The Tor Browser / file comparison
comparison: content/base/src/nsTextFragmentSSE2.cpp
content/base/src/nsTextFragmentSSE2.cpp
- changeset 0
- 6474c204b198
equal
deleted
inserted
replaced
| -1:000000000000 | 0:fcae4fc1cfb2 |
|---|---|
| 1 /* This Source Code Form is subject to the terms of the Mozilla Public | |
| 2 * License, v. 2.0. If a copy of the MPL was not distributed with this | |
| 3 * file, You can obtain one at http://mozilla.org/MPL/2.0/. */ | |
| 4 | |
| 5 // This file should only be compiled if you're on x86 or x86_64. Additionally, | |
| 6 // you'll need to compile this file with -msse2 if you're using gcc. | |
| 7 | |
| 8 #include <emmintrin.h> | |
| 9 #include "nscore.h" | |
| 10 #include "nsAlgorithm.h" | |
| 11 #include "nsTextFragmentImpl.h" | |
| 12 #include <algorithm> | |
| 13 | |
| 14 namespace mozilla { | |
| 15 namespace SSE2 { | |
| 16 | |
| 17 static inline bool | |
| 18 is_zero (__m128i x) | |
| 19 { | |
| 20 return | |
| 21 _mm_movemask_epi8(_mm_cmpeq_epi8(x, _mm_setzero_si128())) == 0xffff; | |
| 22 } | |
| 23 | |
| 24 int32_t | |
| 25 FirstNon8Bit(const char16_t *str, const char16_t *end) | |
| 26 { | |
| 27 const uint32_t numUnicharsPerVector = 8; | |
| 28 typedef Non8BitParameters<sizeof(size_t)> p; | |
| 29 const size_t mask = p::mask(); | |
| 30 const uint32_t numUnicharsPerWord = p::numUnicharsPerWord(); | |
| 31 const int32_t len = end - str; | |
| 32 int32_t i = 0; | |
| 33 | |
| 34 // Align ourselves to a 16-byte boundary, as required by _mm_load_si128 | |
| 35 // (i.e. MOVDQA). | |
| 36 int32_t alignLen = | |
| 37 std::min(len, int32_t(((-NS_PTR_TO_INT32(str)) & 0xf) / sizeof(char16_t))); | |
| 38 for (; i < alignLen; i++) { | |
| 39 if (str[i] > 255) | |
| 40 return i; | |
| 41 } | |
| 42 | |
| 43 // Check one XMM register (16 bytes) at a time. | |
| 44 const int32_t vectWalkEnd = ((len - i) / numUnicharsPerVector) * numUnicharsPerVector; | |
| 45 const uint16_t shortMask = 0xff00; | |
| 46 __m128i vectmask = _mm_set1_epi16(static_cast<int16_t>(shortMask)); | |
| 47 for(; i < vectWalkEnd; i += numUnicharsPerVector) { | |
| 48 const __m128i vect = *reinterpret_cast<const __m128i*>(str + i); | |
| 49 if (!is_zero(_mm_and_si128(vect, vectmask))) | |
| 50 return i; | |
| 51 } | |
| 52 | |
| 53 // Check one word at a time. | |
| 54 const int32_t wordWalkEnd = ((len - i) / numUnicharsPerWord) * numUnicharsPerWord; | |
| 55 for(; i < wordWalkEnd; i += numUnicharsPerWord) { | |
| 56 const size_t word = *reinterpret_cast<const size_t*>(str + i); | |
| 57 if (word & mask) | |
| 58 return i; | |
| 59 } | |
| 60 | |
| 61 // Take care of the remainder one character at a time. | |
| 62 for (; i < len; i++) { | |
| 63 if (str[i] > 255) { | |
| 64 return i; | |
| 65 } | |
| 66 } | |
| 67 | |
| 68 return -1; | |
| 69 } | |
| 70 | |
| 71 } // namespace SSE2 | |
| 72 } // namespace mozilla |
