1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/xpcom/string/src/nsUTF8UtilsSSE2.cpp Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,101 @@
1.4 +/* This Source Code Form is subject to the terms of the Mozilla Public
1.5 + * License, v. 2.0. If a copy of the MPL was not distributed with this
1.6 + * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
1.7 +
1.8 +#include "nscore.h"
1.9 +#include "nsAlgorithm.h"
1.10 +#include <emmintrin.h>
1.11 +#include <nsUTF8Utils.h>
1.12 +
1.13 +void
1.14 +LossyConvertEncoding16to8::write_sse2(const char16_t* aSource,
1.15 + uint32_t aSourceLength)
1.16 +{
1.17 + char* dest = mDestination;
1.18 +
1.19 + // Align source to a 16-byte boundary.
1.20 + uint32_t i = 0;
1.21 + uint32_t alignLen =
1.22 + XPCOM_MIN<uint32_t>(aSourceLength, uint32_t(-NS_PTR_TO_INT32(aSource) & 0xf) / sizeof(char16_t));
1.23 + for (; i < alignLen; i++) {
1.24 + dest[i] = static_cast<unsigned char>(aSource[i]);
1.25 + }
1.26 +
1.27 + // Walk 64 bytes (four XMM registers) at a time.
1.28 + __m128i vectmask = _mm_set1_epi16(0x00ff);
1.29 + for (; aSourceLength - i > 31; i += 32) {
1.30 + __m128i source1 = _mm_load_si128(reinterpret_cast<const __m128i*>(aSource + i));
1.31 + source1 = _mm_and_si128(source1, vectmask);
1.32 +
1.33 + __m128i source2 = _mm_load_si128(reinterpret_cast<const __m128i*>(aSource + i + 8));
1.34 + source2 = _mm_and_si128(source2, vectmask);
1.35 +
1.36 + __m128i source3 = _mm_load_si128(reinterpret_cast<const __m128i*>(aSource + i + 16));
1.37 + source3 = _mm_and_si128(source3, vectmask);
1.38 +
1.39 + __m128i source4 = _mm_load_si128(reinterpret_cast<const __m128i*>(aSource + i + 24));
1.40 + source4 = _mm_and_si128(source4, vectmask);
1.41 +
1.42 +
1.43 + // Pack the source data. SSE2 views this as a saturating uint16_t to
1.44 + // uint8_t conversion, but since we masked off the high-order byte of every
1.45 + // uint16_t, we're really just grabbing the low-order bytes of source1 and
1.46 + // source2.
1.47 + __m128i packed1 = _mm_packus_epi16(source1, source2);
1.48 + __m128i packed2 = _mm_packus_epi16(source3, source4);
1.49 +
1.50 + // This store needs to be unaligned since there's no guarantee that the
1.51 + // alignment we did above for the source will align the destination.
1.52 + _mm_storeu_si128(reinterpret_cast<__m128i*>(dest + i), packed1);
1.53 + _mm_storeu_si128(reinterpret_cast<__m128i*>(dest + i + 16), packed2);
1.54 + }
1.55 +
1.56 + // Finish up the rest.
1.57 + for (; i < aSourceLength; i++) {
1.58 + dest[i] = static_cast<unsigned char>(aSource[i]);
1.59 + }
1.60 +
1.61 + mDestination += i;
1.62 +}
1.63 +
1.64 +void
1.65 +LossyConvertEncoding8to16::write_sse2(const char* aSource,
1.66 + uint32_t aSourceLength)
1.67 +{
1.68 + char16_t *dest = mDestination;
1.69 +
1.70 + // Align source to a 16-byte boundary. We choose to align source rather than
1.71 + // dest because we'd rather have our loads than our stores be fast. You have
1.72 + // to wait for a load to complete, but you can keep on moving after issuing a
1.73 + // store.
1.74 + uint32_t i = 0;
1.75 + uint32_t alignLen = XPCOM_MIN(aSourceLength, uint32_t(-NS_PTR_TO_INT32(aSource) & 0xf));
1.76 + for (; i < alignLen; i++) {
1.77 + dest[i] = static_cast<unsigned char>(aSource[i]);
1.78 + }
1.79 +
1.80 + // Walk 32 bytes (two XMM registers) at a time.
1.81 + for (; aSourceLength - i > 31; i += 32) {
1.82 + __m128i source1 = _mm_load_si128(reinterpret_cast<const __m128i*>(aSource + i));
1.83 + __m128i source2 = _mm_load_si128(reinterpret_cast<const __m128i*>(aSource + i + 16));
1.84 +
1.85 + // Interleave 0s in with the bytes of source to create lo and hi.
1.86 + __m128i lo1 = _mm_unpacklo_epi8(source1, _mm_setzero_si128());
1.87 + __m128i hi1 = _mm_unpackhi_epi8(source1, _mm_setzero_si128());
1.88 + __m128i lo2 = _mm_unpacklo_epi8(source2, _mm_setzero_si128());
1.89 + __m128i hi2 = _mm_unpackhi_epi8(source2, _mm_setzero_si128());
1.90 +
1.91 + // store lo and hi into dest.
1.92 + _mm_storeu_si128(reinterpret_cast<__m128i*>(dest + i), lo1);
1.93 + _mm_storeu_si128(reinterpret_cast<__m128i*>(dest + i + 8), hi1);
1.94 + _mm_storeu_si128(reinterpret_cast<__m128i*>(dest + i + 16), lo2);
1.95 + _mm_storeu_si128(reinterpret_cast<__m128i*>(dest + i + 24), hi2);
1.96 + }
1.97 +
1.98 + // Finish up whatever's left.
1.99 + for (; i < aSourceLength; i++) {
1.100 + dest[i] = static_cast<unsigned char>(aSource[i]);
1.101 + }
1.102 +
1.103 + mDestination += i;
1.104 +}
