sha3 1.0.4 → 1.0.5

data/ext/sha3/lib/low/KeccakP-1600/common/KeccakP-1600-64.macros

@@ -0,0 +1,748 @@
+/*
+The eXtended Keccak Code Package (XKCP)
+https://github.com/XKCP/XKCP
+
+The Keccak-p permutations, designed by Guido Bertoni, Joan Daemen, Michaël Peeters and Gilles Van Assche.
+
+Implementation by Gilles Van Assche and Ronny Van Keer, hereby denoted as "the implementer".
+
+For more information, feedback or questions, please refer to the Keccak Team website:
+https://keccak.team/
+
+To the extent possible under law, the implementer has waived all copyright
+and related or neighboring rights to the source code in this file.
+http://creativecommons.org/publicdomain/zero/1.0/
+*/
+
+#define declareABCDE \
+    uint64_t Aba, Abe, Abi, Abo, Abu; \
+    uint64_t Aga, Age, Agi, Ago, Agu; \
+    uint64_t Aka, Ake, Aki, Ako, Aku; \
+    uint64_t Ama, Ame, Ami, Amo, Amu; \
+    uint64_t Asa, Ase, Asi, Aso, Asu; \
+    uint64_t Bba, Bbe, Bbi, Bbo, Bbu; \
+    uint64_t Bga, Bge, Bgi, Bgo, Bgu; \
+    uint64_t Bka, Bke, Bki, Bko, Bku; \
+    uint64_t Bma, Bme, Bmi, Bmo, Bmu; \
+    uint64_t Bsa, Bse, Bsi, Bso, Bsu; \
+    uint64_t Ca, Ce, Ci, Co, Cu; \
+    uint64_t Da, De, Di, Do, Du; \
+    uint64_t Eba, Ebe, Ebi, Ebo, Ebu; \
+    uint64_t Ega, Ege, Egi, Ego, Egu; \
+    uint64_t Eka, Eke, Eki, Eko, Eku; \
+    uint64_t Ema, Eme, Emi, Emo, Emu; \
+    uint64_t Esa, Ese, Esi, Eso, Esu; \
+
+#define prepareTheta \
+    Ca = Aba^Aga^Aka^Ama^Asa; \
+    Ce = Abe^Age^Ake^Ame^Ase; \
+    Ci = Abi^Agi^Aki^Ami^Asi; \
+    Co = Abo^Ago^Ako^Amo^Aso; \
+    Cu = Abu^Agu^Aku^Amu^Asu; \
+
+#ifdef UseBebigokimisa
+/* --- Code for round, with prepare-theta (lane complementing pattern 'bebigokimisa') */
+/* --- 64-bit lanes mapped to 64-bit words */
+#define thetaRhoPiChiIotaPrepareTheta(i, A, E) \
+    Da = Cu^ROL64(Ce, 1); \
+    De = Ca^ROL64(Ci, 1); \
+    Di = Ce^ROL64(Co, 1); \
+    Do = Ci^ROL64(Cu, 1); \
+    Du = Co^ROL64(Ca, 1); \
+\
+    A##ba ^= Da; \
+    Bba = A##ba; \
+    A##ge ^= De; \
+    Bbe = ROL64(A##ge, 44); \
+    A##ki ^= Di; \
+    Bbi = ROL64(A##ki, 43); \
+    A##mo ^= Do; \
+    Bbo = ROL64(A##mo, 21); \
+    A##su ^= Du; \
+    Bbu = ROL64(A##su, 14); \
+    E##ba =   Bba ^(  Bbe |  Bbi ); \
+    E##ba ^= KeccakF1600RoundConstants[i]; \
+    Ca = E##ba; \
+    E##be =   Bbe ^((~Bbi)|  Bbo ); \
+    Ce = E##be; \
+    E##bi =   Bbi ^(  Bbo &  Bbu ); \
+    Ci = E##bi; \
+    E##bo =   Bbo ^(  Bbu |  Bba ); \
+    Co = E##bo; \
+    E##bu =   Bbu ^(  Bba &  Bbe ); \
+    Cu = E##bu; \
+\
+    A##bo ^= Do; \
+    Bga = ROL64(A##bo, 28); \
+    A##gu ^= Du; \
+    Bge = ROL64(A##gu, 20); \
+    A##ka ^= Da; \
+    Bgi = ROL64(A##ka, 3); \
+    A##me ^= De; \
+    Bgo = ROL64(A##me, 45); \
+    A##si ^= Di; \
+    Bgu = ROL64(A##si, 61); \
+    E##ga =   Bga ^(  Bge |  Bgi ); \
+    Ca ^= E##ga; \
+    E##ge =   Bge ^(  Bgi &  Bgo ); \
+    Ce ^= E##ge; \
+    E##gi =   Bgi ^(  Bgo |(~Bgu)); \
+    Ci ^= E##gi; \
+    E##go =   Bgo ^(  Bgu |  Bga ); \
+    Co ^= E##go; \
+    E##gu =   Bgu ^(  Bga &  Bge ); \
+    Cu ^= E##gu; \
+\
+    A##be ^= De; \
+    Bka = ROL64(A##be, 1); \
+    A##gi ^= Di; \
+    Bke = ROL64(A##gi, 6); \
+    A##ko ^= Do; \
+    Bki = ROL64(A##ko, 25); \
+    A##mu ^= Du; \
+    Bko = ROL64(A##mu, 8); \
+    A##sa ^= Da; \
+    Bku = ROL64(A##sa, 18); \
+    E##ka =   Bka ^(  Bke |  Bki ); \
+    Ca ^= E##ka; \
+    E##ke =   Bke ^(  Bki &  Bko ); \
+    Ce ^= E##ke; \
+    E##ki =   Bki ^((~Bko)&  Bku ); \
+    Ci ^= E##ki; \
+    E##ko = (~Bko)^(  Bku |  Bka ); \
+    Co ^= E##ko; \
+    E##ku =   Bku ^(  Bka &  Bke ); \
+    Cu ^= E##ku; \
+\
+    A##bu ^= Du; \
+    Bma = ROL64(A##bu, 27); \
+    A##ga ^= Da; \
+    Bme = ROL64(A##ga, 36); \
+    A##ke ^= De; \
+    Bmi = ROL64(A##ke, 10); \
+    A##mi ^= Di; \
+    Bmo = ROL64(A##mi, 15); \
+    A##so ^= Do; \
+    Bmu = ROL64(A##so, 56); \
+    E##ma =   Bma ^(  Bme &  Bmi ); \
+    Ca ^= E##ma; \
+    E##me =   Bme ^(  Bmi |  Bmo ); \
+    Ce ^= E##me; \
+    E##mi =   Bmi ^((~Bmo)|  Bmu ); \
+    Ci ^= E##mi; \
+    E##mo = (~Bmo)^(  Bmu &  Bma ); \
+    Co ^= E##mo; \
+    E##mu =   Bmu ^(  Bma |  Bme ); \
+    Cu ^= E##mu; \
+\
+    A##bi ^= Di; \
+    Bsa = ROL64(A##bi, 62); \
+    A##go ^= Do; \
+    Bse = ROL64(A##go, 55); \
+    A##ku ^= Du; \
+    Bsi = ROL64(A##ku, 39); \
+    A##ma ^= Da; \
+    Bso = ROL64(A##ma, 41); \
+    A##se ^= De; \
+    Bsu = ROL64(A##se, 2); \
+    E##sa =   Bsa ^((~Bse)&  Bsi ); \
+    Ca ^= E##sa; \
+    E##se = (~Bse)^(  Bsi |  Bso ); \
+    Ce ^= E##se; \
+    E##si =   Bsi ^(  Bso &  Bsu ); \
+    Ci ^= E##si; \
+    E##so =   Bso ^(  Bsu |  Bsa ); \
+    Co ^= E##so; \
+    E##su =   Bsu ^(  Bsa &  Bse ); \
+    Cu ^= E##su; \
+\
+
+/* --- Code for round (lane complementing pattern 'bebigokimisa') */
+/* --- 64-bit lanes mapped to 64-bit words */
+#define thetaRhoPiChiIota(i, A, E) \
+    Da = Cu^ROL64(Ce, 1); \
+    De = Ca^ROL64(Ci, 1); \
+    Di = Ce^ROL64(Co, 1); \
+    Do = Ci^ROL64(Cu, 1); \
+    Du = Co^ROL64(Ca, 1); \
+\
+    A##ba ^= Da; \
+    Bba = A##ba; \
+    A##ge ^= De; \
+    Bbe = ROL64(A##ge, 44); \
+    A##ki ^= Di; \
+    Bbi = ROL64(A##ki, 43); \
+    A##mo ^= Do; \
+    Bbo = ROL64(A##mo, 21); \
+    A##su ^= Du; \
+    Bbu = ROL64(A##su, 14); \
+    E##ba =   Bba ^(  Bbe |  Bbi ); \
+    E##ba ^= KeccakF1600RoundConstants[i]; \
+    E##be =   Bbe ^((~Bbi)|  Bbo ); \
+    E##bi =   Bbi ^(  Bbo &  Bbu ); \
+    E##bo =   Bbo ^(  Bbu |  Bba ); \
+    E##bu =   Bbu ^(  Bba &  Bbe ); \
+\
+    A##bo ^= Do; \
+    Bga = ROL64(A##bo, 28); \
+    A##gu ^= Du; \
+    Bge = ROL64(A##gu, 20); \
+    A##ka ^= Da; \
+    Bgi = ROL64(A##ka, 3); \
+    A##me ^= De; \
+    Bgo = ROL64(A##me, 45); \
+    A##si ^= Di; \
+    Bgu = ROL64(A##si, 61); \
+    E##ga =   Bga ^(  Bge |  Bgi ); \
+    E##ge =   Bge ^(  Bgi &  Bgo ); \
+    E##gi =   Bgi ^(  Bgo |(~Bgu)); \
+    E##go =   Bgo ^(  Bgu |  Bga ); \
+    E##gu =   Bgu ^(  Bga &  Bge ); \
+\
+    A##be ^= De; \
+    Bka = ROL64(A##be, 1); \
+    A##gi ^= Di; \
+    Bke = ROL64(A##gi, 6); \
+    A##ko ^= Do; \
+    Bki = ROL64(A##ko, 25); \
+    A##mu ^= Du; \
+    Bko = ROL64(A##mu, 8); \
+    A##sa ^= Da; \
+    Bku = ROL64(A##sa, 18); \
+    E##ka =   Bka ^(  Bke |  Bki ); \
+    E##ke =   Bke ^(  Bki &  Bko ); \
+    E##ki =   Bki ^((~Bko)&  Bku ); \
+    E##ko = (~Bko)^(  Bku |  Bka ); \
+    E##ku =   Bku ^(  Bka &  Bke ); \
+\
+    A##bu ^= Du; \
+    Bma = ROL64(A##bu, 27); \
+    A##ga ^= Da; \
+    Bme = ROL64(A##ga, 36); \
+    A##ke ^= De; \
+    Bmi = ROL64(A##ke, 10); \
+    A##mi ^= Di; \
+    Bmo = ROL64(A##mi, 15); \
+    A##so ^= Do; \
+    Bmu = ROL64(A##so, 56); \
+    E##ma =   Bma ^(  Bme &  Bmi ); \
+    E##me =   Bme ^(  Bmi |  Bmo ); \
+    E##mi =   Bmi ^((~Bmo)|  Bmu ); \
+    E##mo = (~Bmo)^(  Bmu &  Bma ); \
+    E##mu =   Bmu ^(  Bma |  Bme ); \
+\
+    A##bi ^= Di; \
+    Bsa = ROL64(A##bi, 62); \
+    A##go ^= Do; \
+    Bse = ROL64(A##go, 55); \
+    A##ku ^= Du; \
+    Bsi = ROL64(A##ku, 39); \
+    A##ma ^= Da; \
+    Bso = ROL64(A##ma, 41); \
+    A##se ^= De; \
+    Bsu = ROL64(A##se, 2); \
+    E##sa =   Bsa ^((~Bse)&  Bsi ); \
+    E##se = (~Bse)^(  Bsi |  Bso ); \
+    E##si =   Bsi ^(  Bso &  Bsu ); \
+    E##so =   Bso ^(  Bsu |  Bsa ); \
+    E##su =   Bsu ^(  Bsa &  Bse ); \
+\
+
+#else /* UseBebigokimisa */
+/* --- Code for round, with prepare-theta */
+/* --- 64-bit lanes mapped to 64-bit words */
+#define thetaRhoPiChiIotaPrepareTheta(i, A, E) \
+    Da = Cu^ROL64(Ce, 1); \
+    De = Ca^ROL64(Ci, 1); \
+    Di = Ce^ROL64(Co, 1); \
+    Do = Ci^ROL64(Cu, 1); \
+    Du = Co^ROL64(Ca, 1); \
+\
+    A##ba ^= Da; \
+    Bba = A##ba; \
+    A##ge ^= De; \
+    Bbe = ROL64(A##ge, 44); \
+    A##ki ^= Di; \
+    Bbi = ROL64(A##ki, 43); \
+    A##mo ^= Do; \
+    Bbo = ROL64(A##mo, 21); \
+    A##su ^= Du; \
+    Bbu = ROL64(A##su, 14); \
+    E##ba =   Bba ^((~Bbe)&  Bbi ); \
+    E##ba ^= KeccakF1600RoundConstants[i]; \
+    Ca = E##ba; \
+    E##be =   Bbe ^((~Bbi)&  Bbo ); \
+    Ce = E##be; \
+    E##bi =   Bbi ^((~Bbo)&  Bbu ); \
+    Ci = E##bi; \
+    E##bo =   Bbo ^((~Bbu)&  Bba ); \
+    Co = E##bo; \
+    E##bu =   Bbu ^((~Bba)&  Bbe ); \
+    Cu = E##bu; \
+\
+    A##bo ^= Do; \
+    Bga = ROL64(A##bo, 28); \
+    A##gu ^= Du; \
+    Bge = ROL64(A##gu, 20); \
+    A##ka ^= Da; \
+    Bgi = ROL64(A##ka, 3); \
+    A##me ^= De; \
+    Bgo = ROL64(A##me, 45); \
+    A##si ^= Di; \
+    Bgu = ROL64(A##si, 61); \
+    E##ga =   Bga ^((~Bge)&  Bgi ); \
+    Ca ^= E##ga; \
+    E##ge =   Bge ^((~Bgi)&  Bgo ); \
+    Ce ^= E##ge; \
+    E##gi =   Bgi ^((~Bgo)&  Bgu ); \
+    Ci ^= E##gi; \
+    E##go =   Bgo ^((~Bgu)&  Bga ); \
+    Co ^= E##go; \
+    E##gu =   Bgu ^((~Bga)&  Bge ); \
+    Cu ^= E##gu; \
+\
+    A##be ^= De; \
+    Bka = ROL64(A##be, 1); \
+    A##gi ^= Di; \
+    Bke = ROL64(A##gi, 6); \
+    A##ko ^= Do; \
+    Bki = ROL64(A##ko, 25); \
+    A##mu ^= Du; \
+    Bko = ROL64(A##mu, 8); \
+    A##sa ^= Da; \
+    Bku = ROL64(A##sa, 18); \
+    E##ka =   Bka ^((~Bke)&  Bki ); \
+    Ca ^= E##ka; \
+    E##ke =   Bke ^((~Bki)&  Bko ); \
+    Ce ^= E##ke; \
+    E##ki =   Bki ^((~Bko)&  Bku ); \
+    Ci ^= E##ki; \
+    E##ko =   Bko ^((~Bku)&  Bka ); \
+    Co ^= E##ko; \
+    E##ku =   Bku ^((~Bka)&  Bke ); \
+    Cu ^= E##ku; \
+\
+    A##bu ^= Du; \
+    Bma = ROL64(A##bu, 27); \
+    A##ga ^= Da; \
+    Bme = ROL64(A##ga, 36); \
+    A##ke ^= De; \
+    Bmi = ROL64(A##ke, 10); \
+    A##mi ^= Di; \
+    Bmo = ROL64(A##mi, 15); \
+    A##so ^= Do; \
+    Bmu = ROL64(A##so, 56); \
+    E##ma =   Bma ^((~Bme)&  Bmi ); \
+    Ca ^= E##ma; \
+    E##me =   Bme ^((~Bmi)&  Bmo ); \
+    Ce ^= E##me; \
+    E##mi =   Bmi ^((~Bmo)&  Bmu ); \
+    Ci ^= E##mi; \
+    E##mo =   Bmo ^((~Bmu)&  Bma ); \
+    Co ^= E##mo; \
+    E##mu =   Bmu ^((~Bma)&  Bme ); \
+    Cu ^= E##mu; \
+\
+    A##bi ^= Di; \
+    Bsa = ROL64(A##bi, 62); \
+    A##go ^= Do; \
+    Bse = ROL64(A##go, 55); \
+    A##ku ^= Du; \
+    Bsi = ROL64(A##ku, 39); \
+    A##ma ^= Da; \
+    Bso = ROL64(A##ma, 41); \
+    A##se ^= De; \
+    Bsu = ROL64(A##se, 2); \
+    E##sa =   Bsa ^((~Bse)&  Bsi ); \
+    Ca ^= E##sa; \
+    E##se =   Bse ^((~Bsi)&  Bso ); \
+    Ce ^= E##se; \
+    E##si =   Bsi ^((~Bso)&  Bsu ); \
+    Ci ^= E##si; \
+    E##so =   Bso ^((~Bsu)&  Bsa ); \
+    Co ^= E##so; \
+    E##su =   Bsu ^((~Bsa)&  Bse ); \
+    Cu ^= E##su; \
+\
+
+/* --- Code for round */
+/* --- 64-bit lanes mapped to 64-bit words */
+#define thetaRhoPiChiIota(i, A, E) \
+    Da = Cu^ROL64(Ce, 1); \
+    De = Ca^ROL64(Ci, 1); \
+    Di = Ce^ROL64(Co, 1); \
+    Do = Ci^ROL64(Cu, 1); \
+    Du = Co^ROL64(Ca, 1); \
+\
+    A##ba ^= Da; \
+    Bba = A##ba; \
+    A##ge ^= De; \
+    Bbe = ROL64(A##ge, 44); \
+    A##ki ^= Di; \
+    Bbi = ROL64(A##ki, 43); \
+    A##mo ^= Do; \
+    Bbo = ROL64(A##mo, 21); \
+    A##su ^= Du; \
+    Bbu = ROL64(A##su, 14); \
+    E##ba =   Bba ^((~Bbe)&  Bbi ); \
+    E##ba ^= KeccakF1600RoundConstants[i]; \
+    E##be =   Bbe ^((~Bbi)&  Bbo ); \
+    E##bi =   Bbi ^((~Bbo)&  Bbu ); \
+    E##bo =   Bbo ^((~Bbu)&  Bba ); \
+    E##bu =   Bbu ^((~Bba)&  Bbe ); \
+\
+    A##bo ^= Do; \
+    Bga = ROL64(A##bo, 28); \
+    A##gu ^= Du; \
+    Bge = ROL64(A##gu, 20); \
+    A##ka ^= Da; \
+    Bgi = ROL64(A##ka, 3); \
+    A##me ^= De; \
+    Bgo = ROL64(A##me, 45); \
+    A##si ^= Di; \
+    Bgu = ROL64(A##si, 61); \
+    E##ga =   Bga ^((~Bge)&  Bgi ); \
+    E##ge =   Bge ^((~Bgi)&  Bgo ); \
+    E##gi =   Bgi ^((~Bgo)&  Bgu ); \
+    E##go =   Bgo ^((~Bgu)&  Bga ); \
+    E##gu =   Bgu ^((~Bga)&  Bge ); \
+\
+    A##be ^= De; \
+    Bka = ROL64(A##be, 1); \
+    A##gi ^= Di; \
+    Bke = ROL64(A##gi, 6); \
+    A##ko ^= Do; \
+    Bki = ROL64(A##ko, 25); \
+    A##mu ^= Du; \
+    Bko = ROL64(A##mu, 8); \
+    A##sa ^= Da; \
+    Bku = ROL64(A##sa, 18); \
+    E##ka =   Bka ^((~Bke)&  Bki ); \
+    E##ke =   Bke ^((~Bki)&  Bko ); \
+    E##ki =   Bki ^((~Bko)&  Bku ); \
+    E##ko =   Bko ^((~Bku)&  Bka ); \
+    E##ku =   Bku ^((~Bka)&  Bke ); \
+\
+    A##bu ^= Du; \
+    Bma = ROL64(A##bu, 27); \
+    A##ga ^= Da; \
+    Bme = ROL64(A##ga, 36); \
+    A##ke ^= De; \
+    Bmi = ROL64(A##ke, 10); \
+    A##mi ^= Di; \
+    Bmo = ROL64(A##mi, 15); \
+    A##so ^= Do; \
+    Bmu = ROL64(A##so, 56); \
+    E##ma =   Bma ^((~Bme)&  Bmi ); \
+    E##me =   Bme ^((~Bmi)&  Bmo ); \
+    E##mi =   Bmi ^((~Bmo)&  Bmu ); \
+    E##mo =   Bmo ^((~Bmu)&  Bma ); \
+    E##mu =   Bmu ^((~Bma)&  Bme ); \
+\
+    A##bi ^= Di; \
+    Bsa = ROL64(A##bi, 62); \
+    A##go ^= Do; \
+    Bse = ROL64(A##go, 55); \
+    A##ku ^= Du; \
+    Bsi = ROL64(A##ku, 39); \
+    A##ma ^= Da; \
+    Bso = ROL64(A##ma, 41); \
+    A##se ^= De; \
+    Bsu = ROL64(A##se, 2); \
+    E##sa =   Bsa ^((~Bse)&  Bsi ); \
+    E##se =   Bse ^((~Bsi)&  Bso ); \
+    E##si =   Bsi ^((~Bso)&  Bsu ); \
+    E##so =   Bso ^((~Bsu)&  Bsa ); \
+    E##su =   Bsu ^((~Bsa)&  Bse ); \
+\
+
+#endif /* UseBebigokimisa */
+
+#define copyFromState(X, state) \
+    X##ba = state[ 0]; \
+    X##be = state[ 1]; \
+    X##bi = state[ 2]; \
+    X##bo = state[ 3]; \
+    X##bu = state[ 4]; \
+    X##ga = state[ 5]; \
+    X##ge = state[ 6]; \
+    X##gi = state[ 7]; \
+    X##go = state[ 8]; \
+    X##gu = state[ 9]; \
+    X##ka = state[10]; \
+    X##ke = state[11]; \
+    X##ki = state[12]; \
+    X##ko = state[13]; \
+    X##ku = state[14]; \
+    X##ma = state[15]; \
+    X##me = state[16]; \
+    X##mi = state[17]; \
+    X##mo = state[18]; \
+    X##mu = state[19]; \
+    X##sa = state[20]; \
+    X##se = state[21]; \
+    X##si = state[22]; \
+    X##so = state[23]; \
+    X##su = state[24]; \
+
+#define copyToState(state, X) \
+    state[ 0] = X##ba; \
+    state[ 1] = X##be; \
+    state[ 2] = X##bi; \
+    state[ 3] = X##bo; \
+    state[ 4] = X##bu; \
+    state[ 5] = X##ga; \
+    state[ 6] = X##ge; \
+    state[ 7] = X##gi; \
+    state[ 8] = X##go; \
+    state[ 9] = X##gu; \
+    state[10] = X##ka; \
+    state[11] = X##ke; \
+    state[12] = X##ki; \
+    state[13] = X##ko; \
+    state[14] = X##ku; \
+    state[15] = X##ma; \
+    state[16] = X##me; \
+    state[17] = X##mi; \
+    state[18] = X##mo; \
+    state[19] = X##mu; \
+    state[20] = X##sa; \
+    state[21] = X##se; \
+    state[22] = X##si; \
+    state[23] = X##so; \
+    state[24] = X##su; \
+
+#define copyStateVariables(X, Y) \
+    X##ba = Y##ba; \
+    X##be = Y##be; \
+    X##bi = Y##bi; \
+    X##bo = Y##bo; \
+    X##bu = Y##bu; \
+    X##ga = Y##ga; \
+    X##ge = Y##ge; \
+    X##gi = Y##gi; \
+    X##go = Y##go; \
+    X##gu = Y##gu; \
+    X##ka = Y##ka; \
+    X##ke = Y##ke; \
+    X##ki = Y##ki; \
+    X##ko = Y##ko; \
+    X##ku = Y##ku; \
+    X##ma = Y##ma; \
+    X##me = Y##me; \
+    X##mi = Y##mi; \
+    X##mo = Y##mo; \
+    X##mu = Y##mu; \
+    X##sa = Y##sa; \
+    X##se = Y##se; \
+    X##si = Y##si; \
+    X##so = Y##so; \
+    X##su = Y##su; \
+
+#if (PLATFORM_BYTE_ORDER == IS_LITTLE_ENDIAN)
+#define HTOLE64(x) (x)
+#else
+#define HTOLE64(x) (\
+  ((x & 0xff00000000000000ull) >> 56) | \
+  ((x & 0x00ff000000000000ull) >> 40) | \
+  ((x & 0x0000ff0000000000ull) >> 24) | \
+  ((x & 0x000000ff00000000ull) >> 8)  | \
+  ((x & 0x00000000ff000000ull) << 8)  | \
+  ((x & 0x0000000000ff0000ull) << 24) | \
+  ((x & 0x000000000000ff00ull) << 40) | \
+  ((x & 0x00000000000000ffull) << 56))
+#endif
+
+#define addInput(X, input, laneCount) \
+    if (laneCount == 21) { \
+        X##ba ^= HTOLE64(input[ 0]); \
+        X##be ^= HTOLE64(input[ 1]); \
+        X##bi ^= HTOLE64(input[ 2]); \
+        X##bo ^= HTOLE64(input[ 3]); \
+        X##bu ^= HTOLE64(input[ 4]); \
+        X##ga ^= HTOLE64(input[ 5]); \
+        X##ge ^= HTOLE64(input[ 6]); \
+        X##gi ^= HTOLE64(input[ 7]); \
+        X##go ^= HTOLE64(input[ 8]); \
+        X##gu ^= HTOLE64(input[ 9]); \
+        X##ka ^= HTOLE64(input[10]); \
+        X##ke ^= HTOLE64(input[11]); \
+        X##ki ^= HTOLE64(input[12]); \
+        X##ko ^= HTOLE64(input[13]); \
+        X##ku ^= HTOLE64(input[14]); \
+        X##ma ^= HTOLE64(input[15]); \
+        X##me ^= HTOLE64(input[16]); \
+        X##mi ^= HTOLE64(input[17]); \
+        X##mo ^= HTOLE64(input[18]); \
+        X##mu ^= HTOLE64(input[19]); \
+        X##sa ^= HTOLE64(input[20]); \
+    } \
+    else if (laneCount < 16) { \
+        if (laneCount < 8) { \
+            if (laneCount < 4) { \
+                if (laneCount < 2) { \
+                    if (laneCount < 1) { \
+                    } \
+                    else { \
+                        X##ba ^= HTOLE64(input[ 0]); \
+                    } \
+                } \
+                else { \
+                    X##ba ^= HTOLE64(input[ 0]); \
+                    X##be ^= HTOLE64(input[ 1]); \
+                    if (laneCount < 3) { \
+                    } \
+                    else { \
+                        X##bi ^= HTOLE64(input[ 2]); \
+                    } \
+                } \
+            } \
+            else { \
+                X##ba ^= HTOLE64(input[ 0]); \
+                X##be ^= HTOLE64(input[ 1]); \
+                X##bi ^= HTOLE64(input[ 2]); \
+                X##bo ^= HTOLE64(input[ 3]); \
+                if (laneCount < 6) { \
+                    if (laneCount < 5) { \
+                    } \
+                    else { \
+                        X##bu ^= HTOLE64(input[ 4]); \
+                    } \
+                } \
+                else { \
+                    X##bu ^= HTOLE64(input[ 4]); \
+                    X##ga ^= HTOLE64(input[ 5]); \
+                    if (laneCount < 7) { \
+                    } \
+                    else { \
+                        X##ge ^= HTOLE64(input[ 6]); \
+                    } \
+                } \
+            } \
+        } \
+        else { \
+            X##ba ^= HTOLE64(input[ 0]); \
+            X##be ^= HTOLE64(input[ 1]); \
+            X##bi ^= HTOLE64(input[ 2]); \
+            X##bo ^= HTOLE64(input[ 3]); \
+            X##bu ^= HTOLE64(input[ 4]); \
+            X##ga ^= HTOLE64(input[ 5]); \
+            X##ge ^= HTOLE64(input[ 6]); \
+            X##gi ^= HTOLE64(input[ 7]); \
+            if (laneCount < 12) { \
+                if (laneCount < 10) { \
+                    if (laneCount < 9) { \
+                    } \
+                    else { \
+                        X##go ^= HTOLE64(input[ 8]); \
+                    } \
+                } \
+                else { \
+                    X##go ^= HTOLE64(input[ 8]); \
+                    X##gu ^= HTOLE64(input[ 9]); \
+                    if (laneCount < 11) { \
+                    } \
+                    else { \
+                        X##ka ^= HTOLE64(input[10]); \
+                    } \
+                } \
+            } \
+            else { \
+                X##go ^= HTOLE64(input[ 8]); \
+                X##gu ^= HTOLE64(input[ 9]); \
+                X##ka ^= HTOLE64(input[10]); \
+                X##ke ^= HTOLE64(input[11]); \
+                if (laneCount < 14) { \
+                    if (laneCount < 13) { \
+                    } \
+                    else { \
+                        X##ki ^= HTOLE64(input[12]); \
+                    } \
+                } \
+                else { \
+                    X##ki ^= HTOLE64(input[12]); \
+                    X##ko ^= HTOLE64(input[13]); \
+                    if (laneCount < 15) { \
+                    } \
+                    else { \
+                        X##ku ^= HTOLE64(input[14]); \
+                    } \
+                } \
+            } \
+        } \
+    } \
+    else { \
+        X##ba ^= HTOLE64(input[ 0]); \
+        X##be ^= HTOLE64(input[ 1]); \
+        X##bi ^= HTOLE64(input[ 2]); \
+        X##bo ^= HTOLE64(input[ 3]); \
+        X##bu ^= HTOLE64(input[ 4]); \
+        X##ga ^= HTOLE64(input[ 5]); \
+        X##ge ^= HTOLE64(input[ 6]); \
+        X##gi ^= HTOLE64(input[ 7]); \
+        X##go ^= HTOLE64(input[ 8]); \
+        X##gu ^= HTOLE64(input[ 9]); \
+        X##ka ^= HTOLE64(input[10]); \
+        X##ke ^= HTOLE64(input[11]); \
+        X##ki ^= HTOLE64(input[12]); \
+        X##ko ^= HTOLE64(input[13]); \
+        X##ku ^= HTOLE64(input[14]); \
+        X##ma ^= HTOLE64(input[15]); \
+        if (laneCount < 24) { \
+            if (laneCount < 20) { \
+                if (laneCount < 18) { \
+                    if (laneCount < 17) { \
+                    } \
+                    else { \
+                        X##me ^= HTOLE64(input[16]); \
+                    } \
+                } \
+                else { \
+                    X##me ^= HTOLE64(input[16]); \
+                    X##mi ^= HTOLE64(input[17]); \
+                    if (laneCount < 19) { \
+                    } \
+                    else { \
+                        X##mo ^= HTOLE64(input[18]); \
+                    } \
+                } \
+            } \
+            else { \
+                X##me ^= HTOLE64(input[16]); \
+                X##mi ^= HTOLE64(input[17]); \
+                X##mo ^= HTOLE64(input[18]); \
+                X##mu ^= HTOLE64(input[19]); \
+                if (laneCount < 22) { \
+                    if (laneCount < 21) { \
+                    } \
+                    else { \
+                        X##sa ^= HTOLE64(input[20]); \
+                    } \
+                } \
+                else { \
+                    X##sa ^= HTOLE64(input[20]); \
+                    X##se ^= HTOLE64(input[21]); \
+                    if (laneCount < 23) { \
+                    } \
+                    else { \
+                        X##si ^= HTOLE64(input[22]); \
+                    } \
+                } \
+            } \
+        } \
+        else { \
+            X##me ^= HTOLE64(input[16]); \
+            X##mi ^= HTOLE64(input[17]); \
+            X##mo ^= HTOLE64(input[18]); \
+            X##mu ^= HTOLE64(input[19]); \
+            X##sa ^= HTOLE64(input[20]); \
+            X##se ^= HTOLE64(input[21]); \
+            X##si ^= HTOLE64(input[22]); \
+            X##so ^= HTOLE64(input[23]); \
+            if (laneCount < 25) { \
+            } \
+            else { \
+                X##su ^= HTOLE64(input[24]); \
+            } \
+        } \
+    }