sleeping_kangaroo12 0.0.1

data/ext/xkcp/lib/low/Xoodoo/plain/Xoodoo-optimized.c

@@ -0,0 +1,399 @@
+/*
+The eXtended Keccak Code Package (XKCP)
+https://github.com/XKCP/XKCP
+
+The Xoodoo permutation, designed by Joan Daemen, Seth Hoffert, Gilles Van Assche and Ronny Van Keer.
+
+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/
+*/
+
+#include <stdio.h>
+#include <string.h>
+#include "Xoodoo.h"
+
+#define VERBOSE         0
+
+#if (VERBOSE > 0)
+    #define    Dump(__t)    printf(__t "\n"); \
+                            printf("a00 %08x, a01 %08x, a02 %08x, a03 %08x\n", a00, a01, a02, a03 ); \
+                            printf("a10 %08x, a11 %08x, a12 %08x, a13 %08x\n", a10, a11, a12, a13 ); \
+                            printf("a20 %08x, a21 %08x, a22 %08x, a23 %08x\n\n", a20, a21, a22, a23 );
+#else
+    #define    Dump(__t)
+#endif
+
+#if (VERBOSE >= 1)
+    #define    Dump1(__t)    Dump(__t)
+#else
+    #define    Dump1(__t)
+#endif
+
+#if (VERBOSE >= 2)
+    #define    Dump2(__t)    Dump(__t)
+#else
+    #define    Dump2(__t)
+#endif
+
+#if (VERBOSE >= 3)
+    #define    Dump3(__t)    Dump(__t)
+#else
+    #define    Dump3(__t)
+#endif
+
+/* ---------------------------------------------------------------- */
+
+void Xoodoo_Initialize(void *state)
+{
+    memset(state, 0, NLANES*sizeof(tXoodooLane));
+}
+
+/* ---------------------------------------------------------------- */
+
+void Xoodoo_AddBytes(void *argState, const unsigned char *argdata, unsigned int offset, unsigned int length)
+{
+#if (PLATFORM_BYTE_ORDER == IS_LITTLE_ENDIAN)
+    if (length == (3*4*4)) {
+        uint32_t *state = (uint32_t *)argState;
+        uint32_t *data = (uint32_t *)argdata;
+        state[0] ^= data[0];
+        state[1] ^= data[1];
+        state[2] ^= data[2];
+        state[3] ^= data[3];
+        state[4] ^= data[4];
+        state[5] ^= data[5];
+        state[6] ^= data[6];
+        state[7] ^= data[7];
+        state[8] ^= data[8];
+        state[9] ^= data[9];
+        state[10] ^= data[10];
+        state[11] ^= data[11];
+    }
+    else {
+        unsigned int sizeLeft = length;
+        unsigned int lanePosition = offset/4;
+        unsigned int offsetInLane = offset%4;
+        const unsigned char *curData = argdata;
+        uint32_t *state = (uint32_t*)argState;
+
+        state += lanePosition;
+        if ((sizeLeft > 0) && (offsetInLane != 0)) {
+            unsigned int bytesInLane = 4 - offsetInLane;
+            uint32_t lane = 0;
+            if (bytesInLane > sizeLeft)
+                bytesInLane = sizeLeft;
+            memcpy((unsigned char*)&lane + offsetInLane, curData, bytesInLane);
+            *state++ ^= lane;
+            sizeLeft -= bytesInLane;
+            curData += bytesInLane;
+        }
+
+        while(sizeLeft >= 4) {
+            *state++ ^= READ32_UNALIGNED( curData );
+            sizeLeft -= 4;
+            curData += 4;
+        }
+
+        if (sizeLeft > 0) {
+            uint32_t lane = 0;
+            memcpy(&lane, curData, sizeLeft);
+            *state ^= lane;
+        }
+    }
+#else
+    #error "Not yet implemented"
+#endif
+}
+
+/* ---------------------------------------------------------------- */
+
+void Xoodoo_OverwriteBytes(void *argstate, const unsigned char *argdata, unsigned int offset, unsigned int length)
+{
+#if (PLATFORM_BYTE_ORDER == IS_LITTLE_ENDIAN)
+    if (length == (3*4*4)) {
+        uint32_t *state = (uint32_t *)argstate;
+        uint32_t *data = (uint32_t *)argdata;
+        state[0] = data[0];
+        state[1] = data[1];
+        state[2] = data[2];
+        state[3] = data[3];
+        state[4] = data[4];
+        state[5] = data[5];
+        state[6] = data[6];
+        state[7] = data[7];
+        state[8] = data[8];
+        state[9] = data[9];
+        state[10] = data[10];
+        state[11] = data[11];
+    }
+    else
+        memcpy((unsigned char*)argstate+offset, argdata, length);
+#else
+    #error "Not yet implemented"
+#endif
+}
+
+/* ---------------------------------------------------------------- */
+
+void Xoodoo_OverwriteWithZeroes(void *argstate, unsigned int byteCount)
+{
+#if (PLATFORM_BYTE_ORDER == IS_LITTLE_ENDIAN)
+    memset(argstate, 0, byteCount);
+#else
+    #error "Not yet implemented"
+#endif
+}
+
+/* ---------------------------------------------------------------- */
+
+void Xoodoo_ExtractBytes(const void *state, unsigned char *data, unsigned int offset, unsigned int length)
+{
+#if (PLATFORM_BYTE_ORDER == IS_LITTLE_ENDIAN)
+    memcpy(data, (unsigned char*)state+offset, length);
+#else
+    #error "Not yet implemented"
+#endif
+}
+
+/* ---------------------------------------------------------------- */
+
+void Xoodoo_ExtractAndAddBytes(const void *argState, const unsigned char *input, unsigned char *output, unsigned int offset, unsigned int length)
+{
+#if (PLATFORM_BYTE_ORDER == IS_LITTLE_ENDIAN)
+    if (length == (3*4*4)) {
+        uint32_t *state = (uint32_t *)argState;
+        const uint32_t *ii = (const uint32_t *)input;
+        uint32_t *oo = (uint32_t *)output;
+
+        oo[0] = state[0] ^ ii[0];
+        oo[1] = state[1] ^ ii[1];
+        oo[2] = state[2] ^ ii[2];
+        oo[3] = state[3] ^ ii[3];
+        oo[4] = state[4] ^ ii[4];
+        oo[5] = state[5] ^ ii[5];
+        oo[6] = state[6] ^ ii[6];
+        oo[7] = state[7] ^ ii[7];
+        oo[8] = state[8] ^ ii[8];
+        oo[9] = state[9] ^ ii[9];
+        oo[10] = state[10] ^ ii[10];
+        oo[11] = state[11] ^ ii[11];
+    }
+    else {
+        unsigned int sizeLeft = length;
+        unsigned int lanePosition = offset/4;
+        unsigned int offsetInLane = offset%4;
+        const unsigned char *curInput = input;
+        unsigned char *curOutput = output;
+        const uint32_t *state = (const uint32_t*)argState;
+
+        state += lanePosition;
+        if ((sizeLeft > 0) && (offsetInLane != 0)) {
+            unsigned int bytesInLane = 4 - offsetInLane;
+            uint32_t  lane = *state++ >> (offsetInLane * 8);
+            if (bytesInLane > sizeLeft)
+                bytesInLane = sizeLeft;
+            sizeLeft -= bytesInLane;
+            do {
+                *curOutput++ = (*curInput++) ^ (unsigned char)lane;
+                lane >>= 8;
+            }
+            while ( --bytesInLane != 0);
+        }
+
+        while(sizeLeft >= 4) {
+            WRITE32_UNALIGNED( curOutput, READ32_UNALIGNED( curInput ) ^ *state++ );
+            sizeLeft -= 4;
+            curInput += 4;
+            curOutput += 4;
+        }
+
+        if (sizeLeft > 0) {
+            uint32_t  lane = *state;
+            do {
+                *curOutput++ = (*curInput++) ^ (unsigned char)lane;
+                lane >>= 8;
+            }
+            while ( --sizeLeft != 0 );
+        }
+    }
+#else
+    #error "Not yet implemented"
+#endif
+}
+
+/* ---------------------------------------------------------------- */
+
+#define    DeclareVars  uint32_t    a00, a01, a02, a03; \
+                        uint32_t    a10, a11, a12, a13; \
+                        uint32_t    a20, a21, a22, a23; \
+                        uint32_t    v1, v2
+
+#define    State2Vars   a00 = state[0+0], a01 = state[0+1], a02 = state[0+2], a03 = state[0+3]; \
+                        a10 = state[4+0], a11 = state[4+1], a12 = state[4+2], a13 = state[4+3]; \
+                        a20 = state[8+0], a21 = state[8+1], a22 = state[8+2], a23 = state[8+3]
+
+#define    Vars2State   state[0+0] = a00, state[0+1] = a01, state[0+2] = a02, state[0+3] = a03; \
+                        state[4+0] = a10, state[4+1] = a11, state[4+2] = a12, state[4+3] = a13; \
+                        state[8+0] = a20, state[8+1] = a21, state[8+2] = a22, state[8+3] = a23
+
+/*
+** Theta: Column Parity Mixer
+*/
+#define Theta()                                             \
+                    v1 = a03 ^ a13 ^ a23;                   \
+                    v2 = a00 ^ a10 ^ a20;                   \
+                    v1 = ROTL32(v1, 5) ^ ROTL32(v1, 14);    \
+                    a00 ^= v1;                              \
+                    a10 ^= v1;                              \
+                    a20 ^= v1;                              \
+                    v1 = a01 ^ a11 ^ a21;                   \
+                    v2 = ROTL32(v2, 5) ^ ROTL32(v2, 14);    \
+                    a01 ^= v2;                              \
+                    a11 ^= v2;                              \
+                    a21 ^= v2;                              \
+                    v2 = a02 ^ a12 ^ a22;                   \
+                    v1 = ROTL32(v1, 5) ^ ROTL32(v1, 14);    \
+                    a02 ^= v1;                              \
+                    a12 ^= v1;                              \
+                    a22 ^= v1;                              \
+                    v2 = ROTL32(v2, 5) ^ ROTL32(v2, 14);    \
+                    a03 ^= v2;                              \
+                    a13 ^= v2;                              \
+                    a23 ^= v2
+
+/*
+** Rho-west: Plane shift
+*/
+#define Rho_west()                          \
+                    a20 = ROTL32(a20, 11);  \
+                    a21 = ROTL32(a21, 11);  \
+                    a22 = ROTL32(a22, 11);  \
+                    a23 = ROTL32(a23, 11);  \
+                    v1 = a13;               \
+                    a13 = a12;              \
+                    a12 = a11;              \
+                    a11 = a10;              \
+                    a10 = v1
+
+/*
+** Iota: Round constants
+*/
+#define Iota(__rc)  a00 ^= __rc
+
+/*
+** Chi: Non linear step, on colums
+*/
+#define Chi()                               \
+                    a00 ^= ~a10 & a20;      \
+                    a10 ^= ~a20 & a00;      \
+                    a20 ^= ~a00 & a10;      \
+                                            \
+                    a01 ^= ~a11 & a21;      \
+                    a11 ^= ~a21 & a01;      \
+                    a21 ^= ~a01 & a11;      \
+                                            \
+                    a02 ^= ~a12 & a22;      \
+                    a12 ^= ~a22 & a02;      \
+                    a22 ^= ~a02 & a12;      \
+                                            \
+                    a03 ^= ~a13 & a23;      \
+                    a13 ^= ~a23 & a03;      \
+                    a23 ^= ~a03 & a13
+
+/*
+** Rho-east: Plane shift
+*/
+#define Rho_east()                          \
+                    a10 = ROTL32(a10, 1);   \
+                    a11 = ROTL32(a11, 1);   \
+                    a12 = ROTL32(a12, 1);   \
+                    a13 = ROTL32(a13, 1);   \
+                    v1  = ROTL32(a23, 8);   \
+                    a23 = ROTL32(a21, 8);   \
+                    a21 = v1;               \
+                    v1  = ROTL32(a22, 8);   \
+                    a22 = ROTL32(a20, 8);   \
+                    a20 = v1
+
+#define Round(__rc)                         \
+                    Theta();                \
+                    Dump3("Theta");         \
+                    Rho_west();             \
+                    Dump3("Rho-west");      \
+                    Iota(__rc);             \
+                    Dump3("Iota");          \
+                    Chi();                  \
+                    Dump3("Chi");           \
+                    Rho_east();             \
+                    Dump3("Rho-east")
+
+static const uint32_t    RC[MAXROUNDS] = {
+    _rc12,
+    _rc11,
+    _rc10,
+    _rc9,
+    _rc8,
+    _rc7,
+    _rc6,
+    _rc5,
+    _rc4,
+    _rc3,
+    _rc2,
+    _rc1
+};
+
+void Xoodoo_Permute_Nrounds( uint32_t * state, uint32_t nr )
+{
+    DeclareVars;
+    uint32_t    i;
+
+    State2Vars;
+    for (i = MAXROUNDS - nr; i < MAXROUNDS; ++i ) {
+        Round(RC[i]);
+        Dump2("Round");
+    }
+    Dump1("Permutation");
+    Vars2State;
+}
+
+void Xoodoo_Permute_6rounds( uint32_t * state)
+{
+    DeclareVars;
+
+    State2Vars;
+    Round(_rc6);
+    Round(_rc5);
+    Round(_rc4);
+    Round(_rc3);
+    Round(_rc2);
+    Round(_rc1);
+    Dump1("Permutation");
+    Vars2State;
+}
+
+void Xoodoo_Permute_12rounds( uint32_t * state)
+{
+    DeclareVars;
+
+    State2Vars;
+    Round(_rc12);
+    Round(_rc11);
+    Round(_rc10);
+    Round(_rc9);
+    Round(_rc8);
+    Round(_rc7);
+    Round(_rc6);
+    Round(_rc5);
+    Round(_rc4);
+    Round(_rc3);
+    Round(_rc2);
+    Round(_rc1);
+    Dump1("Permutation");
+    Vars2State;
+}

data/ext/xkcp/lib/low/Xoodoo/plain/Xoodyak-full-blocks.c

@@ -0,0 +1,127 @@
+/*
+The eXtended Keccak Code Package (XKCP)
+https://github.com/XKCP/XKCP
+
+The Xoodoo permutation, designed by Joan Daemen, Seth Hoffert, Gilles Van Assche and Ronny Van Keer.
+
+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/
+*/
+
+#include <stdint.h>
+#include <string.h>
+#include <stdlib.h>
+#include "Xoodoo-SnP.h"
+#include "Xoodyak-parameters.h"
+
+#ifdef OUTPUT
+#include <stdio.h>
+#endif
+
+#define SnP_AddByte             Xoodoo_AddByte
+#define SnP_AddBytes            Xoodoo_AddBytes
+#define SnP_ExtractBytes        Xoodoo_ExtractBytes
+#define SnP_ExtractAndAddBytes  Xoodoo_ExtractAndAddBytes
+#define SnP_Permute             Xoodoo_Permute_12rounds
+#define SnP_OverwriteBytes      Xoodoo_OverwriteBytes
+
+size_t Xoodyak_AbsorbKeyedFullBlocks(void *state, const uint8_t *X, size_t XLen)
+{
+    size_t  initialLength = XLen;
+
+    do {
+        SnP_Permute(state);                       /* Xoodyak_Up(instance, NULL, 0, 0); */
+        SnP_AddBytes(state, X, 0, Xoodyak_Rkin);  /* Xoodyak_Down(instance, X, Xoodyak_Rkin, 0); */
+        SnP_AddByte(state, 0x01, Xoodyak_Rkin);
+        X       += Xoodyak_Rkin;
+        XLen    -= Xoodyak_Rkin;
+    } while (XLen >= Xoodyak_Rkin);
+
+    return initialLength - XLen;
+}
+
+size_t Xoodyak_AbsorbHashFullBlocks(void *state, const uint8_t *X, size_t XLen)
+{
+    size_t  initialLength = XLen;
+
+    do {
+        SnP_Permute(state);                       /* Xoodyak_Up(instance, NULL, 0, 0); */
+        SnP_AddBytes(state, X, 0, Xoodyak_Rhash); /* Xoodyak_Down(instance, X, Xoodyak_Rhash, 0); */
+        SnP_AddByte(state, 0x01, Xoodyak_Rhash);
+        X       += Xoodyak_Rhash;
+        XLen    -= Xoodyak_Rhash;
+    } while (XLen >= Xoodyak_Rhash);
+
+    return initialLength - XLen;
+}
+
+
+size_t Xoodyak_SqueezeKeyedFullBlocks(void *state, uint8_t *Y, size_t YLen)
+{
+    size_t  initialLength = YLen;
+
+    do {
+        SnP_AddByte(state, 0x01, 0);  /* Xoodyak_Down(instance, NULL, 0, 0); */
+        SnP_Permute(state);           /* Xoodyak_Up(instance, Y, Xoodyak_Rkout, 0); */
+        SnP_ExtractBytes(state, Y, 0, Xoodyak_Rkout);
+        Y    += Xoodyak_Rkout;
+        YLen -= Xoodyak_Rkout;
+    } while (YLen >= Xoodyak_Rkout);
+
+    return initialLength - YLen;
+}
+
+size_t Xoodyak_SqueezeHashFullBlocks(void *state, uint8_t *Y, size_t YLen)
+{
+    size_t  initialLength = YLen;
+
+    do {
+        SnP_AddByte(state, 0x01, 0);  /* Xoodyak_Down(instance, NULL, 0, 0); */
+        SnP_Permute(state);           /* Xoodyak_Up(instance, Y, Xoodyak_Rhash, 0); */
+        SnP_ExtractBytes(state, Y, 0, Xoodyak_Rhash);
+        Y    += Xoodyak_Rhash;
+        YLen -= Xoodyak_Rhash;
+    } while (YLen >= Xoodyak_Rhash);
+
+    return initialLength - YLen;
+}
+
+size_t Xoodyak_EncryptFullBlocks(void *state, const uint8_t *I, uint8_t *O, size_t IOLen)
+{
+    size_t  initialLength = IOLen;
+
+    do {
+        SnP_Permute(state);
+        SnP_ExtractAndAddBytes(state, I, O, 0, Xoodyak_Rkout);
+        SnP_OverwriteBytes(state, O, 0, Xoodyak_Rkout);
+        SnP_AddByte(state, 0x01, Xoodyak_Rkout);
+        I       += Xoodyak_Rkout;
+        O       += Xoodyak_Rkout;
+        IOLen   -= Xoodyak_Rkout;
+    } while (IOLen >= Xoodyak_Rkout);
+
+    return initialLength - IOLen;
+}
+
+size_t Xoodyak_DecryptFullBlocks(void *state, const uint8_t *I, uint8_t *O, size_t IOLen)
+{
+    size_t  initialLength = IOLen;
+
+    do {
+        SnP_Permute(state);
+        SnP_ExtractAndAddBytes(state, I, O, 0, Xoodyak_Rkout);
+        SnP_AddBytes(state, O, 0, Xoodyak_Rkout);
+        SnP_AddByte(state, 0x01, Xoodyak_Rkout);
+        I       += Xoodyak_Rkout;
+        O       += Xoodyak_Rkout;
+        IOLen   -= Xoodyak_Rkout;
+    } while (IOLen >= Xoodyak_Rkout);
+
+    return initialLength - IOLen;
+}

data/ext/xkcp/lib/low/Xoodoo/ref/Xoodoo-SnP.h

@@ -0,0 +1,43 @@
+/*
+The eXtended Keccak Code Package (XKCP)
+https://github.com/XKCP/XKCP
+
+The Xoodoo permutation, designed by Joan Daemen, Seth Hoffert, Gilles Van Assche and Ronny Van Keer.
+
+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/
+*/
+
+#ifndef _Xoodoo_SnP_h_
+#define _Xoodoo_SnP_h_
+
+#include <stddef.h>
+#include <stdint.h>
+
+/** For the documentation, see SnP-documentation.h.
+ */
+
+#define Xoodoo_implementation      "32-bit reference implementation"
+#define Xoodoo_stateSizeInBytes    (3*4*4)
+#define Xoodoo_stateAlignment      4
+#define Xoodoo_HasNround
+
+void Xoodoo_StaticInitialize( void );
+void Xoodoo_Initialize(void *state);
+#define Xoodoo_AddByte(argS, argData, argOffset)    ((uint8_t*)argS)[argOffset] ^= (argData)
+void Xoodoo_AddBytes(void *state, const uint8_t *data, unsigned int offset, unsigned int length);
+void Xoodoo_OverwriteBytes(void *state, const uint8_t *data, unsigned int offset, unsigned int length);
+void Xoodoo_OverwriteWithZeroes(void *state, unsigned int byteCount);
+void Xoodoo_Permute_Nrounds(void *state, unsigned int nrounds);
+void Xoodoo_Permute_6rounds(void *state);
+void Xoodoo_Permute_12rounds(void *state);
+void Xoodoo_ExtractBytes(const void *state, uint8_t *data, unsigned int offset, unsigned int length);
+void Xoodoo_ExtractAndAddBytes(const void *state, const uint8_t *input, uint8_t *output, unsigned int offset, unsigned int length);
+
+#endif