michael@0: /* This Source Code Form is subject to the terms of the Mozilla Public
michael@0:  * License, v. 2.0. If a copy of the MPL was not distributed with this
michael@0:  * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
michael@0: 
michael@0: #include "nscore.h"
michael@0: #include "nsAlgorithm.h"
michael@0: #include <emmintrin.h>
michael@0: #include <nsUTF8Utils.h>
michael@0: 
michael@0: void
michael@0: LossyConvertEncoding16to8::write_sse2(const char16_t* aSource,
michael@0:                                       uint32_t aSourceLength)
michael@0: {
michael@0:   char* dest = mDestination;
michael@0: 
michael@0:   // Align source to a 16-byte boundary.
michael@0:   uint32_t i = 0;
michael@0:   uint32_t alignLen =
michael@0:     XPCOM_MIN<uint32_t>(aSourceLength, uint32_t(-NS_PTR_TO_INT32(aSource) & 0xf) / sizeof(char16_t));
michael@0:   for (; i < alignLen; i++) {
michael@0:     dest[i] = static_cast<unsigned char>(aSource[i]);
michael@0:   }
michael@0: 
michael@0:   // Walk 64 bytes (four XMM registers) at a time.
michael@0:   __m128i vectmask = _mm_set1_epi16(0x00ff);
michael@0:   for (; aSourceLength - i > 31; i += 32) {
michael@0:     __m128i source1 = _mm_load_si128(reinterpret_cast<const __m128i*>(aSource + i));
michael@0:     source1 = _mm_and_si128(source1, vectmask);
michael@0: 
michael@0:     __m128i source2 = _mm_load_si128(reinterpret_cast<const __m128i*>(aSource + i + 8));
michael@0:     source2 = _mm_and_si128(source2, vectmask);
michael@0: 
michael@0:     __m128i source3 = _mm_load_si128(reinterpret_cast<const __m128i*>(aSource + i + 16));
michael@0:     source3 = _mm_and_si128(source3, vectmask);
michael@0: 
michael@0:     __m128i source4 = _mm_load_si128(reinterpret_cast<const __m128i*>(aSource + i + 24));
michael@0:     source4 = _mm_and_si128(source4, vectmask);
michael@0: 
michael@0: 
michael@0:     // Pack the source data.  SSE2 views this as a saturating uint16_t to
michael@0:     // uint8_t conversion, but since we masked off the high-order byte of every
michael@0:     // uint16_t, we're really just grabbing the low-order bytes of source1 and
michael@0:     // source2.
michael@0:     __m128i packed1 = _mm_packus_epi16(source1, source2);
michael@0:     __m128i packed2 = _mm_packus_epi16(source3, source4);
michael@0: 
michael@0:     // This store needs to be unaligned since there's no guarantee that the
michael@0:     // alignment we did above for the source will align the destination.
michael@0:     _mm_storeu_si128(reinterpret_cast<__m128i*>(dest + i),      packed1);
michael@0:     _mm_storeu_si128(reinterpret_cast<__m128i*>(dest + i + 16), packed2);
michael@0:   }
michael@0: 
michael@0:   // Finish up the rest.
michael@0:   for (; i < aSourceLength; i++) {
michael@0:     dest[i] = static_cast<unsigned char>(aSource[i]);
michael@0:   }
michael@0: 
michael@0:   mDestination += i;
michael@0: }
michael@0: 
michael@0: void
michael@0: LossyConvertEncoding8to16::write_sse2(const char* aSource,
michael@0:                                       uint32_t aSourceLength)
michael@0: {
michael@0:   char16_t *dest = mDestination;
michael@0: 
michael@0:   // Align source to a 16-byte boundary.  We choose to align source rather than
michael@0:   // dest because we'd rather have our loads than our stores be fast. You have
michael@0:   // to wait for a load to complete, but you can keep on moving after issuing a
michael@0:   // store.
michael@0:   uint32_t i = 0;
michael@0:   uint32_t alignLen = XPCOM_MIN(aSourceLength, uint32_t(-NS_PTR_TO_INT32(aSource) & 0xf));
michael@0:   for (; i < alignLen; i++) {
michael@0:     dest[i] = static_cast<unsigned char>(aSource[i]);
michael@0:   }
michael@0: 
michael@0:   // Walk 32 bytes (two XMM registers) at a time.
michael@0:   for (; aSourceLength - i > 31; i += 32) {
michael@0:     __m128i source1 = _mm_load_si128(reinterpret_cast<const __m128i*>(aSource + i));
michael@0:     __m128i source2 = _mm_load_si128(reinterpret_cast<const __m128i*>(aSource + i + 16));
michael@0: 
michael@0:     // Interleave 0s in with the bytes of source to create lo and hi.
michael@0:     __m128i lo1 = _mm_unpacklo_epi8(source1, _mm_setzero_si128());
michael@0:     __m128i hi1 = _mm_unpackhi_epi8(source1, _mm_setzero_si128());
michael@0:     __m128i lo2 = _mm_unpacklo_epi8(source2, _mm_setzero_si128());
michael@0:     __m128i hi2 = _mm_unpackhi_epi8(source2, _mm_setzero_si128());
michael@0: 
michael@0:     // store lo and hi into dest.
michael@0:     _mm_storeu_si128(reinterpret_cast<__m128i*>(dest + i),      lo1);
michael@0:     _mm_storeu_si128(reinterpret_cast<__m128i*>(dest + i + 8),  hi1);
michael@0:     _mm_storeu_si128(reinterpret_cast<__m128i*>(dest + i + 16), lo2);
michael@0:     _mm_storeu_si128(reinterpret_cast<__m128i*>(dest + i + 24), hi2);
michael@0:   }
michael@0: 
michael@0:   // Finish up whatever's left.
michael@0:   for (; i < aSourceLength; i++) {
michael@0:     dest[i] = static_cast<unsigned char>(aSource[i]);
michael@0:   }
michael@0: 
michael@0:   mDestination += i;
michael@0: }