Use int64_t internally for metrics

So we can use fetch_add

ConflictSet.cpp

@@ -18,6 +18,7 @@ limitations under the License.
 #include "Internal.h"
 
 #include <algorithm>
+#include <atomic>
 #include <bit>
 #include <cassert>
 #include <compare>
@@ -552,7 +553,7 @@ struct Metric {
   const char *name;
   const char *help;
   ConflictSet::MetricsV1::Type type;
-  std::atomic<double> value;
+  std::atomic<int64_t> value;
 
 protected:
   Metric(ConflictSet::Impl *impl, const char *name, const char *help,
@@ -563,7 +564,7 @@ struct Gauge : private Metric {
   Gauge(ConflictSet::Impl *impl, const char *name, const char *help)
       : Metric(impl, name, help, ConflictSet::MetricsV1::Gauge) {}
 
-  void set(double value) {
+  void set(int64_t value) {
     this->value.store(value, std::memory_order_relaxed);
   }
 };
@@ -573,17 +574,10 @@ struct Counter : private Metric {
       : Metric(impl, name, help, ConflictSet::MetricsV1::Counter) {}
   // Expensive. Accumulate locally and then call add instead of repeatedly
   // calling add.
-  void add(double value) {
+  void add(int64_t value) {
     assert(value >= 0);
-    static_assert(std::atomic<double>::is_always_lock_free);
+    static_assert(std::atomic<int64_t>::is_always_lock_free);
-    double old = this->value.load(std::memory_order_relaxed);
+    this->value.fetch_add(value, std::memory_order_relaxed);
-    for (;;) {
-      double newVal = old + value;
-      if (this->value.compare_exchange_weak(old, newVal,
-                                            std::memory_order_relaxed)) {
-        break;
-      }
-    }
   }
 };
 
@@ -708,29 +702,29 @@ size_t Node::size() const {
 // A type that's plumbed along the check call tree. Lifetime ends after each
 // check call.
 struct ReadContext {
-  double point_read_accum = 0;
+  int64_t point_read_accum = 0;
-  double prefix_read_accum = 0;
+  int64_t prefix_read_accum = 0;
-  double range_read_accum = 0;
+  int64_t range_read_accum = 0;
-  double point_read_short_circuit_accum = 0;
+  int64_t point_read_short_circuit_accum = 0;
-  double prefix_read_short_circuit_accum = 0;
+  int64_t prefix_read_short_circuit_accum = 0;
-  double range_read_short_circuit_accum = 0;
+  int64_t range_read_short_circuit_accum = 0;
-  double point_read_iterations_accum = 0;
+  int64_t point_read_iterations_accum = 0;
-  double prefix_read_iterations_accum = 0;
+  int64_t prefix_read_iterations_accum = 0;
-  double range_read_iterations_accum = 0;
+  int64_t range_read_iterations_accum = 0;
-  double range_read_node_scan_accum = 0;
+  int64_t range_read_node_scan_accum = 0;
   ConflictSet::Impl *impl;
 };
 
 // A type that's plumbed along the non-const call tree. Same lifetime as
 // ConflictSet::Impl
 struct WriteContext {
-  double entries_erased_accum = 0;
+  int64_t entries_erased_accum = 0;
-  double insert_iterations_accum = 0;
+  int64_t insert_iterations_accum = 0;
-  double entries_inserted_accum = 0;
+  int64_t entries_inserted_accum = 0;
-  double nodes_allocated_accum = 0;
+  int64_t nodes_allocated_accum = 0;
-  double nodes_released_accum = 0;
+  int64_t nodes_released_accum = 0;
-  double point_writes_accum = 0;
+  int64_t point_writes_accum = 0;
-  double range_writes_accum = 0;
+  int64_t range_writes_accum = 0;
   template <class T> T *allocate(int c) {
     ++nodes_allocated_accum;
     if constexpr (std::is_same_v<T, Node0>) {
@@ -3122,21 +3116,15 @@ Node *firstGeqPhysical(Node *n, const std::span<const uint8_t> key) {
   }
 }
 
-#define MEASURE_CHECK_CPU_TIME 0
-
 struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
 
   void check(const ReadRange *reads, Result *result, int count) {
-#if MEASURE_CHECK_CPU_TIME
-    timespec ts_begin;
-    clock_gettime(CLOCK_THREAD_CPUTIME_ID, &ts_begin);
-#endif
     ReadContext tls;
     tls.impl = this;
     int commits_accum = 0;
     int conflicts_accum = 0;
     int too_olds_accum = 0;
-    double check_byte_accum = 0;
+    int64_t check_byte_accum = 0;
     for (int i = 0; i < count; ++i) {
       const auto &r = reads[i];
       check_byte_accum += r.begin.len + r.end.len;
@@ -3171,13 +3159,6 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
     conflicts_total.add(conflicts_accum);
     too_olds_total.add(too_olds_accum);
     check_bytes_total.add(check_byte_accum);
-#if MEASURE_CHECK_CPU_TIME
-    timespec ts_end;
-    clock_gettime(CLOCK_THREAD_CPUTIME_ID, &ts_end);
-    check_cpu_seconds_total.add(
-        std::max<double>(0, (ts_end.tv_nsec * 1e-9 + ts_end.tv_sec) -
-                                (ts_begin.tv_nsec * 1e-9 + ts_begin.tv_sec)));
-#endif
   }
 
   void addWrites(const WriteRange *writes, int count, int64_t writeVersion) {
@@ -3203,7 +3184,7 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
       gcScanStep(1000);
     }
 
-    double write_byte_accum = 0;
+    int64_t write_byte_accum = 0;
     for (int i = 0; i < count; ++i) {
       const auto &w = writes[i];
       write_byte_accum += w.begin.len + w.end.len;
@@ -3242,7 +3223,7 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
       rootMaxVersion = std::max(rootMaxVersion, oldestVersion);
       n = nextPhysical(n);
     }
-    double set_oldest_iterations_accum = 0;
+    int64_t set_oldest_iterations_accum = 0;
     for (; fuel > 0 && n != nullptr; ++set_oldest_iterations_accum) {
       rezero(n, oldestVersion);
       // The "make sure gc keeps up with writes" calculations assume that we're
@@ -3443,10 +3424,6 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
   GAUGE(
       oldest_extant_version,
       "A lower bound on the lowest version associated with an existing entry");
-#if MEASURE_CHECK_CPU_TIME
-  COUNTER(check_cpu_seconds_total,
-          "Total cpu seconds spent in a call to check");
-#endif
   // ==================== END METRICS DEFINITIONS ====================
 #undef GAUGE
 #undef COUNTER