Improve qustrlen further by avoiding the alignment prologue

This avoids looping prior to the main 16-byte loop, by performing one
load that may include bytes prior to the start of the string. This is
guaranteed not to fault, since str points to a valid character, but it
may cause Valgrind to print warnings.

Change-Id: I6e9274c1e7444ad48c81fffd14dcae854bba24b2
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>

2 files changed, 50 insertions, 16 deletions

diff --git a/src/corelib/tools/qstring.cpp b/src/corelib/tools/qstring.cpp
index a0c309c1cf..57a43205b7 100644
--- a/src/corelib/tools/qstring.cpp
+++ b/src/corelib/tools/qstring.cpp
@@ -165,28 +165,37 @@ qssize_t qustrlen(const ushort *str) Q_DECL_NOTHROW
     qssize_t result = 0;
 
 #ifdef __SSE2__
-    // progress until we get an aligned pointer
-    const ushort *ptr = str;
-    while (*ptr && quintptr(ptr) % 16)
-        ++ptr;
-    if (*ptr == 0)
-        return ptr - str;
+    // find the 16-byte alignment immediately prior or equal to str
+    quintptr misalignment = quintptr(str) & 0xf;
+    Q_ASSERT((misalignment & 1) == 0);
+    const ushort *ptr = str - (misalignment / 2);
 
     // load 16 bytes and see if we have a null
     // (aligned loads can never segfault)
-    int mask;
     const __m128i zeroes = _mm_setzero_si128();
+    __m128i data = _mm_load_si128(reinterpret_cast<const __m128i *>(ptr));
+    __m128i comparison = _mm_cmpeq_epi16(data, zeroes);
+    quint32 mask = _mm_movemask_epi8(comparison);
+
+    // ignore the result prior to the beginning of str
+    mask >>= misalignment;
+
+    // Have we found something in the first block? Need to handle it now
+    // because of the left shift above.
+    if (mask)
+        return qCountTrailingZeroBits(quint32(mask)) / 2;
+
     do {
-        __m128i data = _mm_load_si128(reinterpret_cast<const __m128i *>(ptr));
         ptr += 8;
+        data = _mm_load_si128(reinterpret_cast<const __m128i *>(ptr));
 
-        __m128i comparison = _mm_cmpeq_epi16(data, zeroes);
+        comparison = _mm_cmpeq_epi16(data, zeroes);
         mask = _mm_movemask_epi8(comparison);
     } while (mask == 0);
 
     // found a null
     uint idx = qCountTrailingZeroBits(quint32(mask));
-    return ptr - str - 8 + idx / 2;
+    return ptr - str + idx / 2;
 #endif
 
     if (sizeof(wchar_t) == sizeof(ushort))
diff --git a/tests/auto/corelib/tools/qstringview/tst_qstringview.cpp b/tests/auto/corelib/tools/qstringview/tst_qstringview.cpp
index 48ea5a794c..4174b85f4c 100644
--- a/tests/auto/corelib/tools/qstringview/tst_qstringview.cpp
+++ b/tests/auto/corelib/tools/qstringview/tst_qstringview.cpp
@@ -364,14 +364,21 @@ void tst_QStringView::basics() const
 void tst_QStringView::literals() const
 {
 #if !defined(Q_OS_WIN) || defined(Q_COMPILER_UNICODE_STRINGS)
-    // the + ensures it's a pointer, not an array
-    QCOMPARE(QStringView(+u"Hello").size(), 5);
-    QStringView sv = u"Hello";
+    const char16_t hello[] = u"Hello";
+    const char16_t longhello[] =
+            u"Hello World. This is a much longer message, to exercise qustrlen.";
+    const char16_t withnull[] = u"a\0zzz";
 #else // storage_type is wchar_t
-    // the + ensures it's a pointer, not an array
-    QCOMPARE(QStringView(+L"Hello").size(), 5);
-    QStringView sv = L"Hello";
+    const wchar_t hello[] = L"Hello";
+    const wchar_t longhello[] =
+            L"Hello World. This is a much longer message, to exercise qustrlen.";
+    const wchar_t withnull[] = L"a\0zzz";
 #endif
+    Q_STATIC_ASSERT(sizeof(longhello) >= 16);
+
+    QCOMPARE(QStringView(hello).size(), 5);
+    QCOMPARE(QStringView(hello + 0).size(), 5); // forces decay to pointer
+    QStringView sv = hello;
     QCOMPARE(sv.size(), 5);
     QVERIFY(!sv.empty());
     QVERIFY(!sv.isEmpty());
@@ -390,6 +397,24 @@ void tst_QStringView::literals() const
     QVERIFY(!sv2.isNull());
     QVERIFY(!sv2.empty());
     QCOMPARE(sv2.size(), 5);
+
+    QStringView sv3(longhello);
+    QCOMPARE(size_t(sv3.size()), sizeof(longhello)/sizeof(longhello[0]) - 1);
+    QCOMPARE(sv3.last(), QLatin1Char('.'));
+    sv3 = longhello;
+    QCOMPARE(size_t(sv3.size()), sizeof(longhello)/sizeof(longhello[0]) - 1);
+
+    for (int i = 0; i < sv3.size(); ++i) {
+        QStringView sv4(longhello + i);
+        QCOMPARE(size_t(sv4.size()), sizeof(longhello)/sizeof(longhello[0]) - 1 - i);
+        QCOMPARE(sv4.last(), QLatin1Char('.'));
+        sv4 = longhello + i;
+        QCOMPARE(size_t(sv4.size()), sizeof(longhello)/sizeof(longhello[0]) - 1 - i);
+    }
+
+    // these are different results
+    QCOMPARE(size_t(QStringView(withnull).size()), sizeof(withnull)/sizeof(withnull[0]) - 1);
+    QCOMPARE(QStringView(withnull + 0).size(), 1);
 }
 
 void tst_QStringView::at() const