Add a once-over unrolled fromLatin1 conversion (32 characters)

This was suggested by Matthias Kretz on my blog. It improves on GCC
(one case up to 5%), but worsens on ICC (up to 5%).

1 files changed, 46 insertions, 0 deletions

diff --git a/tests/benchmarks/corelib/tools/qstring/main.cpp b/tests/benchmarks/corelib/tools/qstring/main.cpp
index 1a38354..96f2c30 100644
--- a/tests/benchmarks/corelib/tools/qstring/main.cpp
+++ b/tests/benchmarks/corelib/tools/qstring/main.cpp
@@ -1538,6 +1538,51 @@ void fromLatin1_sse2_improved(ushort *dst, const char *str, int size)
     fromLatin1_epilog(dst + counter, str + counter, size - counter);
 }
 
+void fromLatin1_sse2_improved2(ushort *dst, const char *str, int size)
+{
+    const __m128i nullMask = _mm_set1_epi32(0);
+    qptrdiff counter = 0;
+    size -= 32;
+    while (size >= counter) {
+        const __m128i chunk1 = _mm_loadu_si128((__m128i*)(str + counter)); // load
+        const __m128i chunk2 = _mm_loadu_si128((__m128i*)(str + counter + 16)); // load
+
+        // unpack the first 8 bytes, padding with zeros
+        const __m128i firstHalf1 = _mm_unpacklo_epi8(chunk1, nullMask);
+        _mm_storeu_si128((__m128i*)(dst + counter), firstHalf1); // store
+
+        // unpack the last 8 bytes, padding with zeros
+        const __m128i secondHalf1 = _mm_unpackhi_epi8(chunk1, nullMask);
+        _mm_storeu_si128((__m128i*)(dst + counter + 8), secondHalf1); // store
+
+        // unpack the first 8 bytes, padding with zeros
+        const __m128i firstHalf2 = _mm_unpacklo_epi8(chunk2, nullMask);
+        _mm_storeu_si128((__m128i*)(dst + counter + 16), firstHalf2); // store
+
+        // unpack the last 8 bytes, padding with zeros
+        const __m128i secondHalf2 = _mm_unpackhi_epi8(chunk2, nullMask);
+        _mm_storeu_si128((__m128i*)(dst + counter + 24), secondHalf2); // store
+
+        counter += 32;
+    }
+    size += 16;
+    if (size >= counter) {
+        const __m128i chunk = _mm_loadu_si128((__m128i*)(str + counter)); // load
+
+        // unpack the first 8 bytes, padding with zeros
+        const __m128i firstHalf = _mm_unpacklo_epi8(chunk, nullMask);
+        _mm_storeu_si128((__m128i*)(dst + counter), firstHalf); // store
+
+        // unpack the last 8 bytes, padding with zeros
+        const __m128i secondHalf = _mm_unpackhi_epi8 (chunk, nullMask);
+        _mm_storeu_si128((__m128i*)(dst + counter + 8), secondHalf); // store
+
+        counter += 16;
+    }
+    size += 16;
+    fromLatin1_epilog(dst + counter, str + counter, size - counter);
+}
+
 void fromLatin1_prolog_unrolled(ushort *dst, const char *str, int size)
 {
     // QString's data pointer is most often ending in 0x2 or 0xa
@@ -1787,6 +1832,7 @@ void tst_QString::fromLatin1Alternatives_data() const
 #ifdef __SSE2__
     QTest::newRow("sse2-qt4.7") << &fromLatin1_sse2_qt47;
     QTest::newRow("sse2-improved") << &fromLatin1_sse2_improved;
+    QTest::newRow("sse2-improved2") << &fromLatin1_sse2_improved2;
     QTest::newRow("sse2-with-prolog-regular") << &fromLatin1_sse2_withprolog<&fromLatin1_regular>;
     QTest::newRow("sse2-with-prolog-unrolled") << &fromLatin1_sse2_withprolog<&fromLatin1_prolog_unrolled>;
     QTest::newRow("sse2-with-prolog-sse2-overcommit") << &fromLatin1_sse2_withprolog<&fromLatin1_prolog_sse2_overcommit>;