- b5 = (tmp >> 34) | (tmp << (64 - 34));
+ b5 = ror64(tmp, 34);
b4 -= b5 + k8;
b5 -= k0 + t1;
tmp = b3 ^ b2;
- b3 = (tmp >> 30) | (tmp << (64 - 30));
+ b3 = ror64(tmp, 30);
b2 -= b3 + k6;
b3 -= k7;
tmp = b1 ^ b0;
- b1 = (tmp >> 39) | (tmp << (64 - 39));
+ b1 = ror64(tmp, 39);
b0 -= b1 + k4;
b1 -= k5;
tmp = b3 ^ b4;
- b3 = (tmp >> 56) | (tmp << (64 - 56));
+ b3 = ror64(tmp, 56);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 54) | (tmp << (64 - 54));
+ b5 = ror64(tmp, 54);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 9) | (tmp << (64 - 9));
+ b7 = ror64(tmp, 9);
b0 -= b7;
tmp = b1 ^ b6;
- b1 = (tmp >> 44) | (tmp << (64 - 44));
+ b1 = ror64(tmp, 44);
b6 -= b1;
tmp = b7 ^ b2;
- b7 = (tmp >> 39) | (tmp << (64 - 39));
+ b7 = ror64(tmp, 39);
b2 -= b7;
tmp = b5 ^ b0;
- b5 = (tmp >> 36) | (tmp << (64 - 36));
+ b5 = ror64(tmp, 36);
b0 -= b5;
tmp = b3 ^ b6;
- b3 = (tmp >> 49) | (tmp << (64 - 49));
+ b3 = ror64(tmp, 49);
b6 -= b3;
tmp = b1 ^ b4;
- b1 = (tmp >> 17) | (tmp << (64 - 17));
+ b1 = ror64(tmp, 17);
b4 -= b1;
tmp = b3 ^ b0;
- b3 = (tmp >> 42) | (tmp << (64 - 42));
+ b3 = ror64(tmp, 42);
b0 -= b3;
tmp = b5 ^ b6;
- b5 = (tmp >> 14) | (tmp << (64 - 14));
+ b5 = ror64(tmp, 14);
b6 -= b5;
tmp = b7 ^ b4;
- b7 = (tmp >> 27) | (tmp << (64 - 27));
+ b7 = ror64(tmp, 27);
b4 -= b7;
tmp = b1 ^ b2;
- b1 = (tmp >> 33) | (tmp << (64 - 33));
+ b1 = ror64(tmp, 33);
b2 -= b1;
tmp = b7 ^ b6;
- b7 = (tmp >> 37) | (tmp << (64 - 37));
+ b7 = ror64(tmp, 37);
b6 -= b7 + k0 + t1;
b7 -= k1 + 12;
tmp = b5 ^ b4;
- b5 = (tmp >> 19) | (tmp << (64 - 19));
+ b5 = ror64(tmp, 19);
b4 -= b5 + k7;
b5 -= k8 + t0;
tmp = b3 ^ b2;
- b3 = (tmp >> 36) | (tmp << (64 - 36));
+ b3 = ror64(tmp, 36);
b2 -= b3 + k5;
b3 -= k6;
tmp = b1 ^ b0;
- b1 = (tmp >> 46) | (tmp << (64 - 46));
+ b1 = ror64(tmp, 46);
b0 -= b1 + k3;
b1 -= k4;
tmp = b3 ^ b4;
- b3 = (tmp >> 22) | (tmp << (64 - 22));
+ b3 = ror64(tmp, 22);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 56) | (tmp << (64 - 56));
+ b5 = ror64(tmp, 56);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 35) | (tmp << (64 - 35));
+ b7 = ror64(tmp, 35);
b0 -= b7;
tmp = b1 ^ b6;
- b1 = (tmp >> 8) | (tmp << (64 - 8));
+ b1 = ror64(tmp, 8);
b6 -= b1;
tmp = b7 ^ b2;
- b7 = (tmp >> 43) | (tmp << (64 - 43));
+ b7 = ror64(tmp, 43);
b2 -= b7;
tmp = b5 ^ b0;
- b5 = (tmp >> 39) | (tmp << (64 - 39));
+ b5 = ror64(tmp, 39);
b0 -= b5;
tmp = b3 ^ b6;
- b3 = (tmp >> 29) | (tmp << (64 - 29));
+ b3 = ror64(tmp, 29);
b6 -= b3;
tmp = b1 ^ b4;
- b1 = (tmp >> 25) | (tmp << (64 - 25));
+ b1 = ror64(tmp, 25);
b4 -= b1;
tmp = b3 ^ b0;
- b3 = (tmp >> 17) | (tmp << (64 - 17));
+ b3 = ror64(tmp, 17);
b0 -= b3;
tmp = b5 ^ b6;
- b5 = (tmp >> 10) | (tmp << (64 - 10));
+ b5 = ror64(tmp, 10);
b6 -= b5;
tmp = b7 ^ b4;
- b7 = (tmp >> 50) | (tmp << (64 - 50));
+ b7 = ror64(tmp, 50);
b4 -= b7;
tmp = b1 ^ b2;
- b1 = (tmp >> 13) | (tmp << (64 - 13));
+ b1 = ror64(tmp, 13);
b2 -= b1;
tmp = b7 ^ b6;
- b7 = (tmp >> 24) | (tmp << (64 - 24));
+ b7 = ror64(tmp, 24);
b6 -= b7 + k8 + t0;
b7 -= k0 + 11;
tmp = b5 ^ b4;
- b5 = (tmp >> 34) | (tmp << (64 - 34));
+ b5 = ror64(tmp, 34);
b4 -= b5 + k6;
b5 -= k7 + t2;
tmp = b3 ^ b2;
- b3 = (tmp >> 30) | (tmp << (64 - 30));
+ b3 = ror64(tmp, 30);
b2 -= b3 + k4;
b3 -= k5;
tmp = b1 ^ b0;
- b1 = (tmp >> 39) | (tmp << (64 - 39));
+ b1 = ror64(tmp, 39);
b0 -= b1 + k2;
b1 -= k3;
tmp = b3 ^ b4;
- b3 = (tmp >> 56) | (tmp << (64 - 56));
+ b3 = ror64(tmp, 56);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 54) | (tmp << (64 - 54));
+ b5 = ror64(tmp, 54);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 9) | (tmp << (64 - 9));
+ b7 = ror64(tmp, 9);
b0 -= b7;
tmp = b1 ^ b6;
- b1 = (tmp >> 44) | (tmp << (64 - 44));
+ b1 = ror64(tmp, 44);
b6 -= b1;
tmp = b7 ^ b2;
- b7 = (tmp >> 39) | (tmp << (64 - 39));
+ b7 = ror64(tmp, 39);
b2 -= b7;
tmp = b5 ^ b0;
- b5 = (tmp >> 36) | (tmp << (64 - 36));
+ b5 = ror64(tmp, 36);
b0 -= b5;
tmp = b3 ^ b6;
- b3 = (tmp >> 49) | (tmp << (64 - 49));
+ b3 = ror64(tmp, 49);
b6 -= b3;
tmp = b1 ^ b4;
- b1 = (tmp >> 17) | (tmp << (64 - 17));
+ b1 = ror64(tmp, 17);
b4 -= b1;
tmp = b3 ^ b0;
- b3 = (tmp >> 42) | (tmp << (64 - 42));
+ b3 = ror64(tmp, 42);
b0 -= b3;
tmp = b5 ^ b6;
- b5 = (tmp >> 14) | (tmp << (64 - 14));
+ b5 = ror64(tmp, 14);
b6 -= b5;
tmp = b7 ^ b4;
- b7 = (tmp >> 27) | (tmp << (64 - 27));
+ b7 = ror64(tmp, 27);
b4 -= b7;
tmp = b1 ^ b2;
- b1 = (tmp >> 33) | (tmp << (64 - 33));
+ b1 = ror64(tmp, 33);
b2 -= b1;
tmp = b7 ^ b6;
- b7 = (tmp >> 37) | (tmp << (64 - 37));
+ b7 = ror64(tmp, 37);
b6 -= b7 + k7 + t2;
b7 -= k8 + 10;
tmp = b5 ^ b4;
- b5 = (tmp >> 19) | (tmp << (64 - 19));
+ b5 = ror64(tmp, 19);
b4 -= b5 + k5;
b5 -= k6 + t1;
tmp = b3 ^ b2;
- b3 = (tmp >> 36) | (tmp << (64 - 36));
+ b3 = ror64(tmp, 36);
b2 -= b3 + k3;
b3 -= k4;
tmp = b1 ^ b0;
- b1 = (tmp >> 46) | (tmp << (64 - 46));
+ b1 = ror64(tmp, 46);
b0 -= b1 + k1;
b1 -= k2;
tmp = b3 ^ b4;
- b3 = (tmp >> 22) | (tmp << (64 - 22));
+ b3 = ror64(tmp, 22);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 56) | (tmp << (64 - 56));
+ b5 = ror64(tmp, 56);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 35) | (tmp << (64 - 35));
+ b7 = ror64(tmp, 35);
b0 -= b7;
tmp = b1 ^ b6;
- b1 = (tmp >> 8) | (tmp << (64 - 8));
+ b1 = ror64(tmp, 8);
b6 -= b1;
tmp = b7 ^ b2;
- b7 = (tmp >> 43) | (tmp << (64 - 43));
+ b7 = ror64(tmp, 43);
b2 -= b7;
tmp = b5 ^ b0;
- b5 = (tmp >> 39) | (tmp << (64 - 39));
+ b5 = ror64(tmp, 39);
b0 -= b5;
tmp = b3 ^ b6;
- b3 = (tmp >> 29) | (tmp << (64 - 29));
+ b3 = ror64(tmp, 29);
b6 -= b3;
tmp = b1 ^ b4;
- b1 = (tmp >> 25) | (tmp << (64 - 25));
+ b1 = ror64(tmp, 25);
b4 -= b1;
tmp = b3 ^ b0;
- b3 = (tmp >> 17) | (tmp << (64 - 17));
+ b3 = ror64(tmp, 17);
b0 -= b3;
tmp = b5 ^ b6;
- b5 = (tmp >> 10) | (tmp << (64 - 10));
+ b5 = ror64(tmp, 10);
b6 -= b5;
tmp = b7 ^ b4;
- b7 = (tmp >> 50) | (tmp << (64 - 50));
+ b7 = ror64(tmp, 50);
b4 -= b7;
tmp = b1 ^ b2;
- b1 = (tmp >> 13) | (tmp << (64 - 13));
+ b1 = ror64(tmp, 13);
b2 -= b1;
tmp = b7 ^ b6;
- b7 = (tmp >> 24) | (tmp << (64 - 24));
+ b7 = ror64(tmp, 24);
b6 -= b7 + k6 + t1;
b7 -= k7 + 9;
tmp = b5 ^ b4;
- b5 = (tmp >> 34) | (tmp << (64 - 34));
+ b5 = ror64(tmp, 34);
b4 -= b5 + k4;
b5 -= k5 + t0;
tmp = b3 ^ b2;
- b3 = (tmp >> 30) | (tmp << (64 - 30));
+ b3 = ror64(tmp, 30);
b2 -= b3 + k2;
b3 -= k3;
tmp = b1 ^ b0;
- b1 = (tmp >> 39) | (tmp << (64 - 39));
+ b1 = ror64(tmp, 39);
b0 -= b1 + k0;
b1 -= k1;
tmp = b3 ^ b4;
- b3 = (tmp >> 56) | (tmp << (64 - 56));
+ b3 = ror64(tmp, 56);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 54) | (tmp << (64 - 54));
+ b5 = ror64(tmp, 54);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 9) | (tmp << (64 - 9));
+ b7 = ror64(tmp, 9);
b0 -= b7;
tmp = b1 ^ b6;
- b1 = (tmp >> 44) | (tmp << (64 - 44));
+ b1 = ror64(tmp, 44);
b6 -= b1;
tmp = b7 ^ b2;
- b7 = (tmp >> 39) | (tmp << (64 - 39));
+ b7 = ror64(tmp, 39);
b2 -= b7;
tmp = b5 ^ b0;
- b5 = (tmp >> 36) | (tmp << (64 - 36));
+ b5 = ror64(tmp, 36);
b0 -= b5;
tmp = b3 ^ b6;
- b3 = (tmp >> 49) | (tmp << (64 - 49));
+ b3 = ror64(tmp, 49);
b6 -= b3;
tmp = b1 ^ b4;
- b1 = (tmp >> 17) | (tmp << (64 - 17));
+ b1 = ror64(tmp, 17);
b4 -= b1;
tmp = b3 ^ b0;
- b3 = (tmp >> 42) | (tmp << (64 - 42));
+ b3 = ror64(tmp, 42);
b0 -= b3;
tmp = b5 ^ b6;
- b5 = (tmp >> 14) | (tmp << (64 - 14));
+ b5 = ror64(tmp, 14);
b6 -= b5;
tmp = b7 ^ b4;
- b7 = (tmp >> 27) | (tmp << (64 - 27));
+ b7 = ror64(tmp, 27);
b4 -= b7;
tmp = b1 ^ b2;
- b1 = (tmp >> 33) | (tmp << (64 - 33));
+ b1 = ror64(tmp, 33);
b2 -= b1;
tmp = b7 ^ b6;
- b7 = (tmp >> 37) | (tmp << (64 - 37));
+ b7 = ror64(tmp, 37);
b6 -= b7 + k5 + t0;
b7 -= k6 + 8;
tmp = b5 ^ b4;
- b5 = (tmp >> 19) | (tmp << (64 - 19));
+ b5 = ror64(tmp, 19);
b4 -= b5 + k3;
b5 -= k4 + t2;
tmp = b3 ^ b2;
- b3 = (tmp >> 36) | (tmp << (64 - 36));
+ b3 = ror64(tmp, 36);
b2 -= b3 + k1;
b3 -= k2;
tmp = b1 ^ b0;
- b1 = (tmp >> 46) | (tmp << (64 - 46));
+ b1 = ror64(tmp, 46);
b0 -= b1 + k8;
b1 -= k0;
tmp = b3 ^ b4;
- b3 = (tmp >> 22) | (tmp << (64 - 22));
+ b3 = ror64(tmp, 22);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 56) | (tmp << (64 - 56));
+ b5 = ror64(tmp, 56);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 35) | (tmp << (64 - 35));
+ b7 = ror64(tmp, 35);
b0 -= b7;
tmp = b1 ^ b6;
- b1 = (tmp >> 8) | (tmp << (64 - 8));
+ b1 = ror64(tmp, 8);
b6 -= b1;
tmp = b7 ^ b2;
- b7 = (tmp >> 43) | (tmp << (64 - 43));
+ b7 = ror64(tmp, 43);
b2 -= b7;
tmp = b5 ^ b0;
- b5 = (tmp >> 39) | (tmp << (64 - 39));
+ b5 = ror64(tmp, 39);
b0 -= b5;
tmp = b3 ^ b6;
- b3 = (tmp >> 29) | (tmp << (64 - 29));
+ b3 = ror64(tmp, 29);
b6 -= b3;
tmp = b1 ^ b4;
- b1 = (tmp >> 25) | (tmp << (64 - 25));
+ b1 = ror64(tmp, 25);
b4 -= b1;
tmp = b3 ^ b0;
- b3 = (tmp >> 17) | (tmp << (64 - 17));
+ b3 = ror64(tmp, 17);
b0 -= b3;
tmp = b5 ^ b6;
- b5 = (tmp >> 10) | (tmp << (64 - 10));
+ b5 = ror64(tmp, 10);
b6 -= b5;
tmp = b7 ^ b4;
- b7 = (tmp >> 50) | (tmp << (64 - 50));
+ b7 = ror64(tmp, 50);
b4 -= b7;
tmp = b1 ^ b2;
- b1 = (tmp >> 13) | (tmp << (64 - 13));
+ b1 = ror64(tmp, 13);
b2 -= b1;
tmp = b7 ^ b6;
- b7 = (tmp >> 24) | (tmp << (64 - 24));
+ b7 = ror64(tmp, 24);
b6 -= b7 + k4 + t2;
b7 -= k5 + 7;
tmp = b5 ^ b4;
- b5 = (tmp >> 34) | (tmp << (64 - 34));
+ b5 = ror64(tmp, 34);
b4 -= b5 + k2;
b5 -= k3 + t1;
tmp = b3 ^ b2;
- b3 = (tmp >> 30) | (tmp << (64 - 30));
+ b3 = ror64(tmp, 30);
b2 -= b3 + k0;
b3 -= k1;
tmp = b1 ^ b0;
- b1 = (tmp >> 39) | (tmp << (64 - 39));
+ b1 = ror64(tmp, 39);
b0 -= b1 + k7;
b1 -= k8;
tmp = b3 ^ b4;
- b3 = (tmp >> 56) | (tmp << (64 - 56));
+ b3 = ror64(tmp, 56);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 54) | (tmp << (64 - 54));
+ b5 = ror64(tmp, 54);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 9) | (tmp << (64 - 9));
+ b7 = ror64(tmp, 9);
b0 -= b7;
tmp = b1 ^ b6;
- b1 = (tmp >> 44) | (tmp << (64 - 44));
+ b1 = ror64(tmp, 44);
b6 -= b1;
tmp = b7 ^ b2;
- b7 = (tmp >> 39) | (tmp << (64 - 39));
+ b7 = ror64(tmp, 39);
b2 -= b7;
tmp = b5 ^ b0;
- b5 = (tmp >> 36) | (tmp << (64 - 36));
+ b5 = ror64(tmp, 36);
b0 -= b5;
tmp = b3 ^ b6;
- b3 = (tmp >> 49) | (tmp << (64 - 49));
+ b3 = ror64(tmp, 49);
b6 -= b3;
tmp = b1 ^ b4;
- b1 = (tmp >> 17) | (tmp << (64 - 17));
+ b1 = ror64(tmp, 17);
b4 -= b1;
tmp = b3 ^ b0;
- b3 = (tmp >> 42) | (tmp << (64 - 42));
+ b3 = ror64(tmp, 42);
b0 -= b3;
tmp = b5 ^ b6;
- b5 = (tmp >> 14) | (tmp << (64 - 14));
+ b5 = ror64(tmp, 14);
b6 -= b5;
tmp = b7 ^ b4;
- b7 = (tmp >> 27) | (tmp << (64 - 27));
+ b7 = ror64(tmp, 27);
b4 -= b7;
tmp = b1 ^ b2;
- b1 = (tmp >> 33) | (tmp << (64 - 33));
+ b1 = ror64(tmp, 33);
b2 -= b1;
tmp = b7 ^ b6;
- b7 = (tmp >> 37) | (tmp << (64 - 37));
+ b7 = ror64(tmp, 37);
b6 -= b7 + k3 + t1;
b7 -= k4 + 6;
tmp = b5 ^ b4;
- b5 = (tmp >> 19) | (tmp << (64 - 19));
+ b5 = ror64(tmp, 19);
b4 -= b5 + k1;
b5 -= k2 + t0;
tmp = b3 ^ b2;
- b3 = (tmp >> 36) | (tmp << (64 - 36));
+ b3 = ror64(tmp, 36);
b2 -= b3 + k8;
b3 -= k0;
tmp = b1 ^ b0;
- b1 = (tmp >> 46) | (tmp << (64 - 46));
+ b1 = ror64(tmp, 46);
b0 -= b1 + k6;
b1 -= k7;
tmp = b3 ^ b4;
- b3 = (tmp >> 22) | (tmp << (64 - 22));
+ b3 = ror64(tmp, 22);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 56) | (tmp << (64 - 56));
+ b5 = ror64(tmp, 56);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 35) | (tmp << (64 - 35));
+ b7 = ror64(tmp, 35);
b0 -= b7;
tmp = b1 ^ b6;
- b1 = (tmp >> 8) | (tmp << (64 - 8));
+ b1 = ror64(tmp, 8);
b6 -= b1;
tmp = b7 ^ b2;
- b7 = (tmp >> 43) | (tmp << (64 - 43));
+ b7 = ror64(tmp, 43);
b2 -= b7;
tmp = b5 ^ b0;
- b5 = (tmp >> 39) | (tmp << (64 - 39));
+ b5 = ror64(tmp, 39);
b0 -= b5;
tmp = b3 ^ b6;
- b3 = (tmp >> 29) | (tmp << (64 - 29));
+ b3 = ror64(tmp, 29);
b6 -= b3;
tmp = b1 ^ b4;
- b1 = (tmp >> 25) | (tmp << (64 - 25));
+ b1 = ror64(tmp, 25);
b4 -= b1;
tmp = b3 ^ b0;
- b3 = (tmp >> 17) | (tmp << (64 - 17));
+ b3 = ror64(tmp, 17);
b0 -= b3;
tmp = b5 ^ b6;
- b5 = (tmp >> 10) | (tmp << (64 - 10));
+ b5 = ror64(tmp, 10);
b6 -= b5;
tmp = b7 ^ b4;
- b7 = (tmp >> 50) | (tmp << (64 - 50));
+ b7 = ror64(tmp, 50);
b4 -= b7;
tmp = b1 ^ b2;
- b1 = (tmp >> 13) | (tmp << (64 - 13));
+ b1 = ror64(tmp, 13);
b2 -= b1;
tmp = b7 ^ b6;
- b7 = (tmp >> 24) | (tmp << (64 - 24));
+ b7 = ror64(tmp, 24);
b6 -= b7 + k2 + t0;
b7 -= k3 + 5;
tmp = b5 ^ b4;
- b5 = (tmp >> 34) | (tmp << (64 - 34));
+ b5 = ror64(tmp, 34);
b4 -= b5 + k0;
b5 -= k1 + t2;
tmp = b3 ^ b2;
- b3 = (tmp >> 30) | (tmp << (64 - 30));
+ b3 = ror64(tmp, 30);
b2 -= b3 + k7;
b3 -= k8;
tmp = b1 ^ b0;
- b1 = (tmp >> 39) | (tmp << (64 - 39));
+ b1 = ror64(tmp, 39);
b0 -= b1 + k5;
b1 -= k6;
tmp = b3 ^ b4;
- b3 = (tmp >> 56) | (tmp << (64 - 56));
+ b3 = ror64(tmp, 56);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 54) | (tmp << (64 - 54));
+ b5 = ror64(tmp, 54);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 9) | (tmp << (64 - 9));
+ b7 = ror64(tmp, 9);
b0 -= b7;
tmp = b1 ^ b6;
- b1 = (tmp >> 44) | (tmp << (64 - 44));
+ b1 = ror64(tmp, 44);
b6 -= b1;
tmp = b7 ^ b2;
- b7 = (tmp >> 39) | (tmp << (64 - 39));
+ b7 = ror64(tmp, 39);
b2 -= b7;
tmp = b5 ^ b0;
- b5 = (tmp >> 36) | (tmp << (64 - 36));
+ b5 = ror64(tmp, 36);
b0 -= b5;
tmp = b3 ^ b6;
- b3 = (tmp >> 49) | (tmp << (64 - 49));
+ b3 = ror64(tmp, 49);
b6 -= b3;
tmp = b1 ^ b4;
- b1 = (tmp >> 17) | (tmp << (64 - 17));
+ b1 = ror64(tmp, 17);
b4 -= b1;
tmp = b3 ^ b0;
- b3 = (tmp >> 42) | (tmp << (64 - 42));
+ b3 = ror64(tmp, 42);
b0 -= b3;
tmp = b5 ^ b6;
- b5 = (tmp >> 14) | (tmp << (64 - 14));
+ b5 = ror64(tmp, 14);
b6 -= b5;
tmp = b7 ^ b4;
- b7 = (tmp >> 27) | (tmp << (64 - 27));
+ b7 = ror64(tmp, 27);
b4 -= b7;
tmp = b1 ^ b2;
- b1 = (tmp >> 33) | (tmp << (64 - 33));
+ b1 = ror64(tmp, 33);
b2 -= b1;
tmp = b7 ^ b6;
- b7 = (tmp >> 37) | (tmp << (64 - 37));
+ b7 = ror64(tmp, 37);
b6 -= b7 + k1 + t2;
b7 -= k2 + 4;
tmp = b5 ^ b4;
- b5 = (tmp >> 19) | (tmp << (64 - 19));
+ b5 = ror64(tmp, 19);
b4 -= b5 + k8;
b5 -= k0 + t1;
tmp = b3 ^ b2;
- b3 = (tmp >> 36) | (tmp << (64 - 36));
+ b3 = ror64(tmp, 36);
b2 -= b3 + k6;
b3 -= k7;
tmp = b1 ^ b0;
- b1 = (tmp >> 46) | (tmp << (64 - 46));
+ b1 = ror64(tmp, 46);
b0 -= b1 + k4;
b1 -= k5;
tmp = b3 ^ b4;
- b3 = (tmp >> 22) | (tmp << (64 - 22));
+ b3 = ror64(tmp, 22);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 56) | (tmp << (64 - 56));
+ b5 = ror64(tmp, 56);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 35) | (tmp << (64 - 35));
+ b7 = ror64(tmp, 35);
b0 -= b7;
tmp = b1 ^ b6;
- b1 = (tmp >> 8) | (tmp << (64 - 8));
+ b1 = ror64(tmp, 8);
b6 -= b1;
tmp = b7 ^ b2;
- b7 = (tmp >> 43) | (tmp << (64 - 43));
+ b7 = ror64(tmp, 43);
b2 -= b7;
tmp = b5 ^ b0;
- b5 = (tmp >> 39) | (tmp << (64 - 39));
+ b5 = ror64(tmp, 39);
b0 -= b5;
tmp = b3 ^ b6;
- b3 = (tmp >> 29) | (tmp << (64 - 29));
+ b3 = ror64(tmp, 29);
b6 -= b3;
tmp = b1 ^ b4;
- b1 = (tmp >> 25) | (tmp << (64 - 25));
+ b1 = ror64(tmp, 25);
b4 -= b1;
tmp = b3 ^ b0;
- b3 = (tmp >> 17) | (tmp << (64 - 17));
+ b3 = ror64(tmp, 17);
b0 -= b3;
tmp = b5 ^ b6;
- b5 = (tmp >> 10) | (tmp << (64 - 10));
+ b5 = ror64(tmp, 10);
b6 -= b5;
tmp = b7 ^ b4;
- b7 = (tmp >> 50) | (tmp << (64 - 50));
+ b7 = ror64(tmp, 50);
b4 -= b7;
tmp = b1 ^ b2;
- b1 = (tmp >> 13) | (tmp << (64 - 13));
+ b1 = ror64(tmp, 13);
b2 -= b1;
tmp = b7 ^ b6;
- b7 = (tmp >> 24) | (tmp << (64 - 24));
+ b7 = ror64(tmp, 24);
b6 -= b7 + k0 + t1;
b7 -= k1 + 3;
tmp = b5 ^ b4;
- b5 = (tmp >> 34) | (tmp << (64 - 34));
+ b5 = ror64(tmp, 34);
b4 -= b5 + k7;
b5 -= k8 + t0;
tmp = b3 ^ b2;
- b3 = (tmp >> 30) | (tmp << (64 - 30));
+ b3 = ror64(tmp, 30);
b2 -= b3 + k5;
b3 -= k6;
tmp = b1 ^ b0;
- b1 = (tmp >> 39) | (tmp << (64 - 39));
+ b1 = ror64(tmp, 39);
b0 -= b1 + k3;
b1 -= k4;
tmp = b3 ^ b4;
- b3 = (tmp >> 56) | (tmp << (64 - 56));
+ b3 = ror64(tmp, 56);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 54) | (tmp << (64 - 54));
+ b5 = ror64(tmp, 54);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 9) | (tmp << (64 - 9));
+ b7 = ror64(tmp, 9);
b0 -= b7;
tmp = b1 ^ b6;
- b1 = (tmp >> 44) | (tmp << (64 - 44));
+ b1 = ror64(tmp, 44);
b6 -= b1;
tmp = b7 ^ b2;
- b7 = (tmp >> 39) | (tmp << (64 - 39));
+ b7 = ror64(tmp, 39);
b2 -= b7;
tmp = b5 ^ b0;
- b5 = (tmp >> 36) | (tmp << (64 - 36));
+ b5 = ror64(tmp, 36);
b0 -= b5;
tmp = b3 ^ b6;
- b3 = (tmp >> 49) | (tmp << (64 - 49));
+ b3 = ror64(tmp, 49);
b6 -= b3;
tmp = b1 ^ b4;
- b1 = (tmp >> 17) | (tmp << (64 - 17));
+ b1 = ror64(tmp, 17);
b4 -= b1;
tmp = b3 ^ b0;
- b3 = (tmp >> 42) | (tmp << (64 - 42));
+ b3 = ror64(tmp, 42);
b0 -= b3;
tmp = b5 ^ b6;
- b5 = (tmp >> 14) | (tmp << (64 - 14));
+ b5 = ror64(tmp, 14);
b6 -= b5;
tmp = b7 ^ b4;
- b7 = (tmp >> 27) | (tmp << (64 - 27));
+ b7 = ror64(tmp, 27);
b4 -= b7;
tmp = b1 ^ b2;
- b1 = (tmp >> 33) | (tmp << (64 - 33));
+ b1 = ror64(tmp, 33);
b2 -= b1;
tmp = b7 ^ b6;
- b7 = (tmp >> 37) | (tmp << (64 - 37));
+ b7 = ror64(tmp, 37);
b6 -= b7 + k8 + t0;
b7 -= k0 + 2;
tmp = b5 ^ b4;
- b5 = (tmp >> 19) | (tmp << (64 - 19));
+ b5 = ror64(tmp, 19);
b4 -= b5 + k6;
b5 -= k7 + t2;
tmp = b3 ^ b2;
- b3 = (tmp >> 36) | (tmp << (64 - 36));
+ b3 = ror64(tmp, 36);
b2 -= b3 + k4;
b3 -= k5;
tmp = b1 ^ b0;
- b1 = (tmp >> 46) | (tmp << (64 - 46));
+ b1 = ror64(tmp, 46);
b0 -= b1 + k2;
b1 -= k3;
tmp = b3 ^ b4;
- b3 = (tmp >> 22) | (tmp << (64 - 22));
+ b3 = ror64(tmp, 22);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 56) | (tmp << (64 - 56));
+ b5 = ror64(tmp, 56);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 35) | (tmp << (64 - 35));
+ b7 = ror64(tmp, 35);
b0 -= b7;
tmp = b1 ^ b6;
- b1 = (tmp >> 8) | (tmp << (64 - 8));
+ b1 = ror64(tmp, 8);
b6 -= b1;
tmp = b7 ^ b2;
- b7 = (tmp >> 43) | (tmp << (64 - 43));
+ b7 = ror64(tmp, 43);
b2 -= b7;
tmp = b5 ^ b0;
- b5 = (tmp >> 39) | (tmp << (64 - 39));
+ b5 = ror64(tmp, 39);
b0 -= b5;
tmp = b3 ^ b6;
- b3 = (tmp >> 29) | (tmp << (64 - 29));
+ b3 = ror64(tmp, 29);
b6 -= b3;
tmp = b1 ^ b4;
- b1 = (tmp >> 25) | (tmp << (64 - 25));
+ b1 = ror64(tmp, 25);
b4 -= b1;
tmp = b3 ^ b0;
- b3 = (tmp >> 17) | (tmp << (64 - 17));
+ b3 = ror64(tmp, 17);
b0 -= b3;
tmp = b5 ^ b6;
- b5 = (tmp >> 10) | (tmp << (64 - 10));
+ b5 = ror64(tmp, 10);
b6 -= b5;
tmp = b7 ^ b4;
- b7 = (tmp >> 50) | (tmp << (64 - 50));
+ b7 = ror64(tmp, 50);
b4 -= b7;
tmp = b1 ^ b2;
- b1 = (tmp >> 13) | (tmp << (64 - 13));
+ b1 = ror64(tmp, 13);
b2 -= b1;
tmp = b7 ^ b6;
- b7 = (tmp >> 24) | (tmp << (64 - 24));
+ b7 = ror64(tmp, 24);
b6 -= b7 + k7 + t2;
b7 -= k8 + 1;
tmp = b5 ^ b4;
- b5 = (tmp >> 34) | (tmp << (64 - 34));
+ b5 = ror64(tmp, 34);
b4 -= b5 + k5;
b5 -= k6 + t1;
tmp = b3 ^ b2;
- b3 = (tmp >> 30) | (tmp << (64 - 30));
+ b3 = ror64(tmp, 30);
b2 -= b3 + k3;
b3 -= k4;
tmp = b1 ^ b0;
- b1 = (tmp >> 39) | (tmp << (64 - 39));
+ b1 = ror64(tmp, 39);
b0 -= b1 + k1;
b1 -= k2;
tmp = b3 ^ b4;
- b3 = (tmp >> 56) | (tmp << (64 - 56));
+ b3 = ror64(tmp, 56);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 54) | (tmp << (64 - 54));
+ b5 = ror64(tmp, 54);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 9) | (tmp << (64 - 9));
+ b7 = ror64(tmp, 9);
b0 -= b7;
tmp = b1 ^ b6;
- b1 = (tmp >> 44) | (tmp << (64 - 44));
+ b1 = ror64(tmp, 44);
b6 -= b1;
tmp = b7 ^ b2;
- b7 = (tmp >> 39) | (tmp << (64 - 39));
+ b7 = ror64(tmp, 39);
b2 -= b7;
tmp = b5 ^ b0;
- b5 = (tmp >> 36) | (tmp << (64 - 36));
+ b5 = ror64(tmp, 36);
b0 -= b5;
tmp = b3 ^ b6;
- b3 = (tmp >> 49) | (tmp << (64 - 49));
+ b3 = ror64(tmp, 49);
b6 -= b3;
tmp = b1 ^ b4;
- b1 = (tmp >> 17) | (tmp << (64 - 17));
+ b1 = ror64(tmp, 17);
b4 -= b1;
tmp = b3 ^ b0;
- b3 = (tmp >> 42) | (tmp << (64 - 42));
+ b3 = ror64(tmp, 42);
b0 -= b3;
tmp = b5 ^ b6;
- b5 = (tmp >> 14) | (tmp << (64 - 14));
+ b5 = ror64(tmp, 14);
b6 -= b5;
tmp = b7 ^ b4;
- b7 = (tmp >> 27) | (tmp << (64 - 27));
+ b7 = ror64(tmp, 27);
b4 -= b7;
tmp = b1 ^ b2;
- b1 = (tmp >> 33) | (tmp << (64 - 33));
+ b1 = ror64(tmp, 33);
b2 -= b1;
tmp = b7 ^ b6;
- b7 = (tmp >> 37) | (tmp << (64 - 37));
+ b7 = ror64(tmp, 37);
b6 -= b7 + k6 + t1;
b7 -= k7;
tmp = b5 ^ b4;
- b5 = (tmp >> 19) | (tmp << (64 - 19));
+ b5 = ror64(tmp, 19);
b4 -= b5 + k4;
b5 -= k5 + t0;
tmp = b3 ^ b2;
- b3 = (tmp >> 36) | (tmp << (64 - 36));
+ b3 = ror64(tmp, 36);
b2 -= b3 + k2;
b3 -= k3;
tmp = b1 ^ b0;
- b1 = (tmp >> 46) | (tmp << (64 - 46));
+ b1 = ror64(tmp, 46);
b0 -= b1 + k0;
b1 -= k1;
@@ -5521,2722 +5521,2722 @@ void threefish_decrypt_1024(struct threefish_key *key_ctx, u64 *input,
b14 -= k0 + t0;
b15 -= k1 + 20;
tmp = b7 ^ b12;
- b7 = (tmp >> 20) | (tmp << (64 - 20));
+ b7 = ror64(tmp, 20);
b12 -= b7;
tmp = b3 ^ b10;
- b3 = (tmp >> 37) | (tmp << (64 - 37));
+ b3 = ror64(tmp, 37);
b10 -= b3;
tmp = b5 ^ b8;
- b5 = (tmp >> 31) | (tmp << (64 - 31));
+ b5 = ror64(tmp, 31);
b8 -= b5;
tmp = b1 ^ b14;
- b1 = (tmp >> 23) | (tmp << (64 - 23));
+ b1 = ror64(tmp, 23);
b14 -= b1;
tmp = b9 ^ b4;
- b9 = (tmp >> 52) | (tmp << (64 - 52));
+ b9 = ror64(tmp, 52);
b4 -= b9;
tmp = b13 ^ b6;
- b13 = (tmp >> 35) | (tmp << (64 - 35));
+ b13 = ror64(tmp, 35);
b6 -= b13;
tmp = b11 ^ b2;
- b11 = (tmp >> 48) | (tmp << (64 - 48));
+ b11 = ror64(tmp, 48);
b2 -= b11;
tmp = b15 ^ b0;
- b15 = (tmp >> 9) | (tmp << (64 - 9));
+ b15 = ror64(tmp, 9);
b0 -= b15;
tmp = b9 ^ b10;
- b9 = (tmp >> 25) | (tmp << (64 - 25));
+ b9 = ror64(tmp, 25);
b10 -= b9;
tmp = b11 ^ b8;
- b11 = (tmp >> 44) | (tmp << (64 - 44));
+ b11 = ror64(tmp, 44);
b8 -= b11;
tmp = b13 ^ b14;
- b13 = (tmp >> 42) | (tmp << (64 - 42));
+ b13 = ror64(tmp, 42);
b14 -= b13;
tmp = b15 ^ b12;
- b15 = (tmp >> 19) | (tmp << (64 - 19));
+ b15 = ror64(tmp, 19);
b12 -= b15;
tmp = b1 ^ b6;
- b1 = (tmp >> 46) | (tmp << (64 - 46));
+ b1 = ror64(tmp, 46);
b6 -= b1;
tmp = b3 ^ b4;
- b3 = (tmp >> 47) | (tmp << (64 - 47));
+ b3 = ror64(tmp, 47);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 44) | (tmp << (64 - 44));
+ b5 = ror64(tmp, 44);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 31) | (tmp << (64 - 31));
+ b7 = ror64(tmp, 31);
b0 -= b7;
tmp = b1 ^ b8;
- b1 = (tmp >> 41) | (tmp << (64 - 41));
+ b1 = ror64(tmp, 41);
b8 -= b1;
tmp = b5 ^ b14;
- b5 = (tmp >> 42) | (tmp << (64 - 42));
+ b5 = ror64(tmp, 42);
b14 -= b5;
tmp = b3 ^ b12;
- b3 = (tmp >> 53) | (tmp << (64 - 53));
+ b3 = ror64(tmp, 53);
b12 -= b3;
tmp = b7 ^ b10;
- b7 = (tmp >> 4) | (tmp << (64 - 4));
+ b7 = ror64(tmp, 4);
b10 -= b7;
tmp = b15 ^ b4;
- b15 = (tmp >> 51) | (tmp << (64 - 51));
+ b15 = ror64(tmp, 51);
b4 -= b15;
tmp = b11 ^ b6;
- b11 = (tmp >> 56) | (tmp << (64 - 56));
+ b11 = ror64(tmp, 56);
b6 -= b11;
tmp = b13 ^ b2;
- b13 = (tmp >> 34) | (tmp << (64 - 34));
+ b13 = ror64(tmp, 34);
b2 -= b13;
tmp = b9 ^ b0;
- b9 = (tmp >> 16) | (tmp << (64 - 16));
+ b9 = ror64(tmp, 16);
b0 -= b9;
tmp = b15 ^ b14;
- b15 = (tmp >> 30) | (tmp << (64 - 30));
+ b15 = ror64(tmp, 30);
b14 -= b15 + k16 + t2;
b15 -= k0 + 19;
tmp = b13 ^ b12;
- b13 = (tmp >> 44) | (tmp << (64 - 44));
+ b13 = ror64(tmp, 44);
b12 -= b13 + k14;
b13 -= k15 + t1;
tmp = b11 ^ b10;
- b11 = (tmp >> 47) | (tmp << (64 - 47));
+ b11 = ror64(tmp, 47);
b10 -= b11 + k12;
b11 -= k13;
tmp = b9 ^ b8;
- b9 = (tmp >> 12) | (tmp << (64 - 12));
+ b9 = ror64(tmp, 12);
b8 -= b9 + k10;
b9 -= k11;
tmp = b7 ^ b6;
- b7 = (tmp >> 31) | (tmp << (64 - 31));
+ b7 = ror64(tmp, 31);
b6 -= b7 + k8;
b7 -= k9;
tmp = b5 ^ b4;
- b5 = (tmp >> 37) | (tmp << (64 - 37));
+ b5 = ror64(tmp, 37);
b4 -= b5 + k6;
b5 -= k7;
tmp = b3 ^ b2;
- b3 = (tmp >> 9) | (tmp << (64 - 9));
+ b3 = ror64(tmp, 9);
b2 -= b3 + k4;
b3 -= k5;
tmp = b1 ^ b0;
- b1 = (tmp >> 41) | (tmp << (64 - 41));
+ b1 = ror64(tmp, 41);
b0 -= b1 + k2;
b1 -= k3;
tmp = b7 ^ b12;
- b7 = (tmp >> 25) | (tmp << (64 - 25));
+ b7 = ror64(tmp, 25);
b12 -= b7;
tmp = b3 ^ b10;
- b3 = (tmp >> 16) | (tmp << (64 - 16));
+ b3 = ror64(tmp, 16);
b10 -= b3;
tmp = b5 ^ b8;
- b5 = (tmp >> 28) | (tmp << (64 - 28));
+ b5 = ror64(tmp, 28);
b8 -= b5;
tmp = b1 ^ b14;
- b1 = (tmp >> 47) | (tmp << (64 - 47));
+ b1 = ror64(tmp, 47);
b14 -= b1;
tmp = b9 ^ b4;
- b9 = (tmp >> 41) | (tmp << (64 - 41));
+ b9 = ror64(tmp, 41);
b4 -= b9;
tmp = b13 ^ b6;
- b13 = (tmp >> 48) | (tmp << (64 - 48));
+ b13 = ror64(tmp, 48);
b6 -= b13;
tmp = b11 ^ b2;
- b11 = (tmp >> 20) | (tmp << (64 - 20));
+ b11 = ror64(tmp, 20);
b2 -= b11;
tmp = b15 ^ b0;
- b15 = (tmp >> 5) | (tmp << (64 - 5));
+ b15 = ror64(tmp, 5);
b0 -= b15;
tmp = b9 ^ b10;
- b9 = (tmp >> 17) | (tmp << (64 - 17));
+ b9 = ror64(tmp, 17);
b10 -= b9;
tmp = b11 ^ b8;
- b11 = (tmp >> 59) | (tmp << (64 - 59));
+ b11 = ror64(tmp, 59);
b8 -= b11;
tmp = b13 ^ b14;
- b13 = (tmp >> 41) | (tmp << (64 - 41));
+ b13 = ror64(tmp, 41);
b14 -= b13;
tmp = b15 ^ b12;
- b15 = (tmp >> 34) | (tmp << (64 - 34));
+ b15 = ror64(tmp, 34);
b12 -= b15;
tmp = b1 ^ b6;
- b1 = (tmp >> 13) | (tmp << (64 - 13));
+ b1 = ror64(tmp, 13);
b6 -= b1;
tmp = b3 ^ b4;
- b3 = (tmp >> 51) | (tmp << (64 - 51));
+ b3 = ror64(tmp, 51);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 4) | (tmp << (64 - 4));
+ b5 = ror64(tmp, 4);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 33) | (tmp << (64 - 33));
+ b7 = ror64(tmp, 33);
b0 -= b7;
tmp = b1 ^ b8;
- b1 = (tmp >> 52) | (tmp << (64 - 52));
+ b1 = ror64(tmp, 52);
b8 -= b1;
tmp = b5 ^ b14;
- b5 = (tmp >> 23) | (tmp << (64 - 23));
+ b5 = ror64(tmp, 23);
b14 -= b5;
tmp = b3 ^ b12;
- b3 = (tmp >> 18) | (tmp << (64 - 18));
+ b3 = ror64(tmp, 18);
b12 -= b3;
tmp = b7 ^ b10;
- b7 = (tmp >> 49) | (tmp << (64 - 49));
+ b7 = ror64(tmp, 49);
b10 -= b7;
tmp = b15 ^ b4;
- b15 = (tmp >> 55) | (tmp << (64 - 55));
+ b15 = ror64(tmp, 55);
b4 -= b15;
tmp = b11 ^ b6;
- b11 = (tmp >> 10) | (tmp << (64 - 10));
+ b11 = ror64(tmp, 10);
b6 -= b11;
tmp = b13 ^ b2;
- b13 = (tmp >> 19) | (tmp << (64 - 19));
+ b13 = ror64(tmp, 19);
b2 -= b13;
tmp = b9 ^ b0;
- b9 = (tmp >> 38) | (tmp << (64 - 38));
+ b9 = ror64(tmp, 38);
b0 -= b9;
tmp = b15 ^ b14;
- b15 = (tmp >> 37) | (tmp << (64 - 37));
+ b15 = ror64(tmp, 37);
b14 -= b15 + k15 + t1;
b15 -= k16 + 18;
tmp = b13 ^ b12;
- b13 = (tmp >> 22) | (tmp << (64 - 22));
+ b13 = ror64(tmp, 22);
b12 -= b13 + k13;
b13 -= k14 + t0;
tmp = b11 ^ b10;
- b11 = (tmp >> 17) | (tmp << (64 - 17));
+ b11 = ror64(tmp, 17);
b10 -= b11 + k11;
b11 -= k12;
tmp = b9 ^ b8;
- b9 = (tmp >> 8) | (tmp << (64 - 8));
+ b9 = ror64(tmp, 8);
b8 -= b9 + k9;
b9 -= k10;
tmp = b7 ^ b6;
- b7 = (tmp >> 47) | (tmp << (64 - 47));
+ b7 = ror64(tmp, 47);
b6 -= b7 + k7;
b7 -= k8;
tmp = b5 ^ b4;
- b5 = (tmp >> 8) | (tmp << (64 - 8));
+ b5 = ror64(tmp, 8);
b4 -= b5 + k5;
b5 -= k6;
tmp = b3 ^ b2;
- b3 = (tmp >> 13) | (tmp << (64 - 13));
+ b3 = ror64(tmp, 13);
b2 -= b3 + k3;
b3 -= k4;
tmp = b1 ^ b0;
- b1 = (tmp >> 24) | (tmp << (64 - 24));
+ b1 = ror64(tmp, 24);
b0 -= b1 + k1;
b1 -= k2;
tmp = b7 ^ b12;
- b7 = (tmp >> 20) | (tmp << (64 - 20));
+ b7 = ror64(tmp, 20);
b12 -= b7;
tmp = b3 ^ b10;
- b3 = (tmp >> 37) | (tmp << (64 - 37));
+ b3 = ror64(tmp, 37);
b10 -= b3;
tmp = b5 ^ b8;
- b5 = (tmp >> 31) | (tmp << (64 - 31));
+ b5 = ror64(tmp, 31);
b8 -= b5;
tmp = b1 ^ b14;
- b1 = (tmp >> 23) | (tmp << (64 - 23));
+ b1 = ror64(tmp, 23);
b14 -= b1;
tmp = b9 ^ b4;
- b9 = (tmp >> 52) | (tmp << (64 - 52));
+ b9 = ror64(tmp, 52);
b4 -= b9;
tmp = b13 ^ b6;
- b13 = (tmp >> 35) | (tmp << (64 - 35));
+ b13 = ror64(tmp, 35);
b6 -= b13;
tmp = b11 ^ b2;
- b11 = (tmp >> 48) | (tmp << (64 - 48));
+ b11 = ror64(tmp, 48);
b2 -= b11;
tmp = b15 ^ b0;
- b15 = (tmp >> 9) | (tmp << (64 - 9));
+ b15 = ror64(tmp, 9);
b0 -= b15;
tmp = b9 ^ b10;
- b9 = (tmp >> 25) | (tmp << (64 - 25));
+ b9 = ror64(tmp, 25);
b10 -= b9;
tmp = b11 ^ b8;
- b11 = (tmp >> 44) | (tmp << (64 - 44));
+ b11 = ror64(tmp, 44);
b8 -= b11;
tmp = b13 ^ b14;
- b13 = (tmp >> 42) | (tmp << (64 - 42));
+ b13 = ror64(tmp, 42);
b14 -= b13;
tmp = b15 ^ b12;
- b15 = (tmp >> 19) | (tmp << (64 - 19));
+ b15 = ror64(tmp, 19);
b12 -= b15;
tmp = b1 ^ b6;
- b1 = (tmp >> 46) | (tmp << (64 - 46));
+ b1 = ror64(tmp, 46);
b6 -= b1;
tmp = b3 ^ b4;
- b3 = (tmp >> 47) | (tmp << (64 - 47));
+ b3 = ror64(tmp, 47);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 44) | (tmp << (64 - 44));
+ b5 = ror64(tmp, 44);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 31) | (tmp << (64 - 31));
+ b7 = ror64(tmp, 31);
b0 -= b7;
tmp = b1 ^ b8;
- b1 = (tmp >> 41) | (tmp << (64 - 41));
+ b1 = ror64(tmp, 41);
b8 -= b1;
tmp = b5 ^ b14;
- b5 = (tmp >> 42) | (tmp << (64 - 42));
+ b5 = ror64(tmp, 42);
b14 -= b5;
tmp = b3 ^ b12;
- b3 = (tmp >> 53) | (tmp << (64 - 53));
+ b3 = ror64(tmp, 53);
b12 -= b3;
tmp = b7 ^ b10;
- b7 = (tmp >> 4) | (tmp << (64 - 4));
+ b7 = ror64(tmp, 4);
b10 -= b7;
tmp = b15 ^ b4;
- b15 = (tmp >> 51) | (tmp << (64 - 51));
+ b15 = ror64(tmp, 51);
b4 -= b15;
tmp = b11 ^ b6;
- b11 = (tmp >> 56) | (tmp << (64 - 56));
+ b11 = ror64(tmp, 56);
b6 -= b11;
tmp = b13 ^ b2;
- b13 = (tmp >> 34) | (tmp << (64 - 34));
+ b13 = ror64(tmp, 34);
b2 -= b13;
tmp = b9 ^ b0;
- b9 = (tmp >> 16) | (tmp << (64 - 16));
+ b9 = ror64(tmp, 16);
b0 -= b9;
tmp = b15 ^ b14;
- b15 = (tmp >> 30) | (tmp << (64 - 30));
+ b15 = ror64(tmp, 30);
b14 -= b15 + k14 + t0;
b15 -= k15 + 17;
tmp = b13 ^ b12;
- b13 = (tmp >> 44) | (tmp << (64 - 44));
+ b13 = ror64(tmp, 44);
b12 -= b13 + k12;
b13 -= k13 + t2;
tmp = b11 ^ b10;
- b11 = (tmp >> 47) | (tmp << (64 - 47));
+ b11 = ror64(tmp, 47);
b10 -= b11 + k10;
b11 -= k11;
tmp = b9 ^ b8;
- b9 = (tmp >> 12) | (tmp << (64 - 12));
+ b9 = ror64(tmp, 12);
b8 -= b9 + k8;
b9 -= k9;
tmp = b7 ^ b6;
- b7 = (tmp >> 31) | (tmp << (64 - 31));
+ b7 = ror64(tmp, 31);
b6 -= b7 + k6;
b7 -= k7;
tmp = b5 ^ b4;
- b5 = (tmp >> 37) | (tmp << (64 - 37));
+ b5 = ror64(tmp, 37);
b4 -= b5 + k4;
b5 -= k5;
tmp = b3 ^ b2;
- b3 = (tmp >> 9) | (tmp << (64 - 9));
+ b3 = ror64(tmp, 9);
b2 -= b3 + k2;
b3 -= k3;
tmp = b1 ^ b0;
- b1 = (tmp >> 41) | (tmp << (64 - 41));
+ b1 = ror64(tmp, 41);
b0 -= b1 + k0;
b1 -= k1;
tmp = b7 ^ b12;
- b7 = (tmp >> 25) | (tmp << (64 - 25));
+ b7 = ror64(tmp, 25);
b12 -= b7;
tmp = b3 ^ b10;
- b3 = (tmp >> 16) | (tmp << (64 - 16));
+ b3 = ror64(tmp, 16);
b10 -= b3;
tmp = b5 ^ b8;
- b5 = (tmp >> 28) | (tmp << (64 - 28));
+ b5 = ror64(tmp, 28);
b8 -= b5;
tmp = b1 ^ b14;
- b1 = (tmp >> 47) | (tmp << (64 - 47));
+ b1 = ror64(tmp, 47);
b14 -= b1;
tmp = b9 ^ b4;
- b9 = (tmp >> 41) | (tmp << (64 - 41));
+ b9 = ror64(tmp, 41);
b4 -= b9;
tmp = b13 ^ b6;
- b13 = (tmp >> 48) | (tmp << (64 - 48));
+ b13 = ror64(tmp, 48);
b6 -= b13;
tmp = b11 ^ b2;
- b11 = (tmp >> 20) | (tmp << (64 - 20));
+ b11 = ror64(tmp, 20);
b2 -= b11;
tmp = b15 ^ b0;
- b15 = (tmp >> 5) | (tmp << (64 - 5));
+ b15 = ror64(tmp, 5);
b0 -= b15;
tmp = b9 ^ b10;
- b9 = (tmp >> 17) | (tmp << (64 - 17));
+ b9 = ror64(tmp, 17);
b10 -= b9;
tmp = b11 ^ b8;
- b11 = (tmp >> 59) | (tmp << (64 - 59));
+ b11 = ror64(tmp, 59);
b8 -= b11;
tmp = b13 ^ b14;
- b13 = (tmp >> 41) | (tmp << (64 - 41));
+ b13 = ror64(tmp, 41);
b14 -= b13;
tmp = b15 ^ b12;
- b15 = (tmp >> 34) | (tmp << (64 - 34));
+ b15 = ror64(tmp, 34);
b12 -= b15;
tmp = b1 ^ b6;
- b1 = (tmp >> 13) | (tmp << (64 - 13));
+ b1 = ror64(tmp, 13);
b6 -= b1;
tmp = b3 ^ b4;
- b3 = (tmp >> 51) | (tmp << (64 - 51));
+ b3 = ror64(tmp, 51);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 4) | (tmp << (64 - 4));
+ b5 = ror64(tmp, 4);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 33) | (tmp << (64 - 33));
+ b7 = ror64(tmp, 33);
b0 -= b7;
tmp = b1 ^ b8;
- b1 = (tmp >> 52) | (tmp << (64 - 52));
+ b1 = ror64(tmp, 52);
b8 -= b1;
tmp = b5 ^ b14;
- b5 = (tmp >> 23) | (tmp << (64 - 23));
+ b5 = ror64(tmp, 23);
b14 -= b5;
tmp = b3 ^ b12;
- b3 = (tmp >> 18) | (tmp << (64 - 18));
+ b3 = ror64(tmp, 18);
b12 -= b3;
tmp = b7 ^ b10;
- b7 = (tmp >> 49) | (tmp << (64 - 49));
+ b7 = ror64(tmp, 49);
b10 -= b7;
tmp = b15 ^ b4;
- b15 = (tmp >> 55) | (tmp << (64 - 55));
+ b15 = ror64(tmp, 55);
b4 -= b15;
tmp = b11 ^ b6;
- b11 = (tmp >> 10) | (tmp << (64 - 10));
+ b11 = ror64(tmp, 10);
b6 -= b11;
tmp = b13 ^ b2;
- b13 = (tmp >> 19) | (tmp << (64 - 19));
+ b13 = ror64(tmp, 19);
b2 -= b13;
tmp = b9 ^ b0;
- b9 = (tmp >> 38) | (tmp << (64 - 38));
+ b9 = ror64(tmp, 38);
b0 -= b9;
tmp = b15 ^ b14;
- b15 = (tmp >> 37) | (tmp << (64 - 37));
+ b15 = ror64(tmp, 37);
b14 -= b15 + k13 + t2;
b15 -= k14 + 16;
tmp = b13 ^ b12;
- b13 = (tmp >> 22) | (tmp << (64 - 22));
+ b13 = ror64(tmp, 22);
b12 -= b13 + k11;
b13 -= k12 + t1;
tmp = b11 ^ b10;
- b11 = (tmp >> 17) | (tmp << (64 - 17));
+ b11 = ror64(tmp, 17);
b10 -= b11 + k9;
b11 -= k10;
tmp = b9 ^ b8;
- b9 = (tmp >> 8) | (tmp << (64 - 8));
+ b9 = ror64(tmp, 8);
b8 -= b9 + k7;
b9 -= k8;
tmp = b7 ^ b6;
- b7 = (tmp >> 47) | (tmp << (64 - 47));
+ b7 = ror64(tmp, 47);
b6 -= b7 + k5;
b7 -= k6;
tmp = b5 ^ b4;
- b5 = (tmp >> 8) | (tmp << (64 - 8));
+ b5 = ror64(tmp, 8);
b4 -= b5 + k3;
b5 -= k4;
tmp = b3 ^ b2;
- b3 = (tmp >> 13) | (tmp << (64 - 13));
+ b3 = ror64(tmp, 13);
b2 -= b3 + k1;
b3 -= k2;
tmp = b1 ^ b0;
- b1 = (tmp >> 24) | (tmp << (64 - 24));
+ b1 = ror64(tmp, 24);
b0 -= b1 + k16;
b1 -= k0;
tmp = b7 ^ b12;
- b7 = (tmp >> 20) | (tmp << (64 - 20));
+ b7 = ror64(tmp, 20);
b12 -= b7;
tmp = b3 ^ b10;
- b3 = (tmp >> 37) | (tmp << (64 - 37));
+ b3 = ror64(tmp, 37);
b10 -= b3;
tmp = b5 ^ b8;
- b5 = (tmp >> 31) | (tmp << (64 - 31));
+ b5 = ror64(tmp, 31);
b8 -= b5;
tmp = b1 ^ b14;
- b1 = (tmp >> 23) | (tmp << (64 - 23));
+ b1 = ror64(tmp, 23);
b14 -= b1;
tmp = b9 ^ b4;
- b9 = (tmp >> 52) | (tmp << (64 - 52));
+ b9 = ror64(tmp, 52);
b4 -= b9;
tmp = b13 ^ b6;
- b13 = (tmp >> 35) | (tmp << (64 - 35));
+ b13 = ror64(tmp, 35);
b6 -= b13;
tmp = b11 ^ b2;
- b11 = (tmp >> 48) | (tmp << (64 - 48));
+ b11 = ror64(tmp, 48);
b2 -= b11;
tmp = b15 ^ b0;
- b15 = (tmp >> 9) | (tmp << (64 - 9));
+ b15 = ror64(tmp, 9);
b0 -= b15;
tmp = b9 ^ b10;
- b9 = (tmp >> 25) | (tmp << (64 - 25));
+ b9 = ror64(tmp, 25);
b10 -= b9;
tmp = b11 ^ b8;
- b11 = (tmp >> 44) | (tmp << (64 - 44));
+ b11 = ror64(tmp, 44);
b8 -= b11;
tmp = b13 ^ b14;
- b13 = (tmp >> 42) | (tmp << (64 - 42));
+ b13 = ror64(tmp, 42);
b14 -= b13;
tmp = b15 ^ b12;
- b15 = (tmp >> 19) | (tmp << (64 - 19));
+ b15 = ror64(tmp, 19);
b12 -= b15;
tmp = b1 ^ b6;
- b1 = (tmp >> 46) | (tmp << (64 - 46));
+ b1 = ror64(tmp, 46);
b6 -= b1;
tmp = b3 ^ b4;
- b3 = (tmp >> 47) | (tmp << (64 - 47));
+ b3 = ror64(tmp, 47);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 44) | (tmp << (64 - 44));
+ b5 = ror64(tmp, 44);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 31) | (tmp << (64 - 31));
+ b7 = ror64(tmp, 31);
b0 -= b7;
tmp = b1 ^ b8;
- b1 = (tmp >> 41) | (tmp << (64 - 41));
+ b1 = ror64(tmp, 41);
b8 -= b1;
tmp = b5 ^ b14;
- b5 = (tmp >> 42) | (tmp << (64 - 42));
+ b5 = ror64(tmp, 42);
b14 -= b5;
tmp = b3 ^ b12;
- b3 = (tmp >> 53) | (tmp << (64 - 53));
+ b3 = ror64(tmp, 53);
b12 -= b3;
tmp = b7 ^ b10;
- b7 = (tmp >> 4) | (tmp << (64 - 4));
+ b7 = ror64(tmp, 4);
b10 -= b7;
tmp = b15 ^ b4;
- b15 = (tmp >> 51) | (tmp << (64 - 51));
+ b15 = ror64(tmp, 51);
b4 -= b15;
tmp = b11 ^ b6;
- b11 = (tmp >> 56) | (tmp << (64 - 56));
+ b11 = ror64(tmp, 56);
b6 -= b11;
tmp = b13 ^ b2;
- b13 = (tmp >> 34) | (tmp << (64 - 34));
+ b13 = ror64(tmp, 34);
b2 -= b13;
tmp = b9 ^ b0;
- b9 = (tmp >> 16) | (tmp << (64 - 16));
+ b9 = ror64(tmp, 16);
b0 -= b9;
tmp = b15 ^ b14;
- b15 = (tmp >> 30) | (tmp << (64 - 30));
+ b15 = ror64(tmp, 30);
b14 -= b15 + k12 + t1;
b15 -= k13 + 15;
tmp = b13 ^ b12;
- b13 = (tmp >> 44) | (tmp << (64 - 44));
+ b13 = ror64(tmp, 44);
b12 -= b13 + k10;
b13 -= k11 + t0;
tmp = b11 ^ b10;
- b11 = (tmp >> 47) | (tmp << (64 - 47));
+ b11 = ror64(tmp, 47);
b10 -= b11 + k8;
b11 -= k9;
tmp = b9 ^ b8;
- b9 = (tmp >> 12) | (tmp << (64 - 12));
+ b9 = ror64(tmp, 12);
b8 -= b9 + k6;
b9 -= k7;
tmp = b7 ^ b6;
- b7 = (tmp >> 31) | (tmp << (64 - 31));
+ b7 = ror64(tmp, 31);
b6 -= b7 + k4;
b7 -= k5;
tmp = b5 ^ b4;
- b5 = (tmp >> 37) | (tmp << (64 - 37));
+ b5 = ror64(tmp, 37);
b4 -= b5 + k2;
b5 -= k3;
tmp = b3 ^ b2;
- b3 = (tmp >> 9) | (tmp << (64 - 9));
+ b3 = ror64(tmp, 9);
b2 -= b3 + k0;
b3 -= k1;
tmp = b1 ^ b0;
- b1 = (tmp >> 41) | (tmp << (64 - 41));
+ b1 = ror64(tmp, 41);
b0 -= b1 + k15;
b1 -= k16;
tmp = b7 ^ b12;
- b7 = (tmp >> 25) | (tmp << (64 - 25));
+ b7 = ror64(tmp, 25);
b12 -= b7;
tmp = b3 ^ b10;
- b3 = (tmp >> 16) | (tmp << (64 - 16));
+ b3 = ror64(tmp, 16);
b10 -= b3;
tmp = b5 ^ b8;
- b5 = (tmp >> 28) | (tmp << (64 - 28));
+ b5 = ror64(tmp, 28);
b8 -= b5;
tmp = b1 ^ b14;
- b1 = (tmp >> 47) | (tmp << (64 - 47));
+ b1 = ror64(tmp, 47);
b14 -= b1;
tmp = b9 ^ b4;
- b9 = (tmp >> 41) | (tmp << (64 - 41));
+ b9 = ror64(tmp, 41);
b4 -= b9;
tmp = b13 ^ b6;
- b13 = (tmp >> 48) | (tmp << (64 - 48));
+ b13 = ror64(tmp, 48);
b6 -= b13;
tmp = b11 ^ b2;
- b11 = (tmp >> 20) | (tmp << (64 - 20));
+ b11 = ror64(tmp, 20);
b2 -= b11;
tmp = b15 ^ b0;
- b15 = (tmp >> 5) | (tmp << (64 - 5));
+ b15 = ror64(tmp, 5);
b0 -= b15;
tmp = b9 ^ b10;
- b9 = (tmp >> 17) | (tmp << (64 - 17));
+ b9 = ror64(tmp, 17);
b10 -= b9;
tmp = b11 ^ b8;
- b11 = (tmp >> 59) | (tmp << (64 - 59));
+ b11 = ror64(tmp, 59);
b8 -= b11;
tmp = b13 ^ b14;
- b13 = (tmp >> 41) | (tmp << (64 - 41));
+ b13 = ror64(tmp, 41);
b14 -= b13;
tmp = b15 ^ b12;
- b15 = (tmp >> 34) | (tmp << (64 - 34));
+ b15 = ror64(tmp, 34);
b12 -= b15;
tmp = b1 ^ b6;
- b1 = (tmp >> 13) | (tmp << (64 - 13));
+ b1 = ror64(tmp, 13);
b6 -= b1;
tmp = b3 ^ b4;
- b3 = (tmp >> 51) | (tmp << (64 - 51));
+ b3 = ror64(tmp, 51);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 4) | (tmp << (64 - 4));
+ b5 = ror64(tmp, 4);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 33) | (tmp << (64 - 33));
+ b7 = ror64(tmp, 33);
b0 -= b7;
tmp = b1 ^ b8;
- b1 = (tmp >> 52) | (tmp << (64 - 52));
+ b1 = ror64(tmp, 52);
b8 -= b1;
tmp = b5 ^ b14;
- b5 = (tmp >> 23) | (tmp << (64 - 23));
+ b5 = ror64(tmp, 23);
b14 -= b5;
tmp = b3 ^ b12;
- b3 = (tmp >> 18) | (tmp << (64 - 18));
+ b3 = ror64(tmp, 18);
b12 -= b3;
tmp = b7 ^ b10;
- b7 = (tmp >> 49) | (tmp << (64 - 49));
+ b7 = ror64(tmp, 49);
b10 -= b7;
tmp = b15 ^ b4;
- b15 = (tmp >> 55) | (tmp << (64 - 55));
+ b15 = ror64(tmp, 55);
b4 -= b15;
tmp = b11 ^ b6;
- b11 = (tmp >> 10) | (tmp << (64 - 10));
+ b11 = ror64(tmp, 10);
b6 -= b11;
tmp = b13 ^ b2;
- b13 = (tmp >> 19) | (tmp << (64 - 19));
+ b13 = ror64(tmp, 19);
b2 -= b13;
tmp = b9 ^ b0;
- b9 = (tmp >> 38) | (tmp << (64 - 38));
+ b9 = ror64(tmp, 38);
b0 -= b9;
tmp = b15 ^ b14;
- b15 = (tmp >> 37) | (tmp << (64 - 37));
+ b15 = ror64(tmp, 37);
b14 -= b15 + k11 + t0;
b15 -= k12 + 14;
tmp = b13 ^ b12;
- b13 = (tmp >> 22) | (tmp << (64 - 22));
+ b13 = ror64(tmp, 22);
b12 -= b13 + k9;
b13 -= k10 + t2;
tmp = b11 ^ b10;
- b11 = (tmp >> 17) | (tmp << (64 - 17));
+ b11 = ror64(tmp, 17);
b10 -= b11 + k7;
b11 -= k8;
tmp = b9 ^ b8;
- b9 = (tmp >> 8) | (tmp << (64 - 8));
+ b9 = ror64(tmp, 8);
b8 -= b9 + k5;
b9 -= k6;
tmp = b7 ^ b6;
- b7 = (tmp >> 47) | (tmp << (64 - 47));
+ b7 = ror64(tmp, 47);
b6 -= b7 + k3;
b7 -= k4;
tmp = b5 ^ b4;
- b5 = (tmp >> 8) | (tmp << (64 - 8));
+ b5 = ror64(tmp, 8);
b4 -= b5 + k1;
b5 -= k2;
tmp = b3 ^ b2;
- b3 = (tmp >> 13) | (tmp << (64 - 13));
+ b3 = ror64(tmp, 13);
b2 -= b3 + k16;
b3 -= k0;
tmp = b1 ^ b0;
- b1 = (tmp >> 24) | (tmp << (64 - 24));
+ b1 = ror64(tmp, 24);
b0 -= b1 + k14;
b1 -= k15;
tmp = b7 ^ b12;
- b7 = (tmp >> 20) | (tmp << (64 - 20));
+ b7 = ror64(tmp, 20);
b12 -= b7;
tmp = b3 ^ b10;
- b3 = (tmp >> 37) | (tmp << (64 - 37));
+ b3 = ror64(tmp, 37);
b10 -= b3;
tmp = b5 ^ b8;
- b5 = (tmp >> 31) | (tmp << (64 - 31));
+ b5 = ror64(tmp, 31);
b8 -= b5;
tmp = b1 ^ b14;
- b1 = (tmp >> 23) | (tmp << (64 - 23));
+ b1 = ror64(tmp, 23);
b14 -= b1;
tmp = b9 ^ b4;
- b9 = (tmp >> 52) | (tmp << (64 - 52));
+ b9 = ror64(tmp, 52);
b4 -= b9;
tmp = b13 ^ b6;
- b13 = (tmp >> 35) | (tmp << (64 - 35));
+ b13 = ror64(tmp, 35);
b6 -= b13;
tmp = b11 ^ b2;
- b11 = (tmp >> 48) | (tmp << (64 - 48));
+ b11 = ror64(tmp, 48);
b2 -= b11;
tmp = b15 ^ b0;
- b15 = (tmp >> 9) | (tmp << (64 - 9));
+ b15 = ror64(tmp, 9);
b0 -= b15;
tmp = b9 ^ b10;
- b9 = (tmp >> 25) | (tmp << (64 - 25));
+ b9 = ror64(tmp, 25);
b10 -= b9;
tmp = b11 ^ b8;
- b11 = (tmp >> 44) | (tmp << (64 - 44));
+ b11 = ror64(tmp, 44);
b8 -= b11;
tmp = b13 ^ b14;
- b13 = (tmp >> 42) | (tmp << (64 - 42));
+ b13 = ror64(tmp, 42);
b14 -= b13;
tmp = b15 ^ b12;
- b15 = (tmp >> 19) | (tmp << (64 - 19));
+ b15 = ror64(tmp, 19);
b12 -= b15;
tmp = b1 ^ b6;
- b1 = (tmp >> 46) | (tmp << (64 - 46));
+ b1 = ror64(tmp, 46);
b6 -= b1;
tmp = b3 ^ b4;
- b3 = (tmp >> 47) | (tmp << (64 - 47));
+ b3 = ror64(tmp, 47);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 44) | (tmp << (64 - 44));
+ b5 = ror64(tmp, 44);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 31) | (tmp << (64 - 31));
+ b7 = ror64(tmp, 31);
b0 -= b7;
tmp = b1 ^ b8;
- b1 = (tmp >> 41) | (tmp << (64 - 41));
+ b1 = ror64(tmp, 41);
b8 -= b1;
tmp = b5 ^ b14;
- b5 = (tmp >> 42) | (tmp << (64 - 42));
+ b5 = ror64(tmp, 42);
b14 -= b5;
tmp = b3 ^ b12;
- b3 = (tmp >> 53) | (tmp << (64 - 53));
+ b3 = ror64(tmp, 53);
b12 -= b3;
tmp = b7 ^ b10;
- b7 = (tmp >> 4) | (tmp << (64 - 4));
+ b7 = ror64(tmp, 4);
b10 -= b7;
tmp = b15 ^ b4;
- b15 = (tmp >> 51) | (tmp << (64 - 51));
+ b15 = ror64(tmp, 51);
b4 -= b15;
tmp = b11 ^ b6;
- b11 = (tmp >> 56) | (tmp << (64 - 56));
+ b11 = ror64(tmp, 56);
b6 -= b11;
tmp = b13 ^ b2;
- b13 = (tmp >> 34) | (tmp << (64 - 34));
+ b13 = ror64(tmp, 34);
b2 -= b13;
tmp = b9 ^ b0;
- b9 = (tmp >> 16) | (tmp << (64 - 16));
+ b9 = ror64(tmp, 16);
b0 -= b9;
tmp = b15 ^ b14;
- b15 = (tmp >> 30) | (tmp << (64 - 30));
+ b15 = ror64(tmp, 30);
b14 -= b15 + k10 + t2;
b15 -= k11 + 13;
tmp = b13 ^ b12;
- b13 = (tmp >> 44) | (tmp << (64 - 44));
+ b13 = ror64(tmp, 44);
b12 -= b13 + k8;
b13 -= k9 + t1;
tmp = b11 ^ b10;
- b11 = (tmp >> 47) | (tmp << (64 - 47));
+ b11 = ror64(tmp, 47);
b10 -= b11 + k6;
b11 -= k7;
tmp = b9 ^ b8;
- b9 = (tmp >> 12) | (tmp << (64 - 12));
+ b9 = ror64(tmp, 12);
b8 -= b9 + k4;
b9 -= k5;
tmp = b7 ^ b6;
- b7 = (tmp >> 31) | (tmp << (64 - 31));
+ b7 = ror64(tmp, 31);
b6 -= b7 + k2;
b7 -= k3;
tmp = b5 ^ b4;
- b5 = (tmp >> 37) | (tmp << (64 - 37));
+ b5 = ror64(tmp, 37);
b4 -= b5 + k0;
b5 -= k1;
tmp = b3 ^ b2;
- b3 = (tmp >> 9) | (tmp << (64 - 9));
+ b3 = ror64(tmp, 9);
b2 -= b3 + k15;
b3 -= k16;
tmp = b1 ^ b0;
- b1 = (tmp >> 41) | (tmp << (64 - 41));
+ b1 = ror64(tmp, 41);
b0 -= b1 + k13;
b1 -= k14;
tmp = b7 ^ b12;
- b7 = (tmp >> 25) | (tmp << (64 - 25));
+ b7 = ror64(tmp, 25);
b12 -= b7;
tmp = b3 ^ b10;
- b3 = (tmp >> 16) | (tmp << (64 - 16));
+ b3 = ror64(tmp, 16);
b10 -= b3;
tmp = b5 ^ b8;
- b5 = (tmp >> 28) | (tmp << (64 - 28));
+ b5 = ror64(tmp, 28);
b8 -= b5;
tmp = b1 ^ b14;
- b1 = (tmp >> 47) | (tmp << (64 - 47));
+ b1 = ror64(tmp, 47);
b14 -= b1;
tmp = b9 ^ b4;
- b9 = (tmp >> 41) | (tmp << (64 - 41));
+ b9 = ror64(tmp, 41);
b4 -= b9;
tmp = b13 ^ b6;
- b13 = (tmp >> 48) | (tmp << (64 - 48));
+ b13 = ror64(tmp, 48);
b6 -= b13;
tmp = b11 ^ b2;
- b11 = (tmp >> 20) | (tmp << (64 - 20));
+ b11 = ror64(tmp, 20);
b2 -= b11;
tmp = b15 ^ b0;
- b15 = (tmp >> 5) | (tmp << (64 - 5));
+ b15 = ror64(tmp, 5);
b0 -= b15;
tmp = b9 ^ b10;
- b9 = (tmp >> 17) | (tmp << (64 - 17));
+ b9 = ror64(tmp, 17);
b10 -= b9;
tmp = b11 ^ b8;
- b11 = (tmp >> 59) | (tmp << (64 - 59));
+ b11 = ror64(tmp, 59);
b8 -= b11;
tmp = b13 ^ b14;
- b13 = (tmp >> 41) | (tmp << (64 - 41));
+ b13 = ror64(tmp, 41);
b14 -= b13;
tmp = b15 ^ b12;
- b15 = (tmp >> 34) | (tmp << (64 - 34));
+ b15 = ror64(tmp, 34);
b12 -= b15;
tmp = b1 ^ b6;
- b1 = (tmp >> 13) | (tmp << (64 - 13));
+ b1 = ror64(tmp, 13);
b6 -= b1;
tmp = b3 ^ b4;
- b3 = (tmp >> 51) | (tmp << (64 - 51));
+ b3 = ror64(tmp, 51);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 4) | (tmp << (64 - 4));
+ b5 = ror64(tmp, 4);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 33) | (tmp << (64 - 33));
+ b7 = ror64(tmp, 33);
b0 -= b7;
tmp = b1 ^ b8;
- b1 = (tmp >> 52) | (tmp << (64 - 52));
+ b1 = ror64(tmp, 52);
b8 -= b1;
tmp = b5 ^ b14;
- b5 = (tmp >> 23) | (tmp << (64 - 23));
+ b5 = ror64(tmp, 23);
b14 -= b5;
tmp = b3 ^ b12;
- b3 = (tmp >> 18) | (tmp << (64 - 18));
+ b3 = ror64(tmp, 18);
b12 -= b3;
tmp = b7 ^ b10;
- b7 = (tmp >> 49) | (tmp << (64 - 49));
+ b7 = ror64(tmp, 49);
b10 -= b7;
tmp = b15 ^ b4;
- b15 = (tmp >> 55) | (tmp << (64 - 55));
+ b15 = ror64(tmp, 55);
b4 -= b15;
tmp = b11 ^ b6;
- b11 = (tmp >> 10) | (tmp << (64 - 10));
+ b11 = ror64(tmp, 10);
b6 -= b11;
tmp = b13 ^ b2;
- b13 = (tmp >> 19) | (tmp << (64 - 19));
+ b13 = ror64(tmp, 19);
b2 -= b13;
tmp = b9 ^ b0;
- b9 = (tmp >> 38) | (tmp << (64 - 38));
+ b9 = ror64(tmp, 38);
b0 -= b9;
tmp = b15 ^ b14;
- b15 = (tmp >> 37) | (tmp << (64 - 37));
+ b15 = ror64(tmp, 37);
b14 -= b15 + k9 + t1;
b15 -= k10 + 12;
tmp = b13 ^ b12;
- b13 = (tmp >> 22) | (tmp << (64 - 22));
+ b13 = ror64(tmp, 22);
b12 -= b13 + k7;
b13 -= k8 + t0;
tmp = b11 ^ b10;
- b11 = (tmp >> 17) | (tmp << (64 - 17));
+ b11 = ror64(tmp, 17);
b10 -= b11 + k5;
b11 -= k6;
tmp = b9 ^ b8;
- b9 = (tmp >> 8) | (tmp << (64 - 8));
+ b9 = ror64(tmp, 8);
b8 -= b9 + k3;
b9 -= k4;
tmp = b7 ^ b6;
- b7 = (tmp >> 47) | (tmp << (64 - 47));
+ b7 = ror64(tmp, 47);
b6 -= b7 + k1;
b7 -= k2;
tmp = b5 ^ b4;
- b5 = (tmp >> 8) | (tmp << (64 - 8));
+ b5 = ror64(tmp, 8);
b4 -= b5 + k16;
b5 -= k0;
tmp = b3 ^ b2;
- b3 = (tmp >> 13) | (tmp << (64 - 13));
+ b3 = ror64(tmp, 13);
b2 -= b3 + k14;
b3 -= k15;
tmp = b1 ^ b0;
- b1 = (tmp >> 24) | (tmp << (64 - 24));
+ b1 = ror64(tmp, 24);
b0 -= b1 + k12;
b1 -= k13;
tmp = b7 ^ b12;
- b7 = (tmp >> 20) | (tmp << (64 - 20));
+ b7 = ror64(tmp, 20);
b12 -= b7;
tmp = b3 ^ b10;
- b3 = (tmp >> 37) | (tmp << (64 - 37));
+ b3 = ror64(tmp, 37);
b10 -= b3;
tmp = b5 ^ b8;
- b5 = (tmp >> 31) | (tmp << (64 - 31));
+ b5 = ror64(tmp, 31);
b8 -= b5;
tmp = b1 ^ b14;
- b1 = (tmp >> 23) | (tmp << (64 - 23));
+ b1 = ror64(tmp, 23);
b14 -= b1;
tmp = b9 ^ b4;
- b9 = (tmp >> 52) | (tmp << (64 - 52));
+ b9 = ror64(tmp, 52);
b4 -= b9;
tmp = b13 ^ b6;
- b13 = (tmp >> 35) | (tmp << (64 - 35));
+ b13 = ror64(tmp, 35);
b6 -= b13;
tmp = b11 ^ b2;
- b11 = (tmp >> 48) | (tmp << (64 - 48));
+ b11 = ror64(tmp, 48);
b2 -= b11;
tmp = b15 ^ b0;
- b15 = (tmp >> 9) | (tmp << (64 - 9));
+ b15 = ror64(tmp, 9);
b0 -= b15;
tmp = b9 ^ b10;
- b9 = (tmp >> 25) | (tmp << (64 - 25));
+ b9 = ror64(tmp, 25);
b10 -= b9;
tmp = b11 ^ b8;
- b11 = (tmp >> 44) | (tmp << (64 - 44));
+ b11 = ror64(tmp, 44);
b8 -= b11;
tmp = b13 ^ b14;
- b13 = (tmp >> 42) | (tmp << (64 - 42));
+ b13 = ror64(tmp, 42);
b14 -= b13;
tmp = b15 ^ b12;
- b15 = (tmp >> 19) | (tmp << (64 - 19));
+ b15 = ror64(tmp, 19);
b12 -= b15;
tmp = b1 ^ b6;
- b1 = (tmp >> 46) | (tmp << (64 - 46));
+ b1 = ror64(tmp, 46);
b6 -= b1;
tmp = b3 ^ b4;
- b3 = (tmp >> 47) | (tmp << (64 - 47));
+ b3 = ror64(tmp, 47);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 44) | (tmp << (64 - 44));
+ b5 = ror64(tmp, 44);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 31) | (tmp << (64 - 31));
+ b7 = ror64(tmp, 31);
b0 -= b7;
tmp = b1 ^ b8;
- b1 = (tmp >> 41) | (tmp << (64 - 41));
+ b1 = ror64(tmp, 41);
b8 -= b1;
tmp = b5 ^ b14;
- b5 = (tmp >> 42) | (tmp << (64 - 42));
+ b5 = ror64(tmp, 42);
b14 -= b5;
tmp = b3 ^ b12;
- b3 = (tmp >> 53) | (tmp << (64 - 53));
+ b3 = ror64(tmp, 53);
b12 -= b3;
tmp = b7 ^ b10;
- b7 = (tmp >> 4) | (tmp << (64 - 4));
+ b7 = ror64(tmp, 4);
b10 -= b7;
tmp = b15 ^ b4;
- b15 = (tmp >> 51) | (tmp << (64 - 51));
+ b15 = ror64(tmp, 51);
b4 -= b15;
tmp = b11 ^ b6;
- b11 = (tmp >> 56) | (tmp << (64 - 56));
+ b11 = ror64(tmp, 56);
b6 -= b11;
tmp = b13 ^ b2;
- b13 = (tmp >> 34) | (tmp << (64 - 34));
+ b13 = ror64(tmp, 34);
b2 -= b13;
tmp = b9 ^ b0;
- b9 = (tmp >> 16) | (tmp << (64 - 16));
+ b9 = ror64(tmp, 16);
b0 -= b9;
tmp = b15 ^ b14;
- b15 = (tmp >> 30) | (tmp << (64 - 30));
+ b15 = ror64(tmp, 30);
b14 -= b15 + k8 + t0;
b15 -= k9 + 11;
tmp = b13 ^ b12;
- b13 = (tmp >> 44) | (tmp << (64 - 44));
+ b13 = ror64(tmp, 44);
b12 -= b13 + k6;
b13 -= k7 + t2;
tmp = b11 ^ b10;
- b11 = (tmp >> 47) | (tmp << (64 - 47));
+ b11 = ror64(tmp, 47);
b10 -= b11 + k4;
b11 -= k5;
tmp = b9 ^ b8;
- b9 = (tmp >> 12) | (tmp << (64 - 12));
+ b9 = ror64(tmp, 12);
b8 -= b9 + k2;
b9 -= k3;
tmp = b7 ^ b6;
- b7 = (tmp >> 31) | (tmp << (64 - 31));
+ b7 = ror64(tmp, 31);
b6 -= b7 + k0;
b7 -= k1;
tmp = b5 ^ b4;
- b5 = (tmp >> 37) | (tmp << (64 - 37));
+ b5 = ror64(tmp, 37);
b4 -= b5 + k15;
b5 -= k16;
tmp = b3 ^ b2;
- b3 = (tmp >> 9) | (tmp << (64 - 9));
+ b3 = ror64(tmp, 9);
b2 -= b3 + k13;
b3 -= k14;
tmp = b1 ^ b0;
- b1 = (tmp >> 41) | (tmp << (64 - 41));
+ b1 = ror64(tmp, 41);
b0 -= b1 + k11;
b1 -= k12;
tmp = b7 ^ b12;
- b7 = (tmp >> 25) | (tmp << (64 - 25));
+ b7 = ror64(tmp, 25);
b12 -= b7;
tmp = b3 ^ b10;
- b3 = (tmp >> 16) | (tmp << (64 - 16));
+ b3 = ror64(tmp, 16);
b10 -= b3;
tmp = b5 ^ b8;
- b5 = (tmp >> 28) | (tmp << (64 - 28));
+ b5 = ror64(tmp, 28);
b8 -= b5;
tmp = b1 ^ b14;
- b1 = (tmp >> 47) | (tmp << (64 - 47));
+ b1 = ror64(tmp, 47);
b14 -= b1;
tmp = b9 ^ b4;
- b9 = (tmp >> 41) | (tmp << (64 - 41));
+ b9 = ror64(tmp, 41);
b4 -= b9;
tmp = b13 ^ b6;
- b13 = (tmp >> 48) | (tmp << (64 - 48));
+ b13 = ror64(tmp, 48);
b6 -= b13;
tmp = b11 ^ b2;
- b11 = (tmp >> 20) | (tmp << (64 - 20));
+ b11 = ror64(tmp, 20);
b2 -= b11;
tmp = b15 ^ b0;
- b15 = (tmp >> 5) | (tmp << (64 - 5));
+ b15 = ror64(tmp, 5);
b0 -= b15;
tmp = b9 ^ b10;
- b9 = (tmp >> 17) | (tmp << (64 - 17));
+ b9 = ror64(tmp, 17);
b10 -= b9;
tmp = b11 ^ b8;
- b11 = (tmp >> 59) | (tmp << (64 - 59));
+ b11 = ror64(tmp, 59);
b8 -= b11;
tmp = b13 ^ b14;
- b13 = (tmp >> 41) | (tmp << (64 - 41));
+ b13 = ror64(tmp, 41);
b14 -= b13;
tmp = b15 ^ b12;
- b15 = (tmp >> 34) | (tmp << (64 - 34));
+ b15 = ror64(tmp, 34);
b12 -= b15;
tmp = b1 ^ b6;
- b1 = (tmp >> 13) | (tmp << (64 - 13));
+ b1 = ror64(tmp, 13);
b6 -= b1;
tmp = b3 ^ b4;
- b3 = (tmp >> 51) | (tmp << (64 - 51));
+ b3 = ror64(tmp, 51);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 4) | (tmp << (64 - 4));
+ b5 = ror64(tmp, 4);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 33) | (tmp << (64 - 33));
+ b7 = ror64(tmp, 33);
b0 -= b7;
tmp = b1 ^ b8;
- b1 = (tmp >> 52) | (tmp << (64 - 52));
+ b1 = ror64(tmp, 52);
b8 -= b1;
tmp = b5 ^ b14;
- b5 = (tmp >> 23) | (tmp << (64 - 23));
+ b5 = ror64(tmp, 23);
b14 -= b5;
tmp = b3 ^ b12;
- b3 = (tmp >> 18) | (tmp << (64 - 18));
+ b3 = ror64(tmp, 18);
b12 -= b3;
tmp = b7 ^ b10;
- b7 = (tmp >> 49) | (tmp << (64 - 49));
+ b7 = ror64(tmp, 49);
b10 -= b7;
tmp = b15 ^ b4;
- b15 = (tmp >> 55) | (tmp << (64 - 55));
+ b15 = ror64(tmp, 55);
b4 -= b15;
tmp = b11 ^ b6;
- b11 = (tmp >> 10) | (tmp << (64 - 10));
+ b11 = ror64(tmp, 10);
b6 -= b11;
tmp = b13 ^ b2;
- b13 = (tmp >> 19) | (tmp << (64 - 19));
+ b13 = ror64(tmp, 19);
b2 -= b13;
tmp = b9 ^ b0;
- b9 = (tmp >> 38) | (tmp << (64 - 38));
+ b9 = ror64(tmp, 38);
b0 -= b9;
tmp = b15 ^ b14;
- b15 = (tmp >> 37) | (tmp << (64 - 37));
+ b15 = ror64(tmp, 37);
b14 -= b15 + k7 + t2;
b15 -= k8 + 10;
tmp = b13 ^ b12;
- b13 = (tmp >> 22) | (tmp << (64 - 22));
+ b13 = ror64(tmp, 22);
b12 -= b13 + k5;
b13 -= k6 + t1;
tmp = b11 ^ b10;
- b11 = (tmp >> 17) | (tmp << (64 - 17));
+ b11 = ror64(tmp, 17);
b10 -= b11 + k3;
b11 -= k4;
tmp = b9 ^ b8;
- b9 = (tmp >> 8) | (tmp << (64 - 8));
+ b9 = ror64(tmp, 8);
b8 -= b9 + k1;
b9 -= k2;
tmp = b7 ^ b6;
- b7 = (tmp >> 47) | (tmp << (64 - 47));
+ b7 = ror64(tmp, 47);
b6 -= b7 + k16;
b7 -= k0;
tmp = b5 ^ b4;
- b5 = (tmp >> 8) | (tmp << (64 - 8));
+ b5 = ror64(tmp, 8);
b4 -= b5 + k14;
b5 -= k15;
tmp = b3 ^ b2;
- b3 = (tmp >> 13) | (tmp << (64 - 13));
+ b3 = ror64(tmp, 13);
b2 -= b3 + k12;
b3 -= k13;
tmp = b1 ^ b0;
- b1 = (tmp >> 24) | (tmp << (64 - 24));
+ b1 = ror64(tmp, 24);
b0 -= b1 + k10;
b1 -= k11;
tmp = b7 ^ b12;
- b7 = (tmp >> 20) | (tmp << (64 - 20));
+ b7 = ror64(tmp, 20);
b12 -= b7;
tmp = b3 ^ b10;
- b3 = (tmp >> 37) | (tmp << (64 - 37));
+ b3 = ror64(tmp, 37);
b10 -= b3;
tmp = b5 ^ b8;
- b5 = (tmp >> 31) | (tmp << (64 - 31));
+ b5 = ror64(tmp, 31);
b8 -= b5;
tmp = b1 ^ b14;
- b1 = (tmp >> 23) | (tmp << (64 - 23));
+ b1 = ror64(tmp, 23);
b14 -= b1;
tmp = b9 ^ b4;
- b9 = (tmp >> 52) | (tmp << (64 - 52));
+ b9 = ror64(tmp, 52);
b4 -= b9;
tmp = b13 ^ b6;
- b13 = (tmp >> 35) | (tmp << (64 - 35));
+ b13 = ror64(tmp, 35);
b6 -= b13;
tmp = b11 ^ b2;
- b11 = (tmp >> 48) | (tmp << (64 - 48));
+ b11 = ror64(tmp, 48);
b2 -= b11;
tmp = b15 ^ b0;
- b15 = (tmp >> 9) | (tmp << (64 - 9));
+ b15 = ror64(tmp, 9);
b0 -= b15;
tmp = b9 ^ b10;
- b9 = (tmp >> 25) | (tmp << (64 - 25));
+ b9 = ror64(tmp, 25);
b10 -= b9;
tmp = b11 ^ b8;
- b11 = (tmp >> 44) | (tmp << (64 - 44));
+ b11 = ror64(tmp, 44);
b8 -= b11;
tmp = b13 ^ b14;
- b13 = (tmp >> 42) | (tmp << (64 - 42));
+ b13 = ror64(tmp, 42);
b14 -= b13;
tmp = b15 ^ b12;
- b15 = (tmp >> 19) | (tmp << (64 - 19));
+ b15 = ror64(tmp, 19);
b12 -= b15;
tmp = b1 ^ b6;
- b1 = (tmp >> 46) | (tmp << (64 - 46));
+ b1 = ror64(tmp, 46);
b6 -= b1;
tmp = b3 ^ b4;
- b3 = (tmp >> 47) | (tmp << (64 - 47));
+ b3 = ror64(tmp, 47);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 44) | (tmp << (64 - 44));
+ b5 = ror64(tmp, 44);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 31) | (tmp << (64 - 31));
+ b7 = ror64(tmp, 31);
b0 -= b7;
tmp = b1 ^ b8;
- b1 = (tmp >> 41) | (tmp << (64 - 41));
+ b1 = ror64(tmp, 41);
b8 -= b1;
tmp = b5 ^ b14;
- b5 = (tmp >> 42) | (tmp << (64 - 42));
+ b5 = ror64(tmp, 42);
b14 -= b5;
tmp = b3 ^ b12;
- b3 = (tmp >> 53) | (tmp << (64 - 53));
+ b3 = ror64(tmp, 53);
b12 -= b3;
tmp = b7 ^ b10;
- b7 = (tmp >> 4) | (tmp << (64 - 4));
+ b7 = ror64(tmp, 4);
b10 -= b7;
tmp = b15 ^ b4;
- b15 = (tmp >> 51) | (tmp << (64 - 51));
+ b15 = ror64(tmp, 51);
b4 -= b15;
tmp = b11 ^ b6;
- b11 = (tmp >> 56) | (tmp << (64 - 56));
+ b11 = ror64(tmp, 56);
b6 -= b11;
tmp = b13 ^ b2;
- b13 = (tmp >> 34) | (tmp << (64 - 34));
+ b13 = ror64(tmp, 34);
b2 -= b13;
tmp = b9 ^ b0;
- b9 = (tmp >> 16) | (tmp << (64 - 16));
+ b9 = ror64(tmp, 16);
b0 -= b9;
tmp = b15 ^ b14;
- b15 = (tmp >> 30) | (tmp << (64 - 30));
+ b15 = ror64(tmp, 30);
b14 -= b15 + k6 + t1;
b15 -= k7 + 9;
tmp = b13 ^ b12;
- b13 = (tmp >> 44) | (tmp << (64 - 44));
+ b13 = ror64(tmp, 44);
b12 -= b13 + k4;
b13 -= k5 + t0;
tmp = b11 ^ b10;
- b11 = (tmp >> 47) | (tmp << (64 - 47));
+ b11 = ror64(tmp, 47);
b10 -= b11 + k2;
b11 -= k3;
tmp = b9 ^ b8;
- b9 = (tmp >> 12) | (tmp << (64 - 12));
+ b9 = ror64(tmp, 12);
b8 -= b9 + k0;
b9 -= k1;
tmp = b7 ^ b6;
- b7 = (tmp >> 31) | (tmp << (64 - 31));
+ b7 = ror64(tmp, 31);
b6 -= b7 + k15;
b7 -= k16;
tmp = b5 ^ b4;
- b5 = (tmp >> 37) | (tmp << (64 - 37));
+ b5 = ror64(tmp, 37);
b4 -= b5 + k13;
b5 -= k14;
tmp = b3 ^ b2;
- b3 = (tmp >> 9) | (tmp << (64 - 9));
+ b3 = ror64(tmp, 9);
b2 -= b3 + k11;
b3 -= k12;
tmp = b1 ^ b0;
- b1 = (tmp >> 41) | (tmp << (64 - 41));
+ b1 = ror64(tmp, 41);
b0 -= b1 + k9;
b1 -= k10;
tmp = b7 ^ b12;
- b7 = (tmp >> 25) | (tmp << (64 - 25));
+ b7 = ror64(tmp, 25);
b12 -= b7;
tmp = b3 ^ b10;
- b3 = (tmp >> 16) | (tmp << (64 - 16));
+ b3 = ror64(tmp, 16);
b10 -= b3;
tmp = b5 ^ b8;
- b5 = (tmp >> 28) | (tmp << (64 - 28));
+ b5 = ror64(tmp, 28);
b8 -= b5;
tmp = b1 ^ b14;
- b1 = (tmp >> 47) | (tmp << (64 - 47));
+ b1 = ror64(tmp, 47);
b14 -= b1;
tmp = b9 ^ b4;
- b9 = (tmp >> 41) | (tmp << (64 - 41));
+ b9 = ror64(tmp, 41);
b4 -= b9;
tmp = b13 ^ b6;
- b13 = (tmp >> 48) | (tmp << (64 - 48));
+ b13 = ror64(tmp, 48);
b6 -= b13;
tmp = b11 ^ b2;
- b11 = (tmp >> 20) | (tmp << (64 - 20));
+ b11 = ror64(tmp, 20);
b2 -= b11;
tmp = b15 ^ b0;
- b15 = (tmp >> 5) | (tmp << (64 - 5));
+ b15 = ror64(tmp, 5);
b0 -= b15;
tmp = b9 ^ b10;
- b9 = (tmp >> 17) | (tmp << (64 - 17));
+ b9 = ror64(tmp, 17);
b10 -= b9;
tmp = b11 ^ b8;
- b11 = (tmp >> 59) | (tmp << (64 - 59));
+ b11 = ror64(tmp, 59);
b8 -= b11;
tmp = b13 ^ b14;
- b13 = (tmp >> 41) | (tmp << (64 - 41));
+ b13 = ror64(tmp, 41);
b14 -= b13;
tmp = b15 ^ b12;
- b15 = (tmp >> 34) | (tmp << (64 - 34));
+ b15 = ror64(tmp, 34);
b12 -= b15;
tmp = b1 ^ b6;
- b1 = (tmp >> 13) | (tmp << (64 - 13));
+ b1 = ror64(tmp, 13);
b6 -= b1;
tmp = b3 ^ b4;
- b3 = (tmp >> 51) | (tmp << (64 - 51));
+ b3 = ror64(tmp, 51);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 4) | (tmp << (64 - 4));
+ b5 = ror64(tmp, 4);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 33) | (tmp << (64 - 33));
+ b7 = ror64(tmp, 33);
b0 -= b7;
tmp = b1 ^ b8;
- b1 = (tmp >> 52) | (tmp << (64 - 52));
+ b1 = ror64(tmp, 52);
b8 -= b1;
tmp = b5 ^ b14;
- b5 = (tmp >> 23) | (tmp << (64 - 23));
+ b5 = ror64(tmp, 23);
b14 -= b5;
tmp = b3 ^ b12;
- b3 = (tmp >> 18) | (tmp << (64 - 18));
+ b3 = ror64(tmp, 18);
b12 -= b3;
tmp = b7 ^ b10;
- b7 = (tmp >> 49) | (tmp << (64 - 49));
+ b7 = ror64(tmp, 49);
b10 -= b7;
tmp = b15 ^ b4;
- b15 = (tmp >> 55) | (tmp << (64 - 55));
+ b15 = ror64(tmp, 55);
b4 -= b15;
tmp = b11 ^ b6;
- b11 = (tmp >> 10) | (tmp << (64 - 10));
+ b11 = ror64(tmp, 10);
b6 -= b11;
tmp = b13 ^ b2;
- b13 = (tmp >> 19) | (tmp << (64 - 19));
+ b13 = ror64(tmp, 19);
b2 -= b13;
tmp = b9 ^ b0;
- b9 = (tmp >> 38) | (tmp << (64 - 38));
+ b9 = ror64(tmp, 38);
b0 -= b9;
tmp = b15 ^ b14;
- b15 = (tmp >> 37) | (tmp << (64 - 37));
+ b15 = ror64(tmp, 37);
b14 -= b15 + k5 + t0;
b15 -= k6 + 8;
tmp = b13 ^ b12;
- b13 = (tmp >> 22) | (tmp << (64 - 22));
+ b13 = ror64(tmp, 22);
b12 -= b13 + k3;
b13 -= k4 + t2;
tmp = b11 ^ b10;
- b11 = (tmp >> 17) | (tmp << (64 - 17));
+ b11 = ror64(tmp, 17);
b10 -= b11 + k1;
b11 -= k2;
tmp = b9 ^ b8;
- b9 = (tmp >> 8) | (tmp << (64 - 8));
+ b9 = ror64(tmp, 8);
b8 -= b9 + k16;
b9 -= k0;
tmp = b7 ^ b6;
- b7 = (tmp >> 47) | (tmp << (64 - 47));
+ b7 = ror64(tmp, 47);
b6 -= b7 + k14;
b7 -= k15;
tmp = b5 ^ b4;
- b5 = (tmp >> 8) | (tmp << (64 - 8));
+ b5 = ror64(tmp, 8);
b4 -= b5 + k12;
b5 -= k13;
tmp = b3 ^ b2;
- b3 = (tmp >> 13) | (tmp << (64 - 13));
+ b3 = ror64(tmp, 13);
b2 -= b3 + k10;
b3 -= k11;
tmp = b1 ^ b0;
- b1 = (tmp >> 24) | (tmp << (64 - 24));
+ b1 = ror64(tmp, 24);
b0 -= b1 + k8;
b1 -= k9;
tmp = b7 ^ b12;
- b7 = (tmp >> 20) | (tmp << (64 - 20));
+ b7 = ror64(tmp, 20);
b12 -= b7;
tmp = b3 ^ b10;
- b3 = (tmp >> 37) | (tmp << (64 - 37));
+ b3 = ror64(tmp, 37);
b10 -= b3;
tmp = b5 ^ b8;
- b5 = (tmp >> 31) | (tmp << (64 - 31));
+ b5 = ror64(tmp, 31);
b8 -= b5;
tmp = b1 ^ b14;
- b1 = (tmp >> 23) | (tmp << (64 - 23));
+ b1 = ror64(tmp, 23);
b14 -= b1;
tmp = b9 ^ b4;
- b9 = (tmp >> 52) | (tmp << (64 - 52));
+ b9 = ror64(tmp, 52);
b4 -= b9;
tmp = b13 ^ b6;
- b13 = (tmp >> 35) | (tmp << (64 - 35));
+ b13 = ror64(tmp, 35);
b6 -= b13;
tmp = b11 ^ b2;
- b11 = (tmp >> 48) | (tmp << (64 - 48));
+ b11 = ror64(tmp, 48);
b2 -= b11;
tmp = b15 ^ b0;
- b15 = (tmp >> 9) | (tmp << (64 - 9));
+ b15 = ror64(tmp, 9);
b0 -= b15;
tmp = b9 ^ b10;
- b9 = (tmp >> 25) | (tmp << (64 - 25));
+ b9 = ror64(tmp, 25);
b10 -= b9;
tmp = b11 ^ b8;
- b11 = (tmp >> 44) | (tmp << (64 - 44));
+ b11 = ror64(tmp, 44);
b8 -= b11;
tmp = b13 ^ b14;
- b13 = (tmp >> 42) | (tmp << (64 - 42));
+ b13 = ror64(tmp, 42);
b14 -= b13;
tmp = b15 ^ b12;
- b15 = (tmp >> 19) | (tmp << (64 - 19));
+ b15 = ror64(tmp, 19);
b12 -= b15;
tmp = b1 ^ b6;
- b1 = (tmp >> 46) | (tmp << (64 - 46));
+ b1 = ror64(tmp, 46);
b6 -= b1;
tmp = b3 ^ b4;
- b3 = (tmp >> 47) | (tmp << (64 - 47));
+ b3 = ror64(tmp, 47);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 44) | (tmp << (64 - 44));
+ b5 = ror64(tmp, 44);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 31) | (tmp << (64 - 31));
+ b7 = ror64(tmp, 31);
b0 -= b7;
tmp = b1 ^ b8;
- b1 = (tmp >> 41) | (tmp << (64 - 41));
+ b1 = ror64(tmp, 41);
b8 -= b1;
tmp = b5 ^ b14;
- b5 = (tmp >> 42) | (tmp << (64 - 42));
+ b5 = ror64(tmp, 42);
b14 -= b5;
tmp = b3 ^ b12;
- b3 = (tmp >> 53) | (tmp << (64 - 53));
+ b3 = ror64(tmp, 53);
b12 -= b3;
tmp = b7 ^ b10;
- b7 = (tmp >> 4) | (tmp << (64 - 4));
+ b7 = ror64(tmp, 4);
b10 -= b7;
tmp = b15 ^ b4;
- b15 = (tmp >> 51) | (tmp << (64 - 51));
+ b15 = ror64(tmp, 51);
b4 -= b15;
tmp = b11 ^ b6;
- b11 = (tmp >> 56) | (tmp << (64 - 56));
+ b11 = ror64(tmp, 56);
b6 -= b11;
tmp = b13 ^ b2;
- b13 = (tmp >> 34) | (tmp << (64 - 34));
+ b13 = ror64(tmp, 34);
b2 -= b13;
tmp = b9 ^ b0;
- b9 = (tmp >> 16) | (tmp << (64 - 16));
+ b9 = ror64(tmp, 16);
b0 -= b9;
tmp = b15 ^ b14;
- b15 = (tmp >> 30) | (tmp << (64 - 30));
+ b15 = ror64(tmp, 30);
b14 -= b15 + k4 + t2;
b15 -= k5 + 7;
tmp = b13 ^ b12;
- b13 = (tmp >> 44) | (tmp << (64 - 44));
+ b13 = ror64(tmp, 44);
b12 -= b13 + k2;
b13 -= k3 + t1;
tmp = b11 ^ b10;
- b11 = (tmp >> 47) | (tmp << (64 - 47));
+ b11 = ror64(tmp, 47);
b10 -= b11 + k0;
b11 -= k1;
tmp = b9 ^ b8;
- b9 = (tmp >> 12) | (tmp << (64 - 12));
+ b9 = ror64(tmp, 12);
b8 -= b9 + k15;
b9 -= k16;
tmp = b7 ^ b6;
- b7 = (tmp >> 31) | (tmp << (64 - 31));
+ b7 = ror64(tmp, 31);
b6 -= b7 + k13;
b7 -= k14;
tmp = b5 ^ b4;
- b5 = (tmp >> 37) | (tmp << (64 - 37));
+ b5 = ror64(tmp, 37);
b4 -= b5 + k11;
b5 -= k12;
tmp = b3 ^ b2;
- b3 = (tmp >> 9) | (tmp << (64 - 9));
+ b3 = ror64(tmp, 9);
b2 -= b3 + k9;
b3 -= k10;
tmp = b1 ^ b0;
- b1 = (tmp >> 41) | (tmp << (64 - 41));
+ b1 = ror64(tmp, 41);
b0 -= b1 + k7;
b1 -= k8;
tmp = b7 ^ b12;
- b7 = (tmp >> 25) | (tmp << (64 - 25));
+ b7 = ror64(tmp, 25);
b12 -= b7;
tmp = b3 ^ b10;
- b3 = (tmp >> 16) | (tmp << (64 - 16));
+ b3 = ror64(tmp, 16);
b10 -= b3;
tmp = b5 ^ b8;
- b5 = (tmp >> 28) | (tmp << (64 - 28));
+ b5 = ror64(tmp, 28);
b8 -= b5;
tmp = b1 ^ b14;
- b1 = (tmp >> 47) | (tmp << (64 - 47));
+ b1 = ror64(tmp, 47);
b14 -= b1;
tmp = b9 ^ b4;
- b9 = (tmp >> 41) | (tmp << (64 - 41));
+ b9 = ror64(tmp, 41);
b4 -= b9;
tmp = b13 ^ b6;
- b13 = (tmp >> 48) | (tmp << (64 - 48));
+ b13 = ror64(tmp, 48);
b6 -= b13;
tmp = b11 ^ b2;
- b11 = (tmp >> 20) | (tmp << (64 - 20));
+ b11 = ror64(tmp, 20);
b2 -= b11;
tmp = b15 ^ b0;
- b15 = (tmp >> 5) | (tmp << (64 - 5));
+ b15 = ror64(tmp, 5);
b0 -= b15;
tmp = b9 ^ b10;
- b9 = (tmp >> 17) | (tmp << (64 - 17));
+ b9 = ror64(tmp, 17);
b10 -= b9;
tmp = b11 ^ b8;
- b11 = (tmp >> 59) | (tmp << (64 - 59));
+ b11 = ror64(tmp, 59);
b8 -= b11;
tmp = b13 ^ b14;
- b13 = (tmp >> 41) | (tmp << (64 - 41));
+ b13 = ror64(tmp, 41);
b14 -= b13;
tmp = b15 ^ b12;
- b15 = (tmp >> 34) | (tmp << (64 - 34));
+ b15 = ror64(tmp, 34);
b12 -= b15;
tmp = b1 ^ b6;
- b1 = (tmp >> 13) | (tmp << (64 - 13));
+ b1 = ror64(tmp, 13);
b6 -= b1;
tmp = b3 ^ b4;
- b3 = (tmp >> 51) | (tmp << (64 - 51));
+ b3 = ror64(tmp, 51);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 4) | (tmp << (64 - 4));
+ b5 = ror64(tmp, 4);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 33) | (tmp << (64 - 33));
+ b7 = ror64(tmp, 33);
b0 -= b7;
tmp = b1 ^ b8;
- b1 = (tmp >> 52) | (tmp << (64 - 52));
+ b1 = ror64(tmp, 52);
b8 -= b1;
tmp = b5 ^ b14;
- b5 = (tmp >> 23) | (tmp << (64 - 23));
+ b5 = ror64(tmp, 23);
b14 -= b5;
tmp = b3 ^ b12;
- b3 = (tmp >> 18) | (tmp << (64 - 18));
+ b3 = ror64(tmp, 18);
b12 -= b3;
tmp = b7 ^ b10;
- b7 = (tmp >> 49) | (tmp << (64 - 49));
+ b7 = ror64(tmp, 49);
b10 -= b7;
tmp = b15 ^ b4;
- b15 = (tmp >> 55) | (tmp << (64 - 55));
+ b15 = ror64(tmp, 55);
b4 -= b15;
tmp = b11 ^ b6;
- b11 = (tmp >> 10) | (tmp << (64 - 10));
+ b11 = ror64(tmp, 10);
b6 -= b11;
tmp = b13 ^ b2;
- b13 = (tmp >> 19) | (tmp << (64 - 19));
+ b13 = ror64(tmp, 19);
b2 -= b13;
tmp = b9 ^ b0;
- b9 = (tmp >> 38) | (tmp << (64 - 38));
+ b9 = ror64(tmp, 38);
b0 -= b9;
tmp = b15 ^ b14;
- b15 = (tmp >> 37) | (tmp << (64 - 37));
+ b15 = ror64(tmp, 37);
b14 -= b15 + k3 + t1;
b15 -= k4 + 6;
tmp = b13 ^ b12;
- b13 = (tmp >> 22) | (tmp << (64 - 22));
+ b13 = ror64(tmp, 22);
b12 -= b13 + k1;
b13 -= k2 + t0;
tmp = b11 ^ b10;
- b11 = (tmp >> 17) | (tmp << (64 - 17));
+ b11 = ror64(tmp, 17);
b10 -= b11 + k16;
b11 -= k0;
tmp = b9 ^ b8;
- b9 = (tmp >> 8) | (tmp << (64 - 8));
+ b9 = ror64(tmp, 8);
b8 -= b9 + k14;
b9 -= k15;
tmp = b7 ^ b6;
- b7 = (tmp >> 47) | (tmp << (64 - 47));
+ b7 = ror64(tmp, 47);
b6 -= b7 + k12;
b7 -= k13;
tmp = b5 ^ b4;
- b5 = (tmp >> 8) | (tmp << (64 - 8));
+ b5 = ror64(tmp, 8);
b4 -= b5 + k10;
b5 -= k11;
tmp = b3 ^ b2;
- b3 = (tmp >> 13) | (tmp << (64 - 13));
+ b3 = ror64(tmp, 13);
b2 -= b3 + k8;
b3 -= k9;
tmp = b1 ^ b0;
- b1 = (tmp >> 24) | (tmp << (64 - 24));
+ b1 = ror64(tmp, 24);
b0 -= b1 + k6;
b1 -= k7;
tmp = b7 ^ b12;
- b7 = (tmp >> 20) | (tmp << (64 - 20));
+ b7 = ror64(tmp, 20);
b12 -= b7;
tmp = b3 ^ b10;
- b3 = (tmp >> 37) | (tmp << (64 - 37));
+ b3 = ror64(tmp, 37);
b10 -= b3;
tmp = b5 ^ b8;
- b5 = (tmp >> 31) | (tmp << (64 - 31));
+ b5 = ror64(tmp, 31);
b8 -= b5;
tmp = b1 ^ b14;
- b1 = (tmp >> 23) | (tmp << (64 - 23));
+ b1 = ror64(tmp, 23);
b14 -= b1;
tmp = b9 ^ b4;
- b9 = (tmp >> 52) | (tmp << (64 - 52));
+ b9 = ror64(tmp, 52);
b4 -= b9;
tmp = b13 ^ b6;
- b13 = (tmp >> 35) | (tmp << (64 - 35));
+ b13 = ror64(tmp, 35);
b6 -= b13;
tmp = b11 ^ b2;
- b11 = (tmp >> 48) | (tmp << (64 - 48));
+ b11 = ror64(tmp, 48);
b2 -= b11;
tmp = b15 ^ b0;
- b15 = (tmp >> 9) | (tmp << (64 - 9));
+ b15 = ror64(tmp, 9);
b0 -= b15;
tmp = b9 ^ b10;
- b9 = (tmp >> 25) | (tmp << (64 - 25));
+ b9 = ror64(tmp, 25);
b10 -= b9;
tmp = b11 ^ b8;
- b11 = (tmp >> 44) | (tmp << (64 - 44));
+ b11 = ror64(tmp, 44);
b8 -= b11;
tmp = b13 ^ b14;
- b13 = (tmp >> 42) | (tmp << (64 - 42));
+ b13 = ror64(tmp, 42);
b14 -= b13;
tmp = b15 ^ b12;
- b15 = (tmp >> 19) | (tmp << (64 - 19));
+ b15 = ror64(tmp, 19);
b12 -= b15;
tmp = b1 ^ b6;
- b1 = (tmp >> 46) | (tmp << (64 - 46));
+ b1 = ror64(tmp, 46);
b6 -= b1;
tmp = b3 ^ b4;
- b3 = (tmp >> 47) | (tmp << (64 - 47));
+ b3 = ror64(tmp, 47);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 44) | (tmp << (64 - 44));
+ b5 = ror64(tmp, 44);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 31) | (tmp << (64 - 31));
+ b7 = ror64(tmp, 31);
b0 -= b7;
tmp = b1 ^ b8;
- b1 = (tmp >> 41) | (tmp << (64 - 41));
+ b1 = ror64(tmp, 41);
b8 -= b1;
tmp = b5 ^ b14;
- b5 = (tmp >> 42) | (tmp << (64 - 42));
+ b5 = ror64(tmp, 42);
b14 -= b5;
tmp = b3 ^ b12;
- b3 = (tmp >> 53) | (tmp << (64 - 53));
+ b3 = ror64(tmp, 53);
b12 -= b3;
tmp = b7 ^ b10;
- b7 = (tmp >> 4) | (tmp << (64 - 4));
+ b7 = ror64(tmp, 4);
b10 -= b7;
tmp = b15 ^ b4;
- b15 = (tmp >> 51) | (tmp << (64 - 51));
+ b15 = ror64(tmp, 51);
b4 -= b15;
tmp = b11 ^ b6;
- b11 = (tmp >> 56) | (tmp << (64 - 56));
+ b11 = ror64(tmp, 56);
b6 -= b11;
tmp = b13 ^ b2;
- b13 = (tmp >> 34) | (tmp << (64 - 34));
+ b13 = ror64(tmp, 34);
b2 -= b13;
tmp = b9 ^ b0;
- b9 = (tmp >> 16) | (tmp << (64 - 16));
+ b9 = ror64(tmp, 16);
b0 -= b9;
tmp = b15 ^ b14;
- b15 = (tmp >> 30) | (tmp << (64 - 30));
+ b15 = ror64(tmp, 30);
b14 -= b15 + k2 + t0;
b15 -= k3 + 5;
tmp = b13 ^ b12;
- b13 = (tmp >> 44) | (tmp << (64 - 44));
+ b13 = ror64(tmp, 44);
b12 -= b13 + k0;
b13 -= k1 + t2;
tmp = b11 ^ b10;
- b11 = (tmp >> 47) | (tmp << (64 - 47));
+ b11 = ror64(tmp, 47);
b10 -= b11 + k15;
b11 -= k16;
tmp = b9 ^ b8;
- b9 = (tmp >> 12) | (tmp << (64 - 12));
+ b9 = ror64(tmp, 12);
b8 -= b9 + k13;
b9 -= k14;
tmp = b7 ^ b6;
- b7 = (tmp >> 31) | (tmp << (64 - 31));
+ b7 = ror64(tmp, 31);
b6 -= b7 + k11;
b7 -= k12;
tmp = b5 ^ b4;
- b5 = (tmp >> 37) | (tmp << (64 - 37));
+ b5 = ror64(tmp, 37);
b4 -= b5 + k9;
b5 -= k10;
tmp = b3 ^ b2;
- b3 = (tmp >> 9) | (tmp << (64 - 9));
+ b3 = ror64(tmp, 9);
b2 -= b3 + k7;
b3 -= k8;
tmp = b1 ^ b0;
- b1 = (tmp >> 41) | (tmp << (64 - 41));
+ b1 = ror64(tmp, 41);
b0 -= b1 + k5;
b1 -= k6;
tmp = b7 ^ b12;
- b7 = (tmp >> 25) | (tmp << (64 - 25));
+ b7 = ror64(tmp, 25);
b12 -= b7;
tmp = b3 ^ b10;
- b3 = (tmp >> 16) | (tmp << (64 - 16));
+ b3 = ror64(tmp, 16);
b10 -= b3;
tmp = b5 ^ b8;
- b5 = (tmp >> 28) | (tmp << (64 - 28));
+ b5 = ror64(tmp, 28);
b8 -= b5;
tmp = b1 ^ b14;
- b1 = (tmp >> 47) | (tmp << (64 - 47));
+ b1 = ror64(tmp, 47);
b14 -= b1;
tmp = b9 ^ b4;
- b9 = (tmp >> 41) | (tmp << (64 - 41));
+ b9 = ror64(tmp, 41);
b4 -= b9;
tmp = b13 ^ b6;
- b13 = (tmp >> 48) | (tmp << (64 - 48));
+ b13 = ror64(tmp, 48);
b6 -= b13;
tmp = b11 ^ b2;
- b11 = (tmp >> 20) | (tmp << (64 - 20));
+ b11 = ror64(tmp, 20);
b2 -= b11;
tmp = b15 ^ b0;
- b15 = (tmp >> 5) | (tmp << (64 - 5));
+ b15 = ror64(tmp, 5);
b0 -= b15;
tmp = b9 ^ b10;
- b9 = (tmp >> 17) | (tmp << (64 - 17));
+ b9 = ror64(tmp, 17);
b10 -= b9;
tmp = b11 ^ b8;
- b11 = (tmp >> 59) | (tmp << (64 - 59));
+ b11 = ror64(tmp, 59);
b8 -= b11;
tmp = b13 ^ b14;
- b13 = (tmp >> 41) | (tmp << (64 - 41));
+ b13 = ror64(tmp, 41);
b14 -= b13;
tmp = b15 ^ b12;
- b15 = (tmp >> 34) | (tmp << (64 - 34));
+ b15 = ror64(tmp, 34);
b12 -= b15;
tmp = b1 ^ b6;
- b1 = (tmp >> 13) | (tmp << (64 - 13));
+ b1 = ror64(tmp, 13);
b6 -= b1;
tmp = b3 ^ b4;
- b3 = (tmp >> 51) | (tmp << (64 - 51));
+ b3 = ror64(tmp, 51);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 4) | (tmp << (64 - 4));
+ b5 = ror64(tmp, 4);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 33) | (tmp << (64 - 33));
+ b7 = ror64(tmp, 33);
b0 -= b7;
tmp = b1 ^ b8;
- b1 = (tmp >> 52) | (tmp << (64 - 52));
+ b1 = ror64(tmp, 52);
b8 -= b1;
tmp = b5 ^ b14;
- b5 = (tmp >> 23) | (tmp << (64 - 23));
+ b5 = ror64(tmp, 23);
b14 -= b5;
tmp = b3 ^ b12;
- b3 = (tmp >> 18) | (tmp << (64 - 18));
+ b3 = ror64(tmp, 18);
b12 -= b3;
tmp = b7 ^ b10;
- b7 = (tmp >> 49) | (tmp << (64 - 49));
+ b7 = ror64(tmp, 49);
b10 -= b7;
tmp = b15 ^ b4;
- b15 = (tmp >> 55) | (tmp << (64 - 55));
+ b15 = ror64(tmp, 55);
b4 -= b15;
tmp = b11 ^ b6;
- b11 = (tmp >> 10) | (tmp << (64 - 10));
+ b11 = ror64(tmp, 10);
b6 -= b11;
tmp = b13 ^ b2;
- b13 = (tmp >> 19) | (tmp << (64 - 19));
+ b13 = ror64(tmp, 19);
b2 -= b13;
tmp = b9 ^ b0;
- b9 = (tmp >> 38) | (tmp << (64 - 38));
+ b9 = ror64(tmp, 38);
b0 -= b9;
tmp = b15 ^ b14;
- b15 = (tmp >> 37) | (tmp << (64 - 37));
+ b15 = ror64(tmp, 37);
b14 -= b15 + k1 + t2;
b15 -= k2 + 4;
tmp = b13 ^ b12;
- b13 = (tmp >> 22) | (tmp << (64 - 22));
+ b13 = ror64(tmp, 22);
b12 -= b13 + k16;
b13 -= k0 + t1;
tmp = b11 ^ b10;
- b11 = (tmp >> 17) | (tmp << (64 - 17));
+ b11 = ror64(tmp, 17);
b10 -= b11 + k14;
b11 -= k15;
tmp = b9 ^ b8;
- b9 = (tmp >> 8) | (tmp << (64 - 8));
+ b9 = ror64(tmp, 8);
b8 -= b9 + k12;
b9 -= k13;
tmp = b7 ^ b6;
- b7 = (tmp >> 47) | (tmp << (64 - 47));
+ b7 = ror64(tmp, 47);
b6 -= b7 + k10;
b7 -= k11;
tmp = b5 ^ b4;
- b5 = (tmp >> 8) | (tmp << (64 - 8));
+ b5 = ror64(tmp, 8);
b4 -= b5 + k8;
b5 -= k9;
tmp = b3 ^ b2;
- b3 = (tmp >> 13) | (tmp << (64 - 13));
+ b3 = ror64(tmp, 13);
b2 -= b3 + k6;
b3 -= k7;
tmp = b1 ^ b0;
- b1 = (tmp >> 24) | (tmp << (64 - 24));
+ b1 = ror64(tmp, 24);
b0 -= b1 + k4;
b1 -= k5;
tmp = b7 ^ b12;
- b7 = (tmp >> 20) | (tmp << (64 - 20));
+ b7 = ror64(tmp, 20);
b12 -= b7;
tmp = b3 ^ b10;
- b3 = (tmp >> 37) | (tmp << (64 - 37));
+ b3 = ror64(tmp, 37);
b10 -= b3;
tmp = b5 ^ b8;
- b5 = (tmp >> 31) | (tmp << (64 - 31));
+ b5 = ror64(tmp, 31);
b8 -= b5;
tmp = b1 ^ b14;
- b1 = (tmp >> 23) | (tmp << (64 - 23));
+ b1 = ror64(tmp, 23);
b14 -= b1;
tmp = b9 ^ b4;
- b9 = (tmp >> 52) | (tmp << (64 - 52));
+ b9 = ror64(tmp, 52);
b4 -= b9;
tmp = b13 ^ b6;
- b13 = (tmp >> 35) | (tmp << (64 - 35));
+ b13 = ror64(tmp, 35);
b6 -= b13;
tmp = b11 ^ b2;
- b11 = (tmp >> 48) | (tmp << (64 - 48));
+ b11 = ror64(tmp, 48);
b2 -= b11;
tmp = b15 ^ b0;
- b15 = (tmp >> 9) | (tmp << (64 - 9));
+ b15 = ror64(tmp, 9);
b0 -= b15;
tmp = b9 ^ b10;
- b9 = (tmp >> 25) | (tmp << (64 - 25));
+ b9 = ror64(tmp, 25);
b10 -= b9;
tmp = b11 ^ b8;
- b11 = (tmp >> 44) | (tmp << (64 - 44));
+ b11 = ror64(tmp, 44);
b8 -= b11;
tmp = b13 ^ b14;
- b13 = (tmp >> 42) | (tmp << (64 - 42));
+ b13 = ror64(tmp, 42);
b14 -= b13;
tmp = b15 ^ b12;
- b15 = (tmp >> 19) | (tmp << (64 - 19));
+ b15 = ror64(tmp, 19);
b12 -= b15;
tmp = b1 ^ b6;
- b1 = (tmp >> 46) | (tmp << (64 - 46));
+ b1 = ror64(tmp, 46);
b6 -= b1;
tmp = b3 ^ b4;
- b3 = (tmp >> 47) | (tmp << (64 - 47));
+ b3 = ror64(tmp, 47);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 44) | (tmp << (64 - 44));
+ b5 = ror64(tmp, 44);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 31) | (tmp << (64 - 31));
+ b7 = ror64(tmp, 31);
b0 -= b7;
tmp = b1 ^ b8;
- b1 = (tmp >> 41) | (tmp << (64 - 41));
+ b1 = ror64(tmp, 41);
b8 -= b1;
tmp = b5 ^ b14;
- b5 = (tmp >> 42) | (tmp << (64 - 42));
+ b5 = ror64(tmp, 42);
b14 -= b5;
tmp = b3 ^ b12;
- b3 = (tmp >> 53) | (tmp << (64 - 53));
+ b3 = ror64(tmp, 53);
b12 -= b3;
tmp = b7 ^ b10;
- b7 = (tmp >> 4) | (tmp << (64 - 4));
+ b7 = ror64(tmp, 4);
b10 -= b7;
tmp = b15 ^ b4;
- b15 = (tmp >> 51) | (tmp << (64 - 51));
+ b15 = ror64(tmp, 51);
b4 -= b15;
tmp = b11 ^ b6;
- b11 = (tmp >> 56) | (tmp << (64 - 56));
+ b11 = ror64(tmp, 56);
b6 -= b11;
tmp = b13 ^ b2;
- b13 = (tmp >> 34) | (tmp << (64 - 34));
+ b13 = ror64(tmp, 34);
b2 -= b13;
tmp = b9 ^ b0;
- b9 = (tmp >> 16) | (tmp << (64 - 16));
+ b9 = ror64(tmp, 16);
b0 -= b9;
tmp = b15 ^ b14;
- b15 = (tmp >> 30) | (tmp << (64 - 30));
+ b15 = ror64(tmp, 30);
b14 -= b15 + k0 + t1;
b15 -= k1 + 3;
tmp = b13 ^ b12;
- b13 = (tmp >> 44) | (tmp << (64 - 44));
+ b13 = ror64(tmp, 44);
b12 -= b13 + k15;
b13 -= k16 + t0;
tmp = b11 ^ b10;
- b11 = (tmp >> 47) | (tmp << (64 - 47));
+ b11 = ror64(tmp, 47);
b10 -= b11 + k13;
b11 -= k14;
tmp = b9 ^ b8;
- b9 = (tmp >> 12) | (tmp << (64 - 12));
+ b9 = ror64(tmp, 12);
b8 -= b9 + k11;
b9 -= k12;
tmp = b7 ^ b6;
- b7 = (tmp >> 31) | (tmp << (64 - 31));
+ b7 = ror64(tmp, 31);
b6 -= b7 + k9;
b7 -= k10;
tmp = b5 ^ b4;
- b5 = (tmp >> 37) | (tmp << (64 - 37));
+ b5 = ror64(tmp, 37);
b4 -= b5 + k7;
b5 -= k8;
tmp = b3 ^ b2;
- b3 = (tmp >> 9) | (tmp << (64 - 9));
+ b3 = ror64(tmp, 9);
b2 -= b3 + k5;
b3 -= k6;
tmp = b1 ^ b0;
- b1 = (tmp >> 41) | (tmp << (64 - 41));
+ b1 = ror64(tmp, 41);
b0 -= b1 + k3;
b1 -= k4;
tmp = b7 ^ b12;
- b7 = (tmp >> 25) | (tmp << (64 - 25));
+ b7 = ror64(tmp, 25);
b12 -= b7;
tmp = b3 ^ b10;
- b3 = (tmp >> 16) | (tmp << (64 - 16));
+ b3 = ror64(tmp, 16);
b10 -= b3;
tmp = b5 ^ b8;
- b5 = (tmp >> 28) | (tmp << (64 - 28));
+ b5 = ror64(tmp, 28);
b8 -= b5;
tmp = b1 ^ b14;
- b1 = (tmp >> 47) | (tmp << (64 - 47));
+ b1 = ror64(tmp, 47);
b14 -= b1;
tmp = b9 ^ b4;
- b9 = (tmp >> 41) | (tmp << (64 - 41));
+ b9 = ror64(tmp, 41);
b4 -= b9;
tmp = b13 ^ b6;
- b13 = (tmp >> 48) | (tmp << (64 - 48));
+ b13 = ror64(tmp, 48);
b6 -= b13;
tmp = b11 ^ b2;
- b11 = (tmp >> 20) | (tmp << (64 - 20));
+ b11 = ror64(tmp, 20);
b2 -= b11;
tmp = b15 ^ b0;
- b15 = (tmp >> 5) | (tmp << (64 - 5));
+ b15 = ror64(tmp, 5);
b0 -= b15;
tmp = b9 ^ b10;
- b9 = (tmp >> 17) | (tmp << (64 - 17));
+ b9 = ror64(tmp, 17);
b10 -= b9;
tmp = b11 ^ b8;
- b11 = (tmp >> 59) | (tmp << (64 - 59));
+ b11 = ror64(tmp, 59);
b8 -= b11;
tmp = b13 ^ b14;
- b13 = (tmp >> 41) | (tmp << (64 - 41));
+ b13 = ror64(tmp, 41);
b14 -= b13;
tmp = b15 ^ b12;
- b15 = (tmp >> 34) | (tmp << (64 - 34));
+ b15 = ror64(tmp, 34);
b12 -= b15;
tmp = b1 ^ b6;
- b1 = (tmp >> 13) | (tmp << (64 - 13));
+ b1 = ror64(tmp, 13);
b6 -= b1;
tmp = b3 ^ b4;
- b3 = (tmp >> 51) | (tmp << (64 - 51));
+ b3 = ror64(tmp, 51);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 4) | (tmp << (64 - 4));
+ b5 = ror64(tmp, 4);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 33) | (tmp << (64 - 33));
+ b7 = ror64(tmp, 33);
b0 -= b7;
tmp = b1 ^ b8;
- b1 = (tmp >> 52) | (tmp << (64 - 52));
+ b1 = ror64(tmp, 52);
b8 -= b1;
tmp = b5 ^ b14;
- b5 = (tmp >> 23) | (tmp << (64 - 23));
+ b5 = ror64(tmp, 23);
b14 -= b5;
tmp = b3 ^ b12;
- b3 = (tmp >> 18) | (tmp << (64 - 18));
+ b3 = ror64(tmp, 18);
b12 -= b3;
tmp = b7 ^ b10;
- b7 = (tmp >> 49) | (tmp << (64 - 49));
+ b7 = ror64(tmp, 49);
b10 -= b7;
tmp = b15 ^ b4;
- b15 = (tmp >> 55) | (tmp << (64 - 55));
+ b15 = ror64(tmp, 55);
b4 -= b15;
tmp = b11 ^ b6;
- b11 = (tmp >> 10) | (tmp << (64 - 10));
+ b11 = ror64(tmp, 10);
b6 -= b11;
tmp = b13 ^ b2;
- b13 = (tmp >> 19) | (tmp << (64 - 19));
+ b13 = ror64(tmp, 19);
b2 -= b13;
tmp = b9 ^ b0;
- b9 = (tmp >> 38) | (tmp << (64 - 38));
+ b9 = ror64(tmp, 38);
b0 -= b9;
tmp = b15 ^ b14;
- b15 = (tmp >> 37) | (tmp << (64 - 37));
+ b15 = ror64(tmp, 37);
b14 -= b15 + k16 + t0;
b15 -= k0 + 2;
tmp = b13 ^ b12;
- b13 = (tmp >> 22) | (tmp << (64 - 22));
+ b13 = ror64(tmp, 22);
b12 -= b13 + k14;
b13 -= k15 + t2;
tmp = b11 ^ b10;
- b11 = (tmp >> 17) | (tmp << (64 - 17));
+ b11 = ror64(tmp, 17);
b10 -= b11 + k12;
b11 -= k13;
tmp = b9 ^ b8;
- b9 = (tmp >> 8) | (tmp << (64 - 8));
+ b9 = ror64(tmp, 8);
b8 -= b9 + k10;
b9 -= k11;
tmp = b7 ^ b6;
- b7 = (tmp >> 47) | (tmp << (64 - 47));
+ b7 = ror64(tmp, 47);
b6 -= b7 + k8;
b7 -= k9;
tmp = b5 ^ b4;
- b5 = (tmp >> 8) | (tmp << (64 - 8));
+ b5 = ror64(tmp, 8);
b4 -= b5 + k6;
b5 -= k7;
tmp = b3 ^ b2;
- b3 = (tmp >> 13) | (tmp << (64 - 13));
+ b3 = ror64(tmp, 13);
b2 -= b3 + k4;
b3 -= k5;
tmp = b1 ^ b0;
- b1 = (tmp >> 24) | (tmp << (64 - 24));
+ b1 = ror64(tmp, 24);
b0 -= b1 + k2;
b1 -= k3;
tmp = b7 ^ b12;
- b7 = (tmp >> 20) | (tmp << (64 - 20));
+ b7 = ror64(tmp, 20);
b12 -= b7;
tmp = b3 ^ b10;
- b3 = (tmp >> 37) | (tmp << (64 - 37));
+ b3 = ror64(tmp, 37);
b10 -= b3;
tmp = b5 ^ b8;
- b5 = (tmp >> 31) | (tmp << (64 - 31));
+ b5 = ror64(tmp, 31);
b8 -= b5;
tmp = b1 ^ b14;
- b1 = (tmp >> 23) | (tmp << (64 - 23));
+ b1 = ror64(tmp, 23);
b14 -= b1;
tmp = b9 ^ b4;
- b9 = (tmp >> 52) | (tmp << (64 - 52));
+ b9 = ror64(tmp, 52);
b4 -= b9;
tmp = b13 ^ b6;
- b13 = (tmp >> 35) | (tmp << (64 - 35));
+ b13 = ror64(tmp, 35);
b6 -= b13;
tmp = b11 ^ b2;
- b11 = (tmp >> 48) | (tmp << (64 - 48));
+ b11 = ror64(tmp, 48);
b2 -= b11;
tmp = b15 ^ b0;
- b15 = (tmp >> 9) | (tmp << (64 - 9));
+ b15 = ror64(tmp, 9);
b0 -= b15;
tmp = b9 ^ b10;
- b9 = (tmp >> 25) | (tmp << (64 - 25));
+ b9 = ror64(tmp, 25);
b10 -= b9;
tmp = b11 ^ b8;
- b11 = (tmp >> 44) | (tmp << (64 - 44));
+ b11 = ror64(tmp, 44);
b8 -= b11;
tmp = b13 ^ b14;
- b13 = (tmp >> 42) | (tmp << (64 - 42));
+ b13 = ror64(tmp, 42);
b14 -= b13;
tmp = b15 ^ b12;
- b15 = (tmp >> 19) | (tmp << (64 - 19));
+ b15 = ror64(tmp, 19);
b12 -= b15;
tmp = b1 ^ b6;
- b1 = (tmp >> 46) | (tmp << (64 - 46));
+ b1 = ror64(tmp, 46);
b6 -= b1;
tmp = b3 ^ b4;
- b3 = (tmp >> 47) | (tmp << (64 - 47));
+ b3 = ror64(tmp, 47);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 44) | (tmp << (64 - 44));
+ b5 = ror64(tmp, 44);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 31) | (tmp << (64 - 31));
+ b7 = ror64(tmp, 31);
b0 -= b7;
tmp = b1 ^ b8;
- b1 = (tmp >> 41) | (tmp << (64 - 41));
+ b1 = ror64(tmp, 41);
b8 -= b1;
tmp = b5 ^ b14;
- b5 = (tmp >> 42) | (tmp << (64 - 42));
+ b5 = ror64(tmp, 42);
b14 -= b5;
tmp = b3 ^ b12;
- b3 = (tmp >> 53) | (tmp << (64 - 53));
+ b3 = ror64(tmp, 53);
b12 -= b3;
tmp = b7 ^ b10;
- b7 = (tmp >> 4) | (tmp << (64 - 4));
+ b7 = ror64(tmp, 4);
b10 -= b7;
tmp = b15 ^ b4;
- b15 = (tmp >> 51) | (tmp << (64 - 51));
+ b15 = ror64(tmp, 51);
b4 -= b15;
tmp = b11 ^ b6;
- b11 = (tmp >> 56) | (tmp << (64 - 56));
+ b11 = ror64(tmp, 56);
b6 -= b11;
tmp = b13 ^ b2;
- b13 = (tmp >> 34) | (tmp << (64 - 34));
+ b13 = ror64(tmp, 34);
b2 -= b13;
tmp = b9 ^ b0;
- b9 = (tmp >> 16) | (tmp << (64 - 16));
+ b9 = ror64(tmp, 16);
b0 -= b9;
tmp = b15 ^ b14;
- b15 = (tmp >> 30) | (tmp << (64 - 30));
+ b15 = ror64(tmp, 30);
b14 -= b15 + k15 + t2;
b15 -= k16 + 1;
tmp = b13 ^ b12;
- b13 = (tmp >> 44) | (tmp << (64 - 44));
+ b13 = ror64(tmp, 44);
b12 -= b13 + k13;
b13 -= k14 + t1;
tmp = b11 ^ b10;
- b11 = (tmp >> 47) | (tmp << (64 - 47));
+ b11 = ror64(tmp, 47);
b10 -= b11 + k11;
b11 -= k12;
tmp = b9 ^ b8;
- b9 = (tmp >> 12) | (tmp << (64 - 12));
+ b9 = ror64(tmp, 12);
b8 -= b9 + k9;
b9 -= k10;
tmp = b7 ^ b6;
- b7 = (tmp >> 31) | (tmp << (64 - 31));
+ b7 = ror64(tmp, 31);
b6 -= b7 + k7;
b7 -= k8;
tmp = b5 ^ b4;
- b5 = (tmp >> 37) | (tmp << (64 - 37));
+ b5 = ror64(tmp, 37);
b4 -= b5 + k5;
b5 -= k6;
tmp = b3 ^ b2;
- b3 = (tmp >> 9) | (tmp << (64 - 9));
+ b3 = ror64(tmp, 9);
b2 -= b3 + k3;
b3 -= k4;
tmp = b1 ^ b0;
- b1 = (tmp >> 41) | (tmp << (64 - 41));
+ b1 = ror64(tmp, 41);
b0 -= b1 + k1;
b1 -= k2;
tmp = b7 ^ b12;
- b7 = (tmp >> 25) | (tmp << (64 - 25));
+ b7 = ror64(tmp, 25);
b12 -= b7;
tmp = b3 ^ b10;
- b3 = (tmp >> 16) | (tmp << (64 - 16));
+ b3 = ror64(tmp, 16);
b10 -= b3;
tmp = b5 ^ b8;
- b5 = (tmp >> 28) | (tmp << (64 - 28));
+ b5 = ror64(tmp, 28);
b8 -= b5;
tmp = b1 ^ b14;
- b1 = (tmp >> 47) | (tmp << (64 - 47));
+ b1 = ror64(tmp, 47);
b14 -= b1;
tmp = b9 ^ b4;
- b9 = (tmp >> 41) | (tmp << (64 - 41));
+ b9 = ror64(tmp, 41);
b4 -= b9;
tmp = b13 ^ b6;
- b13 = (tmp >> 48) | (tmp << (64 - 48));
+ b13 = ror64(tmp, 48);
b6 -= b13;
tmp = b11 ^ b2;
- b11 = (tmp >> 20) | (tmp << (64 - 20));
+ b11 = ror64(tmp, 20);
b2 -= b11;
tmp = b15 ^ b0;
- b15 = (tmp >> 5) | (tmp << (64 - 5));
+ b15 = ror64(tmp, 5);
b0 -= b15;
tmp = b9 ^ b10;
- b9 = (tmp >> 17) | (tmp << (64 - 17));
+ b9 = ror64(tmp, 17);
b10 -= b9;
tmp = b11 ^ b8;
- b11 = (tmp >> 59) | (tmp << (64 - 59));
+ b11 = ror64(tmp, 59);
b8 -= b11;
tmp = b13 ^ b14;
- b13 = (tmp >> 41) | (tmp << (64 - 41));
+ b13 = ror64(tmp, 41);
b14 -= b13;
tmp = b15 ^ b12;
- b15 = (tmp >> 34) | (tmp << (64 - 34));
+ b15 = ror64(tmp, 34);
b12 -= b15;
tmp = b1 ^ b6;
- b1 = (tmp >> 13) | (tmp << (64 - 13));
+ b1 = ror64(tmp, 13);
b6 -= b1;
tmp = b3 ^ b4;
- b3 = (tmp >> 51) | (tmp << (64 - 51));
+ b3 = ror64(tmp, 51);
b4 -= b3;
tmp = b5 ^ b2;
- b5 = (tmp >> 4) | (tmp << (64 - 4));
+ b5 = ror64(tmp, 4);
b2 -= b5;
tmp = b7 ^ b0;
- b7 = (tmp >> 33) | (tmp << (64 - 33));
+ b7 = ror64(tmp, 33);
b0 -= b7;
tmp = b1 ^ b8;
- b1 = (tmp >> 52) | (tmp << (64 - 52));
+ b1 = ror64(tmp, 52);
b8 -= b1;
tmp = b5 ^ b14;
- b5 = (tmp >> 23) | (tmp << (64 - 23));
+ b5 = ror64(tmp, 23);
b14 -= b5;
tmp = b3 ^ b12;
- b3 = (tmp >> 18) | (tmp << (64 - 18));
+ b3 = ror64(tmp, 18);
b12 -= b3;
tmp = b7 ^ b10;
- b7 = (tmp >> 49) | (tmp << (64 - 49));
+ b7 = ror64(tmp, 49);
b10 -= b7;
tmp = b15 ^ b4;
- b15 = (tmp >> 55) | (tmp << (64 - 55));
+ b15 = ror64(tmp, 55);
b4 -= b15;
tmp = b11 ^ b6;
- b11 = (tmp >> 10) | (tmp << (64 - 10));
+ b11 = ror64(tmp, 10);
b6 -= b11;
tmp = b13 ^ b2;
- b13 = (tmp >> 19) | (tmp << (64 - 19));
+ b13 = ror64(tmp, 19);
b2 -= b13;
tmp = b9 ^ b0;
- b9 = (tmp >> 38) | (tmp << (64 - 38));
+ b9 = ror64(tmp, 38);
b0 -= b9;
tmp = b15 ^ b14;
- b15 = (tmp >> 37) | (tmp << (64 - 37));
+ b15 = ror64(tmp, 37);
b14 -= b15 + k14 + t1;
b15 -= k15;
tmp = b13 ^ b12;
- b13 = (tmp >> 22) | (tmp << (64 - 22));
+ b13 = ror64(tmp, 22);
b12 -= b13 + k12;
b13 -= k13 + t0;
tmp = b11 ^ b10;
- b11 = (tmp >> 17) | (tmp << (64 - 17));
+ b11 = ror64(tmp, 17);
b10 -= b11 + k10;
b11 -= k11;
tmp = b9 ^ b8;
- b9 = (tmp >> 8) | (tmp << (64 - 8));
+ b9 = ror64(tmp, 8);
b8 -= b9 + k8;
b9 -= k9;
tmp = b7 ^ b6;
- b7 = (tmp >> 47) | (tmp << (64 - 47));
+ b7 = ror64(tmp, 47);
b6 -= b7 + k6;
b7 -= k7;
tmp = b5 ^ b4;
- b5 = (tmp >> 8) | (tmp << (64 - 8));
+ b5 = ror64(tmp, 8);
b4 -= b5 + k4;
b5 -= k5;
tmp = b3 ^ b2;
- b3 = (tmp >> 13) | (tmp << (64 - 13));
+ b3 = ror64(tmp, 13);
b2 -= b3 + k2;
b3 -= k3;
tmp = b1 ^ b0;
- b1 = (tmp >> 24) | (tmp << (64 - 24));
+ b1 = ror64(tmp, 24);
b0 -= b1 + k0;
b1 -= k1;