[STATS] separate noTotal bit flag from onlyInMaster and noUnits.

[jlpeyton]

Summary

This change logically separates the stats_flags_e::noTotal bit flag from the stats_flags_e::onlyInMaster and stats_flags_e::noUnits bit flags. If no TOTAL_foo output is wanted for a particular statistic, the flag must be explicitly included in that statistic's flags.

[jcownie]

LGTM

openmp/trunk/runtime/src/kmp_stats.h

 // #define KMP_DEVELOPER_STATS 1
 
 /*!
  * @ingroup STATS_GATHERING
  * \brief flags to describe the statistic (timer or counter)
  *
  */
 enum stats_flags_e {
-    noTotal      = 1<<0,             //!< do not show a TOTAL_aggregation for this statistic
-    onlyInMaster = (1<<1) | noTotal, //!< statistic is valid only for master
-    noUnits      = (1<<2) | noTotal, //!< statistic doesn't need units printed next to it in output
-    notInMaster  = 1<<3,             //!< statistic is valid only for non-master threads
-    logEvent     = 1<<4              //!< statistic can be logged when KMP_STATS_EVENTS is on (valid only for timers)
+    noTotal      = 1<<0,     //!< do not show a TOTAL_aggregation for this statistic
+    onlyInMaster = 1<<1,     //!< statistic is valid only for master
+    noUnits      = 1<<2,     //!< statistic doesn't need units printed next to it in output
+    notInMaster  = 1<<3,     //!< statistic is valid only for non-master threads
+    logEvent     = 1<<4      //!< statistic can be logged on the event timeline when KMP_STATS_EVENTS is on (valid only for timers)
 };
 
 /*!
  * \brief Add new counters under KMP_FOREACH_COUNTER() macro in kmp_stats.h
  *
  * @param macro a user defined macro that takes three arguments - macro(COUNTER_NAME, flags, arg)
  * @param arg a user defined argument to send to the user defined macro
  *
  * \details A counter counts the occurrence of some event.
  * Each thread accumulates its own count, at the end of execution the counts are aggregated treating each thread
  * as a separate measurement. (Unless onlyInMaster is set, in which case there's only a single measurement).
  * The min,mean,max are therefore the values for the threads.
  * Adding the counter here and then putting a KMP_BLOCK_COUNTER(name) at the point you want to count is all you need to do.
  * All of the tables and printing is generated from this macro.
  * Format is "macro(name, flags, arg)"
  *
  * @ingroup STATS_GATHERING
  */
 #define KMP_FOREACH_COUNTER(macro, arg)                         \
-    macro (OMP_PARALLEL, stats_flags_e::onlyInMaster, arg)      \
+    macro (OMP_PARALLEL, stats_flags_e::onlyInMaster | stats_flags_e::noTotal, arg) \
     macro (OMP_NESTED_PARALLEL, 0, arg)                         \
     macro (OMP_FOR_static, 0, arg)                              \
     macro (OMP_FOR_dynamic, 0, arg)                             \
     macro (OMP_DISTRIBUTE, 0, arg)                              \
     macro (OMP_BARRIER, 0, arg)                                 \
     macro (OMP_CRITICAL,0, arg)                                 \
     macro (OMP_SINGLE, 0, arg)                                  \
     macro (OMP_MASTER, 0, arg)                                  \

  * For most timers the printing code also provides an aggregation over the thread totals. These are printed as TOTAL_foo.
  * The count is normally a time (in ticks), hence the name "timer". (But can be any value, so we use this for "number of arguments passed to fork"
  * as well).
  * For timers the threads are not significant, it's the individual observations that count, so the statistics are at that level.
  * Format is "macro(name, flags, arg)"
  *
  * @ingroup STATS_GATHERING2
  */
-#define KMP_FOREACH_TIMER(macro, arg)                           \
-    macro (OMP_start_end, stats_flags_e::onlyInMaster, arg)     \
-    macro (OMP_serial, stats_flags_e::onlyInMaster, arg)        \
-    macro (OMP_work, 0, arg)                                    \
-    macro (OMP_barrier, 0, arg)                                 \
-    macro (FOR_static_scheduling, 0, arg)                       \
-    macro (FOR_dynamic_scheduling, 0, arg)                      \
-    macro (OMP_task,   0, arg)                                  \
-    macro (OMP_single, 0, arg)                                  \
-    macro (OMP_master, 0, arg)                                  \
-    macro (OMP_set_numthreads, stats_flags_e::noUnits, arg)     \
-    macro (OMP_PARALLEL_args,  stats_flags_e::noUnits, arg)     \
-    macro (FOR_static_iterations, stats_flags_e::noUnits, arg)  \
-    macro (FOR_dynamic_iterations, stats_flags_e::noUnits, arg) \
-    KMP_FOREACH_DEVELOPER_TIMER(macro, arg)                     \
+#define KMP_FOREACH_TIMER(macro, arg)                                   \
+    macro (OMP_start_end, stats_flags_e::onlyInMaster | stats_flags_e::noTotal, arg) \
+    macro (OMP_serial,    stats_flags_e::onlyInMaster | stats_flags_e::noTotal, arg) \
+    macro (OMP_work,      0, arg)                                       \
+    macro (OMP_barrier,   0, arg)                                       \
+    macro (FOR_static_scheduling, 0, arg)                               \
+    macro (FOR_dynamic_scheduling, 0, arg)                              \
+    macro (OMP_task,      0, arg)                                       \
+    macro (OMP_single,    0, arg)                                       \
+    macro (OMP_master,    0, arg)                                       \
+    macro (OMP_set_numthreads,    stats_flags_e::noUnits | stats_flags_e::noTotal, arg) \
+    macro (OMP_PARALLEL_args,     stats_flags_e::noUnits | stats_flags_e::noTotal, arg) \
+    macro (FOR_static_iterations, stats_flags_e::noUnits | stats_flags_e::noTotal, arg) \
+    macro (FOR_dynamic_iterations,stats_flags_e::noUnits | stats_flags_e::noTotal, arg) \
+    KMP_FOREACH_DEVELOPER_TIMER(macro, arg)                             \
     macro (LAST,0, arg)
 
 
 // OMP_start_end          -- Time from when OpenMP is initialized until the stats are printed at exit
 // OMP_serial             -- Thread zero time executing serial code
 // OMP_work               -- Elapsed time in code dispatched by a fork (measured in the thread)
 // OMP_barrier            -- Time at "real" barriers (includes task time)
 // FOR_static_scheduling  -- Time spent doing scheduling for a static "for"

 
 class timeStat : public statistic
 {
     static statInfo timerInfo[];
 
  public:
     timeStat() : statistic() {}
     static const char * name(timer_e e) { return timerInfo[e].name; }
-    static bool  noTotal    (timer_e e) { return timerInfo[e].flags & stats_flags_e::noTotal; }
+    static bool  noTotal    (timer_e e) { return timerInfo[e].flags & stats_flags_e::noTotal;      }
     static bool  masterOnly (timer_e e) { return timerInfo[e].flags & stats_flags_e::onlyInMaster; }
     static bool  workerOnly (timer_e e) { return timerInfo[e].flags & stats_flags_e::notInMaster;  }
     static bool  noUnits    (timer_e e) { return timerInfo[e].flags & stats_flags_e::noUnits;      }
     static bool  logEvent   (timer_e e) { return timerInfo[e].flags & stats_flags_e::logEvent;     }
     static void  clearEventFlags()      {
         for(int i=0;i<TIMER_LAST;i++) {
             timerInfo[i].flags &= (~(stats_flags_e::logEvent));
         }