digest-kangarootwelve 0.4.7 → 0.5.0

data/ext/digest/kangarootwelve/XKCP/lib/low/KeccakP-1600/AVX2/KeccakP-1600-AVX2.s

@@ -1014,6 +1014,9 @@ KeccakP1600_12rounds_FastLoop_Absorb_LanesAddLoop:
 .size   KeccakP1600_12rounds_FastLoop_Absorb,.-KeccakP1600_12rounds_FastLoop_Absorb
 .endif
 
+.ifndef old_gas_syntax
+.section .rodata
+.endif
 .equ    ALLON,        0xFFFFFFFFFFFFFFFF
 
 .balign 64

data/ext/digest/kangarootwelve/XKCP/lib/low/KeccakP-1600/AVX2/KeccakP-1600-SnP.h

@@ -22,6 +22,7 @@ Please refer to SnP-documentation.h for more details.
 #define _KeccakP_1600_SnP_h_
 
 #include <stddef.h>
+#include <stdint.h>
 
 #ifdef __MINGW32__
 #define FORCE_SYSV __attribute__((sysv_abi))
@@ -29,24 +30,33 @@ Please refer to SnP-documentation.h for more details.
 #define FORCE_SYSV
 #endif
 
+typedef struct {
+    uint64_t Aba, Abe, Abi, Abo, Abu;
+    uint64_t Aka, Asa, Aga, Ama;
+    uint64_t Ame, Agi, Aso, Aku;
+    uint64_t Ake, Asi, Ago, Amu;
+    uint64_t Ase, Ami, Ako, Agu;
+    uint64_t Age, Aki, Amo, Asu;
+} KeccakP1600_AVX2_state;
+
+typedef KeccakP1600_AVX2_state KeccakP1600_state;
+
 #define KeccakP1600_implementation      "AVX2 optimized implementation"
-#define KeccakP1600_stateSizeInBytes    200
-#define KeccakP1600_stateAlignment      8
 #define KeccakF1600_FastLoop_supported
 #define KeccakP1600_12rounds_FastLoop_supported
 
 #define KeccakP1600_StaticInitialize()
-FORCE_SYSV void KeccakP1600_Initialize(void *state);
-FORCE_SYSV void KeccakP1600_AddByte(void *state, unsigned char data, unsigned int offset);
-FORCE_SYSV void KeccakP1600_AddBytes(void *state, const unsigned char *data, unsigned int offset, unsigned int length);
-FORCE_SYSV void KeccakP1600_OverwriteBytes(void *state, const unsigned char *data, unsigned int offset, unsigned int length);
-FORCE_SYSV void KeccakP1600_OverwriteWithZeroes(void *state, unsigned int byteCount);
-FORCE_SYSV void KeccakP1600_Permute_Nrounds(void *state, unsigned int nrounds);
-FORCE_SYSV void KeccakP1600_Permute_12rounds(void *state);
-FORCE_SYSV void KeccakP1600_Permute_24rounds(void *state);
-FORCE_SYSV void KeccakP1600_ExtractBytes(const void *state, unsigned char *data, unsigned int offset, unsigned int length);
-FORCE_SYSV void KeccakP1600_ExtractAndAddBytes(const void *state, const unsigned char *input, unsigned char *output, unsigned int offset, unsigned int length);
-FORCE_SYSV size_t KeccakF1600_FastLoop_Absorb(void *state, unsigned int laneCount, const unsigned char *data, size_t dataByteLen);
-FORCE_SYSV size_t KeccakP1600_12rounds_FastLoop_Absorb(void *state, unsigned int laneCount, const unsigned char *data, size_t dataByteLen);
+FORCE_SYSV void KeccakP1600_Initialize(KeccakP1600_AVX2_state *state);
+FORCE_SYSV void KeccakP1600_AddByte(KeccakP1600_AVX2_state *state, unsigned char data, unsigned int offset);
+FORCE_SYSV void KeccakP1600_AddBytes(KeccakP1600_AVX2_state *state, const unsigned char *data, unsigned int offset, unsigned int length);
+FORCE_SYSV void KeccakP1600_OverwriteBytes(KeccakP1600_AVX2_state *state, const unsigned char *data, unsigned int offset, unsigned int length);
+FORCE_SYSV void KeccakP1600_OverwriteWithZeroes(KeccakP1600_AVX2_state *state, unsigned int byteCount);
+FORCE_SYSV void KeccakP1600_Permute_Nrounds(KeccakP1600_AVX2_state *state, unsigned int nrounds);
+FORCE_SYSV void KeccakP1600_Permute_12rounds(KeccakP1600_AVX2_state *state);
+FORCE_SYSV void KeccakP1600_Permute_24rounds(KeccakP1600_AVX2_state *state);
+FORCE_SYSV void KeccakP1600_ExtractBytes(const KeccakP1600_AVX2_state *state, unsigned char *data, unsigned int offset, unsigned int length);
+FORCE_SYSV void KeccakP1600_ExtractAndAddBytes(const KeccakP1600_AVX2_state *state, const unsigned char *input, unsigned char *output, unsigned int offset, unsigned int length);
+FORCE_SYSV size_t KeccakF1600_FastLoop_Absorb(KeccakP1600_AVX2_state *state, unsigned int laneCount, const unsigned char *data, size_t dataByteLen);
+FORCE_SYSV size_t KeccakP1600_12rounds_FastLoop_Absorb(KeccakP1600_AVX2_state *state, unsigned int laneCount, const unsigned char *data, size_t dataByteLen);
 
 #endif

data/ext/digest/kangarootwelve/XKCP/lib/low/KeccakP-1600/AVX512/C/KeccakP-1600-AVX512.c

@@ -37,6 +37,7 @@ We would like to thank Vladimir Sedach, we have used parts of his Keccak AVX-512
 #include "align.h"
 #include "brg_endian.h"
 #include "KeccakP-1600-AVX512-config.h"
+#include "KeccakP-1600-SnP.h"
 
 #if (PLATFORM_BYTE_ORDER != IS_LITTLE_ENDIAN)
 #error Expecting a little-endian platform
@@ -226,19 +227,19 @@ typedef __m512i     V512;
 
 /* ---------------------------------------------------------------- */
 
-void KeccakP1600_Initialize(void *state)
+void KeccakP1600_Initialize(KeccakP1600_plain64_state *state)
 {
-    memset(state, 0, 1600/8);
+    memset(state->A, 0, sizeof(KeccakP1600_plain64_state));
 }
 
 /* ---------------------------------------------------------------- */
 
-void KeccakP1600_AddBytes(void *state, const unsigned char *data, unsigned int offset, unsigned int length)
+void KeccakP1600_AddBytes(KeccakP1600_plain64_state *state, const unsigned char *data, unsigned int offset, unsigned int length)
 {
     uint8_t  *stateAsBytes;
     uint64_t *stateAsLanes;
 
-    for( stateAsBytes = (uint8_t*)state; ((offset % 8) != 0) && (length != 0); ++offset, --length)
+    for( stateAsBytes = (uint8_t*)state->A; ((offset % 8) != 0) && (length != 0); ++offset, --length)
         stateAsBytes[offset] ^= *(data++);
     for (stateAsLanes = (uint64_t*)(stateAsBytes + offset); length >= 8*8; stateAsLanes += 8, data += 8*8, length -= 8*8)
         STORE_8Lanes( stateAsLanes, XOR(LOAD_8Lanes(stateAsLanes), LOAD_8Lanes((const uint64_t*)data)));
@@ -250,33 +251,33 @@ void KeccakP1600_AddBytes(void *state, const unsigned char *data, unsigned int o
 
 /* ---------------------------------------------------------------- */
 
-void KeccakP1600_OverwriteBytes(void *state, const unsigned char *data, unsigned int offset, unsigned int length)
+void KeccakP1600_OverwriteBytes(KeccakP1600_plain64_state *state, const unsigned char *data, unsigned int offset, unsigned int length)
 {
-    memcpy((unsigned char*)state+offset, data, length);
+    memcpy((unsigned char*)state->A + offset, data, length);
 }
 
 /* ---------------------------------------------------------------- */
 
-void KeccakP1600_OverwriteWithZeroes(void *state, unsigned int byteCount)
+void KeccakP1600_OverwriteWithZeroes(KeccakP1600_plain64_state *state, unsigned int byteCount)
 {
-    memset(state, 0, byteCount);
+    memset(state->A, 0, byteCount);
 }
 
 /* ---------------------------------------------------------------- */
 
-void KeccakP1600_ExtractBytes(const void *state, unsigned char *data, unsigned int offset, unsigned int length)
+void KeccakP1600_ExtractBytes(const KeccakP1600_plain64_state *state, unsigned char *data, unsigned int offset, unsigned int length)
 {
-    memcpy(data, (unsigned char*)state+offset, length);
+    memcpy(data, (unsigned char*)state->A + offset, length);
 }
 
 /* ---------------------------------------------------------------- */
 
-void KeccakP1600_ExtractAndAddBytes(const void *state, const unsigned char *input, unsigned char *output, unsigned int offset, unsigned int length)
+void KeccakP1600_ExtractAndAddBytes(const KeccakP1600_plain64_state *state, const unsigned char *input, unsigned char *output, unsigned int offset, unsigned int length)
 {
     uint8_t  *stateAsBytes;
     uint64_t *stateAsLanes;
 
-    for( stateAsBytes = (uint8_t*)state; ((offset % 8) != 0) && (length != 0); ++offset, --length)
+    for( stateAsBytes = (uint8_t*)state->A; ((offset % 8) != 0) && (length != 0); ++offset, --length)
         *(output++) = stateAsBytes[offset] ^ *(input++);
     for (stateAsLanes = (uint64_t*)(stateAsBytes + offset); length >= 8*8; stateAsLanes += 8, input += 8*8, output += 8*8, length -= 8*8)
         STORE_8Lanes( (uint64_t*)output, XOR(LOAD_8Lanes(stateAsLanes), LOAD_8Lanes((const uint64_t*)input)));
@@ -495,11 +496,11 @@ const uint64_t KeccakP1600RoundConstants[24] = {
 #error "Unrolling is not correctly specified!"
 #endif
 
-void KeccakP1600_Permute_Nrounds(void *state, unsigned int nrounds)
+void KeccakP1600_Permute_Nrounds(KeccakP1600_plain64_state *state, unsigned int nrounds)
 {
     KeccakP_DeclareVars
     unsigned int i;
-    uint64_t *stateAsLanes = (uint64_t*)state;
+    uint64_t *stateAsLanes = state->A;
 
     copyFromState(stateAsLanes);
     if ((nrounds & 1) != 0) {
@@ -522,13 +523,13 @@ void KeccakP1600_Permute_Nrounds(void *state, unsigned int nrounds)
 
 /* ---------------------------------------------------------------- */
 
-void KeccakP1600_Permute_12rounds(void *state)
+void KeccakP1600_Permute_12rounds(KeccakP1600_plain64_state *state)
 {
     KeccakP_DeclareVars
     #ifndef KeccakP1600_fullUnrolling
     unsigned int i;
     #endif
-    uint64_t *stateAsLanes = (uint64_t*)state;
+    uint64_t *stateAsLanes = state->A;
 
     copyFromState(stateAsLanes);
     rounds12;
@@ -537,20 +538,20 @@ void KeccakP1600_Permute_12rounds(void *state)
 
 /* ---------------------------------------------------------------- */
 
-void KeccakP1600_Permute_24rounds(void *state)
+void KeccakP1600_Permute_24rounds(KeccakP1600_plain64_state *state)
 {
     KeccakP_DeclareVars
     #ifndef KeccakP1600_fullUnrolling
     unsigned int i;
     #endif
-    uint64_t *stateAsLanes = (uint64_t*)state;
+    uint64_t *stateAsLanes = state->A;
 
     copyFromState(stateAsLanes);
     rounds24;
     copyToState(stateAsLanes);
 }
 
-size_t KeccakF1600_FastLoop_Absorb(void *state, unsigned int laneCount, const unsigned char *data, size_t dataByteLen)
+size_t KeccakF1600_FastLoop_Absorb(KeccakP1600_plain64_state *state, unsigned int laneCount, const unsigned char *data, size_t dataByteLen)
 {
     size_t originalDataByteLen = dataByteLen;
 
@@ -559,7 +560,7 @@ size_t KeccakF1600_FastLoop_Absorb(void *state, unsigned int laneCount, const un
         #ifndef KeccakP1600_fullUnrolling
         unsigned int i;
         #endif
-        uint64_t *stateAsLanes = (uint64_t*)state;
+        uint64_t *stateAsLanes = state->A;
         uint64_t *inDataAsLanes = (uint64_t*)data;
 
         copyFromState(stateAsLanes);
@@ -586,7 +587,7 @@ size_t KeccakF1600_FastLoop_Absorb(void *state, unsigned int laneCount, const un
     return originalDataByteLen - dataByteLen;
 }
 
-size_t KeccakP1600_12rounds_FastLoop_Absorb(void *state, unsigned int laneCount, const unsigned char *data, size_t dataByteLen)
+size_t KeccakP1600_12rounds_FastLoop_Absorb(KeccakP1600_plain64_state *state, unsigned int laneCount, const unsigned char *data, size_t dataByteLen)
 {
     size_t originalDataByteLen = dataByteLen;
 
@@ -595,7 +596,7 @@ size_t KeccakP1600_12rounds_FastLoop_Absorb(void *state, unsigned int laneCount,
         #if !defined(KeccakP1600_fullUnrolling) && (KeccakP1600_unrolling < 12)
         unsigned int i;
         #endif
-        uint64_t *stateAsLanes = (uint64_t*)state;
+        uint64_t *stateAsLanes = state->A;
         uint64_t *inDataAsLanes = (uint64_t*)data;
 
         copyFromState(stateAsLanes);

data/ext/digest/kangarootwelve/XKCP/lib/low/KeccakP-1600/AVX512/C/KeccakP-1600-SnP.h

@@ -22,26 +22,31 @@ Please refer to SnP-documentation.h for more details.
 #define _KeccakP_1600_SnP_h_
 
 #include <stddef.h>
+#include <stdint.h>
 #include "KeccakP-1600-AVX512-config.h"
 
+typedef struct {
+    uint64_t A[25];
+} KeccakP1600_plain64_state;
+
+typedef KeccakP1600_plain64_state KeccakP1600_state;
+
 #define KeccakP1600_implementation      "AVX-512 optimized implementation (" KeccakP1600_implementation_config ")"
-#define KeccakP1600_stateSizeInBytes    200
-#define KeccakP1600_stateAlignment      8
 #define KeccakF1600_FastLoop_supported
 #define KeccakP1600_12rounds_FastLoop_supported
 
 #define KeccakP1600_StaticInitialize()
-void KeccakP1600_Initialize(void *state);
+void KeccakP1600_Initialize(KeccakP1600_plain64_state *state);
 #define KeccakP1600_AddByte(state, byte, offset) ((unsigned char*)(state))[offset] ^= (byte)
-void KeccakP1600_AddBytes(void *state, const unsigned char *data, unsigned int offset, unsigned int length);
-void KeccakP1600_OverwriteBytes(void *state, const unsigned char *data, unsigned int offset, unsigned int length);
-void KeccakP1600_OverwriteWithZeroes(void *state, unsigned int byteCount);
-void KeccakP1600_Permute_Nrounds(void *state, unsigned int nrounds);
-void KeccakP1600_Permute_12rounds(void *state);
-void KeccakP1600_Permute_24rounds(void *state);
-void KeccakP1600_ExtractBytes(const void *state, unsigned char *data, unsigned int offset, unsigned int length);
-void KeccakP1600_ExtractAndAddBytes(const void *state, const unsigned char *input, unsigned char *output, unsigned int offset, unsigned int length);
-size_t KeccakF1600_FastLoop_Absorb(void *state, unsigned int laneCount, const unsigned char *data, size_t dataByteLen);
-size_t KeccakP1600_12rounds_FastLoop_Absorb(void *state, unsigned int laneCount, const unsigned char *data, size_t dataByteLen);
+void KeccakP1600_AddBytes(KeccakP1600_plain64_state *state, const unsigned char *data, unsigned int offset, unsigned int length);
+void KeccakP1600_OverwriteBytes(KeccakP1600_plain64_state *state, const unsigned char *data, unsigned int offset, unsigned int length);
+void KeccakP1600_OverwriteWithZeroes(KeccakP1600_plain64_state *state, unsigned int byteCount);
+void KeccakP1600_Permute_Nrounds(KeccakP1600_plain64_state *state, unsigned int nrounds);
+void KeccakP1600_Permute_12rounds(KeccakP1600_plain64_state *state);
+void KeccakP1600_Permute_24rounds(KeccakP1600_plain64_state *state);
+void KeccakP1600_ExtractBytes(const KeccakP1600_plain64_state *state, unsigned char *data, unsigned int offset, unsigned int length);
+void KeccakP1600_ExtractAndAddBytes(const KeccakP1600_plain64_state *state, const unsigned char *input, unsigned char *output, unsigned int offset, unsigned int length);
+size_t KeccakF1600_FastLoop_Absorb(KeccakP1600_plain64_state *state, unsigned int laneCount, const unsigned char *data, size_t dataByteLen);
+size_t KeccakP1600_12rounds_FastLoop_Absorb(KeccakP1600_plain64_state *state, unsigned int laneCount, const unsigned char *data, size_t dataByteLen);
 
 #endif

data/ext/digest/kangarootwelve/XKCP/lib/low/KeccakP-1600/AVX512/KeccakP-1600-AVX512.s

@@ -20,7 +20,7 @@
 
 # -----------------------------------------------------------------------------
 #
-# void KeccakP1600_Initialize(void *state);
+# void KeccakP1600_Initialize(KeccakP1600_plain64_state *state);
 #
 .globl  KeccakP1600_Initialize
 .globl _KeccakP1600_Initialize
@@ -42,7 +42,7 @@ _KeccakP1600_Initialize:
 
 # -----------------------------------------------------------------------------
 #
-# void KeccakP1600_AddByte(void *state, unsigned char data, unsigned int offset);
+# void KeccakP1600_AddByte(KeccakP1600_plain64_state *state, unsigned char data, unsigned int offset);
 #                                %rdi                 %rsi               %rdx
 #!!
 #.globl  KeccakP1600_AddByte
@@ -61,7 +61,7 @@ _KeccakP1600_Initialize:
 
 # -----------------------------------------------------------------------------
 #
-# void KeccakP1600_AddBytes(void *state, const unsigned char *data, unsigned int offset, unsigned int length);
+# void KeccakP1600_AddBytes(KeccakP1600_plain64_state *state, const unsigned char *data, unsigned int offset, unsigned int length);
 #                                %rdi                         %rsi               %rdx                 %rcx
 #
 .globl  KeccakP1600_AddBytes
@@ -116,7 +116,7 @@ KeccakP1600_AddBytes_Exit:
 
 # -----------------------------------------------------------------------------
 #
-# void KeccakP1600_OverwriteBytes(void *state, const unsigned char *data, unsigned int offset, unsigned int length);
+# void KeccakP1600_OverwriteBytes(KeccakP1600_plain64_state *state, const unsigned char *data, unsigned int offset, unsigned int length);
 #                                       %rdi                        %rsi               %rdx                 %rcx
 #
 .globl  KeccakP1600_OverwriteBytes
@@ -171,7 +171,7 @@ KeccakP1600_OverwriteBytes_Exit:
 
 # -----------------------------------------------------------------------------
 #
-# void KeccakP1600_OverwriteWithZeroes(void *state, unsigned int byteCount);
+# void KeccakP1600_OverwriteWithZeroes(KeccakP1600_plain64_state *state, unsigned int byteCount);
 #                                            %rdi                %rsi
 #
 .globl  KeccakP1600_OverwriteWithZeroes
@@ -207,7 +207,7 @@ KeccakP1600_OverwriteWithZeroes_Exit:
 
 # -----------------------------------------------------------------------------
 #
-# void KeccakP1600_ExtractBytes(const void *state, unsigned char *data, unsigned int offset, unsigned int length);
+# void KeccakP1600_ExtractBytes(const KeccakP1600_plain64_state *state, unsigned char *data, unsigned int offset, unsigned int length);
 #                                           %rdi                  %rsi               %rdx                 %rcx
 #
 .globl  KeccakP1600_ExtractBytes
@@ -262,7 +262,7 @@ KeccakP1600_ExtractBytes_Exit:
 
 # -----------------------------------------------------------------------------
 #
-# void KeccakP1600_ExtractAndAddBytes(const void *state, const unsigned char *input, unsigned char *output, unsigned int offset, unsigned int length);
+# void KeccakP1600_ExtractAndAddBytes(const KeccakP1600_plain64_state *state, const unsigned char *input, unsigned char *output, unsigned int offset, unsigned int length);
 #                                                 %rdi                        %rsi                  %rdx                 %rcx                  %r8
 #
 .globl  KeccakP1600_ExtractAndAddBytes
@@ -445,7 +445,7 @@ __KeccakF1600:
 
 # -----------------------------------------------------------------------------
 #
-# void KeccakP1600_Permute_24rounds(void *state);
+# void KeccakP1600_Permute_24rounds(KeccakP1600_plain64_state *state);
 #                                        %rdi
 #
 .globl  KeccakP1600_Permute_24rounds
@@ -507,7 +507,7 @@ _KeccakP1600_Permute_24rounds:
 
 # -----------------------------------------------------------------------------
 #
-# void KeccakP1600_Permute_12rounds(void *state);
+# void KeccakP1600_Permute_12rounds(KeccakP1600_plain64_state *state);
 #                                        %rdi
 #
 .globl  KeccakP1600_Permute_12rounds
@@ -569,7 +569,7 @@ _KeccakP1600_Permute_12rounds:
 
 # -----------------------------------------------------------------------------
 #
-# void KeccakP1600_Permute_Nrounds(void *state, unsigned int nrounds);
+# void KeccakP1600_Permute_Nrounds(KeccakP1600_plain64_state *state, unsigned int nrounds);
 #                                        %rdi                %rsi
 #
 .globl  KeccakP1600_Permute_Nrounds
@@ -702,7 +702,7 @@ _KeccakP1600_Permute_Nrounds:
 
 # -----------------------------------------------------------------------------
 #
-# size_t KeccakF1600_FastLoop_Absorb(void *state, unsigned int laneCount, const unsigned char *data, size_t dataByteLen);
+# size_t KeccakF1600_FastLoop_Absorb(KeccakP1600_plain64_state *state, unsigned int laneCount, const unsigned char *data, size_t dataByteLen);
 #                                          %rdi                %rsi                            %rdx         %rcx
 #
 .globl  KeccakF1600_FastLoop_Absorb
@@ -841,7 +841,7 @@ KeccakF1600_FastLoop_Absorb_LanesAddLoop:
 
 # -----------------------------------------------------------------------------
 #
-# size_t KeccakP1600_12rounds_FastLoop_Absorb(void *state, unsigned int laneCount, const unsigned char *data, size_t dataByteLen);
+# size_t KeccakP1600_12rounds_FastLoop_Absorb(KeccakP1600_plain64_state *state, unsigned int laneCount, const unsigned char *data, size_t dataByteLen);
 #                                          %rdi                %rsi                            %rdx         %rcx
 #
 .globl  KeccakP1600_12rounds_FastLoop_Absorb

data/ext/digest/kangarootwelve/XKCP/lib/low/KeccakP-1600/AVX512/KeccakP-1600-SnP.h

@@ -22,6 +22,7 @@ Please refer to SnP-documentation.h for more details.
 #define _KeccakP_1600_SnP_h_
 
 #include <stddef.h>
+#include <stdint.h>
 
 #ifdef __MINGW32__
 #define FORCE_SYSV __attribute__((sysv_abi))
@@ -29,25 +30,28 @@ Please refer to SnP-documentation.h for more details.
 #define FORCE_SYSV
 #endif
 
+typedef struct {
+    uint64_t A[25];
+} KeccakP1600_plain64_state;
+
+typedef KeccakP1600_plain64_state KeccakP1600_state;
+
 #define KeccakP1600_implementation      "AVX-512 optimized implementation"
-#define KeccakP1600_stateSizeInBytes    200
-#define KeccakP1600_stateAlignment      8
 #define KeccakF1600_FastLoop_supported
 #define KeccakP1600_12rounds_FastLoop_supported
 
 #define KeccakP1600_StaticInitialize()
-FORCE_SYSV void KeccakP1600_Initialize(void *state);
-//void KeccakP1600_AddByte(void *state, unsigned char data, unsigned int offset);
+FORCE_SYSV void KeccakP1600_Initialize(KeccakP1600_plain64_state *state);
 #define KeccakP1600_AddByte(state, byte, offset) ((unsigned char*)(state))[(offset)] ^= (byte)
-FORCE_SYSV void KeccakP1600_AddBytes(void *state, const unsigned char *data, unsigned int offset, unsigned int length);
-FORCE_SYSV void KeccakP1600_OverwriteBytes(void *state, const unsigned char *data, unsigned int offset, unsigned int length);
-FORCE_SYSV void KeccakP1600_OverwriteWithZeroes(void *state, unsigned int byteCount);
-FORCE_SYSV void KeccakP1600_Permute_Nrounds(void *state, unsigned int nrounds);
-FORCE_SYSV void KeccakP1600_Permute_12rounds(void *state);
-FORCE_SYSV void KeccakP1600_Permute_24rounds(void *state);
-FORCE_SYSV void KeccakP1600_ExtractBytes(const void *state, unsigned char *data, unsigned int offset, unsigned int length);
-FORCE_SYSV void KeccakP1600_ExtractAndAddBytes(const void *state, const unsigned char *input, unsigned char *output, unsigned int offset, unsigned int length);
-FORCE_SYSV size_t KeccakF1600_FastLoop_Absorb(void *state, unsigned int laneCount, const unsigned char *data, size_t dataByteLen);
-FORCE_SYSV size_t KeccakP1600_12rounds_FastLoop_Absorb(void *state, unsigned int laneCount, const unsigned char *data, size_t dataByteLen);
+FORCE_SYSV void KeccakP1600_AddBytes(KeccakP1600_plain64_state *state, const unsigned char *data, unsigned int offset, unsigned int length);
+FORCE_SYSV void KeccakP1600_OverwriteBytes(KeccakP1600_plain64_state *state, const unsigned char *data, unsigned int offset, unsigned int length);
+FORCE_SYSV void KeccakP1600_OverwriteWithZeroes(KeccakP1600_plain64_state *state, unsigned int byteCount);
+FORCE_SYSV void KeccakP1600_Permute_Nrounds(KeccakP1600_plain64_state *state, unsigned int nrounds);
+FORCE_SYSV void KeccakP1600_Permute_12rounds(KeccakP1600_plain64_state *state);
+FORCE_SYSV void KeccakP1600_Permute_24rounds(KeccakP1600_plain64_state *state);
+FORCE_SYSV void KeccakP1600_ExtractBytes(const KeccakP1600_plain64_state *state, unsigned char *data, unsigned int offset, unsigned int length);
+FORCE_SYSV void KeccakP1600_ExtractAndAddBytes(const KeccakP1600_plain64_state *state, const unsigned char *input, unsigned char *output, unsigned int offset, unsigned int length);
+FORCE_SYSV size_t KeccakF1600_FastLoop_Absorb(KeccakP1600_plain64_state *state, unsigned int laneCount, const unsigned char *data, size_t dataByteLen);
+FORCE_SYSV size_t KeccakP1600_12rounds_FastLoop_Absorb(KeccakP1600_plain64_state *state, unsigned int laneCount, const unsigned char *data, size_t dataByteLen);
 
 #endif

data/ext/digest/kangarootwelve/XKCP/lib/low/KeccakP-1600/XOP/KeccakP-1600-SnP.h

@@ -21,24 +21,30 @@ Please refer to SnP-documentation.h for more details.
 #ifndef _KeccakP_1600_SnP_h_
 #define _KeccakP_1600_SnP_h_
 
+#include <stddef.h>
+#include <stdint.h>
+#include "align.h"
 #include "KeccakP-1600-XOP-config.h"
 
+typedef struct {
+    ALIGN(16) uint64_t A[25];
+} KeccakP1600_align128plain64_state;
+
+typedef KeccakP1600_align128plain64_state KeccakP1600_state;
+
 #define KeccakP1600_implementation      "implementation optimized for XOP (" KeccakP1600_implementation_config ")"
-#define KeccakP1600_stateSizeInBytes    200
 #define KeccakP1600_stateAlignment      16
 
-#include <stddef.h>
-
 #define KeccakP1600_StaticInitialize()
-void KeccakP1600_Initialize(void *state);
-void KeccakP1600_AddByte(void *state, unsigned char data, unsigned int offset);
-void KeccakP1600_AddBytes(void *state, const unsigned char *data, unsigned int offset, unsigned int length);
-void KeccakP1600_OverwriteBytes(void *state, const unsigned char *data, unsigned int offset, unsigned int length);
-void KeccakP1600_OverwriteWithZeroes(void *state, unsigned int byteCount);
-void KeccakP1600_Permute_Nrounds(void *state, unsigned int nrounds);
-void KeccakP1600_Permute_12rounds(void *state);
-void KeccakP1600_Permute_24rounds(void *state);
-void KeccakP1600_ExtractBytes(const void *state, unsigned char *data, unsigned int offset, unsigned int length);
-void KeccakP1600_ExtractAndAddBytes(const void *state, const unsigned char *input, unsigned char *output, unsigned int offset, unsigned int length);
+void KeccakP1600_Initialize(KeccakP1600_align128plain64_state *state);
+void KeccakP1600_AddByte(KeccakP1600_align128plain64_state *state, unsigned char data, unsigned int offset);
+void KeccakP1600_AddBytes(KeccakP1600_align128plain64_state *state, const unsigned char *data, unsigned int offset, unsigned int length);
+void KeccakP1600_OverwriteBytes(KeccakP1600_align128plain64_state *state, const unsigned char *data, unsigned int offset, unsigned int length);
+void KeccakP1600_OverwriteWithZeroes(KeccakP1600_align128plain64_state *state, unsigned int byteCount);
+void KeccakP1600_Permute_Nrounds(KeccakP1600_align128plain64_state *state, unsigned int nrounds);
+void KeccakP1600_Permute_12rounds(KeccakP1600_align128plain64_state *state);
+void KeccakP1600_Permute_24rounds(KeccakP1600_align128plain64_state *state);
+void KeccakP1600_ExtractBytes(const KeccakP1600_align128plain64_state *state, unsigned char *data, unsigned int offset, unsigned int length);
+void KeccakP1600_ExtractAndAddBytes(const KeccakP1600_align128plain64_state *state, const unsigned char *input, unsigned char *output, unsigned int offset, unsigned int length);
 
 #endif