|
35 | 35 | #include "v8-fast-api-calls.h"
|
36 | 36 | #include "v8.h"
|
37 | 37 |
|
| 38 | +#include <stdint.h> |
38 | 39 | #include <climits>
|
39 | 40 | #include <cstring>
|
40 | 41 | #include "nbytes.h"
|
@@ -752,13 +753,40 @@ uint32_t FastByteLengthUtf8(Local<Value> receiver,
|
752 | 753 | if (source.length > 128) {
|
753 | 754 | return simdutf::utf8_length_from_latin1(source.data, source.length);
|
754 | 755 | }
|
| 756 | + |
755 | 757 | uint32_t length = source.length;
|
756 |
| - uint32_t result = length; |
757 |
| - const uint8_t* data = reinterpret_cast<const uint8_t*>(source.data); |
758 |
| - for (uint32_t i = 0; i < length; ++i) { |
759 |
| - result += (data[i] >> 7); |
| 758 | + const auto input = reinterpret_cast<const uint8_t*>(source.data); |
| 759 | + |
| 760 | + uint32_t answer = length; |
| 761 | + uint32_t i = 0; |
| 762 | + |
| 763 | + auto pop = [](uint64_t v) { |
| 764 | + return static_cast<size_t>(((v >> 7) & UINT64_C(0x0101010101010101)) * |
| 765 | + UINT64_C(0x0101010101010101) >> |
| 766 | + 56); |
| 767 | + }; |
| 768 | + |
| 769 | + for (; i + 32 <= length; i += 32) { |
| 770 | + uint64_t v; |
| 771 | + memcpy(&v, input + i, 8); |
| 772 | + answer += pop(v); |
| 773 | + memcpy(&v, input + i + 8, 8); |
| 774 | + answer += pop(v); |
| 775 | + memcpy(&v, input + i + 16, 8); |
| 776 | + answer += pop(v); |
| 777 | + memcpy(&v, input + i + 24, 8); |
| 778 | + answer += pop(v); |
| 779 | + } |
| 780 | + for (; i + 8 <= length; i += 8) { |
| 781 | + uint64_t v; |
| 782 | + memcpy(&v, input + i, 8); |
| 783 | + answer += pop(v); |
760 | 784 | }
|
761 |
| - return result; |
| 785 | + for (; i + 1 <= length; i += 1) { |
| 786 | + answer += input[i] >> 7; |
| 787 | + } |
| 788 | + |
| 789 | + return answer; |
762 | 790 | }
|
763 | 791 |
|
764 | 792 | static v8::CFunction fast_byte_length_utf8(
|
|
0 commit comments