Order the accesses to A[] in the Keccak permutation.

Consistent speed up of about ~5% on my machine.

Also tried moving all the A[] values to local registers, was slower.

1 files changed, 17 insertions, 17 deletions

diff --git a/src/lib/hash/keccak/keccak.cpp b/src/lib/hash/keccak/keccak.cpp
index c0b430a8f..44297dcef 100644
--- a/src/lib/hash/keccak/keccak.cpp
+++ b/src/lib/hash/keccak/keccak.cpp
@@ -41,30 +41,30 @@ void keccak_f_1600(u64bit A[25])
       const u64bit D4 = rotate_left(C4, 1) ^ C2;
 
       const u64bit B00 = A[ 0] ^ D1;
-      const u64bit B01 = rotate_left(A[ 6] ^ D2, 44);
-      const u64bit B02 = rotate_left(A[12] ^ D3, 43);
-      const u64bit B03 = rotate_left(A[18] ^ D4, 21);
-      const u64bit B04 = rotate_left(A[24] ^ D0, 14);
-      const u64bit B05 = rotate_left(A[ 3] ^ D4, 28);
-      const u64bit B06 = rotate_left(A[ 9] ^ D0, 20);
-      const u64bit B07 = rotate_left(A[10] ^ D1, 3);
-      const u64bit B08 = rotate_left(A[16] ^ D2, 45);
-      const u64bit B09 = rotate_left(A[22] ^ D3, 61);
       const u64bit B10 = rotate_left(A[ 1] ^ D2, 1);
-      const u64bit B11 = rotate_left(A[ 7] ^ D3, 6);
-      const u64bit B12 = rotate_left(A[13] ^ D4, 25);
-      const u64bit B13 = rotate_left(A[19] ^ D0, 8);
-      const u64bit B14 = rotate_left(A[20] ^ D1, 18);
+      const u64bit B20 = rotate_left(A[ 2] ^ D3, 62);
+      const u64bit B05 = rotate_left(A[ 3] ^ D4, 28);
       const u64bit B15 = rotate_left(A[ 4] ^ D0, 27);
       const u64bit B16 = rotate_left(A[ 5] ^ D1, 36);
-      const u64bit B17 = rotate_left(A[11] ^ D2, 10);
-      const u64bit B18 = rotate_left(A[17] ^ D3, 15);
-      const u64bit B19 = rotate_left(A[23] ^ D4, 56);
-      const u64bit B20 = rotate_left(A[ 2] ^ D3, 62);
+      const u64bit B01 = rotate_left(A[ 6] ^ D2, 44);
+      const u64bit B11 = rotate_left(A[ 7] ^ D3, 6);
       const u64bit B21 = rotate_left(A[ 8] ^ D4, 55);
+      const u64bit B06 = rotate_left(A[ 9] ^ D0, 20);
+      const u64bit B07 = rotate_left(A[10] ^ D1, 3);
+      const u64bit B17 = rotate_left(A[11] ^ D2, 10);
+      const u64bit B02 = rotate_left(A[12] ^ D3, 43);
+      const u64bit B12 = rotate_left(A[13] ^ D4, 25);
       const u64bit B22 = rotate_left(A[14] ^ D0, 39);
       const u64bit B23 = rotate_left(A[15] ^ D1, 41);
+      const u64bit B08 = rotate_left(A[16] ^ D2, 45);
+      const u64bit B18 = rotate_left(A[17] ^ D3, 15);
+      const u64bit B03 = rotate_left(A[18] ^ D4, 21);
+      const u64bit B13 = rotate_left(A[19] ^ D0, 8);
+      const u64bit B14 = rotate_left(A[20] ^ D1, 18);
       const u64bit B24 = rotate_left(A[21] ^ D2, 2);
+      const u64bit B09 = rotate_left(A[22] ^ D3, 61);
+      const u64bit B19 = rotate_left(A[23] ^ D4, 56);
+      const u64bit B04 = rotate_left(A[24] ^ D0, 14);
 
       A[ 0] = B00 ^ (~B01 & B02);
       A[ 1] = B01 ^ (~B02 & B03);