x86-64 transform optimizations (Mikko T.)

[mesa.git] / src / mesa / math / m_debug_util.h

diff --git a/src/mesa/math/m_debug_util.h b/src/mesa/math/m_debug_util.h
index c07cdcf7ba7520dc56daec49955e6ab90ba089ea..765f54dfb5a4d6df0442c8df79bc6cce93b2f3a2 100644 (file)
--- a/src/mesa/math/m_debug_util.h
+++ b/src/mesa/math/m_debug_util.h
@@ -185,6 +185,44 @@ extern char *mesa_profile;
 
 #endif
 
+#elif defined(__amd64__)
+
+#define rdtscll(val) do { \
+     unsigned int a,d; \
+     __asm__ volatile("rdtsc" : "=a" (a), "=d" (d)); \
+     (val) = ((unsigned long)a) | (((unsigned long)d)<<32); \
+} while(0) 
+
+/* Copied from i386 PIII version */
+#define  INIT_COUNTER()                                                        \
+   do {                                                                        \
+      int cycle_i;                                                     \
+      counter_overhead = LONG_MAX;                                     \
+      for ( cycle_i = 0 ; cycle_i < 16 ; cycle_i++ ) {                 \
+        unsigned long cycle_tmp1, cycle_tmp2;                          \
+        rdtscll(cycle_tmp1);                                           \
+        rdtscll(cycle_tmp2);                                           \
+        if ( counter_overhead > (cycle_tmp2 - cycle_tmp1) ) {          \
+           counter_overhead = cycle_tmp2 - cycle_tmp1;                 \
+        }                                                              \
+      }                                                                        \
+   } while (0)
+
+
+#define  BEGIN_RACE(x)                                                 \
+   x = LONG_MAX;                                                       \
+   for ( cycle_i = 0 ; cycle_i < 10 ; cycle_i++ ) {                    \
+      unsigned long cycle_tmp1, cycle_tmp2;                            \
+      rdtscll(cycle_tmp1);                                             \
+
+#define END_RACE(x)                                                    \
+      rdtscll(cycle_tmp2);                                             \
+      if ( x > (cycle_tmp2 - cycle_tmp1) ) {                           \
+        x = cycle_tmp2 - cycle_tmp1;                                   \
+      }                                                                        \
+   }                                                                   \
+   x -= counter_overhead;
+
 #elif defined(__sparc__)
 
 #define  INIT_COUNTER()        \