1 files changed, 35 insertions, 29 deletions

diff --git a/src/corelib/tools/qstring.cpp b/src/corelib/tools/qstring.cpp
index e27afb8..f8f4bdc 100644
--- a/src/corelib/tools/qstring.cpp
+++ b/src/corelib/tools/qstring.cpp
@@ -3575,6 +3575,38 @@ bool QString::endsWith(const QChar &c, Qt::CaseSensitivity cs) const
     Use toLocal8Bit() instead.
 */
 
+#if defined(QT_ALWAYS_HAVE_SSE2)
+static inline __m128i mergeQuestionMarks(__m128i chunk)
+{
+    const __m128i questionMark = _mm_set1_epi16('?');
+
+    // SSE has no compare instruction for unsigned comparison.
+    // The variables must be shiffted + 0x8000 to be compared
+    const __m128i signedBitOffset = _mm_set1_epi16(0x8000);
+    const __m128i thresholdMask = _mm_set1_epi16(0xff + 0x8000);
+
+    const __m128i signedChunk = _mm_add_epi16(chunk, signedBitOffset);
+    const __m128i offLimitMask = _mm_cmpgt_epi16(signedChunk, thresholdMask);
+
+#  ifdef __SSE4_1__
+    // replace the non-Latin 1 characters in the chunk with question marks
+    chunk = _mm_blendv_epi8(chunk, questionMark, offLimitMask);
+#  else
+    // offLimitQuestionMark contains '?' for each 16 bits that was off-limit
+    // the 16 bits that were correct contains zeros
+    const __m128i offLimitQuestionMark = _mm_and_si128(offLimitMask, questionMark);
+
+    // correctBytes contains the bytes that were in limit
+    // the 16 bits that were off limits contains zeros
+    const __m128i correctBytes = _mm_andnot_si128(offLimitMask, chunk);
+
+    // merge offLimitQuestionMark and correctBytes to have the result
+    chunk = _mm_or_si128(correctBytes, offLimitQuestionMark);
+#  endif
+    return chunk;
+}
+#endif
+
 static QByteArray toLatin1_helper(const QChar *data, int length)
 {
     QByteArray ba;
@@ -3585,41 +3617,15 @@ static QByteArray toLatin1_helper(const QChar *data, int length)
 #if defined(QT_ALWAYS_HAVE_SSE2)
         if (length >= 16) {
             const int chunkCount = length >> 4; // divided by 16
-            const __m128i questionMark = _mm_set1_epi16('?');
-            // SSE has no compare instruction for unsigned comparison.
-            // The variables must be shiffted + 0x8000 to be compared
-            const __m128i signedBitOffset = _mm_set1_epi16(0x8000);
-            const __m128i thresholdMask = _mm_set1_epi16(0xff + 0x8000);
+
             for (int i = 0; i < chunkCount; ++i) {
                 __m128i chunk1 = _mm_loadu_si128((__m128i*)src); // load
+                chunk1 = mergeQuestionMarks(chunk1);
                 src += 8;
-                {
-                    // each 16 bit is equal to 0xFF if the source is outside latin 1 (>0xff)
-                    const __m128i signedChunk = _mm_add_epi16(chunk1, signedBitOffset);
-                    const __m128i offLimitMask = _mm_cmpgt_epi16(signedChunk, thresholdMask);
-
-                    // offLimitQuestionMark contains '?' for each 16 bits that was off-limit
-                    // the 16 bits that were correct contains zeros
-                    const __m128i offLimitQuestionMark = _mm_and_si128(offLimitMask, questionMark);
-
-                    // correctBytes contains the bytes that were in limit
-                    // the 16 bits that were off limits contains zeros
-                    const __m128i correctBytes = _mm_andnot_si128(offLimitMask, chunk1);
-
-                    // merge offLimitQuestionMark and correctBytes to have the result
-                    chunk1 = _mm_or_si128(correctBytes, offLimitQuestionMark);
-                }
 
                 __m128i chunk2 = _mm_loadu_si128((__m128i*)src); // load
+                chunk2 = mergeQuestionMarks(chunk2);
                 src += 8;
-                {
-                    // exactly the same operations as for the previous chunk of data
-                    const __m128i signedChunk = _mm_add_epi16(chunk2, signedBitOffset);
-                    const __m128i offLimitMask = _mm_cmpgt_epi16(signedChunk, thresholdMask);
-                    const __m128i offLimitQuestionMark = _mm_and_si128(offLimitMask, questionMark);
-                    const __m128i correctBytes = _mm_andnot_si128(offLimitMask, chunk2);
-                    chunk2 = _mm_or_si128(correctBytes, offLimitQuestionMark);
-                }
 
                 // pack the two vector to 16 x 8bits elements
                 const __m128i result = _mm_packus_epi16(chunk1, chunk2);