diff --git a/src/corelib/global/qglobal.h b/src/corelib/global/qglobal.h
index cbc11893087..c6b5c5a96f1 100644
--- a/src/corelib/global/qglobal.h
+++ b/src/corelib/global/qglobal.h
@@ -626,6 +626,29 @@ using qsizetype = QIntegerForSizeof<std::size_t>::Signed;
 template <typename T>
 constexpr inline T qAbs(const T &t) { return t >= 0 ? t : -t; }
 
+// gcc < 10 doesn't have __has_builtin
+#if defined(Q_PROCESSOR_ARM_64) && (__has_builtin(__builtin_round) || defined(Q_CC_GNU)) && !defined(Q_CC_CLANG)
+// ARM64 has a single instruction that can do C++ rounding with conversion to integer.
+// Note current clang versions have non-constexpr __builtin_round, ### allow clang this path when they fix it.
+constexpr inline int qRound(double d)
+{ return int(__builtin_round(d)); }
+constexpr inline int qRound(float f)
+{ return int(__builtin_roundf(f)); }
+constexpr inline qint64 qRound64(double d)
+{ return qint64(__builtin_round(d)); }
+constexpr inline qint64 qRound64(float f)
+{ return qint64(__builtin_roundf(f)); }
+#elif defined(__SSE2__) && (__has_builtin(__builtin_copysign) || defined(Q_CC_GNU))
+// SSE has binary operations directly on floating point making copysign fast
+constexpr inline int qRound(double d)
+{ return int(d + __builtin_copysign(0.5, d)); }
+constexpr inline int qRound(float f)
+{ return int(f + __builtin_copysignf(0.5f, f)); }
+constexpr inline qint64 qRound64(double d)
+{ return qint64(d + __builtin_copysign(0.5, d)); }
+constexpr inline qint64 qRound64(float f)
+{ return qint64(f + __builtin_copysignf(0.5f, f)); }
+#else
 constexpr inline int qRound(double d)
 { return d >= 0.0 ? int(d + 0.5) : int(d - 0.5); }
 constexpr inline int qRound(float d)
@@ -635,6 +658,7 @@ constexpr inline qint64 qRound64(double d)
 { return d >= 0.0 ? qint64(d + 0.5) : qint64(d - 0.5); }
 constexpr inline qint64 qRound64(float d)
 { return d >= 0.0f ? qint64(d + 0.5f) : qint64(d - 0.5f); }
+#endif
 
 namespace QTypeTraits {