20 files changed, 70 insertions, 69 deletions

diff --git a/nist/make-package.sh b/nist/make-package.sh
index 455dcde..6fa7e7c 100755
--- a/nist/make-package.sh
+++ b/nist/make-package.sh
@@ -53,7 +53,8 @@ add-variant ()
 
     implementations=(
         ref
-        add_tabulatedtweakey
+        add_tweakeyloop
+        add_tweakeysequences
     )
 
     for implem in ${implementations[@]}
diff --git a/src/add_developedtweakey/ae-common.h b/src/add_tweakeyloop/ae-common.h
index 561854e..561854e 100644
--- a/src/add_developedtweakey/ae-common.h
+++ b/src/add_tweakeyloop/ae-common.h
diff --git a/src/add_developedtweakey/cipher.c b/src/add_tweakeyloop/cipher.c
index e5ccd15..e5ccd15 100644
--- a/src/add_developedtweakey/cipher.c
+++ b/src/add_tweakeyloop/cipher.c
diff --git a/src/add_developedtweakey/cipher.h b/src/add_tweakeyloop/cipher.h
index 06dfde5..06dfde5 100644
--- a/src/add_developedtweakey/cipher.h
+++ b/src/add_tweakeyloop/cipher.h
diff --git a/src/add_developedtweakey/lilliput-ae-i.c b/src/add_tweakeyloop/lilliput-ae-i.c
index b1758c9..b1758c9 100644
--- a/src/add_developedtweakey/lilliput-ae-i.c
+++ b/src/add_tweakeyloop/lilliput-ae-i.c
diff --git a/src/add_developedtweakey/lilliput-ae-ii.c b/src/add_tweakeyloop/lilliput-ae-ii.c
index 26885e5..26885e5 100644
--- a/src/add_developedtweakey/lilliput-ae-ii.c
+++ b/src/add_tweakeyloop/lilliput-ae-ii.c
diff --git a/src/add_developedtweakey/lilliput-ae.h b/src/add_tweakeyloop/lilliput-ae.h
index e2d5051..e2d5051 100644
--- a/src/add_developedtweakey/lilliput-ae.h
+++ b/src/add_tweakeyloop/lilliput-ae.h
diff --git a/src/add_developedtweakey/parameters.h b/src/add_tweakeyloop/parameters.h
index 681a152..681a152 100644
--- a/src/add_developedtweakey/parameters.h
+++ b/src/add_tweakeyloop/parameters.h
diff --git a/src/add_developedtweakey/tweakey.c b/src/add_tweakeyloop/tweakey.c
index 0588c1b..54b3036 100644
--- a/src/add_developedtweakey/tweakey.c
+++ b/src/add_tweakeyloop/tweakey.c
@@ -43,26 +43,6 @@ void tweakey_state_extract(
 }
 
 
-static uint8_t _M1(uint8_t x)
-{
-    return x<<3 ^ x>>3;
-}
-
-static uint8_t _M2(uint8_t x)
-{
-    return x<<6 ^ (x & 0xf8) ^ x>>6;
-}
-
-static uint8_t _M3(uint8_t x)
-{
-    return (uint8_t)(x<<3) >> 3;
-}
-
-static uint8_t _M4(uint8_t x)
-{
-    return (uint8_t)(x<<2) >> 3;
-}
-
 static void _multiply_M(const uint8_t X[LANE_BYTES], uint8_t Y[LANE_BYTES])
 {
     Y[7] = X[6];
@@ -77,26 +57,18 @@ static void _multiply_M(const uint8_t X[LANE_BYTES], uint8_t Y[LANE_BYTES])
 
 static void _multiply_M2(const uint8_t X[LANE_BYTES], uint8_t Y[LANE_BYTES])
 {
-    Y[7] = X[5];
-    Y[6] = X[5]<<3 ^ X[4];
-    Y[5] = X[5]<<6 ^ _M1(X[4]) ^ X[3];
-    Y[4] = X[4]>>6 ^ X[3]>>3   ^ X[2];
-    Y[3] = X[6]<<2 ^ X[1];
-    Y[2] = X[5]<<2 ^ X[0];
-    Y[1] = X[7];
-    Y[0] = X[6];
+    uint8_t M_X[LANE_BYTES];
+    _multiply_M(X, M_X);
+    _multiply_M(M_X, Y);
 }
 
 static void _multiply_M3(const uint8_t X[LANE_BYTES], uint8_t Y[LANE_BYTES])
 {
-    Y[7] = X[5]<<3   ^ X[4];
-    Y[6] = X[5]<<6   ^ _M1(X[4]) ^ X[3];
-    Y[5] = _M2(X[4]) ^ _M1(X[3]) ^ X[2];
-    Y[4] = X[6]<<2   ^ X[3]>>6   ^ X[2]>>3 ^ X[1];
-    Y[3] = X[5]<<2   ^ X[0];
-    Y[2] = X[7]      ^ X[5]<<5   ^ X[4]<<2;
-    Y[1] = X[6];
-    Y[0] = X[5];
+    uint8_t M_X[LANE_BYTES];
+    uint8_t M2_X[LANE_BYTES];
+    _multiply_M(X, M_X);
+    _multiply_M(M_X, M2_X);
+    _multiply_M(M2_X, Y);
 }
 
 static void _multiply_MR(const uint8_t X[LANE_BYTES], uint8_t Y[LANE_BYTES])
@@ -113,26 +85,18 @@ static void _multiply_MR(const uint8_t X[LANE_BYTES], uint8_t Y[LANE_BYTES])
 
 static void _multiply_MR2(const uint8_t X[LANE_BYTES], uint8_t Y[LANE_BYTES])
 {
-    Y[0] = X[2];
-    Y[1] = X[3]    ^ X[4]>>3;
-    Y[2] = X[4]    ^ X[5]>>3 ^ _M3(X[6]);
-    Y[3] = X[5]    ^ X[6]<<3;
-    Y[4] = X[3]<<2 ^ X[6]    ^ X[7]<<3;
-    Y[5] = X[4]<<2 ^ X[7];
-    Y[6] = X[0];
-    Y[7] = X[1];
+    uint8_t MR_X[LANE_BYTES];
+    _multiply_MR(X, MR_X);
+    _multiply_MR(MR_X, Y);
 }
 
 static void _multiply_MR3(const uint8_t X[LANE_BYTES], uint8_t Y[LANE_BYTES])
 {
-    Y[0] = X[3]      ^ X[4]>>3;
-    Y[1] = X[4]      ^ X[5]>>3 ^ _M3(X[6]);
-    Y[2] = _M4(X[3]) ^ X[5]    ^ _M1(X[6]) ^ _M3(X[7]);
-    Y[3] = X[3]<<2   ^ X[6]    ^ X[7]<<3;
-    Y[4] = X[0]<<3   ^ X[4]<<2 ^ X[7];
-    Y[5] = X[0]      ^ X[5]<<2 ^ X[6]<<5;
-    Y[6] = X[1];
-    Y[7] = X[2];
+    uint8_t MR_X[LANE_BYTES];
+    uint8_t MR2_X[LANE_BYTES];
+    _multiply_MR(X, MR_X);
+    _multiply_MR(MR_X, MR2_X);
+    _multiply_MR(MR2_X, Y);
 }
 
 typedef void (*matrix_multiplication)(const uint8_t X[LANE_BYTES], uint8_t Y[LANE_BYTES]);
diff --git a/src/add_developedtweakey/tweakey.h b/src/add_tweakeyloop/tweakey.h
index 5470bc8..5470bc8 100644
--- a/src/add_developedtweakey/tweakey.h
+++ b/src/add_tweakeyloop/tweakey.h
diff --git a/src/add_tabulatedtweakey/ae-common.h b/src/add_tweakeysequences/ae-common.h
index 561854e..561854e 100644
--- a/src/add_tabulatedtweakey/ae-common.h
+++ b/src/add_tweakeysequences/ae-common.h
diff --git a/src/add_tabulatedtweakey/cipher.c b/src/add_tweakeysequences/cipher.c
index e5ccd15..e5ccd15 100644
--- a/src/add_tabulatedtweakey/cipher.c
+++ b/src/add_tweakeysequences/cipher.c
diff --git a/src/add_tabulatedtweakey/cipher.h b/src/add_tweakeysequences/cipher.h
index 06dfde5..06dfde5 100644
--- a/src/add_tabulatedtweakey/cipher.h
+++ b/src/add_tweakeysequences/cipher.h
diff --git a/src/add_tabulatedtweakey/lilliput-ae-i.c b/src/add_tweakeysequences/lilliput-ae-i.c
index b1758c9..b1758c9 100644
--- a/src/add_tabulatedtweakey/lilliput-ae-i.c
+++ b/src/add_tweakeysequences/lilliput-ae-i.c
diff --git a/src/add_tabulatedtweakey/lilliput-ae-ii.c b/src/add_tweakeysequences/lilliput-ae-ii.c
index 26885e5..26885e5 100644
--- a/src/add_tabulatedtweakey/lilliput-ae-ii.c
+++ b/src/add_tweakeysequences/lilliput-ae-ii.c
diff --git a/src/add_tabulatedtweakey/lilliput-ae.h b/src/add_tweakeysequences/lilliput-ae.h
index e2d5051..e2d5051 100644
--- a/src/add_tabulatedtweakey/lilliput-ae.h
+++ b/src/add_tweakeysequences/lilliput-ae.h
diff --git a/src/add_tabulatedtweakey/parameters.h b/src/add_tweakeysequences/parameters.h
index 681a152..681a152 100644
--- a/src/add_tabulatedtweakey/parameters.h
+++ b/src/add_tweakeysequences/parameters.h
diff --git a/src/add_tabulatedtweakey/tweakey.c b/src/add_tweakeysequences/tweakey.c
index 776352f..776352f 100644
--- a/src/add_tabulatedtweakey/tweakey.c
+++ b/src/add_tweakeysequences/tweakey.c
diff --git a/src/add_tabulatedtweakey/tweakey.h b/src/add_tweakeysequences/tweakey.h
index 5470bc8..5470bc8 100644
--- a/src/add_tabulatedtweakey/tweakey.h
+++ b/src/add_tweakeysequences/tweakey.h
diff --git a/src/ref/tweakey.c b/src/ref/tweakey.c
index 54b3036..0588c1b 100644
--- a/src/ref/tweakey.c
+++ b/src/ref/tweakey.c
@@ -43,6 +43,26 @@ void tweakey_state_extract(
 }
 
 
+static uint8_t _M1(uint8_t x)
+{
+    return x<<3 ^ x>>3;
+}
+
+static uint8_t _M2(uint8_t x)
+{
+    return x<<6 ^ (x & 0xf8) ^ x>>6;
+}
+
+static uint8_t _M3(uint8_t x)
+{
+    return (uint8_t)(x<<3) >> 3;
+}
+
+static uint8_t _M4(uint8_t x)
+{
+    return (uint8_t)(x<<2) >> 3;
+}
+
 static void _multiply_M(const uint8_t X[LANE_BYTES], uint8_t Y[LANE_BYTES])
 {
     Y[7] = X[6];
@@ -57,18 +77,26 @@ static void _multiply_M(const uint8_t X[LANE_BYTES], uint8_t Y[LANE_BYTES])
 
 static void _multiply_M2(const uint8_t X[LANE_BYTES], uint8_t Y[LANE_BYTES])
 {
-    uint8_t M_X[LANE_BYTES];
-    _multiply_M(X, M_X);
-    _multiply_M(M_X, Y);
+    Y[7] = X[5];
+    Y[6] = X[5]<<3 ^ X[4];
+    Y[5] = X[5]<<6 ^ _M1(X[4]) ^ X[3];
+    Y[4] = X[4]>>6 ^ X[3]>>3   ^ X[2];
+    Y[3] = X[6]<<2 ^ X[1];
+    Y[2] = X[5]<<2 ^ X[0];
+    Y[1] = X[7];
+    Y[0] = X[6];
 }
 
 static void _multiply_M3(const uint8_t X[LANE_BYTES], uint8_t Y[LANE_BYTES])
 {
-    uint8_t M_X[LANE_BYTES];
-    uint8_t M2_X[LANE_BYTES];
-    _multiply_M(X, M_X);
-    _multiply_M(M_X, M2_X);
-    _multiply_M(M2_X, Y);
+    Y[7] = X[5]<<3   ^ X[4];
+    Y[6] = X[5]<<6   ^ _M1(X[4]) ^ X[3];
+    Y[5] = _M2(X[4]) ^ _M1(X[3]) ^ X[2];
+    Y[4] = X[6]<<2   ^ X[3]>>6   ^ X[2]>>3 ^ X[1];
+    Y[3] = X[5]<<2   ^ X[0];
+    Y[2] = X[7]      ^ X[5]<<5   ^ X[4]<<2;
+    Y[1] = X[6];
+    Y[0] = X[5];
 }
 
 static void _multiply_MR(const uint8_t X[LANE_BYTES], uint8_t Y[LANE_BYTES])
@@ -85,18 +113,26 @@ static void _multiply_MR(const uint8_t X[LANE_BYTES], uint8_t Y[LANE_BYTES])
 
 static void _multiply_MR2(const uint8_t X[LANE_BYTES], uint8_t Y[LANE_BYTES])
 {
-    uint8_t MR_X[LANE_BYTES];
-    _multiply_MR(X, MR_X);
-    _multiply_MR(MR_X, Y);
+    Y[0] = X[2];
+    Y[1] = X[3]    ^ X[4]>>3;
+    Y[2] = X[4]    ^ X[5]>>3 ^ _M3(X[6]);
+    Y[3] = X[5]    ^ X[6]<<3;
+    Y[4] = X[3]<<2 ^ X[6]    ^ X[7]<<3;
+    Y[5] = X[4]<<2 ^ X[7];
+    Y[6] = X[0];
+    Y[7] = X[1];
 }
 
 static void _multiply_MR3(const uint8_t X[LANE_BYTES], uint8_t Y[LANE_BYTES])
 {
-    uint8_t MR_X[LANE_BYTES];
-    uint8_t MR2_X[LANE_BYTES];
-    _multiply_MR(X, MR_X);
-    _multiply_MR(MR_X, MR2_X);
-    _multiply_MR(MR2_X, Y);
+    Y[0] = X[3]      ^ X[4]>>3;
+    Y[1] = X[4]      ^ X[5]>>3 ^ _M3(X[6]);
+    Y[2] = _M4(X[3]) ^ X[5]    ^ _M1(X[6]) ^ _M3(X[7]);
+    Y[3] = X[3]<<2   ^ X[6]    ^ X[7]<<3;
+    Y[4] = X[0]<<3   ^ X[4]<<2 ^ X[7];
+    Y[5] = X[0]      ^ X[5]<<2 ^ X[6]<<5;
+    Y[6] = X[1];
+    Y[7] = X[2];
 }
 
 typedef void (*matrix_multiplication)(const uint8_t X[LANE_BYTES], uint8_t Y[LANE_BYTES]);