Add LSH dynamic dispatch (PR #1032)

This commit adds dynamic dispatch to LSH. The implementation pivots on AVX2 and SSSE3.

1 files changed, 223 insertions, 581 deletions

diff --git a/lsh256.cpp b/lsh256.cpp
index 6fcc58c6..0f895bb4 100644
--- a/lsh256.cpp
+++ b/lsh256.cpp
@@ -4,149 +4,20 @@
 //           see https://seed.kisa.or.kr/kisa/algorithm/EgovLSHInfo.do

 //           and https://seed.kisa.or.kr/kisa/Board/22/detailView.do.

 

-// The source file below uses GCC's function multiversioning to

-// speed up a rotate. When the rotate is performed with the SSE

-// unit there's a 2.5 to 3.0 cpb profit. AVX and AVX2 code paths

-// slow down with multiversioning. It looks like GCC inserts calls

-// to zeroupper() in each AVX function rather than deferring until

-// the end of Restart(), Update() or Final(). That mistake costs

-// about 3 cpb.

-

-// Function multiversioning does not work with Clang. Enabling it for

-// LLVM Clang 7.0 and above resulted in linker errors. Also see

-// https://bugs.llvm.org/show_bug.cgi?id=50025.

-

-// We are hitting some sort of GCC bug in the LSH256 AVX2 code path.

-// Clang is OK on the AVX2 code path. When we enable AVX2 for

-// rotate_msg_gamma, msg_exp_even and msg_exp_odd, then GCC arrives

-// at the wrong result. Making any one of the functions SSE2 clears

-// the problem. See CRYPTOPP_WORKAROUND_AVX2_BUG below.

-

-// TODO: cut-over to a *_simd.cpp file for proper runtime dispatching.

+// We are hitting some sort of GCC bug in the LSH AVX2 code path.

+// Clang is OK on the AVX2 code path. We believe it is GCC Issue

+// 82735, https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82735. It

+// makes using zeroupper a little tricky.

 

 #include "pch.h"

 #include "config.h"

 

 #include "lsh.h"

+#include "cpu.h"

 #include "misc.h"

 

-// Only enable the intrinsics for 64-bit machines

-#ifndef CRYPTOPP_DISABLE_ASM

-# if (defined(__SSE2__) && defined(__amd64__)) || (defined(_MSC_VER) && defined(_M_X64))

-#  define CRYPTOPP_LSH256_SSE2_AVAILABLE 1

-# endif

-# if defined(__SSSE3__) && defined(__amd64__)

-#  define CRYPTOPP_LSH256_SSSE3_AVAILABLE 1

-# endif

-# if defined(__XOP__) && defined(__amd64__)

-#  define CRYPTOPP_LSH256_XOP_AVAILABLE 1

-# endif

-# if defined(__AVX__) && defined(__amd64__)

-#  define CRYPTOPP_LSH256_AVX_AVAILABLE 1

-# endif

-# if defined(__AVX2__) && defined(__amd64__)

-#  define CRYPTOPP_LSH256_AVX2_AVAILABLE 1

-# endif

-#endif

-

-#if defined(CRYPTOPP_LSH256_SSE2_AVAILABLE)

-# include <emmintrin.h>

-#endif

-

-#if defined(CRYPTOPP_LSH256_SSSE3_AVAILABLE)

-# include <tmmintrin.h>

-#endif

-

-#if defined(CRYPTOPP_LSH256_XOP_AVAILABLE)

-# include <ammintrin.h>

-#endif

-

-#if defined(CRYPTOPP_LSH256_AVX_AVAILABLE)

-# include <immintrin.h>

-#endif

-

-#if defined(CRYPTOPP_HAVE_ATTRIBUTE_TARGET)

-# include <x86intrin.h>

-#endif

-

-#if defined(__GNUC__) && defined(__amd64__)

-# include <x86intrin.h>

-#endif

-

-// Use GCC_VERSION to avoid Clang, ICC and other impostors

-#if defined(CRYPTOPP_GCC_VERSION)

-# define CRYPTOPP_WORKAROUND_AVX2_BUG 1

-#endif

-

 ANONYMOUS_NAMESPACE_BEGIN

 

-using CryptoPP::byte;

-using CryptoPP::word32;

-using CryptoPP::rotlFixed;

-using CryptoPP::rotlConstant;

-

-using CryptoPP::GetBlock;

-using CryptoPP::LittleEndian;

-using CryptoPP::ConditionalByteReverse;

-using CryptoPP::LITTLE_ENDIAN_ORDER;

-

-typedef byte lsh_u8;

-typedef word32 lsh_u32;

-typedef word32 lsh_uint;

-typedef word32 lsh_err;

-typedef word32 lsh_type;

-

-struct LSH256_Context

-{

-	LSH256_Context(word32* state, word32 algType, word32& remainingBitLength) :

-		cv_l(state+0), cv_r(state+8), sub_msgs(state+16),

-		last_block(reinterpret_cast<byte*>(state+48)) ,

-		remain_databitlen(remainingBitLength), algtype(algType) {}

-

-	lsh_u32* cv_l;  // start of our state block

-	lsh_u32* cv_r;

-	lsh_u32* sub_msgs;

-	lsh_u8*  last_block;

-	lsh_u32& remain_databitlen;

-	lsh_type algtype;

-};

-

-struct LSH256_Internal

-{

-	LSH256_Internal(word32* state) :

-		submsg_e_l(state+16), submsg_e_r(state+24),

-		submsg_o_l(state+32), submsg_o_r(state+40) { }

-

-	lsh_u32* submsg_e_l; /* even left sub-message  */

-	lsh_u32* submsg_e_r; /* even right sub-message */

-	lsh_u32* submsg_o_l; /* odd left sub-message   */

-	lsh_u32* submsg_o_r; /* odd right sub-message  */

-};

-

-#if defined(CRYPTOPP_LSH256_AVX_AVAILABLE)

-// Zero the upper 128 bits of all YMM registers

-// on entry and exit. It avoids AVX state

-// transition penalties when saving state.

-struct AVX_Cleanup

-{

-	AVX_Cleanup() {

-		_mm256_zeroupper();

-	}

-	~AVX_Cleanup() {

-		_mm256_zeroupper();

-	}

-};

-#endif

-

-// error: '_mm256_set_m128i' was not declared in this scope?

-#if defined(CRYPTOPP_LSH256_AVX_AVAILABLE)

-inline __m256i _MM256_SET_M128I(__m128i hi, __m128i lo)

-{

-	return _mm256_insertf128_si256 (

-		_mm256_castsi128_si256(lo), hi, 1);

-}

-#endif

-

 /* LSH Constants */

 

 const unsigned int LSH256_MSG_BLK_BYTE_LEN = 128;

@@ -180,55 +51,44 @@ const unsigned int LSH_SUCCESS = 0x0;
 const unsigned int LSH_ERR_INVALID_DATABITLEN = 0x2403;

 const unsigned int LSH_ERR_INVALID_STATE = 0x2404;

 

-/* LSH AlgType Macro */

+/* Index into our state array */

 

-inline bool LSH_IS_LSH512(lsh_uint val) {

-	return (val & 0xf0000) == 0;

-}

-

-inline lsh_uint LSH_GET_SMALL_HASHBIT(lsh_uint val) {

-	return val >> 24;

-}

+const unsigned int AlgorithmType = 80;

+const unsigned int RemainingBits = 81;

 

-inline lsh_uint LSH_GET_HASHBYTE(lsh_uint val) {

-	return val & 0xffff;

-}

-

-inline lsh_uint LSH_GET_HASHBIT(lsh_uint val) {

-	return (LSH_GET_HASHBYTE(val) << 3) - LSH_GET_SMALL_HASHBIT(val);

-}

-

-inline lsh_u32 loadLE32(lsh_u32 v) {

-	return ConditionalByteReverse(LITTLE_ENDIAN_ORDER, v);

-}

+NAMESPACE_END

 

-lsh_u32 ROTL(lsh_u32 x, lsh_u32 r) {

-	return rotlFixed(x, r);

-}

+NAMESPACE_BEGIN(CryptoPP)

+NAMESPACE_BEGIN(LSH)

 

 /* -------------------------------------------------------- *

 * LSH: iv

 * -------------------------------------------------------- */

 

+//extern const word32 LSH256_IV224[CV_WORD_LEN];

+//extern const word32 LSH256_IV256[CV_WORD_LEN];

+//extern const word32 LSH256_StepConstants[CONST_WORD_LEN * NUM_STEPS];

+

 CRYPTOPP_ALIGN_DATA(32)

-const lsh_u32 g_IV224[CV_WORD_LEN] = {

+extern

+const word32 LSH256_IV224[CV_WORD_LEN] = {

 	0x068608D3, 0x62D8F7A7, 0xD76652AB, 0x4C600A43, 0xBDC40AA8, 0x1ECA0B68, 0xDA1A89BE, 0x3147D354,

-	0x707EB4F9, 0xF65B3862, 0x6B0B2ABE, 0x56B8EC0A, 0xCF237286, 0xEE0D1727, 0x33636595, 0x8BB8D05F,

+	0x707EB4F9, 0xF65B3862, 0x6B0B2ABE, 0x56B8EC0A, 0xCF237286, 0xEE0D1727, 0x33636595, 0x8BB8D05F

 };

 

 CRYPTOPP_ALIGN_DATA(32)

-const lsh_u32 g_IV256[CV_WORD_LEN] = {

+extern

+const word32 LSH256_IV256[CV_WORD_LEN] = {

 	0x46a10f1f, 0xfddce486, 0xb41443a8, 0x198e6b9d, 0x3304388d, 0xb0f5a3c7, 0xb36061c4, 0x7adbd553,

 	0x105d5378, 0x2f74de54, 0x5c2f2d95, 0xf2553fbe, 0x8051357a, 0x138668c8, 0x47aa4484, 0xe01afb41

 };

 

-const lsh_uint g_gamma256[8] = { 0, 8, 16, 24, 24, 16, 8, 0 };

-

 /* -------------------------------------------------------- *

 * LSH: step constants

 * -------------------------------------------------------- */

 

-const lsh_u32 g_StepConstants[CONST_WORD_LEN * NUM_STEPS] = {

+extern

+const word32 LSH256_StepConstants[CONST_WORD_LEN * NUM_STEPS] = {

 	0x917caf90, 0x6c1b10a2, 0x6f352943, 0xcf778243, 0x2ceb7472, 0x29e96ff2, 0x8a9ba428, 0x2eeb2642,

 	0x0e2c4021, 0x872bb30e, 0xa45e6cb2, 0x46f9c612, 0x185fe69e, 0x1359621b, 0x263fccb2, 0x1a116870,

 	0x3a6c612f, 0xb2dec195, 0x02cb1f56, 0x40bfd858, 0x784684b6, 0x6cbb7d2e, 0x660c7ed8, 0x2b79d88a,

@@ -257,6 +117,87 @@ const lsh_u32 g_StepConstants[CONST_WORD_LEN * NUM_STEPS] = {
 	0x592c0f3b, 0x947c5f77, 0x6fff49b9, 0xf71a7e5a, 0x1de8c0f5, 0xc2569600, 0xc4e4ac8c, 0x823c9ce1

 };

 

+NAMESPACE_END  // LSH

+NAMESPACE_END  // Crypto++

+

+ANONYMOUS_NAMESPACE_BEGIN

+

+using CryptoPP::byte;

+using CryptoPP::word32;

+using CryptoPP::rotlFixed;

+using CryptoPP::rotlConstant;

+

+using CryptoPP::GetBlock;

+using CryptoPP::LittleEndian;

+using CryptoPP::ConditionalByteReverse;

+using CryptoPP::LITTLE_ENDIAN_ORDER;

+

+using CryptoPP::LSH::LSH256_IV224;

+using CryptoPP::LSH::LSH256_IV256;

+using CryptoPP::LSH::LSH256_StepConstants;

+

+typedef byte lsh_u8;

+typedef word32 lsh_u32;

+typedef word32 lsh_uint;

+typedef word32 lsh_err;

+typedef word32 lsh_type;

+

+struct LSH256_Context

+{

+	LSH256_Context(word32* state, word32 algType, word32& remainingBitLength) :

+		cv_l(state+0), cv_r(state+8), sub_msgs(state+16),

+		last_block(reinterpret_cast<byte*>(state+48)),

+		remain_databitlen(remainingBitLength),

+		alg_type(static_cast<lsh_type>(algType)) {}

+

+	lsh_u32* cv_l;  // start of our state block

+	lsh_u32* cv_r;

+	lsh_u32* sub_msgs;

+	lsh_u8*  last_block;

+	lsh_u32& remain_databitlen;

+	lsh_type alg_type;

+};

+

+struct LSH256_Internal

+{

+	LSH256_Internal(word32* state) :

+		submsg_e_l(state+16), submsg_e_r(state+24),

+		submsg_o_l(state+32), submsg_o_r(state+40) { }

+

+	lsh_u32* submsg_e_l; /* even left sub-message  */

+	lsh_u32* submsg_e_r; /* even right sub-message */

+	lsh_u32* submsg_o_l; /* odd left sub-message   */

+	lsh_u32* submsg_o_r; /* odd right sub-message  */

+};

+

+const word32 g_gamma256[8] = { 0, 8, 16, 24, 24, 16, 8, 0 };

+

+/* LSH AlgType Macro */

+

+inline bool LSH_IS_LSH512(lsh_uint val) {

+	return (val & 0xf0000) == 0;

+}

+

+inline lsh_uint LSH_GET_SMALL_HASHBIT(lsh_uint val) {

+	return val >> 24;

+}

+

+inline lsh_uint LSH_GET_HASHBYTE(lsh_uint val) {

+	return val & 0xffff;

+}

+

+inline lsh_uint LSH_GET_HASHBIT(lsh_uint val) {

+	return (LSH_GET_HASHBYTE(val) << 3) - LSH_GET_SMALL_HASHBIT(val);

+}

+

+inline lsh_u32 loadLE32(lsh_u32 v) {

+	return ConditionalByteReverse(LITTLE_ENDIAN_ORDER, v);

+}

+

+lsh_u32 ROTL(lsh_u32 x, lsh_u32 r) {

+	return rotlFixed(x, r);

+}

+

 // Original code relied upon unaligned lsh_u32 buffer

 inline void load_msg_blk(LSH256_Internal* i_state, const lsh_u8 msgblk[LSH256_MSG_BLK_BYTE_LEN])

 {

@@ -267,34 +208,6 @@ inline void load_msg_blk(LSH256_Internal* i_state, const lsh_u8 msgblk[LSH256_MS
 	lsh_u32* submsg_o_l = i_state->submsg_o_l;

 	lsh_u32* submsg_o_r = i_state->submsg_o_r;

 

-#if defined(CRYPTOPP_LSH256_AVX_AVAILABLE)

-	_mm256_storeu_si256(M256_CAST(submsg_e_l+0),

-		_mm256_loadu_si256(CONST_M256_CAST(msgblk+0)));

-	_mm256_storeu_si256(M256_CAST(submsg_e_r+0),

-		_mm256_loadu_si256(CONST_M256_CAST(msgblk+32)));

-	_mm256_storeu_si256(M256_CAST(submsg_o_l+0),

-		_mm256_loadu_si256(CONST_M256_CAST(msgblk+64)));

-	_mm256_storeu_si256(M256_CAST(submsg_o_r+0),

-		_mm256_loadu_si256(CONST_M256_CAST(msgblk+96)));

-

-#elif defined(CRYPTOPP_LSH256_SSE2_AVAILABLE)

-	_mm_storeu_si128(M128_CAST(submsg_e_l+0),

-		_mm_loadu_si128(CONST_M128_CAST(msgblk+0)));

-	_mm_storeu_si128(M128_CAST(submsg_e_l+4),

-		_mm_loadu_si128(CONST_M128_CAST(msgblk+16)));

-	_mm_storeu_si128(M128_CAST(submsg_e_r+0),

-		_mm_loadu_si128(CONST_M128_CAST(msgblk+32)));

-	_mm_storeu_si128(M128_CAST(submsg_e_r+4),

-		_mm_loadu_si128(CONST_M128_CAST(msgblk+48)));

-	_mm_storeu_si128(M128_CAST(submsg_o_l+0),

-		_mm_loadu_si128(CONST_M128_CAST(msgblk+64)));

-	_mm_storeu_si128(M128_CAST(submsg_o_l+4),

-		_mm_loadu_si128(CONST_M128_CAST(msgblk+80)));

-	_mm_storeu_si128(M128_CAST(submsg_o_r+0),

-		_mm_loadu_si128(CONST_M128_CAST(msgblk+96)));

-	_mm_storeu_si128(M128_CAST(submsg_o_r+4),

-		_mm_loadu_si128(CONST_M128_CAST(msgblk+112)));

-#else

 	typedef GetBlock<word32, LittleEndian, false> InBlock;

 

 	InBlock input(msgblk);

@@ -306,7 +219,6 @@ inline void load_msg_blk(LSH256_Internal* i_state, const lsh_u8 msgblk[LSH256_MS
 		(submsg_o_l[4])(submsg_o_l[5])(submsg_o_l[6])(submsg_o_l[7])

 		(submsg_o_r[0])(submsg_o_r[1])(submsg_o_r[2])(submsg_o_r[3])

 		(submsg_o_r[4])(submsg_o_r[5])(submsg_o_r[6])(submsg_o_r[7]);

-#endif

 }

 

 inline void msg_exp_even(LSH256_Internal* i_state)

@@ -318,44 +230,6 @@ inline void msg_exp_even(LSH256_Internal* i_state)
 	lsh_u32* submsg_o_l = i_state->submsg_o_l;

 	lsh_u32* submsg_o_r = i_state->submsg_o_r;

 

-#if defined(CRYPTOPP_LSH256_AVX2_AVAILABLE)

-	const __m256i mask = _mm256_set_epi32(0x1b1a1918, 0x17161514,

-		0x13121110, 0x1f1e1d1c, 0x07060504, 0x03020100, 0x0b0a0908, 0x0f0e0d0c);

-

-	_mm256_storeu_si256(M256_CAST(submsg_e_l+0), _mm256_add_epi32(

-		_mm256_loadu_si256(CONST_M256_CAST(submsg_o_l+0)),

-		_mm256_shuffle_epi8(

-			_mm256_loadu_si256(CONST_M256_CAST(submsg_e_l+0)), mask)));

-	_mm256_storeu_si256(M256_CAST(submsg_e_r+0), _mm256_add_epi32(

-		_mm256_loadu_si256(CONST_M256_CAST(submsg_o_r+0)),

-		_mm256_shuffle_epi8(

-			_mm256_loadu_si256(CONST_M256_CAST(submsg_e_r+0)), mask)));

-

-#elif defined(CRYPTOPP_LSH256_SSE2_AVAILABLE)

-	_mm_storeu_si128(M128_CAST(submsg_e_l+0), _mm_add_epi32(

-		_mm_shuffle_epi32(

-			_mm_loadu_si128(CONST_M128_CAST(submsg_o_l+0)), _MM_SHUFFLE(3,2,1,0)),

-		_mm_shuffle_epi32(

-			_mm_loadu_si128(CONST_M128_CAST(submsg_e_l+0)), _MM_SHUFFLE(1,0,2,3))));

-

-	_mm_storeu_si128(M128_CAST(submsg_e_l+4), _mm_add_epi32(

-		_mm_shuffle_epi32(

-			_mm_loadu_si128(CONST_M128_CAST(submsg_o_l+4)), _MM_SHUFFLE(3,2,1,0)),

-		_mm_shuffle_epi32(

-			_mm_loadu_si128(CONST_M128_CAST(submsg_e_l+4)), _MM_SHUFFLE(2,1,0,3))));

-

-	_mm_storeu_si128(M128_CAST(submsg_e_r+0), _mm_add_epi32(

-		_mm_shuffle_epi32(

-			_mm_loadu_si128(CONST_M128_CAST(submsg_o_r+0)), _MM_SHUFFLE(3,2,1,0)),

-		_mm_shuffle_epi32(

-			_mm_loadu_si128(CONST_M128_CAST(submsg_e_r+0)), _MM_SHUFFLE(1,0,2,3))));

-

-	_mm_storeu_si128(M128_CAST(submsg_e_r+4), _mm_add_epi32(

-		_mm_shuffle_epi32(

-			_mm_loadu_si128(CONST_M128_CAST(submsg_o_r+4)), _MM_SHUFFLE(3,2,1,0)),

-		_mm_shuffle_epi32(

-			_mm_loadu_si128(CONST_M128_CAST(submsg_e_r+4)), _MM_SHUFFLE(2,1,0,3))));

-#else

 	lsh_u32 temp;

 	temp = submsg_e_l[0];

 	submsg_e_l[0] = submsg_o_l[0] + submsg_e_l[3];

@@ -377,7 +251,6 @@ inline void msg_exp_even(LSH256_Internal* i_state)
 	submsg_e_r[7] = submsg_o_r[7] + submsg_e_r[6];

 	submsg_e_r[6] = submsg_o_r[6] + submsg_e_r[5];

 	submsg_e_r[5] = submsg_o_r[5] + temp;

-#endif

 }

 

 inline void msg_exp_odd(LSH256_Internal* i_state)

@@ -389,44 +262,6 @@ inline void msg_exp_odd(LSH256_Internal* i_state)
 	lsh_u32* submsg_o_l = i_state->submsg_o_l;

 	lsh_u32* submsg_o_r = i_state->submsg_o_r;

 

-#if defined(CRYPTOPP_LSH256_AVX2_AVAILABLE)

-	const __m256i mask = _mm256_set_epi32(0x1b1a1918, 0x17161514,

-		0x13121110, 0x1f1e1d1c, 0x07060504, 0x03020100, 0x0b0a0908, 0x0f0e0d0c);

-

-	_mm256_storeu_si256(M256_CAST(submsg_o_l+0), _mm256_add_epi32(

-		_mm256_loadu_si256(CONST_M256_CAST(submsg_e_l+0)),

-		_mm256_shuffle_epi8(

-			_mm256_loadu_si256(CONST_M256_CAST(submsg_o_l+0)), mask)));

-	_mm256_storeu_si256(M256_CAST(submsg_o_r+0), _mm256_add_epi32(

-		_mm256_loadu_si256(CONST_M256_CAST(submsg_e_r+0)),

-		_mm256_shuffle_epi8(

-			_mm256_loadu_si256(CONST_M256_CAST(submsg_o_r+0)), mask)));

-

-#elif defined(CRYPTOPP_LSH256_SSE2_AVAILABLE)

-	_mm_storeu_si128(M128_CAST(submsg_o_l+0), _mm_add_epi32(

-		_mm_shuffle_epi32(

-			_mm_loadu_si128(CONST_M128_CAST(submsg_e_l+0)), _MM_SHUFFLE(3,2,1,0)),

-		_mm_shuffle_epi32(

-			_mm_loadu_si128(CONST_M128_CAST(submsg_o_l+0)), _MM_SHUFFLE(1,0,2,3))));

-

-	_mm_storeu_si128(M128_CAST(submsg_o_l+4), _mm_add_epi32(

-		_mm_shuffle_epi32(

-			_mm_loadu_si128(CONST_M128_CAST(submsg_e_l+4)), _MM_SHUFFLE(3,2,1,0)),

-		_mm_shuffle_epi32(

-			_mm_loadu_si128(CONST_M128_CAST(submsg_o_l+4)), _MM_SHUFFLE(2,1,0,3))));

-

-	_mm_storeu_si128(M128_CAST(submsg_o_r+0), _mm_add_epi32(

-		_mm_shuffle_epi32(

-			_mm_loadu_si128(CONST_M128_CAST(submsg_e_r+0)), _MM_SHUFFLE(3,2,1,0)),

-		_mm_shuffle_epi32(

-			_mm_loadu_si128(CONST_M128_CAST(submsg_o_r+0)), _MM_SHUFFLE(1,0,2,3))));

-

-	_mm_storeu_si128(M128_CAST(submsg_o_r+4), _mm_add_epi32(

-		_mm_shuffle_epi32(

-			_mm_loadu_si128(CONST_M128_CAST(submsg_e_r+4)), _MM_SHUFFLE(3,2,1,0)),

-		_mm_shuffle_epi32(

-			_mm_loadu_si128(CONST_M128_CAST(submsg_o_r+4)), _MM_SHUFFLE(2,1,0,3))));

-#else

 	lsh_u32 temp;

 	temp = submsg_o_l[0];

 	submsg_o_l[0] = submsg_e_l[0] + submsg_o_l[3];

@@ -448,14 +283,13 @@ inline void msg_exp_odd(LSH256_Internal* i_state)
 	submsg_o_r[7] = submsg_e_r[7] + submsg_o_r[6];

 	submsg_o_r[6] = submsg_e_r[6] + submsg_o_r[5];

 	submsg_o_r[5] = submsg_e_r[5] + temp;

-#endif

 }

 

 inline void load_sc(const lsh_u32** p_const_v, size_t i)

 {

 	CRYPTOPP_ASSERT(p_const_v != NULLPTR);

 

-	*p_const_v = &g_StepConstants[i];

+	*p_const_v = &LSH256_StepConstants[i];

 }

 

 inline void msg_add_even(lsh_u32 cv_l[8], lsh_u32 cv_r[8], LSH256_Internal* i_state)

@@ -465,28 +299,6 @@ inline void msg_add_even(lsh_u32 cv_l[8], lsh_u32 cv_r[8], LSH256_Internal* i_st
 	lsh_u32* submsg_e_l = i_state->submsg_e_l;

 	lsh_u32* submsg_e_r = i_state->submsg_e_r;

 

-#if defined(CRYPTOPP_LSH256_AVX2_AVAILABLE)

-	_mm256_storeu_si256(M256_CAST(cv_l+0), _mm256_xor_si256(

-		_mm256_loadu_si256(CONST_M256_CAST(cv_l+0)),

-		_mm256_loadu_si256(CONST_M256_CAST(submsg_e_l+0))));

-	_mm256_storeu_si256(M256_CAST(cv_r+0), _mm256_xor_si256(

-		_mm256_loadu_si256(CONST_M256_CAST(cv_r+0)),

-		_mm256_loadu_si256(CONST_M256_CAST(submsg_e_r+0))));

-

-#elif defined(CRYPTOPP_LSH256_SSE2_AVAILABLE)

-	_mm_storeu_si128(M128_CAST(cv_l+0), _mm_xor_si128(

-		_mm_loadu_si128(CONST_M128_CAST(cv_l+0)),

-		_mm_loadu_si128(CONST_M128_CAST(submsg_e_l+0))));

-	_mm_storeu_si128(M128_CAST(cv_l+4), _mm_xor_si128(

-		_mm_loadu_si128(CONST_M128_CAST(cv_l+4)),

-		_mm_loadu_si128(CONST_M128_CAST(submsg_e_l+4))));

-	_mm_storeu_si128(M128_CAST(cv_r+0), _mm_xor_si128(

-		_mm_loadu_si128(CONST_M128_CAST(cv_r+0)),

-		_mm_loadu_si128(CONST_M128_CAST(submsg_e_r+0))));

-	_mm_storeu_si128(M128_CAST(cv_r+4), _mm_xor_si128(

-		_mm_loadu_si128(CONST_M128_CAST(cv_r+4)),

-		_mm_loadu_si128(CONST_M128_CAST(submsg_e_r+4))));

-#else

 	cv_l[0] ^= submsg_e_l[0];  cv_l[1] ^= submsg_e_l[1];

 	cv_l[2] ^= submsg_e_l[2];  cv_l[3] ^= submsg_e_l[3];

 	cv_l[4] ^= submsg_e_l[4];  cv_l[5] ^= submsg_e_l[5];

@@ -495,7 +307,6 @@ inline void msg_add_even(lsh_u32 cv_l[8], lsh_u32 cv_r[8], LSH256_Internal* i_st
 	cv_r[2] ^= submsg_e_r[2];  cv_r[3] ^= submsg_e_r[3];

 	cv_r[4] ^= submsg_e_r[4];  cv_r[5] ^= submsg_e_r[5];

 	cv_r[6] ^= submsg_e_r[6];  cv_r[7] ^= submsg_e_r[7];

-#endif

 }

 

 inline void msg_add_odd(lsh_u32 cv_l[8], lsh_u32 cv_r[8], LSH256_Internal* i_state)

@@ -505,28 +316,6 @@ inline void msg_add_odd(lsh_u32 cv_l[8], lsh_u32 cv_r[8], LSH256_Internal* i_sta
 	lsh_u32* submsg_o_l = i_state->submsg_o_l;

 	lsh_u32* submsg_o_r = i_state->submsg_o_r;

 

-#if defined(CRYPTOPP_LSH256_AVX2_AVAILABLE)

-	_mm256_storeu_si256(M256_CAST(cv_l), _mm256_xor_si256(

-		_mm256_loadu_si256(CONST_M256_CAST(cv_l)),

-		_mm256_loadu_si256(CONST_M256_CAST(submsg_o_l))));

-	_mm256_storeu_si256(M256_CAST(cv_r), _mm256_xor_si256(

-		_mm256_loadu_si256(CONST_M256_CAST(cv_r)),

-		_mm256_loadu_si256(CONST_M256_CAST(submsg_o_r))));

-

-#elif defined(CRYPTOPP_LSH256_SSE2_AVAILABLE)

-	_mm_storeu_si128(M128_CAST(cv_l), _mm_xor_si128(

-		_mm_loadu_si128(CONST_M128_CAST(cv_l)),

-		_mm_loadu_si128(CONST_M128_CAST(submsg_o_l))));

-	_mm_storeu_si128(M128_CAST(cv_l+4), _mm_xor_si128(

-		_mm_loadu_si128(CONST_M128_CAST(cv_l+4)),

-		_mm_loadu_si128(CONST_M128_CAST(submsg_o_l+4))));

-	_mm_storeu_si128(M128_CAST(cv_r), _mm_xor_si128(

-		_mm_loadu_si128(CONST_M128_CAST(cv_r)),

-		_mm_loadu_si128(CONST_M128_CAST(submsg_o_r))));

-	_mm_storeu_si128(M128_CAST(cv_r+4), _mm_xor_si128(

-		_mm_loadu_si128(CONST_M128_CAST(cv_r+4)),

-		_mm_loadu_si128(CONST_M128_CAST(submsg_o_r+4))));

-#else

 	cv_l[0] ^= submsg_o_l[0];  cv_l[1] ^= submsg_o_l[1];

 	cv_l[2] ^= submsg_o_l[2];  cv_l[3] ^= submsg_o_l[3];

 	cv_l[4] ^= submsg_o_l[4];  cv_l[5] ^= submsg_o_l[5];

@@ -535,24 +324,10 @@ inline void msg_add_odd(lsh_u32 cv_l[8], lsh_u32 cv_r[8], LSH256_Internal* i_sta
 	cv_r[2] ^= submsg_o_r[2];  cv_r[3] ^= submsg_o_r[3];

 	cv_r[4] ^= submsg_o_r[4];  cv_r[5] ^= submsg_o_r[5];

 	cv_r[6] ^= submsg_o_r[6];  cv_r[7] ^= submsg_o_r[7];

-#endif

 }

 

-inline void add_blk(lsh_u32 cv_l[8], const lsh_u32 cv_r[8])

+inline void add_blk(lsh_u32 cv_l[8], lsh_u32 cv_r[8])

 {

-#if defined(CRYPTOPP_LSH256_AVX2_AVAILABLE)

-	_mm256_storeu_si256(M256_CAST(cv_l), _mm256_add_epi32(

-		_mm256_loadu_si256(CONST_M256_CAST(cv_l)),

-		_mm256_loadu_si256(CONST_M256_CAST(cv_r))));

-

-#elif defined(CRYPTOPP_LSH256_SSE2_AVAILABLE)

-	_mm_storeu_si128(M128_CAST(cv_l), _mm_add_epi32(

-		_mm_loadu_si128(CONST_M128_CAST(cv_l)),

-		_mm_loadu_si128(CONST_M128_CAST(cv_r))));

-	_mm_storeu_si128(M128_CAST(cv_l+4), _mm_add_epi32(

-		_mm_loadu_si128(CONST_M128_CAST(cv_l+4)),

-		_mm_loadu_si128(CONST_M128_CAST(cv_r+4))));

-#else

 	cv_l[0] += cv_r[0];

 	cv_l[1] += cv_r[1];

 	cv_l[2] += cv_r[2];

@@ -561,31 +336,11 @@ inline void add_blk(lsh_u32 cv_l[8], const lsh_u32 cv_r[8])
 	cv_l[5] += cv_r[5];

 	cv_l[6] += cv_r[6];

 	cv_l[7] += cv_r[7];

-#endif

 }

 

 template <unsigned int R>

 inline void rotate_blk(lsh_u32 cv[8])

 {

-#if defined(CRYPTOPP_LSH256_AVX2_AVAILABLE)

-	_mm256_storeu_si256(M256_CAST(cv), _mm256_or_si256(

-		_mm256_slli_epi32(_mm256_loadu_si256(CONST_M256_CAST(cv)), R),

-		_mm256_srli_epi32(_mm256_loadu_si256(CONST_M256_CAST(cv)), 32-R)));

-

-#elif defined(CRYPTOPP_LSH256_XOP_AVAILABLE)

-	_mm_storeu_si128(M128_CAST(cv),

-		_mm_roti_epi32(_mm_loadu_si128(CONST_M128_CAST(cv)), R));

-	_mm_storeu_si128(M128_CAST(cv+4),

-		_mm_roti_epi32(_mm_loadu_si128(CONST_M128_CAST(cv+4)), R));

-

-#elif defined(CRYPTOPP_LSH256_SSE2_AVAILABLE)

-	_mm_storeu_si128(M128_CAST(cv), _mm_or_si128(

-		_mm_slli_epi32(_mm_loadu_si128(CONST_M128_CAST(cv)), R),

-		_mm_srli_epi32(_mm_loadu_si128(CONST_M128_CAST(cv)), 32-R)));

-	_mm_storeu_si128(M128_CAST(cv+4), _mm_or_si128(

-		_mm_slli_epi32(_mm_loadu_si128(CONST_M128_CAST(cv+4)), R),

-		_mm_srli_epi32(_mm_loadu_si128(CONST_M128_CAST(cv+4)), 32-R)));

-#else

 	cv[0] = rotlConstant<R>(cv[0]);

 	cv[1] = rotlConstant<R>(cv[1]);

 	cv[2] = rotlConstant<R>(cv[2]);

@@ -594,24 +349,10 @@ inline void rotate_blk(lsh_u32 cv[8])
 	cv[5] = rotlConstant<R>(cv[5]);

 	cv[6] = rotlConstant<R>(cv[6]);

 	cv[7] = rotlConstant<R>(cv[7]);

-#endif

 }

 

 inline void xor_with_const(lsh_u32 cv_l[8], const lsh_u32 const_v[8])

 {

-#if defined(CRYPTOPP_LSH256_AVX2_AVAILABLE)

-	_mm256_storeu_si256(M256_CAST(cv_l), _mm256_xor_si256(

-		_mm256_loadu_si256(CONST_M256_CAST(cv_l)),

-		_mm256_loadu_si256(CONST_M256_CAST(const_v))));

-

-#elif defined(CRYPTOPP_LSH256_SSE2_AVAILABLE)

-	_mm_storeu_si128(M128_CAST(cv_l), _mm_xor_si128(

-		_mm_loadu_si128(CONST_M128_CAST(cv_l)),

-		_mm_loadu_si128(CONST_M128_CAST(const_v))));

-	_mm_storeu_si128(M128_CAST(cv_l+4), _mm_xor_si128(

-		_mm_loadu_si128(CONST_M128_CAST(cv_l+4)),

-		_mm_loadu_si128(CONST_M128_CAST(const_v+4))));

-#else

 	cv_l[0] ^= const_v[0];

 	cv_l[1] ^= const_v[1];

 	cv_l[2] ^= const_v[2];

@@ -620,92 +361,20 @@ inline void xor_with_const(lsh_u32 cv_l[8], const lsh_u32 const_v[8])
 	cv_l[5] ^= const_v[5];

 	cv_l[6] ^= const_v[6];

 	cv_l[7] ^= const_v[7];

-#endif

-}

-

-#if defined(CRYPTOPP_LSH256_AVX2_AVAILABLE)  && !defined(CRYPTOPP_WORKAROUND_AVX2_BUG)

-inline void rotate_msg_gamma(lsh_u32 cv_r[8])

-{

-	// g_gamma256[8] = { 0, 8, 16, 24, 24, 16, 8, 0 };

-	_mm256_storeu_si256(M256_CAST(cv_r+0),

-		_mm256_shuffle_epi8(_mm256_loadu_si256(CONST_M256_CAST(cv_r+0)),

-			_mm256_set_epi8(

-				/* hi lane */ 15,14,13,12, 10,9,8,11, 5,4,7,6, 0,3,2,1,

-				/* lo lane */ 12,15,14,13, 9,8,11,10, 6,5,4,7, 3,2,1,0)));

-}

-#else  // CRYPTOPP_LSH256_AVX2_AVAILABLE

-# if defined(CRYPTOPP_HAVE_ATTRIBUTE_TARGET)

-CRYPTOPP_TARGET_SSSE3

-inline void rotate_msg_gamma(lsh_u32 cv_r[8])

-{

-	// g_gamma256[8] = { 0, 8, 16, 24, 24, 16, 8, 0 };

-	_mm_storeu_si128(M128_CAST(cv_r+0),

-		_mm_shuffle_epi8(_mm_loadu_si128(CONST_M128_CAST(cv_r+0)),

-			_mm_set_epi8(12,15,14,13, 9,8,11,10, 6,5,4,7, 3,2,1,0)));

-	_mm_storeu_si128(M128_CAST(cv_r+4),

-		_mm_shuffle_epi8(_mm_loadu_si128(CONST_M128_CAST(cv_r+4)),

-			_mm_set_epi8(15,14,13,12, 10,9,8,11, 5,4,7,6, 0,3,2,1)));

 }

-# endif

 

-CRYPTOPP_TARGET_DEFAULT

 inline void rotate_msg_gamma(lsh_u32 cv_r[8])

 {

-#if defined(CRYPTOPP_LSH256_SSSE3_AVAILABLE)

-	_mm_storeu_si128(M128_CAST(cv_r+0),

-		_mm_shuffle_epi8(_mm_loadu_si128(CONST_M128_CAST(cv_r+0)),

-			_mm_set_epi8(12,15,14,13, 9,8,11,10, 6,5,4,7, 3,2,1,0)));

-	_mm_storeu_si128(M128_CAST(cv_r+4),

-		_mm_shuffle_epi8(_mm_loadu_si128(CONST_M128_CAST(cv_r+4)),

-			_mm_set_epi8(15,14,13,12, 10,9,8,11, 5,4,7,6, 0,3,2,1)));

-

-#else

 	cv_r[1] = rotlFixed(cv_r[1], g_gamma256[1]);

 	cv_r[2] = rotlFixed(cv_r[2], g_gamma256[2]);

 	cv_r[3] = rotlFixed(cv_r[3], g_gamma256[3]);

 	cv_r[4] = rotlFixed(cv_r[4], g_gamma256[4]);

 	cv_r[5] = rotlFixed(cv_r[5], g_gamma256[5]);

 	cv_r[6] = rotlFixed(cv_r[6], g_gamma256[6]);

-#endif

 }

-#endif  // CRYPTOPP_LSH256_AVX2_AVAILABLE

 

 inline void word_perm(lsh_u32 cv_l[8], lsh_u32 cv_r[8])

 {

-#if defined(CRYPTOPP_LSH256_AVX2_AVAILABLE)

-	__m256i temp;

-	temp = _mm256_shuffle_epi32(

-		_mm256_loadu_si256(CONST_M256_CAST(cv_l)), _MM_SHUFFLE(3,1,0,2));

-	_mm256_storeu_si256(M256_CAST(cv_r),

-		_mm256_shuffle_epi32(

-			_mm256_loadu_si256(CONST_M256_CAST(cv_r)), _MM_SHUFFLE(1,2,3,0)));

-	_mm256_storeu_si256(M256_CAST(cv_l),

-		_mm256_permute2x128_si256(temp,

-			_mm256_loadu_si256(CONST_M256_CAST(cv_r)), _MM_SHUFFLE(0,3,0,1)));

-	_mm256_storeu_si256(M256_CAST(cv_r),

-		_mm256_permute2x128_si256(temp,

-			_mm256_loadu_si256(CONST_M256_CAST(cv_r)), _MM_SHUFFLE(0,2,0,0)));

-

-#elif defined(CRYPTOPP_LSH256_SSE2_AVAILABLE)

-	_mm_storeu_si128(M128_CAST(cv_l+0), _mm_shuffle_epi32(

-		_mm_loadu_si128(CONST_M128_CAST(cv_l+0)), _MM_SHUFFLE(3,1,0,2)));

-	_mm_storeu_si128(M128_CAST(cv_l+4), _mm_shuffle_epi32(

-		_mm_loadu_si128(CONST_M128_CAST(cv_l+4)), _MM_SHUFFLE(3,1,0,2)));

-	_mm_storeu_si128(M128_CAST(cv_r+0), _mm_shuffle_epi32(

-		_mm_loadu_si128(CONST_M128_CAST(cv_r+0)), _MM_SHUFFLE(1,2,3,0)));

-	_mm_storeu_si128(M128_CAST(cv_r+4), _mm_shuffle_epi32(

-		_mm_loadu_si128(CONST_M128_CAST(cv_r+4)), _MM_SHUFFLE(1,2,3,0)));

-

-	__m128i temp = _mm_loadu_si128(CONST_M128_CAST(cv_l+0));

-	_mm_storeu_si128(M128_CAST(cv_l+0),

-		_mm_loadu_si128(CONST_M128_CAST(cv_l+4)));

-	_mm_storeu_si128(M128_CAST(cv_l+4),

-		_mm_loadu_si128(CONST_M128_CAST(cv_r+4)));

-	_mm_storeu_si128(M128_CAST(cv_r+4),

-		_mm_loadu_si128(CONST_M128_CAST(cv_r+0)));

-	_mm_storeu_si128(M128_CAST(cv_r+0), temp);

-

-#else

 	lsh_u32 temp;

 	temp = cv_l[0];

 	cv_l[0] = cv_l[6];

@@ -725,8 +394,6 @@ inline void word_perm(lsh_u32 cv_l[8], lsh_u32 cv_r[8])
 	cv_l[7] = cv_r[5];

 	cv_r[5] = cv_r[3];

 	cv_r[3] = temp;

-

-#endif

 };

 

 /* -------------------------------------------------------- *

@@ -736,8 +403,6 @@ inline void word_perm(lsh_u32 cv_l[8], lsh_u32 cv_r[8])
 template <unsigned int Alpha, unsigned int Beta>

 inline void mix(lsh_u32 cv_l[8], lsh_u32 cv_r[8], const lsh_u32 const_v[8])

 {

-	CRYPTOPP_ASSERT(const_v != NULLPTR);

-

 	add_blk(cv_l, cv_r);

 	rotate_blk<Alpha>(cv_l);

 	xor_with_const(cv_l, const_v);

@@ -762,10 +427,6 @@ inline void compress(LSH256_Context* ctx, const lsh_u8 pdMsgBlk[LSH256_MSG_BLK_B
 	lsh_u32* cv_l = ctx->cv_l;

 	lsh_u32* cv_r = ctx->cv_r;

 

-#if defined(CRYPTOPP_LSH256_AVX_AVAILABLE)

-	AVX_Cleanup cleanup;

-#endif

-

 	load_msg_blk(i_state, pdMsgBlk);

 

 	msg_add_even(cv_l, cv_r, i_state);

@@ -801,23 +462,6 @@ inline void compress(LSH256_Context* ctx, const lsh_u8 pdMsgBlk[LSH256_MSG_BLK_B
 

 inline void load_iv(lsh_u32 cv_l[8], lsh_u32 cv_r[8], const lsh_u32 iv[16])

 {

-	// The IV's are 32-byte aligned so we can use aligned loads.

-#if defined(CRYPTOPP_LSH256_AVX_AVAILABLE)

-	_mm256_storeu_si256(M256_CAST(cv_l+0),

-		_mm256_load_si256(CONST_M256_CAST(iv+0)));

-	_mm256_storeu_si256(M256_CAST(cv_r+0),

-		_mm256_load_si256(CONST_M256_CAST(iv+8)));

-

-#elif defined(CRYPTOPP_LSH256_SSE2_AVAILABLE)

-	_mm_storeu_si128(M128_CAST(cv_l+ 0),

-		_mm_load_si128(CONST_M128_CAST(iv+ 0)));

-	_mm_storeu_si128(M128_CAST(cv_l+ 4),

-		_mm_load_si128(CONST_M128_CAST(iv+ 4)));

-	_mm_storeu_si128(M128_CAST(cv_r+ 0),

-		_mm_load_si128(CONST_M128_CAST(iv+ 8)));

-	_mm_storeu_si128(M128_CAST(cv_r+ 4),

-		_mm_load_si128(CONST_M128_CAST(iv+12)));

-#else

 	cv_l[0] = iv[0];

 	cv_l[1] = iv[1];

 	cv_l[2] = iv[2];

@@ -834,73 +478,36 @@ inline void load_iv(lsh_u32 cv_l[8], lsh_u32 cv_r[8], const lsh_u32 iv[16])
 	cv_r[5] = iv[13];

 	cv_r[6] = iv[14];

 	cv_r[7] = iv[15];

-#endif

 }

 

 inline void zero_iv(lsh_u32 cv_l[8], lsh_u32 cv_r[8])

 {

-#if defined(CRYPTOPP_LSH256_AVX_AVAILABLE)

-	_mm256_storeu_si256(M256_CAST(cv_l+0), _mm256_setzero_si256());

-	_mm256_storeu_si256(M256_CAST(cv_r+0), _mm256_setzero_si256());

-

-#elif defined(CRYPTOPP_LSH256_SSE2_AVAILABLE)

-	_mm_storeu_si128(M128_CAST(cv_l+0), _mm_setzero_si128());

-	_mm_storeu_si128(M128_CAST(cv_l+4), _mm_setzero_si128());

-	_mm_storeu_si128(M128_CAST(cv_r+0), _mm_setzero_si128());

-	_mm_storeu_si128(M128_CAST(cv_r+4), _mm_setzero_si128());

-#else

 	memset(cv_l, 0x00, 8*sizeof(lsh_u32));

 	memset(cv_r, 0x00, 8*sizeof(lsh_u32));

-#endif

 }

 

 inline void zero_submsgs(LSH256_Context* ctx)

 {

-	lsh_u32* sub_msgs = ctx->sub_msgs;

+	CRYPTOPP_ASSERT(ctx != NULLPTR);

 

-#if defined(CRYPTOPP_LSH256_AVX_AVAILABLE)

-	_mm256_storeu_si256(M256_CAST(sub_msgs+ 0), _mm256_setzero_si256());

-	_mm256_storeu_si256(M256_CAST(sub_msgs+ 8), _mm256_setzero_si256());

-	_mm256_storeu_si256(M256_CAST(sub_msgs+16), _mm256_setzero_si256());

-	_mm256_storeu_si256(M256_CAST(sub_msgs+24), _mm256_setzero_si256());

-

-#elif defined(CRYPTOPP_LSH256_SSE2_AVAILABLE)

-	_mm_storeu_si128(M128_CAST(sub_msgs+ 0), _mm_setzero_si128());

-	_mm_storeu_si128(M128_CAST(sub_msgs+ 4), _mm_setzero_si128());

-	_mm_storeu_si128(M128_CAST(sub_msgs+ 8), _mm_setzero_si128());

-	_mm_storeu_si128(M128_CAST(sub_msgs+12), _mm_setzero_si128());

-	_mm_storeu_si128(M128_CAST(sub_msgs+16), _mm_setzero_si128());

-	_mm_storeu_si128(M128_CAST(sub_msgs+20), _mm_setzero_si128());

-	_mm_storeu_si128(M128_CAST(sub_msgs+24), _mm_setzero_si128());

-	_mm_storeu_si128(M128_CAST(sub_msgs+28), _mm_setzero_si128());

-

-#else

+	lsh_u32* sub_msgs = ctx->sub_msgs;

 	memset(sub_msgs, 0x00, 32*sizeof(lsh_u32));

-#endif

 }

 

 inline void init224(LSH256_Context* ctx)

 {

 	CRYPTOPP_ASSERT(ctx != NULLPTR);

 

-#if defined(CRYPTOPP_LSH256_AVX_AVAILABLE)

-	AVX_Cleanup cleanup;

-#endif

-

 	zero_submsgs(ctx);

-	load_iv(ctx->cv_l, ctx->cv_r, g_IV224);

+	load_iv(ctx->cv_l, ctx->cv_r, LSH256_IV224);

 }

 

 inline void init256(LSH256_Context* ctx)

 {

 	CRYPTOPP_ASSERT(ctx != NULLPTR);

 

-#if defined(CRYPTOPP_LSH256_AVX_AVAILABLE)

-	AVX_Cleanup cleanup;

-#endif

-

 	zero_submsgs(ctx);

-	load_iv(ctx->cv_l, ctx->cv_r, g_IV256);

+	load_iv(ctx->cv_l, ctx->cv_r, LSH256_IV256);

 }

 

 /* -------------------------------------------------------- */

@@ -909,27 +516,9 @@ inline void fin(LSH256_Context* ctx)
 {

 	CRYPTOPP_ASSERT(ctx != NULLPTR);

 

-#if defined(CRYPTOPP_LSH256_AVX2_AVAILABLE)

-	AVX_Cleanup cleanup;

-#endif

-

-#if defined(CRYPTOPP_LSH256_AVX2_AVAILABLE)

-	_mm256_storeu_si256(M256_CAST(ctx->cv_l+0), _mm256_xor_si256(

-		_mm256_loadu_si256(CONST_M256_CAST(ctx->cv_l+0)),

-		_mm256_loadu_si256(CONST_M256_CAST(ctx->cv_r+0))));

-

-#elif defined(CRYPTOPP_LSH256_SSE2_AVAILABLE)

-	_mm_storeu_si128(M128_CAST(ctx->cv_l+0), _mm_xor_si128(

-		_mm_loadu_si128(CONST_M128_CAST(ctx->cv_l+0)),

-		_mm_loadu_si128(CONST_M128_CAST(ctx->cv_r+0))));

-	_mm_storeu_si128(M128_CAST(ctx->cv_l+4), _mm_xor_si128(

-		_mm_loadu_si128(CONST_M128_CAST(ctx->cv_l+4)),

-		_mm_loadu_si128(CONST_M128_CAST(ctx->cv_r+4))));

-#else

 	for (size_t i = 0; i < HASH_VAL_MAX_WORD_LEN; i++){

 		ctx->cv_l[i] = loadLE32(ctx->cv_l[i] ^ ctx->cv_r[i]);

 	}

-#endif

 }

 

 /* -------------------------------------------------------- */

@@ -937,14 +526,14 @@ inline void fin(LSH256_Context* ctx)
 inline void get_hash(LSH256_Context* ctx, lsh_u8* pbHashVal)

 {

 	CRYPTOPP_ASSERT(ctx != NULLPTR);

-	CRYPTOPP_ASSERT(ctx->algtype != 0);

+	CRYPTOPP_ASSERT(ctx->alg_type != 0);

 	CRYPTOPP_ASSERT(pbHashVal != NULLPTR);

 

-	lsh_uint algtype = ctx->algtype;

-	lsh_uint hash_val_byte_len = LSH_GET_HASHBYTE(algtype);

-	lsh_uint hash_val_bit_len = LSH_GET_SMALL_HASHBIT(algtype);

+	lsh_uint alg_type = ctx->alg_type;

+	lsh_uint hash_val_byte_len = LSH_GET_HASHBYTE(alg_type);

+	lsh_uint hash_val_bit_len = LSH_GET_SMALL_HASHBIT(alg_type);

 

-	// Multiplying by sizeof(lsh_u8) looks odd...

+	// Multiplying by looks odd...

 	memcpy(pbHashVal, ctx->cv_l, hash_val_byte_len);

 	if (hash_val_bit_len){

 		pbHashVal[hash_val_byte_len-1] &= (((lsh_u8)0xff) << hash_val_bit_len);

@@ -956,13 +545,13 @@ inline void get_hash(LSH256_Context* ctx, lsh_u8* pbHashVal)
 lsh_err lsh256_init(LSH256_Context* ctx)

 {

 	CRYPTOPP_ASSERT(ctx != NULLPTR);

-	CRYPTOPP_ASSERT(ctx->algtype != 0);

+	CRYPTOPP_ASSERT(ctx->alg_type != 0);

 

-	lsh_u32 algtype = ctx->algtype;

+	lsh_u32 alg_type = ctx->alg_type;

 	const lsh_u32* const_v = NULL;

 	ctx->remain_databitlen = 0;

 

-	switch (algtype)

+	switch (alg_type)

 	{

 	case LSH_TYPE_256_256:

 		init256(ctx);

@@ -977,13 +566,9 @@ lsh_err lsh256_init(LSH256_Context* ctx)
 	lsh_u32* cv_l = ctx->cv_l;

 	lsh_u32* cv_r = ctx->cv_r;

 

-#if defined(CRYPTOPP_LSH256_AVX_AVAILABLE)

-	AVX_Cleanup cleanup;

-#endif

-

 	zero_iv(cv_l, cv_r);

 	cv_l[0] = LSH256_HASH_VAL_MAX_BYTE_LEN;

-	cv_l[1] = LSH_GET_HASHBIT(algtype);

+	cv_l[1] = LSH_GET_HASHBIT(alg_type);

 

 	for (size_t i = 0; i < NUM_STEPS / 2; i++)

 	{

@@ -1005,19 +590,20 @@ lsh_err lsh256_update(LSH256_Context* ctx, const lsh_u8* data, size_t databitlen
 	CRYPTOPP_ASSERT(ctx != NULLPTR);

 	CRYPTOPP_ASSERT(data != NULLPTR);

 	CRYPTOPP_ASSERT(databitlen % 8 == 0);

-	CRYPTOPP_ASSERT(ctx->algtype != 0);

+	CRYPTOPP_ASSERT(ctx->alg_type != 0);

 

 	if (databitlen == 0){

 		return LSH_SUCCESS;

 	}

 

+	// We are byte oriented. tail bits will always be 0.

 	size_t databytelen = databitlen >> 3;

-	lsh_uint pos2 = databitlen & 0x7;

+	// lsh_uint pos2 = databitlen & 0x7;

+	const size_t pos2 = 0;

 

-	// We are byte oriented. remain_msg_bit will always be 0.

-	lsh_uint remain_msg_byte = ctx->remain_databitlen >> 3;

+	size_t remain_msg_byte = ctx->remain_databitlen >> 3;

 	// lsh_uint remain_msg_bit = ctx->remain_databitlen & 7;

-	const lsh_uint remain_msg_bit = 0;

+	const size_t remain_msg_bit = 0;

 

 	if (remain_msg_byte >= LSH256_MSG_BLK_BYTE_LEN){

 		return LSH_ERR_INVALID_STATE;

@@ -1038,7 +624,7 @@ lsh_err lsh256_update(LSH256_Context* ctx, const lsh_u8* data, size_t databitlen
 	}

 

 	if (remain_msg_byte > 0){

-		lsh_uint more_byte = LSH256_MSG_BLK_BYTE_LEN - remain_msg_byte;

+		size_t more_byte = LSH256_MSG_BLK_BYTE_LEN - remain_msg_byte;

 		memcpy(ctx->last_block + remain_msg_byte, data, more_byte);

 		compress(ctx, ctx->last_block);

 		data += more_byte;

@@ -1075,10 +661,10 @@ lsh_err lsh256_final(LSH256_Context* ctx, lsh_u8* hashval)
 	CRYPTOPP_ASSERT(ctx != NULLPTR);

 	CRYPTOPP_ASSERT(hashval != NULLPTR);

 

-	// We are byte oriented. remain_msg_bit will always be 0.

-	lsh_uint remain_msg_byte = ctx->remain_databitlen >> 3;

+	// We are byte oriented. tail bits will always be 0.

+	size_t remain_msg_byte = ctx->remain_databitlen >> 3;

 	// lsh_uint remain_msg_bit = ctx->remain_databitlen & 7;

-	const lsh_uint remain_msg_bit = 0;

+	const size_t remain_msg_bit = 0;

 

 	if (remain_msg_byte >= LSH256_MSG_BLK_BYTE_LEN){

 		return LSH_ERR_INVALID_STATE;

@@ -1104,42 +690,98 @@ ANONYMOUS_NAMESPACE_END
 

 NAMESPACE_BEGIN(CryptoPP)

 

+#if defined(CRYPTOPP_ENABLE_64BIT_SSE)

+# if defined(CRYPTOPP_AVX2_AVAILABLE)

+	extern void LSH256_Base_Restart_AVX2(word32* state);

+	extern void LSH256_Base_Update_AVX2(word32* state, const byte *input, size_t size);

+	extern void LSH256_Base_TruncatedFinal_AVX2(word32* state, byte *hash, size_t size);

+# endif

+# if defined(CRYPTOPP_SSSE3_AVAILABLE)

+	extern void LSH256_Base_Restart_SSSE3(word32* state);

+	extern void LSH256_Base_Update_SSSE3(word32* state, const byte *input, size_t size);

+	extern void LSH256_Base_TruncatedFinal_SSSE3(word32* state, byte *hash, size_t size);

+# endif

+#endif

+

+void LSH256_Base_Restart_CXX(word32* state)

+{

+	state[RemainingBits] = 0;

+	LSH256_Context ctx(state, state[AlgorithmType], state[RemainingBits]);

+	lsh_err err = lsh256_init(&ctx);

+

+	if (err != LSH_SUCCESS)

+		throw Exception(Exception::OTHER_ERROR, "LSH256_Base: lsh256_init failed");

+}

+

+void LSH256_Base_Update_CXX(word32* state, const byte *input, size_t size)

+{

+	LSH256_Context ctx(state, state[AlgorithmType], state[RemainingBits]);

+	lsh_err err = lsh256_update(&ctx, input, 8*size);

+

+	if (err != LSH_SUCCESS)

+		throw Exception(Exception::OTHER_ERROR, "LSH256_Base: lsh256_update failed");

+}

+

+void LSH256_Base_TruncatedFinal_CXX(word32* state, byte *hash, size_t)

+{

+	LSH256_Context ctx(state, state[AlgorithmType], state[RemainingBits]);

+	lsh_err err = lsh256_final(&ctx, hash);

+

+	if (err != LSH_SUCCESS)

+		throw Exception(Exception::OTHER_ERROR, "LSH256_Base: lsh256_final failed");

+}

+

 std::string LSH256_Base::AlgorithmProvider() const

 {

-#if defined(CRYPTOPP_LSH256_AVX2_AVAILABLE)

-	return "AVX2";

-#elif defined(CRYPTOPP_LSH256_AVX_AVAILABLE)

-	return "AVX";

-#elif defined(CRYPTOPP_LSH256_SSSE3_AVAILABLE)

-	return "SSSE3";

-#elif defined(CRYPTOPP_LSH256_SSE2_AVAILABLE)

-	return "SSE2";

-#else

-	return "C++";

+#if defined(CRYPTOPP_ENABLE_64BIT_SSE)

+#if defined(CRYPTOPP_AVX2_AVAILABLE)

+	if (HasAVX2())

+		return "AVX2";

+	else

+#endif

+#if defined(CRYPTOPP_SSSE3_AVAILABLE)

+	if (HasSSSE3())

+		return "SSSE3";

+	else

 #endif

+#endif  // CRYPTOPP_ENABLE_64BIT_SSE

+

+	return "C++";

 }

 

 void LSH256_Base::Restart()

 {

-	m_remainingBitLength = 0;

-

-	LSH256_Context ctx(m_state, m_algType, m_remainingBitLength);

-	lsh_err err = lsh256_init(&ctx);

+#if defined(CRYPTOPP_AVX2_AVAILABLE) && defined(CRYPTOPP_ENABLE_64BIT_SSE)

+	if (HasAVX2())

+		LSH256_Base_Restart_AVX2(m_state);

+	else

+#endif

+#if defined(CRYPTOPP_SSSE3_AVAILABLE) && defined(CRYPTOPP_ENABLE_64BIT_SSE)

+	if (HasSSSE3())

+		LSH256_Base_Restart_SSSE3(m_state);

+	else

+#endif

 

-	if (err != LSH_SUCCESS)

-		throw Exception(Exception::OTHER_ERROR, "LSH256_Base: lsh256_init failed");

+	LSH256_Base_Restart_CXX(m_state);

 }

 

-void LSH256_Base::Update(const byte *input, size_t length)

+void LSH256_Base::Update(const byte *input, size_t size)

 {

 	CRYPTOPP_ASSERT(input != NULLPTR);

-	CRYPTOPP_ASSERT(length);

+	CRYPTOPP_ASSERT(size);

 

-	LSH256_Context ctx(m_state, m_algType, m_remainingBitLength);

-	lsh_err err = lsh256_update(&ctx, input, 8*length);

+#if defined(CRYPTOPP_AVX2_AVAILABLE) && defined(CRYPTOPP_ENABLE_64BIT_SSE)

+	if (HasAVX2())

+		LSH256_Base_Update_AVX2(m_state, input, size);

+	else

+#endif

+#if defined(CRYPTOPP_SSSE3_AVAILABLE) && defined(CRYPTOPP_ENABLE_64BIT_SSE)

+	if (HasSSSE3())

+		LSH256_Base_Update_SSSE3(m_state, input, size);

+	else

+#endif

 

-	if (err != LSH_SUCCESS)

-		throw Exception(Exception::OTHER_ERROR, "LSH256_Base: lsh256_update failed");

+	LSH256_Base_Update_CXX(m_state, input, size);

 }

 

 void LSH256_Base::TruncatedFinal(byte *hash, size_t size)

@@ -1147,27 +789,27 @@ void LSH256_Base::TruncatedFinal(byte *hash, size_t size)
 	CRYPTOPP_ASSERT(hash != NULLPTR);

 	ThrowIfInvalidTruncatedSize(size);

 

-	LSH256_Context ctx(m_state, m_algType, m_remainingBitLength);

-	lsh_err err;

+	// TODO: determine if LSH256 supports truncated hashes. See the code

+	// in get_hash(), where a bit-length is added to the last output

+	// byte of the hash function.

+	byte fullHash[LSH256_HASH_VAL_MAX_BYTE_LEN];

+	bool copyOut = (size < DigestSize());

 

-	if (size >= DigestSize())

-	{

-		err = lsh256_final(&ctx, hash);

-	}

+#if defined(CRYPTOPP_AVX2_AVAILABLE) && defined(CRYPTOPP_ENABLE_64BIT_SSE)

+	if (HasAVX2())

+		LSH256_Base_TruncatedFinal_AVX2(m_state, copyOut ? fullHash : hash, size);

 	else

-	{

-		// TODO: determine if LSH256 supports truncated hashes. See the code

-		// in get_hash(), where a bit-length is added to the last output

-		// byte of the hash function.

-		// CRYPTOPP_ASSERT(0);

+#endif

+#if defined(CRYPTOPP_SSSE3_AVAILABLE) && defined(CRYPTOPP_ENABLE_64BIT_SSE)

+	if (HasSSSE3())

+		LSH256_Base_TruncatedFinal_SSSE3(m_state, copyOut ? fullHash : hash, size);

+	else

+#endif

 

-		byte fullHash[HASH_VAL_MAX_WORD_LEN * sizeof(lsh_u32)];

-		err = lsh256_final(&ctx, fullHash);

-		memcpy(hash, fullHash, size);

-	}

+	LSH256_Base_TruncatedFinal_CXX(m_state, copyOut ? fullHash : hash, size);

 

-	if (err != LSH_SUCCESS)

-		throw Exception(Exception::OTHER_ERROR, "LSH256_Base: lsh256_final failed");

+	if (copyOut)

+		memcpy(hash, fullHash, size);

 

     Restart();

 }