util: Optimize util_is_power_of_two_nonzero

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Suggested-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Eduardo Lima Mitev <elima@igalia.com>

diff --git a/src/util/bitscan.h b/src/util/bitscan.h
index a3f2d414bd6ec535c0f66d9d15729a2880ac2fbc..5cc75f0beba0f7cf029bc85ad62bc948916c1b84 100644 (file)
--- a/src/util/bitscan.h
+++ b/src/util/bitscan.h
@@ -38,6 +38,10 @@
 #include <intrin.h>
 #endif
 
+#if defined(__POPCNT__)
+#include <popcntintrin.h>
+#endif
+
 #include "c99_compat.h"
 
 #ifdef __cplusplus
@@ -127,7 +131,20 @@ util_is_power_of_two_or_zero(unsigned v)
 static inline bool
 util_is_power_of_two_nonzero(unsigned v)
 {
+   /* __POPCNT__ is different from HAVE___BUILTIN_POPCOUNT.  The latter
+    * indicates the existence of the __builtin_popcount function.  The former
+    * indicates that _mm_popcnt_u32 exists and is a native instruction.
+    *
+    * The other alternative is to use SSE 4.2 compile-time flags.  This has
+    * two drawbacks.  First, there is currently no build infrastructure for
+    * SSE 4.2 (only 4.1), so that would have to be added.  Second, some AMD
+    * CPUs support POPCNT but not SSE 4.2 (e.g., Barcelona).
+    */
+#ifdef __POPCNT__
+   return _mm_popcnt_u32(v) == 1;
+#else
    return v != 0 && (v & (v - 1)) == 0;
+#endif
 }
 
 /* For looping over a bitmask when you want to loop over consecutive bits