diff --git a/benchmarks/BenchOps/BenchOps.cpp b/benchmarks/BenchOps/BenchOps.cpp
index fe44e4a4c..34d477116 100644
--- a/benchmarks/BenchOps/BenchOps.cpp
+++ b/benchmarks/BenchOps/BenchOps.cpp
@@ -14,6 +14,19 @@ Data src[16], dst[16];
 
 #define N 10000000
 
+int count;
+
+force_inline void SSEZero16(void *t)
+{
+	_mm_storeu_si128((__m128i*)t, _mm_setzero_si128());
+}
+
+force_inline void SSEZero32(void *t)
+{
+	_mm_storeu_si128((__m128i*)t, _mm_setzero_si128());
+	_mm_storeu_si128((__m128i*)t + 1, _mm_setzero_si128());
+}
+
 CONSOLE_APP_MAIN
 {
 	{
@@ -23,6 +36,17 @@ CONSOLE_APP_MAIN
 			AtomicDec(val);
 		}
 	}
+	{
+		RTIMING("Atomic and test");
+		for(int i = 0; i < N; i++) {
+			if(i & 1) {
+				AtomicInc(val);
+				AtomicInc(val);
+			}
+			if(AtomicDec(val))
+				count++;
+		}
+	}
 	{
 		static StaticMutex mtx;
 		RTIMING("StaticMutex");
@@ -50,6 +74,12 @@ CONSOLE_APP_MAIN
 			delete[] new byte[32];
 		}
 	}
+	{
+		RTIMING("Alloc32/Free32");
+		for(int i = 0; i < N; i++) {
+			MemoryFree32(MemoryAlloc32());
+		}
+	}
 	{
 		RTIMING("Alloc/Free large");
 		for(int i = 0; i < N; i++) {
@@ -74,6 +104,18 @@ CONSOLE_APP_MAIN
 			dst[i & 15].Zero();
 		}
 	}
+	{
+		RTIMING("SSEZero32");
+		for(int i = 0; i < N; i++) {
+			SSEZero32(&dst[i & 15]);
+		}
+	}
+	{
+		RTIMING("SSEZero16");
+		for(int i = 0; i < N; i++) {
+			SSEZero16(&dst[i & 15]);
+		}
+	}
 	{
 		RTIMING("Copy32");
 		for(int i = 0; i < N; i++) {
diff --git a/benchmarks/BenchOps/BenchOps.upp b/benchmarks/BenchOps/BenchOps.upp
index 7f14d1880..d4f5979e6 100644
--- a/benchmarks/BenchOps/BenchOps.upp
+++ b/benchmarks/BenchOps/BenchOps.upp
@@ -2,7 +2,7 @@ uses
 	Core;
 
 file
-	BenchOps.cpp;
+	BenchOps.cpp optimize_speed;
 
 mainconfig
 	"" = "MT";