Chris@102: /*
Chris@102:  *            Copyright Andrey Semashev 2013.
Chris@102:  * Distributed under the Boost Software License, Version 1.0.
Chris@102:  *    (See accompanying file LICENSE_1_0.txt or copy at
Chris@102:  *          http://www.boost.org/LICENSE_1_0.txt)
Chris@102:  */
Chris@102: /*!
Chris@102:  * \file   uuid/detail/uuid_x86.hpp
Chris@102:  *
Chris@102:  * \brief  This header contains optimized SSE implementation of \c boost::uuid operations.
Chris@102:  */
Chris@102: 
Chris@102: #ifndef BOOST_UUID_DETAIL_UUID_X86_HPP_INCLUDED_
Chris@102: #define BOOST_UUID_DETAIL_UUID_X86_HPP_INCLUDED_
Chris@102: 
Chris@102: // MSVC does not always have immintrin.h (at least, not up to MSVC 10), so include the appropriate header for each instruction set
Chris@102: #if defined(BOOST_UUID_USE_SSE41)
Chris@102: #include <smmintrin.h>
Chris@102: #elif defined(BOOST_UUID_USE_SSE3)
Chris@102: #include <pmmintrin.h>
Chris@102: #else
Chris@102: #include <emmintrin.h>
Chris@102: #endif
Chris@102: 
Chris@102: namespace boost {
Chris@102: namespace uuids {
Chris@102: namespace detail {
Chris@102: 
Chris@102: BOOST_FORCEINLINE __m128i load_unaligned_si128(const uint8_t* p) BOOST_NOEXCEPT
Chris@102: {
Chris@102: #if defined(BOOST_UUID_USE_SSE3)
Chris@102:     return _mm_lddqu_si128(reinterpret_cast< const __m128i* >(p));
Chris@102: #else
Chris@102:     return _mm_loadu_si128(reinterpret_cast< const __m128i* >(p));
Chris@102: #endif
Chris@102: }
Chris@102: 
Chris@102: } // namespace detail
Chris@102: 
Chris@102: inline bool uuid::is_nil() const BOOST_NOEXCEPT
Chris@102: {
Chris@102:     register __m128i mm = uuids::detail::load_unaligned_si128(data);
Chris@102: #if defined(BOOST_UUID_USE_SSE41)
Chris@102:     return _mm_test_all_zeros(mm, mm) != 0;
Chris@102: #else
Chris@102:     mm = _mm_cmpeq_epi8(mm, _mm_setzero_si128());
Chris@102:     return _mm_movemask_epi8(mm) == 0xFFFF;
Chris@102: #endif
Chris@102: }
Chris@102: 
Chris@102: inline void uuid::swap(uuid& rhs) BOOST_NOEXCEPT
Chris@102: {
Chris@102:     register __m128i mm_this = uuids::detail::load_unaligned_si128(data);
Chris@102:     register __m128i mm_rhs = uuids::detail::load_unaligned_si128(rhs.data);
Chris@102:     _mm_storeu_si128(reinterpret_cast< __m128i* >(rhs.data), mm_this);
Chris@102:     _mm_storeu_si128(reinterpret_cast< __m128i* >(data), mm_rhs);
Chris@102: }
Chris@102: 
Chris@102: inline bool operator== (uuid const& lhs, uuid const& rhs) BOOST_NOEXCEPT
Chris@102: {
Chris@102:     register __m128i mm_left = uuids::detail::load_unaligned_si128(lhs.data);
Chris@102:     register __m128i mm_right = uuids::detail::load_unaligned_si128(rhs.data);
Chris@102: 
Chris@102:     register __m128i mm_cmp = _mm_cmpeq_epi32(mm_left, mm_right);
Chris@102: #if defined(BOOST_UUID_USE_SSE41)
Chris@102:     return _mm_test_all_ones(mm_cmp);
Chris@102: #else
Chris@102:     return _mm_movemask_epi8(mm_cmp) == 0xFFFF;
Chris@102: #endif
Chris@102: }
Chris@102: 
Chris@102: inline bool operator< (uuid const& lhs, uuid const& rhs) BOOST_NOEXCEPT
Chris@102: {
Chris@102:     register __m128i mm_left = uuids::detail::load_unaligned_si128(lhs.data);
Chris@102:     register __m128i mm_right = uuids::detail::load_unaligned_si128(rhs.data);
Chris@102: 
Chris@102:     // To emulate lexicographical_compare behavior we have to perform two comparisons - the forward and reverse one.
Chris@102:     // Then we know which bytes are equivalent and which ones are different, and for those different the comparison results
Chris@102:     // will be opposite. Then we'll be able to find the first differing comparison result (for both forward and reverse ways),
Chris@102:     // and depending on which way it is for, this will be the result of the operation. There are a few notes to consider:
Chris@102:     //
Chris@102:     // 1. Due to little endian byte order the first bytes go into the lower part of the xmm registers,
Chris@102:     //    so the comparison results in the least significant bits will actually be the most signigicant for the final operation result.
Chris@102:     //    This means we have to determine which of the comparison results have the least significant bit on, and this is achieved with
Chris@102:     //    the "(x - 1) ^ x" trick.
Chris@102:     // 2. Because there is only signed comparison in SSE/AVX, we have to invert byte comparison results whenever signs of the corresponding
Chris@102:     //    bytes are different. I.e. in signed comparison it's -1 < 1, but in unsigned it is the opposite (255 > 1). To do that we XOR left and right,
Chris@102:     //    making the most significant bit of each byte 1 if the signs are different, and later apply this mask with another XOR to the comparison results.
Chris@102:     // 3. pcmpgtw compares for "greater" relation, so we swap the arguments to get what we need.
Chris@102: 
Chris@102:     const __m128i mm_signs_mask = _mm_xor_si128(mm_left, mm_right);
Chris@102: 
Chris@102:     __m128i mm_cmp = _mm_cmpgt_epi8(mm_right, mm_left), mm_rcmp = _mm_cmpgt_epi8(mm_left, mm_right);
Chris@102: 
Chris@102:     mm_cmp = _mm_xor_si128(mm_signs_mask, mm_cmp);
Chris@102:     mm_rcmp = _mm_xor_si128(mm_signs_mask, mm_rcmp);
Chris@102: 
Chris@102:     uint32_t cmp = static_cast< uint32_t >(_mm_movemask_epi8(mm_cmp)), rcmp = static_cast< uint32_t >(_mm_movemask_epi8(mm_rcmp));
Chris@102: 
Chris@102:     cmp = (cmp - 1u) ^ cmp;
Chris@102:     rcmp = (rcmp - 1u) ^ rcmp;
Chris@102: 
Chris@102:     return static_cast< uint16_t >(cmp) < static_cast< uint16_t >(rcmp);
Chris@102: }
Chris@102: 
Chris@102: } // namespace uuids
Chris@102: } // namespace boost
Chris@102: 
Chris@102: #endif // BOOST_UUID_DETAIL_UUID_X86_HPP_INCLUDED_