From 83ba0d56f878409d3f758549c72e5d099cc71a07 Mon Sep 17 00:00:00 2001 From: Thiago Macieira Date: Wed, 22 Dec 2010 19:52:48 +0100 Subject: [PATCH] Add an SSE4.2 even simpler version of toLatin1 MIME-Version: 1.0 Content-Type: text/plain; charset=utf8 Content-Transfer-Encoding: 8bit Use the new PCMPESTRM instruction (Parallel CoMPare Explicit-length STRings with result in a Mask) which is added in SSE4.2 for facilitating string operations. The "compare ranges" mode allows us to search for characters outside the Latin 1 range and then use the SSE4.1 PBLENDVB instruction to replace those with question marks. Unlike previous SSE compare instructions, the PCMPxSTRx family allows us to operate on unsigned 16-bit values. This saves us another parallel add. Reviewed-By: Samuel Rødal (cherry picked from commit 45d2d36c9dbcbce403c78838ea52acd1ab111b68) Change-Id: I0f9d864f9d19c0f0da43ccb6918dc28295074496 Reviewed-on: http://codereview.qt-project.org/4468 Reviewed-by: Qt Sanity Bot Reviewed-by: Oswald Buddenhagen Reviewed-by: Samuel Rødal --- src/corelib/tools/qstring.cpp | 23 +++++++++++++++++++++++ tests/auto/qstring/tst_qstring.cpp | 4 ++++ 2 files changed, 27 insertions(+) diff --git a/src/corelib/tools/qstring.cpp b/src/corelib/tools/qstring.cpp index b34d638..3a2e35c 100644 --- a/src/corelib/tools/qstring.cpp +++ b/src/corelib/tools/qstring.cpp @@ -3540,6 +3540,28 @@ static inline __m128i mergeQuestionMarks(__m128i chunk) { const __m128i questionMark = _mm_set1_epi16('?'); +# ifdef __SSE4_2__ + // compare the unsigned shorts for the range 0x0100-0xFFFF + // note on the use of _mm_cmpestrm: + // The MSDN documentation online (http://technet.microsoft.com/en-us/library/bb514080.aspx) + // says for range search the following: + // For each character c in a, determine whether b0 <= c <= b1 or b2 <= c <= b3 + // + // However, all examples on the Internet, including from Intel + // (see http://software.intel.com/en-us/articles/xml-parsing-accelerator-with-intel-streaming-simd-extensions-4-intel-sse4/) + // put the range to be searched first + // + // Disassembly and instruction-level debugging with GCC and ICC show + // that they are doing the right thing. Inverting the arguments in the + // instruction does cause a bunch of test failures. + + const int mode = _SIDD_UWORD_OPS | _SIDD_CMP_RANGES | _SIDD_UNIT_MASK; + const __m128i rangeMatch = _mm_cvtsi32_si128(0xffff0100); + const __m128i offLimitMask = _mm_cmpestrm(rangeMatch, 2, chunk, 8, mode); + + // replace the non-Latin 1 characters in the chunk with question marks + chunk = _mm_blendv_epi8(chunk, questionMark, offLimitMask); +# else // SSE has no compare instruction for unsigned comparison. // The variables must be shiffted + 0x8000 to be compared const __m128i signedBitOffset = _mm_set1_epi16(0x8000); @@ -3563,6 +3585,7 @@ static inline __m128i mergeQuestionMarks(__m128i chunk) // merge offLimitQuestionMark and correctBytes to have the result chunk = _mm_or_si128(correctBytes, offLimitQuestionMark); # endif +# endif return chunk; } #endif diff --git a/tests/auto/qstring/tst_qstring.cpp b/tests/auto/qstring/tst_qstring.cpp index cf68bda..844624a 100644 --- a/tests/auto/qstring/tst_qstring.cpp +++ b/tests/auto/qstring/tst_qstring.cpp @@ -3462,6 +3462,10 @@ void tst_QString::toLatin1Roundtrip_data() static const ushort unicode6[] = { 0x180, 0x1ff, 0x8001, 0x8080, 0xfffc }; QTest::newRow("non-latin1b") << QByteArray("?????") << QString::fromUtf16(unicode6, 5) << questionmarks; + + static const ushort unicode7[] = { 'H', 'e', 'l', 'l', 'o', 0x100, 0x17f, 0x180, 0x8080, 0xfffc }; + static const ushort unicode7q[] = { 'H', 'e', 'l', 'l', 'o', '?', '?', '?', '?', '?' }; + QTest::newRow("mixed") << QByteArray("Hello?????") << QString::fromUtf16(unicode7, 10) << QString::fromUtf16(unicode7q, 10); } void tst_QString::toLatin1Roundtrip() -- 2.7.4