sleeping_kangaroo12 0.0.1

data/ext/xkcp/lib/low/KeccakP-800/plain/KeccakP-800-opt32.macros

@@ -0,0 +1,459 @@
+/*
+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 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 _b  0
+#define _g  5
+#define _k  10
+#define _m  15
+#define _s  20
+
+#define _a  0
+#define _e  1
+#define _i  2
+#define _o  3
+#define _u  4
+
+#define declareBCDE \
+    uint32_t Ba, Be, Bi, Bo, Bu; \
+    uint32_t Ca, Ce, Ci, Co, Cu; \
+    uint32_t Da, De, Di, Do, Du; \
+    uint32_t Estate[25]; \
+
+#define prepareTheta \
+    Ca = Astate[_b+_a]^Astate[_g+_a]^Astate[_k+_a]^Astate[_m+_a]^Astate[_s+_a]; \
+    Ce = Astate[_b+_e]^Astate[_g+_e]^Astate[_k+_e]^Astate[_m+_e]^Astate[_s+_e]; \
+    Ci = Astate[_b+_i]^Astate[_g+_i]^Astate[_k+_i]^Astate[_m+_i]^Astate[_s+_i]; \
+    Co = Astate[_b+_o]^Astate[_g+_o]^Astate[_k+_o]^Astate[_m+_o]^Astate[_s+_o]; \
+    Cu = Astate[_b+_u]^Astate[_g+_u]^Astate[_k+_u]^Astate[_m+_u]^Astate[_s+_u]; \
+
+#ifdef UseBebigokimisa
+/* --- Code for round, with prepare-theta (lane complementing pattern 'bebigokimisa') */
+/* --- 32-bit lanes mapped to 32-bit words */
+#define thetaRhoPiChiIotaPrepareTheta(i, A, E) \
+    Da = Cu^ROL32(Ce, 1); \
+    De = Ca^ROL32(Ci, 1); \
+    Di = Ce^ROL32(Co, 1); \
+    Do = Ci^ROL32(Cu, 1); \
+    Du = Co^ROL32(Ca, 1); \
+\
+    A##state[_b+_a] ^= Da; \
+    Ba = A##state[_b+_a]; \
+    A##state[_g+_e] ^= De; \
+    Be = ROL32(A##state[_g+_e], 12); \
+    A##state[_k+_i] ^= Di; \
+    Bi = ROL32(A##state[_k+_i], 11); \
+    A##state[_m+_o] ^= Do; \
+    Bo = ROL32(A##state[_m+_o], 21); \
+    A##state[_s+_u] ^= Du; \
+    Bu = ROL32(A##state[_s+_u], 14); \
+    E##state[_b+_a] =   Ba ^(  Be |  Bi ); \
+    E##state[_b+_a] ^= KeccakF800RoundConstants[i]; \
+    Ca = E##state[_b+_a]; \
+    E##state[_b+_e] =   Be ^((~Bi)|  Bo ); \
+    Ce = E##state[_b+_e]; \
+    E##state[_b+_i] =   Bi ^(  Bo &  Bu ); \
+    Ci = E##state[_b+_i]; \
+    E##state[_b+_o] =   Bo ^(  Bu |  Ba ); \
+    Co = E##state[_b+_o]; \
+    E##state[_b+_u] =   Bu ^(  Ba &  Be ); \
+    Cu = E##state[_b+_u]; \
+\
+    A##state[_b+_o] ^= Do; \
+    Ba = ROL32(A##state[_b+_o], 28); \
+    A##state[_g+_u] ^= Du; \
+    Be = ROL32(A##state[_g+_u], 20); \
+    A##state[_k+_a] ^= Da; \
+    Bi = ROL32(A##state[_k+_a], 3); \
+    A##state[_m+_e] ^= De; \
+    Bo = ROL32(A##state[_m+_e], 13); \
+    A##state[_s+_i] ^= Di; \
+    Bu = ROL32(A##state[_s+_i], 29); \
+    E##state[_g+_a] =   Ba ^(  Be |  Bi ); \
+    Ca ^= E##state[_g+_a]; \
+    E##state[_g+_e] =   Be ^(  Bi &  Bo ); \
+    Ce ^= E##state[_g+_e]; \
+    E##state[_g+_i] =   Bi ^(  Bo |(~Bu)); \
+    Ci ^= E##state[_g+_i]; \
+    E##state[_g+_o] =   Bo ^(  Bu |  Ba ); \
+    Co ^= E##state[_g+_o]; \
+    E##state[_g+_u] =   Bu ^(  Ba &  Be ); \
+    Cu ^= E##state[_g+_u]; \
+\
+    A##state[_b+_e] ^= De; \
+    Ba = ROL32(A##state[_b+_e], 1); \
+    A##state[_g+_i] ^= Di; \
+    Be = ROL32(A##state[_g+_i], 6); \
+    A##state[_k+_o] ^= Do; \
+    Bi = ROL32(A##state[_k+_o], 25); \
+    A##state[_m+_u] ^= Du; \
+    Bo = ROL32(A##state[_m+_u], 8); \
+    A##state[_s+_a] ^= Da; \
+    Bu = ROL32(A##state[_s+_a], 18); \
+    E##state[_k+_a] =   Ba ^(  Be |  Bi ); \
+    Ca ^= E##state[_k+_a]; \
+    E##state[_k+_e] =   Be ^(  Bi &  Bo ); \
+    Ce ^= E##state[_k+_e]; \
+    E##state[_k+_i] =   Bi ^((~Bo)&  Bu ); \
+    Ci ^= E##state[_k+_i]; \
+    E##state[_k+_o] = (~Bo)^(  Bu |  Ba ); \
+    Co ^= E##state[_k+_o]; \
+    E##state[_k+_u] =   Bu ^(  Ba &  Be ); \
+    Cu ^= E##state[_k+_u]; \
+\
+    A##state[_b+_u] ^= Du; \
+    Ba = ROL32(A##state[_b+_u], 27); \
+    A##state[_g+_a] ^= Da; \
+    Be = ROL32(A##state[_g+_a], 4); \
+    A##state[_k+_e] ^= De; \
+    Bi = ROL32(A##state[_k+_e], 10); \
+    A##state[_m+_i] ^= Di; \
+    Bo = ROL32(A##state[_m+_i], 15); \
+    A##state[_s+_o] ^= Do; \
+    Bu = ROL32(A##state[_s+_o], 24); \
+    E##state[_m+_a] =   Ba ^(  Be &  Bi ); \
+    Ca ^= E##state[_m+_a]; \
+    E##state[_m+_e] =   Be ^(  Bi |  Bo ); \
+    Ce ^= E##state[_m+_e]; \
+    E##state[_m+_i] =   Bi ^((~Bo)|  Bu ); \
+    Ci ^= E##state[_m+_i]; \
+    E##state[_m+_o] = (~Bo)^(  Bu &  Ba ); \
+    Co ^= E##state[_m+_o]; \
+    E##state[_m+_u] =   Bu ^(  Ba |  Be ); \
+    Cu ^= E##state[_m+_u]; \
+\
+    A##state[_b+_i] ^= Di; \
+    Ba = ROL32(A##state[_b+_i], 30); \
+    A##state[_g+_o] ^= Do; \
+    Be = ROL32(A##state[_g+_o], 23); \
+    A##state[_k+_u] ^= Du; \
+    Bi = ROL32(A##state[_k+_u], 7); \
+    A##state[_m+_a] ^= Da; \
+    Bo = ROL32(A##state[_m+_a], 9); \
+    A##state[_s+_e] ^= De; \
+    Bu = ROL32(A##state[_s+_e], 2); \
+    E##state[_s+_a] =   Ba ^((~Be)&  Bi ); \
+    Ca ^= E##state[_s+_a]; \
+    E##state[_s+_e] = (~Be)^(  Bi |  Bo ); \
+    Ce ^= E##state[_s+_e]; \
+    E##state[_s+_i] =   Bi ^(  Bo &  Bu ); \
+    Ci ^= E##state[_s+_i]; \
+    E##state[_s+_o] =   Bo ^(  Bu |  Ba ); \
+    Co ^= E##state[_s+_o]; \
+    E##state[_s+_u] =   Bu ^(  Ba &  Be ); \
+    Cu ^= E##state[_s+_u]; \
+\
+
+/* --- Code for round (lane complementing pattern 'bebigokimisa') */
+/* --- 32-bit lanes mapped to 32-bit words */
+#define thetaRhoPiChiIota(i, A, E) \
+    Da = Cu^ROL32(Ce, 1); \
+    De = Ca^ROL32(Ci, 1); \
+    Di = Ce^ROL32(Co, 1); \
+    Do = Ci^ROL32(Cu, 1); \
+    Du = Co^ROL32(Ca, 1); \
+\
+    A##state[_b+_a] ^= Da; \
+    Ba = A##state[_b+_a]; \
+    A##state[_g+_e] ^= De; \
+    Be = ROL32(A##state[_g+_e], 12); \
+    A##state[_k+_i] ^= Di; \
+    Bi = ROL32(A##state[_k+_i], 11); \
+    A##state[_m+_o] ^= Do; \
+    Bo = ROL32(A##state[_m+_o], 21); \
+    A##state[_s+_u] ^= Du; \
+    Bu = ROL32(A##state[_s+_u], 14); \
+    E##state[_b+_a] =   Ba ^(  Be |  Bi ); \
+    E##state[_b+_a] ^= KeccakF800RoundConstants[i]; \
+    E##state[_b+_e] =   Be ^((~Bi)|  Bo ); \
+    E##state[_b+_i] =   Bi ^(  Bo &  Bu ); \
+    E##state[_b+_o] =   Bo ^(  Bu |  Ba ); \
+    E##state[_b+_u] =   Bu ^(  Ba &  Be ); \
+\
+    A##state[_b+_o] ^= Do; \
+    Ba = ROL32(A##state[_b+_o], 28); \
+    A##state[_g+_u] ^= Du; \
+    Be = ROL32(A##state[_g+_u], 20); \
+    A##state[_k+_a] ^= Da; \
+    Bi = ROL32(A##state[_k+_a], 3); \
+    A##state[_m+_e] ^= De; \
+    Bo = ROL32(A##state[_m+_e], 13); \
+    A##state[_s+_i] ^= Di; \
+    Bu = ROL32(A##state[_s+_i], 29); \
+    E##state[_g+_a] =   Ba ^(  Be |  Bi ); \
+    E##state[_g+_e] =   Be ^(  Bi &  Bo ); \
+    E##state[_g+_i] =   Bi ^(  Bo |(~Bu)); \
+    E##state[_g+_o] =   Bo ^(  Bu |  Ba ); \
+    E##state[_g+_u] =   Bu ^(  Ba &  Be ); \
+\
+    A##state[_b+_e] ^= De; \
+    Ba = ROL32(A##state[_b+_e], 1); \
+    A##state[_g+_i] ^= Di; \
+    Be = ROL32(A##state[_g+_i], 6); \
+    A##state[_k+_o] ^= Do; \
+    Bi = ROL32(A##state[_k+_o], 25); \
+    A##state[_m+_u] ^= Du; \
+    Bo = ROL32(A##state[_m+_u], 8); \
+    A##state[_s+_a] ^= Da; \
+    Bu = ROL32(A##state[_s+_a], 18); \
+    E##state[_k+_a] =   Ba ^(  Be |  Bi ); \
+    E##state[_k+_e] =   Be ^(  Bi &  Bo ); \
+    E##state[_k+_i] =   Bi ^((~Bo)&  Bu ); \
+    E##state[_k+_o] = (~Bo)^(  Bu |  Ba ); \
+    E##state[_k+_u] =   Bu ^(  Ba &  Be ); \
+\
+    A##state[_b+_u] ^= Du; \
+    Ba = ROL32(A##state[_b+_u], 27); \
+    A##state[_g+_a] ^= Da; \
+    Be = ROL32(A##state[_g+_a], 4); \
+    A##state[_k+_e] ^= De; \
+    Bi = ROL32(A##state[_k+_e], 10); \
+    A##state[_m+_i] ^= Di; \
+    Bo = ROL32(A##state[_m+_i], 15); \
+    A##state[_s+_o] ^= Do; \
+    Bu = ROL32(A##state[_s+_o], 24); \
+    E##state[_m+_a] =   Ba ^(  Be &  Bi ); \
+    E##state[_m+_e] =   Be ^(  Bi |  Bo ); \
+    E##state[_m+_i] =   Bi ^((~Bo)|  Bu ); \
+    E##state[_m+_o] = (~Bo)^(  Bu &  Ba ); \
+    E##state[_m+_u] =   Bu ^(  Ba |  Be ); \
+\
+    A##state[_b+_i] ^= Di; \
+    Ba = ROL32(A##state[_b+_i], 30); \
+    A##state[_g+_o] ^= Do; \
+    Be = ROL32(A##state[_g+_o], 23); \
+    A##state[_k+_u] ^= Du; \
+    Bi = ROL32(A##state[_k+_u], 7); \
+    A##state[_m+_a] ^= Da; \
+    Bo = ROL32(A##state[_m+_a], 9); \
+    A##state[_s+_e] ^= De; \
+    Bu = ROL32(A##state[_s+_e], 2); \
+    E##state[_s+_a] =   Ba ^((~Be)&  Bi ); \
+    E##state[_s+_e] = (~Be)^(  Bi |  Bo ); \
+    E##state[_s+_i] =   Bi ^(  Bo &  Bu ); \
+    E##state[_s+_o] =   Bo ^(  Bu |  Ba ); \
+    E##state[_s+_u] =   Bu ^(  Ba &  Be ); \
+\
+
+#else /* UseBebigokimisa */
+/* --- Code for round, with prepare-theta */
+/* --- 32-bit lanes mapped to 32-bit words */
+#define thetaRhoPiChiIotaPrepareTheta(i, A, E) \
+    Da = Cu^ROL32(Ce, 1); \
+    De = Ca^ROL32(Ci, 1); \
+    Di = Ce^ROL32(Co, 1); \
+    Do = Ci^ROL32(Cu, 1); \
+    Du = Co^ROL32(Ca, 1); \
+\
+    A##state[_b+_a] ^= Da; \
+    Ba = A##state[_b+_a]; \
+    A##state[_g+_e] ^= De; \
+    Be = ROL32(A##state[_g+_e], 12); \
+    A##state[_k+_i] ^= Di; \
+    Bi = ROL32(A##state[_k+_i], 11); \
+    A##state[_m+_o] ^= Do; \
+    Bo = ROL32(A##state[_m+_o], 21); \
+    A##state[_s+_u] ^= Du; \
+    Bu = ROL32(A##state[_s+_u], 14); \
+    E##state[_b+_a] =   Ba ^((~Be)&  Bi ); \
+    E##state[_b+_a] ^= KeccakF800RoundConstants[i]; \
+    Ca = E##state[_b+_a]; \
+    E##state[_b+_e] =   Be ^((~Bi)&  Bo ); \
+    Ce = E##state[_b+_e]; \
+    E##state[_b+_i] =   Bi ^((~Bo)&  Bu ); \
+    Ci = E##state[_b+_i]; \
+    E##state[_b+_o] =   Bo ^((~Bu)&  Ba ); \
+    Co = E##state[_b+_o]; \
+    E##state[_b+_u] =   Bu ^((~Ba)&  Be ); \
+    Cu = E##state[_b+_u]; \
+\
+    A##state[_b+_o] ^= Do; \
+    Ba = ROL32(A##state[_b+_o], 28); \
+    A##state[_g+_u] ^= Du; \
+    Be = ROL32(A##state[_g+_u], 20); \
+    A##state[_k+_a] ^= Da; \
+    Bi = ROL32(A##state[_k+_a], 3); \
+    A##state[_m+_e] ^= De; \
+    Bo = ROL32(A##state[_m+_e], 13); \
+    A##state[_s+_i] ^= Di; \
+    Bu = ROL32(A##state[_s+_i], 29); \
+    E##state[_g+_a] =   Ba ^((~Be)&  Bi ); \
+    Ca ^= E##state[_g+_a]; \
+    E##state[_g+_e] =   Be ^((~Bi)&  Bo ); \
+    Ce ^= E##state[_g+_e]; \
+    E##state[_g+_i] =   Bi ^((~Bo)&  Bu ); \
+    Ci ^= E##state[_g+_i]; \
+    E##state[_g+_o] =   Bo ^((~Bu)&  Ba ); \
+    Co ^= E##state[_g+_o]; \
+    E##state[_g+_u] =   Bu ^((~Ba)&  Be ); \
+    Cu ^= E##state[_g+_u]; \
+\
+    A##state[_b+_e] ^= De; \
+    Ba = ROL32(A##state[_b+_e], 1); \
+    A##state[_g+_i] ^= Di; \
+    Be = ROL32(A##state[_g+_i], 6); \
+    A##state[_k+_o] ^= Do; \
+    Bi = ROL32(A##state[_k+_o], 25); \
+    A##state[_m+_u] ^= Du; \
+    Bo = ROL32(A##state[_m+_u], 8); \
+    A##state[_s+_a] ^= Da; \
+    Bu = ROL32(A##state[_s+_a], 18); \
+    E##state[_k+_a] =   Ba ^((~Be)&  Bi ); \
+    Ca ^= E##state[_k+_a]; \
+    E##state[_k+_e] =   Be ^((~Bi)&  Bo ); \
+    Ce ^= E##state[_k+_e]; \
+    E##state[_k+_i] =   Bi ^((~Bo)&  Bu ); \
+    Ci ^= E##state[_k+_i]; \
+    E##state[_k+_o] =   Bo ^((~Bu)&  Ba ); \
+    Co ^= E##state[_k+_o]; \
+    E##state[_k+_u] =   Bu ^((~Ba)&  Be ); \
+    Cu ^= E##state[_k+_u]; \
+\
+    A##state[_b+_u] ^= Du; \
+    Ba = ROL32(A##state[_b+_u], 27); \
+    A##state[_g+_a] ^= Da; \
+    Be = ROL32(A##state[_g+_a], 4); \
+    A##state[_k+_e] ^= De; \
+    Bi = ROL32(A##state[_k+_e], 10); \
+    A##state[_m+_i] ^= Di; \
+    Bo = ROL32(A##state[_m+_i], 15); \
+    A##state[_s+_o] ^= Do; \
+    Bu = ROL32(A##state[_s+_o], 24); \
+    E##state[_m+_a] =   Ba ^((~Be)&  Bi ); \
+    Ca ^= E##state[_m+_a]; \
+    E##state[_m+_e] =   Be ^((~Bi)&  Bo ); \
+    Ce ^= E##state[_m+_e]; \
+    E##state[_m+_i] =   Bi ^((~Bo)&  Bu ); \
+    Ci ^= E##state[_m+_i]; \
+    E##state[_m+_o] =   Bo ^((~Bu)&  Ba ); \
+    Co ^= E##state[_m+_o]; \
+    E##state[_m+_u] =   Bu ^((~Ba)&  Be ); \
+    Cu ^= E##state[_m+_u]; \
+\
+    A##state[_b+_i] ^= Di; \
+    Ba = ROL32(A##state[_b+_i], 30); \
+    A##state[_g+_o] ^= Do; \
+    Be = ROL32(A##state[_g+_o], 23); \
+    A##state[_k+_u] ^= Du; \
+    Bi = ROL32(A##state[_k+_u], 7); \
+    A##state[_m+_a] ^= Da; \
+    Bo = ROL32(A##state[_m+_a], 9); \
+    A##state[_s+_e] ^= De; \
+    Bu = ROL32(A##state[_s+_e], 2); \
+    E##state[_s+_a] =   Ba ^((~Be)&  Bi ); \
+    Ca ^= E##state[_s+_a]; \
+    E##state[_s+_e] =   Be ^((~Bi)&  Bo ); \
+    Ce ^= E##state[_s+_e]; \
+    E##state[_s+_i] =   Bi ^((~Bo)&  Bu ); \
+    Ci ^= E##state[_s+_i]; \
+    E##state[_s+_o] =   Bo ^((~Bu)&  Ba ); \
+    Co ^= E##state[_s+_o]; \
+    E##state[_s+_u] =   Bu ^((~Ba)&  Be ); \
+    Cu ^= E##state[_s+_u]; \
+\
+
+/* --- Code for round */
+/* --- 32-bit lanes mapped to 32-bit words */
+#define thetaRhoPiChiIota(i, A, E) \
+    Da = Cu^ROL32(Ce, 1); \
+    De = Ca^ROL32(Ci, 1); \
+    Di = Ce^ROL32(Co, 1); \
+    Do = Ci^ROL32(Cu, 1); \
+    Du = Co^ROL32(Ca, 1); \
+\
+    A##state[_b+_a] ^= Da; \
+    Ba = A##state[_b+_a]; \
+    A##state[_g+_e] ^= De; \
+    Be = ROL32(A##state[_g+_e], 12); \
+    A##state[_k+_i] ^= Di; \
+    Bi = ROL32(A##state[_k+_i], 11); \
+    A##state[_m+_o] ^= Do; \
+    Bo = ROL32(A##state[_m+_o], 21); \
+    A##state[_s+_u] ^= Du; \
+    Bu = ROL32(A##state[_s+_u], 14); \
+    E##state[_b+_a] =   Ba ^((~Be)&  Bi ); \
+    E##state[_b+_a] ^= KeccakF800RoundConstants[i]; \
+    E##state[_b+_e] =   Be ^((~Bi)&  Bo ); \
+    E##state[_b+_i] =   Bi ^((~Bo)&  Bu ); \
+    E##state[_b+_o] =   Bo ^((~Bu)&  Ba ); \
+    E##state[_b+_u] =   Bu ^((~Ba)&  Be ); \
+\
+    A##state[_b+_o] ^= Do; \
+    Ba = ROL32(A##state[_b+_o], 28); \
+    A##state[_g+_u] ^= Du; \
+    Be = ROL32(A##state[_g+_u], 20); \
+    A##state[_k+_a] ^= Da; \
+    Bi = ROL32(A##state[_k+_a], 3); \
+    A##state[_m+_e] ^= De; \
+    Bo = ROL32(A##state[_m+_e], 13); \
+    A##state[_s+_i] ^= Di; \
+    Bu = ROL32(A##state[_s+_i], 29); \
+    E##state[_g+_a] =   Ba ^((~Be)&  Bi ); \
+    E##state[_g+_e] =   Be ^((~Bi)&  Bo ); \
+    E##state[_g+_i] =   Bi ^((~Bo)&  Bu ); \
+    E##state[_g+_o] =   Bo ^((~Bu)&  Ba ); \
+    E##state[_g+_u] =   Bu ^((~Ba)&  Be ); \
+\
+    A##state[_b+_e] ^= De; \
+    Ba = ROL32(A##state[_b+_e], 1); \
+    A##state[_g+_i] ^= Di; \
+    Be = ROL32(A##state[_g+_i], 6); \
+    A##state[_k+_o] ^= Do; \
+    Bi = ROL32(A##state[_k+_o], 25); \
+    A##state[_m+_u] ^= Du; \
+    Bo = ROL32(A##state[_m+_u], 8); \
+    A##state[_s+_a] ^= Da; \
+    Bu = ROL32(A##state[_s+_a], 18); \
+    E##state[_k+_a] =   Ba ^((~Be)&  Bi ); \
+    E##state[_k+_e] =   Be ^((~Bi)&  Bo ); \
+    E##state[_k+_i] =   Bi ^((~Bo)&  Bu ); \
+    E##state[_k+_o] =   Bo ^((~Bu)&  Ba ); \
+    E##state[_k+_u] =   Bu ^((~Ba)&  Be ); \
+\
+    A##state[_b+_u] ^= Du; \
+    Ba = ROL32(A##state[_b+_u], 27); \
+    A##state[_g+_a] ^= Da; \
+    Be = ROL32(A##state[_g+_a], 4); \
+    A##state[_k+_e] ^= De; \
+    Bi = ROL32(A##state[_k+_e], 10); \
+    A##state[_m+_i] ^= Di; \
+    Bo = ROL32(A##state[_m+_i], 15); \
+    A##state[_s+_o] ^= Do; \
+    Bu = ROL32(A##state[_s+_o], 24); \
+    E##state[_m+_a] =   Ba ^((~Be)&  Bi ); \
+    E##state[_m+_e] =   Be ^((~Bi)&  Bo ); \
+    E##state[_m+_i] =   Bi ^((~Bo)&  Bu ); \
+    E##state[_m+_o] =   Bo ^((~Bu)&  Ba ); \
+    E##state[_m+_u] =   Bu ^((~Ba)&  Be ); \
+\
+    A##state[_b+_i] ^= Di; \
+    Ba = ROL32(A##state[_b+_i], 30); \
+    A##state[_g+_o] ^= Do; \
+    Be = ROL32(A##state[_g+_o], 23); \
+    A##state[_k+_u] ^= Du; \
+    Bi = ROL32(A##state[_k+_u], 7); \
+    A##state[_m+_a] ^= Da; \
+    Bo = ROL32(A##state[_m+_a], 9); \
+    A##state[_s+_e] ^= De; \
+    Bu = ROL32(A##state[_s+_e], 2); \
+    E##state[_s+_a] =   Ba ^((~Be)&  Bi ); \
+    E##state[_s+_e] =   Be ^((~Bi)&  Bo ); \
+    E##state[_s+_i] =   Bi ^((~Bo)&  Bu ); \
+    E##state[_s+_o] =   Bo ^((~Bu)&  Ba ); \
+    E##state[_s+_u] =   Bu ^((~Ba)&  Be ); \
+\
+
+#endif /* UseBebigokimisa */
+

data/ext/xkcp/lib/low/KeccakP-800/plain/KeccakP-800-unrolling-bis.macros

@@ -0,0 +1,83 @@
+/*
+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 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/
+*/
+
+#if (defined(KeccakP800_fullUnrolling))
+#define rounds22 \
+    thetaRhoPiChiIota( 0, A, E) \
+    thetaRhoPiChiIota( 1, E, A) \
+    thetaRhoPiChiIota( 2, A, E) \
+    thetaRhoPiChiIota( 3, E, A) \
+    thetaRhoPiChiIota( 4, A, E) \
+    thetaRhoPiChiIota( 5, E, A) \
+    thetaRhoPiChiIota( 6, A, E) \
+    thetaRhoPiChiIota( 7, E, A) \
+    thetaRhoPiChiIota( 8, A, E) \
+    thetaRhoPiChiIota( 9, E, A) \
+    thetaRhoPiChiIota(10, A, E) \
+    thetaRhoPiChiIota(11, E, A) \
+    thetaRhoPiChiIota(12, A, E) \
+    thetaRhoPiChiIota(13, E, A) \
+    thetaRhoPiChiIota(14, A, E) \
+    thetaRhoPiChiIota(15, E, A) \
+    thetaRhoPiChiIota(16, A, E) \
+    thetaRhoPiChiIota(17, E, A) \
+    thetaRhoPiChiIota(18, A, E) \
+    thetaRhoPiChiIota(19, E, A) \
+    thetaRhoPiChiIota(20, A, E) \
+    thetaRhoPiChiIota(21, E, A) \
+
+#define rounds12 \
+    thetaRhoPiChiIota(10, A, E) \
+    thetaRhoPiChiIota(11, E, A) \
+    thetaRhoPiChiIota(12, A, E) \
+    thetaRhoPiChiIota(13, E, A) \
+    thetaRhoPiChiIota(14, A, E) \
+    thetaRhoPiChiIota(15, E, A) \
+    thetaRhoPiChiIota(16, A, E) \
+    thetaRhoPiChiIota(17, E, A) \
+    thetaRhoPiChiIota(18, A, E) \
+    thetaRhoPiChiIota(19, E, A) \
+    thetaRhoPiChiIota(20, A, E) \
+    thetaRhoPiChiIota(21, E, A) \
+
+#elif (KeccakP800_unrolling == 2)
+#define rounds22 \
+    for(i=0; i<22; i+=2) { \
+        thetaRhoPiChiIota(i  , A, E) \
+        thetaRhoPiChiIota(i+1, E, A) \
+    } \
+
+#define rounds12 \
+    for(i=10; i<22; i+=2) { \
+        thetaRhoPiChiIota(i  , A, E) \
+        thetaRhoPiChiIota(i+1, E, A) \
+    } \
+
+#else
+#error "KeccakP800_unrolling is not correctly specified!"
+#endif
+
+#define roundsN(A,E,__nrounds) \
+    i = 22 - (__nrounds); \
+    if ((i&1) != 0) { \
+        memcpy( (void*)Estate, (void*)Astate, 25*4); \
+        thetaRhoPiChiIota(i, E, A) \
+        ++i; \
+    } \
+    for( /* empty */; i<22; i+=2) { \
+        thetaRhoPiChiIota(i  , A, E) \
+        thetaRhoPiChiIota(i+1, E, A) \
+    }

data/ext/xkcp/lib/low/KeccakP-800/plain/KeccakP-800-unrolling.macros

@@ -0,0 +1,88 @@
+/*
+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 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/
+*/
+
+#if (defined(KeccakP800_fullUnrolling))
+#define rounds22 \
+    prepareTheta \
+    thetaRhoPiChiIotaPrepareTheta( 0, A, E) \
+    thetaRhoPiChiIotaPrepareTheta( 1, E, A) \
+    thetaRhoPiChiIotaPrepareTheta( 2, A, E) \
+    thetaRhoPiChiIotaPrepareTheta( 3, E, A) \
+    thetaRhoPiChiIotaPrepareTheta( 4, A, E) \
+    thetaRhoPiChiIotaPrepareTheta( 5, E, A) \
+    thetaRhoPiChiIotaPrepareTheta( 6, A, E) \
+    thetaRhoPiChiIotaPrepareTheta( 7, E, A) \
+    thetaRhoPiChiIotaPrepareTheta( 8, A, E) \
+    thetaRhoPiChiIotaPrepareTheta( 9, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(10, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(11, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(12, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(13, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(14, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(15, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(16, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(17, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(18, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(19, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(20, A, E) \
+    thetaRhoPiChiIota(21, E, A) \
+
+#define rounds12 \
+    prepareTheta \
+    thetaRhoPiChiIotaPrepareTheta(10, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(11, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(12, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(13, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(14, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(15, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(16, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(17, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(18, A, E) \
+    thetaRhoPiChiIotaPrepareTheta(19, E, A) \
+    thetaRhoPiChiIotaPrepareTheta(20, A, E) \
+    thetaRhoPiChiIota(21, E, A) \
+
+#elif (KeccakP800_unrolling == 2)
+#define rounds22 \
+    prepareTheta \
+    for(i=0; i<22; i+=2) { \
+        thetaRhoPiChiIotaPrepareTheta(i  , A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+1, E, A) \
+    } \
+
+#define rounds12 \
+    prepareTheta \
+    for(i=10; i<22; i+=2) { \
+        thetaRhoPiChiIotaPrepareTheta(i  , A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+1, E, A) \
+    } \
+
+#else
+#error "KeccakP800_unrolling is not correctly specified!"
+#endif
+
+#define roundsN(A,E,__nrounds) \
+    prepareTheta \
+    i = 22 - (__nrounds); \
+    if ((i&1) != 0) { \
+        memcpy( (void*)Estate, (void*)Astate, 25*4); \
+        thetaRhoPiChiIotaPrepareTheta(i, E, A) \
+        ++i; \
+    } \
+    for( /* empty */; i<22; i+=2) { \
+        thetaRhoPiChiIotaPrepareTheta(i  , A, E) \
+        thetaRhoPiChiIotaPrepareTheta(i+1, E, A) \
+    }

data/ext/xkcp/lib/low/KeccakP-800/plain/lcu2/KeccakP-800-opt32-config.h

@@ -0,0 +1,7 @@
+/*
+This file defines some parameters of the implementation in the parent directory.
+*/
+
+#define KeccakP800_implementation_config "lane complementing, 2 rounds unrolled"
+#define KeccakP800_unrolling 2
+#define KeccakP800_useLaneComplementing

data/ext/xkcp/lib/low/KeccakP-800/plain/lcua/KeccakP-800-opt32-config.h

@@ -0,0 +1,7 @@
+/*
+This file defines some parameters of the implementation in the parent directory.
+*/
+
+#define KeccakP800_implementation_config "lane complementing, all rounds unrolled"
+#define KeccakP800_fullUnrolling
+#define KeccakP800_useLaneComplementing

data/ext/xkcp/lib/low/KeccakP-800/plain/u2/KeccakP-800-opt32-config.h

@@ -0,0 +1,7 @@
+/*
+This file defines some parameters of the implementation in the parent directory.
+*/
+
+#define KeccakP800_implementation_config "alternate flavor, 2 rounds unrolled"
+#define KeccakP800_unrolling 2
+#define KeccakP800_useFlavorBis

data/ext/xkcp/lib/low/KeccakP-800/plain/ua/KeccakP-800-opt32-config.h

@@ -0,0 +1,7 @@
+/*
+This file defines some parameters of the implementation in the parent directory.
+*/
+
+#define KeccakP800_implementation_config "alternate flavor, all rounds unrolled"
+#define KeccakP800_fullUnrolling
+#define KeccakP800_useFlavorBis