thevar1able/ClickHouse
1
0
Fork 0
mirror of https://github.com/ClickHouse/ClickHouse.git synced 2024-10-09 01:50:50 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fixed compilation issues and fixed several bugs in SplitTokenExtractor::next

Browse Source 
* Handling all characters above 0x80 as symbols (fixes UTF8 tokens)
* Properly handling tokens that end exactly on haystack boundary.
This commit is contained in:
Vasily Nemkov 2020-04-02 00:28:02 +03:00
parent 9d6c88c78e
commit 90cb6a25cf
1 changed files with 21 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

31
dbms/src/Storages/MergeTree/MergeTreeIndexFullText.cpp
View File

@ -19,6 +19,10 @@
#include <boost/algorithm/string.hpp> #include <boost/algorithm/string.hpp>
#include <immintrin.h>
#include <nmmintrin.h>
#include <emmintrin.h>
namespace DB namespace DB
{ {
@ -609,16 +613,17 @@ bool SplitTokenExtractor::next(const char * data, size_t len, size_t * pos, size
while (*pos < len) while (*pos < len)
{ {
#if __SSE2__ #if defined(__SSE2__)
// NOTE: we assume that `data` string is padded from the right with 15 zero-bytes. // NOTE: we assume that `data` string is padded from the right with 15 zero-bytes.
const __m128i haystack = _mm_loadu_si128(reinterpret_cast<const __m128i *>(data + *pos)); const __m128i haystack = _mm_loadu_si128(reinterpret_cast<const __m128i *>(data + *pos));
const size_t haystack_length = 16; const size_t haystack_length = 16;
#if __SSE4_2__ #if defined(__SSE4_2__)
// With the help of https://www.strchr.com/strcmp_and_strlen_using_sse_4.2 // With the help of https://www.strchr.com/strcmp_and_strlen_using_sse_4.2
static const auto alnum_chars_ranges = _mm_set_epi8(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 'Z', 'A', 'z', 'a', '9', '0'); static const auto alnum_chars_ranges = _mm_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
'\xFF', '\x80', 'z', 'a', 'Z', 'A', '9', '0');
// Every bit represents if `haystack` character is in the ranges (1) or not(0) // Every bit represents if `haystack` character is in the ranges (1) or not(0)
const auto result_bitmask = _mm_cvtsi128_si32(_mm_cmpestrm(alnum_chars_ranges, 6, haystack, haystack_length, _SIDD_CMP_RANGES | _SIDD_UBYTE_OPS)); const int result_bitmask = _mm_cvtsi128_si32(_mm_cmpestrm(alnum_chars_ranges, 8, haystack, haystack_length, _SIDD_CMP_RANGES));
#else #else
// NOTE: -1 and +1 required since SSE2 has no `>=` and `<=` instructions on packed 8-bit integers (epi8). // NOTE: -1 and +1 required since SSE2 has no `>=` and `<=` instructions on packed 8-bit integers (epi8).
static const auto number_begin = _mm_set1_epi8('0' - 1); static const auto number_begin = _mm_set1_epi8('0' - 1);
@ -627,13 +632,16 @@ bool SplitTokenExtractor::next(const char * data, size_t len, size_t * pos, size
static const auto alpha_lower_end = _mm_set1_epi8('z' + 1); static const auto alpha_lower_end = _mm_set1_epi8('z' + 1);
static const auto alpha_upper_begin = _mm_set1_epi8('A' - 1); static const auto alpha_upper_begin = _mm_set1_epi8('A' - 1);
static const auto alpha_upper_end = _mm_set1_epi8('Z' + 1); static const auto alpha_upper_end = _mm_set1_epi8('Z' + 1);
static const auto zero = _mm_set1_epi8(0);
// every bit represents if `haystack` character `c` statisfies condition: // every bit represents if `haystack` character `c` statisfies condition:
// (c > '0' - 1 && c < '9' + 1) || (c > 'a' - 1 && c < 'z' + 1) || (c > 'A' - 1 && c < 'Z' + 1) // (c < 0) || (c > '0' - 1 && c < '9' + 1) || (c > 'a' - 1 && c < 'z' + 1) || (c > 'A' - 1 && c < 'Z' + 1)
const int result_bitmask = _mm_movemask_epi8(_mm_or_si128(_mm_or_si128( // < 0 since _mm_cmplt_epi8 threats chars as SIGNED, and hence all chars > 0x80 are negative.
_mm_and_si128(_mm_cmpgt_epi8(haystack, number_begin), _mm_cmplt_epi8(haystack, number_end)), const int result_bitmask = _mm_movemask_epi8(_mm_or_si128(_mm_or_si128(_mm_or_si128(
_mm_and_si128(_mm_cmpgt_epi8(haystack, alpha_lower_begin), _mm_cmplt_epi8(haystack, alpha_lower_end))), _mm_cmplt_epi8(haystack, zero),
_mm_and_si128(_mm_cmpgt_epi8(haystack, alpha_upper_begin), _mm_cmplt_epi8(haystack, alpha_upper_end)))); _mm_and_si128(_mm_cmpgt_epi8(haystack, number_begin), _mm_cmplt_epi8(haystack, number_end))),
_mm_and_si128(_mm_cmpgt_epi8(haystack, alpha_lower_begin), _mm_cmplt_epi8(haystack, alpha_lower_end))),
_mm_and_si128(_mm_cmpgt_epi8(haystack, alpha_upper_begin), _mm_cmplt_epi8(haystack, alpha_upper_end))));
#endif #endif
// NOTE: __builtin_ctz family explicitly state that result is UNDEFINED if argument is 0 // NOTE: __builtin_ctz family explicitly state that result is UNDEFINED if argument is 0
if (result_bitmask == 0) if (result_bitmask == 0)
@ -649,12 +657,15 @@ bool SplitTokenExtractor::next(const char * data, size_t len, size_t * pos, size
const auto start = getTrailingZeroBits(result_bitmask); const auto start = getTrailingZeroBits(result_bitmask);
if (*token_len == 0) if (*token_len == 0)
*token_start = *pos + start; *token_start = *pos + start;
else if (start != 0)
// token is not continued in this haystack
return true;
const auto l = getTrailingZeroBits(~(result_bitmask >> start)); const auto l = getTrailingZeroBits(~(result_bitmask >> start));
*token_len += l; *token_len += l;
*pos += start + l; *pos += start + l;
if (start + l == 16) if (start + l == haystack_length)
// check if there are leftovers in next `haystack` // check if there are leftovers in next `haystack`
continue; continue;