bilt | A build tool for NPM monorepos | Build Tool library

Your benchmark is not benchmarking max. It is completely limited by the time the input operations take. Input operations take many times longer than std::max or your mmax implementation.

Furthermore, the mmax or std::max calls will be optimized out by any optimizing compiler, because their results are never used and they don't have any other side effect. See e.g. here on godbolt. So you are likely not benchmarking them at all.

Assuming your claims were even true:

Your function has undefined behavior for some arguments, which std::max doesn't have, e.g. a+b has undefined behavior if it causes an overflow. Comparing the speed is therefore really unfair, since your implementation doesn't even always work.

Here is a quick-bench with a better (though not rigorously verified) benchmark.

• "std_impl" uses std::max
• "naive_impl" uses a simple branch to get the maximum
• "op_impl" is the implementation in the question
• "only_iter" just passes through the first value without any calculation

As you can see in the plot, your implementation is worse than the naive one or std::max, both of which are equal in performance.

The naive and standard library implementations do however, contrary to yours, actually work with all possible input values (I restricted the test case values in the vector to a range that works with your implementation.)

In a previous version of this benchmark I made a mistake (doing benchmarks right is difficult!) that made it seem as if the naive implementation was much worse than std::max and OP's implementation, which turned out to be an artifact of how google benchmark's DoNotOptimize works (at least on Clang, maybe this is a bug in google benchmark or maybe I was using it wrong). If someone finds another flaw, let me know please!

Benchmark code: