diff options
| author | Thiago Macieira <thiago.macieira@intel.com> | 2018-09-12 18:10:11 -0700 |
|---|---|---|
| committer | Thiago Macieira <thiago.macieira@intel.com> | 2018-11-08 15:19:50 +0000 |
| commit | 570ef11c28b885817a69523835fac40d9e0d1f4e (patch) | |
| tree | dcc3047732d0822476b7a9ac60d98eef339daebb /src/corelib/codecs | |
| parent | a0907e6ac1deee086c04a8987ad76fc2c78f65ab (diff) | |
QUtf8Codec: Use one 32-byte load instead of two 16-byte ones on AVX2
The number of instructions is the same. But if the CPU can issue
32-byte-wide loads, this will be faster. For CPUs that would do two
16-byte loads, this is no worse than current code.
Change-Id: I8f261579aad648fdb4f0fffd1553d060b4fc852f
Reviewed-by: Allan Sandfeld Jensen <allan.jensen@qt.io>
Diffstat (limited to 'src/corelib/codecs')
| -rw-r--r-- | src/corelib/codecs/qutfcodec.cpp | 7 |
1 files changed, 6 insertions, 1 deletions
diff --git a/src/corelib/codecs/qutfcodec.cpp b/src/corelib/codecs/qutfcodec.cpp index 26c68cdee5..8bc1294c49 100644 --- a/src/corelib/codecs/qutfcodec.cpp +++ b/src/corelib/codecs/qutfcodec.cpp @@ -70,9 +70,14 @@ static inline bool simdEncodeAscii(uchar *&dst, const ushort *&nextAscii, const { // do sixteen characters at a time for ( ; end - src >= 16; src += 16, dst += 16) { +# ifdef __AVX2__ + __m256i data = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(src)); + __m128i data1 = _mm256_castsi256_si128(data); + __m128i data2 = _mm256_extracti128_si256(data, 1); +# else __m128i data1 = _mm_loadu_si128((const __m128i*)src); __m128i data2 = _mm_loadu_si128(1+(const __m128i*)src); - +# endif // check if everything is ASCII // the highest ASCII value is U+007F |