5 files changed, 65 insertions, 6 deletions
diff --git a/src/corelib/json/qjson_p.h b/src/corelib/json/qjson_p.h
index 0b3f517990..59d0c91785 100644
--- a/src/corelib/json/qjson_p.h
+++ b/src/corelib/json/qjson_p.h
@@ -402,7 +402,7 @@ public:
             // pack with itself, we'll discard the high part anyway
             chunk = _mm_packus_epi16(chunk, chunk);
             // unaligned 64-bit store
-            *(quint64*)&l[i] = _mm_cvtsi128_si64(chunk);
+            qUnalignedStore(l + i, _mm_cvtsi128_si64(chunk));
             i += 8;
         }
 #  endif
diff --git a/src/corelib/tools/qhash.cpp b/src/corelib/tools/qhash.cpp
index 775e1364a1..ac9e08de8b 100644
--- a/src/corelib/tools/qhash.cpp
+++ b/src/corelib/tools/qhash.cpp
@@ -106,24 +106,24 @@ static uint crc32(const Char *ptr, size_t len, uint h)
 
     p += 8;
     for ( ; p <= e; p += 8)
-        h2 = _mm_crc32_u64(h2, *reinterpret_cast<const qlonglong *>(p - 8));
+        h2 = _mm_crc32_u64(h2, qUnalignedLoad<qlonglong>(p - 8));
     h = h2;
     p -= 8;
 
     len = e - p;
     if (len & 4) {
-        h = _mm_crc32_u32(h, *reinterpret_cast<const uint *>(p));
+        h = _mm_crc32_u32(h, qUnalignedLoad<uint>(p));
         p += 4;
     }
 #  else
     p += 4;
     for ( ; p <= e; p += 4)
-        h = _mm_crc32_u32(h, *reinterpret_cast<const uint *>(p - 4));
+        h = _mm_crc32_u32(h, qUnalignedLoad<uint>(p - 4));
     p -= 4;
     len = e - p;
 #  endif
     if (len & 2) {
-        h = _mm_crc32_u16(h, *reinterpret_cast<const ushort *>(p));
+        h = _mm_crc32_u16(h, qUnalignedLoad<ushort>(p));
         p += 2;
     }
     if (sizeof(Char) == 1 && len & 1)
diff --git a/src/corelib/tools/qsimd.cpp b/src/corelib/tools/qsimd.cpp
index f07eb098f2..5ca2ce4c6f 100644
--- a/src/corelib/tools/qsimd.cpp
+++ b/src/corelib/tools/qsimd.cpp
@@ -716,4 +716,26 @@ void qDumpCPUFeatures()
     puts("");
 }
 
+/*!
+    \internal
+    \fn T qUnalignedLoad(const void *ptr)
+    \since 5.6.1
+
+    Loads a \c{T} from address \a ptr, which may be misaligned.
+
+    Use of this function avoid the undefined behavior that the C++ standard
+    otherwise attributes to unaligned loads.
+*/
+
+/*!
+    \internal
+    \fn void qUnalignedStore(void *ptr, T t)
+    \since 5.6.1
+
+    Stores \a t to address \a ptr, which may be misaligned.
+
+    Use of this function avoid the undefined behavior that the C++ standard
+    otherwise attributes to unaligned stores.
+*/
+
 QT_END_NAMESPACE
diff --git a/src/corelib/tools/qsimd_p.h b/src/corelib/tools/qsimd_p.h
index 12a329f36c..8171184ad2 100644
--- a/src/corelib/tools/qsimd_p.h
+++ b/src/corelib/tools/qsimd_p.h
@@ -470,6 +470,43 @@ unsigned _bit_scan_forward(unsigned val)
 #define ALIGNMENT_PROLOGUE_16BYTES(ptr, i, length) \
     for (; i < static_cast<int>(qMin(static_cast<quintptr>(length), ((4 - ((reinterpret_cast<quintptr>(ptr) >> 2) & 0x3)) & 0x3))); ++i)
 
+// these defines are copied from qendian.h
+// in Qt 5.7, they have been moved to qglobal.h
+// drop them when merging this to 5.7
+#ifdef __has_builtin
+#  define QT_HAS_BUILTIN(x)     __has_builtin(x)
+#else
+#  define QT_HAS_BUILTIN(x)     0
+#endif
+
+template <typename T>
+Q_ALWAYS_INLINE
+T qUnalignedLoad(const void *ptr) Q_DECL_NOTHROW
+{
+    T result;
+#if QT_HAS_BUILTIN(__builtin_memcpy)
+    __builtin_memcpy
+#else
+    memcpy
+#endif
+    /*memcpy*/(&result, ptr, sizeof result);
+    return result;
+}
+
+template <typename T>
+Q_ALWAYS_INLINE
+void qUnalignedStore(void *ptr, T t) Q_DECL_NOTHROW
+{
+#if QT_HAS_BUILTIN(__builtin_memcpy)
+    __builtin_memcpy
+#else
+    memcpy
+#endif
+    /*memcpy*/(ptr, &t, sizeof t);
+}
+
+#undef QT_HAS_BUILTIN
+
 QT_END_NAMESPACE
 
 #endif // QSIMD_P_H
diff --git a/src/corelib/tools/qstring.cpp b/src/corelib/tools/qstring.cpp
index 39ec66c7f1..9924b606c5 100644
--- a/src/corelib/tools/qstring.cpp
+++ b/src/corelib/tools/qstring.cpp
@@ -577,7 +577,7 @@ static int ucstrncmp(const QChar *a, const uchar *c, int l)
     // we'll read uc[offset..offset+7] (16 bytes) and c[offset..offset+7] (8 bytes)
     if (uc + offset + 7 < e) {
         // same, but we're using an 8-byte load
-        __m128i chunk = _mm_cvtsi64_si128(*(const long long *)(c + offset));
+        __m128i chunk = _mm_cvtsi64_si128(qUnalignedLoad<long long>(c + offset));
         __m128i secondHalf = _mm_unpacklo_epi8(chunk, nullmask);
 
         __m128i ucdata = _mm_loadu_si128((const __m128i*)(uc + offset));