ardupilot/benchmarks/AP_gbenchmark.h

50 lines
1.6 KiB
C

/*
* Utility header for benchmarks with Google Benchmark.
*/
#include <benchmark/benchmark.h>
/* The two functions below are an approach proposed by Chandler Carruth in
* CPPCON 2015: CppCon 2015: "Tuning C++: Benchmarks, and CPUs, and Compilers!
* Oh My!" in order keep the compiler from optimizing the use of a variable
* (gbenchmark_escape) or whole memory (gbenchmark_clobber).
*
* The compiler optimizer may sometimes remove code when it sees it isn't
* necessary. For example, when a variable isn't used, the optimizer removes
* the code that computes the value for that variable - that's not good for
* benchmarks. The function gbenchmark_escape(void *p) makes the compiler think
* that that p is being used in a code that might have "unknowable side
* effects", which keeps it from removing the variable. The "side effects" in
* the case here would be the benchmark numbers.
*
* Here is an example that would give wrong benchmark values:
*
* static void BM_Test(benchmark::State& state)
* {
* while (state.KeepRunning()) {
* float a = expensive_operation();
* }
* }
*
* Since variable a isn't used, the call to expensive_operation() is removed
* from the compiled program. The benchmark would show that
* expensive_operation() is extremely fast. The following code would fix that:
*
* static void BM_Test(benchmark::State& state)
* {
* while (state.KeepRunning()) {
* float a = expensive_operation();
* gbenchmark_escape(&a);
* }
* }
*/
inline void gbenchmark_escape(void* p)
{
asm volatile("" : : "g"(p) : "memory");
}
inline void gbenchmark_clobber()
{
asm volatile("" : : : "memory");
}