diff --git a/makefile b/makefile
index 9805cae..7ac27c0 100644
--- a/makefile
+++ b/makefile
@@ -16,6 +16,9 @@ release: test.cpp magyarsort.h
 release_ypsu: ypsu.cpp magyarsort.h
 	g++ ypsu.cpp -DNDEBUG -std=c++17 -O2 -o ypsu.out
 
+release3_ypsu: ypsu.cpp magyarsort.h
+	g++ ypsu.cpp -DNDEBUG -std=c++17 -O3 -o ypsu.out
+
 release_ypsu_assert: ypsu.cpp magyarsort.h
 	g++ ypsu.cpp -std=c++17 -O2 -o ypsu.out
 
diff --git a/thiersort2.h b/thiersort2.h
index 04d06ee..b47b04c 100644
--- a/thiersort2.h
+++ b/thiersort2.h
@@ -24,8 +24,12 @@ static inline uint32_t witch_bucket(uint32_t key) {
 	/* This approach uses 12 bits from a 32 bit float to map onto a byte bucket index */
 	th2_fu as;
 	as.f = (float) key;
+
 	uint32_t witch_base = (key <= 2) ? 0 : (as.u >> 23) - 128; // 0, [127..159] -> [0..31]
 	return witch_base * 8 + ((as.u >> (23 - 3)) & 7);
+	/* Alternative (but I measure it being worse):
+	return (as.u >> 23);
+	*/
 }
 
 /**