feat!(google_benchmark): improve location of valgrind trapdoors

Author: GuillaumeLagrange

core/BUILD

@@ -7,7 +7,7 @@ CODSPEED_VERSION = "1.0.0"
 cc_library(
     name = "codspeed",
     srcs = glob(["src/**/*.cpp"] + ["src/**/*.h"] + ["src/**/*.hpp"]),
-    hdrs = ["include/codspeed.h"],
+    hdrs = glob(["include/**/*.h"] + ["include/**/*.hpp"]),
     includes = ["include"],
     defines = [
         "CODSPEED_VERSION=\\\"{}\\\"".format(CODSPEED_VERSION),

core/src/callgrind.h renamed to core/include/callgrind.h

@@ -20,12 +20,13 @@ inline void measurement_set_metadata() {
   CALLGRIND_DUMP_STATS_AT(metadata.c_str());
 }
 
-inline void measurement_start() {
+__attribute__((always_inline)) inline void measurement_start() {
   CALLGRIND_ZERO_STATS;
   CALLGRIND_START_INSTRUMENTATION;
 }
 
-inline void measurement_stop(const std::string &name) {
+__attribute__((always_inline)) inline void measurement_stop(
+    const std::string &name) {
   CALLGRIND_STOP_INSTRUMENTATION;
   CALLGRIND_DUMP_STATS_AT(name.c_str());
 };

core/src/measurement.hpp renamed to core/include/measurement.hpp

@@ -82,7 +82,6 @@ void CodSpeed::start_benchmark(const std::string &name) {
   }
 
   current_benchmark = uri;
-  measurement_start();
 }
 
 void CodSpeed::end_benchmark() {

core/src/valgrind.h renamed to core/include/valgrind.h

@@ -185,6 +185,7 @@ BENCHMARK(BM_test)->Unit(benchmark::kMillisecond);
 
 #ifdef CODSPEED_ENABLED
 #include <codspeed.h>
+#include <measurement.hpp>>
 
 #include <filesystem>
 #endif // CODSPEED_ENABLED
@@ -945,6 +946,9 @@ class BENCHMARK_EXPORT BENCHMARK_INTERNAL_CACHELINE_ALIGNED State {
 
  public:
   const IterationCount max_iterations;
+#ifdef CODSPEED_INSTRUMENTATION
+  codspeed::CodSpeed *codspeed_;
+#endif
 
  private:
   bool started_;
@@ -965,7 +969,12 @@ class BENCHMARK_EXPORT BENCHMARK_INTERNAL_CACHELINE_ALIGNED State {
         const std::vector<int64_t>& ranges, int thread_i, int n_threads,
         internal::ThreadTimer* timer, internal::ThreadManager* manager,
         internal::PerfCountersMeasurement* perf_counters_measurement,
-        ProfilerManager* profiler_manager);
+        ProfilerManager* profiler_manager
+  #ifdef CODSPEED_INSTRUMENTATION
+        ,
+        codspeed::CodSpeed *codspeed = NULL
+  #endif
+        );
 
   void StartKeepRunning();
   // Implementation of KeepRunning() and KeepRunningBatch().
@@ -1055,6 +1064,11 @@ struct State::StateIterator {
   BENCHMARK_ALWAYS_INLINE
   bool operator!=(StateIterator const&) const {
     if (BENCHMARK_BUILTIN_EXPECT(cached_ != 0, true)) return true;
+#ifdef CODSPEED_INSTRUMENTATION
+    if (parent_->codspeed_ != NULL) {
+      parent_->codspeed_->end_benchmark();
+    }
+#endif
     parent_->FinishKeepRunning();
     return false;
   }
@@ -1065,15 +1079,15 @@ struct State::StateIterator {
 };
 
 inline BENCHMARK_ALWAYS_INLINE State::StateIterator State::begin() {
-#ifdef CODSPEED_INSTRUMENTATION
-  codspeed::CodSpeed::getInstance()->start_benchmark(name());
-#endif
   return StateIterator(this);
 }
 inline BENCHMARK_ALWAYS_INLINE State::StateIterator State::end() {
   StartKeepRunning();
 #ifdef CODSPEED_INSTRUMENTATION
-  codspeed::CodSpeed::getInstance()->end_benchmark();
+  if (this->codspeed_ != NULL) {
+    this->codspeed_->start_benchmark(name_);
+    measurement_start();
+  }
 #endif
   return StateIterator();
 }

core/src/codspeed.cpp

@@ -174,10 +174,18 @@ State::State(std::string name, IterationCount max_iters,
              const std::vector<int64_t>& ranges, int thread_i, int n_threads,
              internal::ThreadTimer* timer, internal::ThreadManager* manager,
              internal::PerfCountersMeasurement* perf_counters_measurement,
-             ProfilerManager* profiler_manager)
+             ProfilerManager* profiler_manager
+#ifdef CODSPEED_INSTRUMENTATION
+             ,
+             codspeed::CodSpeed* codspeed
+#endif
+             )
     : total_iterations_(0),
       batch_leftover_(0),
       max_iterations(max_iters),
+#ifdef CODSPEED_INSTRUMENTATION
+      codspeed_(codspeed),
+#endif
       started_(false),
       finished_(false),
       skipped_(internal::NotSkipped),

google_benchmark/include/benchmark/benchmark.h

@@ -102,11 +102,13 @@ State BenchmarkInstance::RunInstrumented(
 
   internal::ThreadTimer warmup_timer = internal::ThreadTimer::Create();
   State warmup_state(name_.function_name, 1, args_, 0, 1, &warmup_timer,
-                     manager, perf_counters_measurement, profiler_manager);
+                     manager, perf_counters_measurement, profiler_manager,
+                     NULL);
   benchmark_.Run(warmup_state);
 
   State st(name().str(), 1, args_, 0, 1, timer, manager,
-           perf_counters_measurement, profiler_manager);
+           perf_counters_measurement, profiler_manager, codspeed);
+  ;
   benchmark_.Run(st);
   return st;
 }