Google Benchmark runs live on Compiler Explorer
Google Benchmark is a Compiler Explorer library. Add benchmark from the libraries panel, turn on the execution pane, and BENCHMARK_MAIN() runs on CE’s servers and prints the timing table back to you. No local checkout, no CMake, no linking to work out. That alone is worth knowing, because the friction of setting up a benchmark harness is usually what stops people from measuring at all.
The demo doubles as the first thing a benchmark will teach you the hard way. benchmark::DoNotOptimize is what keeps the optimizer from deleting the code you are measuring.
Both loops call std::sqrt(2.0). The only difference is what happens to the result:
BM_Unfenced 0.000 ns 0.000 ns 1000000000000
BM_Fenced 0.326 ns 0.189 ns 3631323115BM_Unfenced throws its result away, so the optimizer correctly reasons that the whole loop body is dead and removes it. Benchmark then measures an empty loop and reports 0.000 ns over a trillion “iterations”, a number that looks precise and means nothing. BM_Fenced passes the result through DoNotOptimize, which forces the compiler to actually compute and keep it, and now you get an honest sub-nanosecond timing.
What DoNotOptimize actually does
DoNotOptimize(x) is not a function call in the normal sense. It expands to an inline-assembly barrier that tells the compiler this value is observed and may not be discarded, without emitting a single instruction of its own. Its sibling benchmark::ClobberMemory() does the same for stores, marking all of memory as read so a write you are timing cannot be optimized away either.
The rule that falls out of this is simple: every microbenchmark must consume its result. A benchmark that computes a value and drops it is measuring nothing, and the giveaway is exactly what you see above, a suspiciously round time over an astronomically large iteration count.
One caveat the output makes honest: Compiler Explorer’s benchmark library is a debug build (“Library was built as DEBUG. Timings may be affected”), and CE’s shared executors are noisy. Treat the live numbers as a teaching aid, not a measurement of record. The lesson about DoNotOptimize holds identically on a release build on your own hardware.
Sources: Google Benchmark and its user guide (see DoNotOptimize and ClobberMemory).