1berry/qtbase
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

qt_is_ascii: improve isAscii a little further (QUrl, QLatin1String)

Browse Source 
Turns out that the non-AVX2 code was beating the performance of the AVX2
because the simdTestMask function did a little too much. So just use the
same VPMOVMSKB technique for it.

Change-Id: I0825ff5b5f6f4c85939ffffd152f3b636ab998db
Reviewed-by: Allan Sandfeld Jensen <allan.jensen@qt.io>
This commit is contained in:
Thiago Macieira 2018-05-16 13:54:56 -07:00
parent 8d19afcc81
commit c780434ba2
1 changed files with 12 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
src/corelib/tools/qstring.cpp
View File

@ -311,12 +311,19 @@ bool QtPrivate::isAscii(QLatin1String s) Q_DECL_NOTHROW
const char *ptr = s.begin(); const char *ptr = s.begin();
const char *end = s.end(); const char *end = s.end();
#if defined(__AVX2__) #if defined(__SSE2__)
if (!simdTestMask(ptr, end, 0x80808080))
return false;
#elif defined(__SSE2__)
// Testing for the high bit can be done efficiently with just PMOVMSKB // Testing for the high bit can be done efficiently with just PMOVMSKB
while (ptr + 16 < end) { # if defined(__AVX2__)
while (ptr + 32 <= end) {
__m256i data = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(ptr));
quint32 mask = _mm256_movemask_epi8(data);
if (mask)
return false;
ptr += 32;
}
# endif
while (ptr + 16 <= end) {
__m128i data = _mm_loadu_si128(reinterpret_cast<const __m128i *>(ptr)); __m128i data = _mm_loadu_si128(reinterpret_cast<const __m128i *>(ptr));
quint32 mask = _mm_movemask_epi8(data); quint32 mask = _mm_movemask_epi8(data);
if (mask) if (mask)