Improve pbitops performance by enabling GCC intrinsics when available

panda/src/putil/config_util.cxx

@@ -218,13 +218,4 @@ init_libputil() {
 
   BamCacheIndex::register_with_read_factory();
   BamCacheRecord::register_with_read_factory();
-
-  // Initialize the num_bits_on table, for BitMask::get_num_on_bits().
-  for (int bit = 0; bit < 16; ++bit) {
-    int w = (1 << bit);
-    for (int i = 0; i < w; ++i) {
-      num_bits_on[i + w] = num_bits_on[i] + 1;
-    }
-  }
 }
-

panda/src/putil/pbitops.I

@@ -28,7 +28,11 @@ count_bits_in_word(PN_uint16 x) {
 ////////////////////////////////////////////////////////////////////
 INLINE int
 count_bits_in_word(PN_uint32 x) {
+#if defined(__GNUC__) && defined(__POPCNT__)
+  return __builtin_popcount((unsigned int) x);
+#else
   return (int)num_bits_on[x & 0xffff] + (int)num_bits_on[(x >> 16) & 0xffff];
+#endif
 }
 
 ////////////////////////////////////////////////////////////////////
@@ -37,7 +41,11 @@ count_bits_in_word(PN_uint32 x) {
 ////////////////////////////////////////////////////////////////////
 INLINE int
 count_bits_in_word(PN_uint64 x) {
+#if defined(__GNUC__) && defined(__POPCNT__)
+  return __builtin_popcountll((unsigned long long) x);
+#else
   return count_bits_in_word((PN_uint32)x) + count_bits_in_word((PN_uint32)(x >> 32));
+#endif
 }
 
 ////////////////////////////////////////////////////////////////////
@@ -137,12 +145,16 @@ flood_bits_up(PN_uint64 x) {
 ////////////////////////////////////////////////////////////////////
 INLINE int
 get_lowest_on_bit(PN_uint16 x) {
+#ifdef __GNUC__
+  return __builtin_ffs((unsigned int) x) - 1;
+#else
   if (x == 0) {
     return -1;
   }
 
   PN_uint16 w = (x & (~x + 1));
   return (int)num_bits_on[w - 1];
+#endif
 }
 
 ////////////////////////////////////////////////////////////////////
@@ -152,12 +164,16 @@ get_lowest_on_bit(PN_uint16 x) {
 ////////////////////////////////////////////////////////////////////
 INLINE int
 get_lowest_on_bit(PN_uint32 x) {
+#ifdef __GNUC__
+  return __builtin_ffs((unsigned int) x) - 1;
+#else
   if (x == 0) {
     return -1;
   }
 
   PN_uint32 w = (x & (~x + 1));
   return count_bits_in_word(w - 1);
+#endif
 }
 
 ////////////////////////////////////////////////////////////////////
@@ -167,12 +183,16 @@ get_lowest_on_bit(PN_uint32 x) {
 ////////////////////////////////////////////////////////////////////
 INLINE int
 get_lowest_on_bit(PN_uint64 x) {
+#ifdef __GNUC__
+  return __builtin_ffsll((unsigned long long) x) - 1;
+#else
   if (x == 0) {
     return -1;
   }
 
   PN_uint64 w = (x & (~x + 1));
   return count_bits_in_word(w - 1);
+#endif
 }
 
 ////////////////////////////////////////////////////////////////////
@@ -182,8 +202,12 @@ get_lowest_on_bit(PN_uint64 x) {
 ////////////////////////////////////////////////////////////////////
 INLINE int
 get_highest_on_bit(PN_uint16 x) {
+#ifdef __GNUC__
+  return (x == 0) ? -1 : 31 - __builtin_clz((unsigned int) x);
+#else
   PN_uint16 w = flood_bits_down(x);
   return count_bits_in_word(w) - 1;
+#endif
 }
 
 ////////////////////////////////////////////////////////////////////
@@ -193,8 +217,12 @@ get_highest_on_bit(PN_uint16 x) {
 ////////////////////////////////////////////////////////////////////
 INLINE int
 get_highest_on_bit(PN_uint32 x) {
+#ifdef __GNUC__
+  return (x == 0) ? -1 : 31 - __builtin_clz((unsigned int) x);
+#else
   PN_uint32 w = flood_bits_down(x);
   return count_bits_in_word(w) - 1;
+#endif
 }
 
 ////////////////////////////////////////////////////////////////////
@@ -204,8 +232,12 @@ get_highest_on_bit(PN_uint32 x) {
 ////////////////////////////////////////////////////////////////////
 INLINE int
 get_highest_on_bit(PN_uint64 x) {
+#ifdef __GNUC__
+  return (x == 0) ? -1 : 63 - __builtin_clzll((unsigned long long) x);
+#else
   PN_uint64 w = flood_bits_down(x);
   return count_bits_in_word(w) - 1;
+#endif
 }
 
 ////////////////////////////////////////////////////////////////////

panda/src/putil/pbitops.h

@@ -46,7 +46,7 @@ INLINE int get_next_higher_bit(PN_uint32 x);
 INLINE int get_next_higher_bit(PN_uint64 x);
 
 // This table precomputes the number of on bits in each 16-bit word.
-extern EXPCL_PANDA_PUTIL unsigned char num_bits_on[65536];
+extern EXPCL_PANDA_PUTIL const unsigned char num_bits_on[65536];
 
 #include "pbitops.I"