[libcxx-commits] [libcxx] [libc++] Optimize search_n (PR #171389)

[Louis Dionne via libcxx-commits]

================
@@ -81,6 +81,54 @@ int main(int argc, char** argv) {
     bm.operator()<std::list<int>>("rng::search_n(list<int>, pred) (no match)", ranges_search_n_pred);
   }
 
+  // Benchmark {std,ranges}::search_n where the needle almost matches a lot.
+  {
+    auto bm = []<class Container>(std::string name, auto search_n) {
+      benchmark::RegisterBenchmark(
+          name,
+          [search_n](auto& st) {
+            std::size_t const size = st.range(0);
+            using ValueType        = typename Container::value_type;
+            ValueType x            = Generate<ValueType>::random();
+            ValueType y            = random_different_from({x});
+            Container haystack(size, x);
+            std::size_t n = size / 10; // needle size is 10% of the haystack
+
+            // Make sure there are no actual matches
+            for (size_t i = 0; i < size; i += getRandomInteger<size_t>(1, n - 1)) {
+              *std::next(haystack.begin(), i) = y;
+            }
+
+            for ([[maybe_unused]] auto _ : st) {
+              benchmark::DoNotOptimize(haystack);
+              benchmark::DoNotOptimize(n);
+              benchmark::DoNotOptimize(y);
+              auto result = search_n(haystack.begin(), haystack.end(), n, x);
+              benchmark::DoNotOptimize(result);
+            }
+          })
+          ->Arg(1000) // non power-of-two
+          ->Arg(1024)
+          ->Arg(8192)
+          ->Arg(1 << 20);
----------------
ldionne wrote:

Let's pick some sizes that make sense to reduce the cost of running these benchmarks (and apply the change everywhere in this file).

```
->Arg(32)
->Arg(1024)
->Arg(8192)
```


https://github.com/llvm/llvm-project/pull/171389