GH-48897: [C++] Add benchmark for CountSetBits

Author: pitrou

cpp/src/arrow/util/bit_util_benchmark.cc

@@ -439,6 +439,17 @@ static void SetBitsTo(benchmark::State& state) {
   state.SetBytesProcessed(state.iterations() * nbytes);
 }
 
+static void CountSetBits(benchmark::State& state) {
+  int64_t nbytes = state.range(0);
+  std::shared_ptr<Buffer> buffer = CreateRandomBuffer(nbytes);
+
+  for (auto _ : state) {
+    auto count = internal::CountSetBits(buffer->data(), /*bit_offset=*/0, nbytes * 8);
+    benchmark::DoNotOptimize(count);
+  }
+  state.SetBytesProcessed(state.iterations() * nbytes);
+}
+
 template <int64_t OffsetSrc, int64_t OffsetDest = 0>
 static void CopyBitmap(benchmark::State& state) {  // NOLINT non-const reference
   const int64_t buffer_size = state.range(0);
@@ -519,6 +530,7 @@ BENCHMARK(ReverseSetBitRunReader)->Apply(SetBitRunReaderPercentageArg);
 BENCHMARK(VisitBits)->Arg(kBufferSize);
 BENCHMARK(VisitBitsUnrolled)->Arg(kBufferSize);
 BENCHMARK(SetBitsTo)->Arg(2)->Arg(1 << 4)->Arg(1 << 10)->Arg(1 << 17);
+BENCHMARK(CountSetBits)->Arg(1 << 4)->Arg(1 << 10)->Arg(1 << 17);
 
 #ifdef ARROW_WITH_BENCHMARKS_REFERENCE
 static void ReferenceNaiveBitmapWriter(benchmark::State& state) {

cpp/src/arrow/util/bitmap_ops.cc

@@ -17,6 +17,7 @@
 
 #include "arrow/util/bitmap_ops.h"
 
+#include <array>
 #include <cstdint>
 #include <cstring>
 #include <functional>
@@ -55,13 +56,15 @@ int64_t CountSetBits(const uint8_t* data, int64_t bit_offset, int64_t length) {
     constexpr int64_t kCountUnrollFactor = 4;
     const int64_t words_rounded =
         bit_util::RoundDown(p.aligned_words, kCountUnrollFactor);
-    int64_t count_unroll[kCountUnrollFactor] = {0};
+    std::array<int64_t, kCountUnrollFactor> count_unroll{};
 
     // Unroll the loop for better performance
     for (int64_t i = 0; i < words_rounded; i += kCountUnrollFactor) {
-      for (int64_t k = 0; k < kCountUnrollFactor; k++) {
-        count_unroll[k] += bit_util::PopCount(u64_data[k]);
-      }
+      // (hand-unrolled as some gcc versions would unnest a nested `for` loop)
+      count_unroll[0] += bit_util::PopCount(u64_data[0]);
+      count_unroll[1] += bit_util::PopCount(u64_data[1]);
+      count_unroll[2] += bit_util::PopCount(u64_data[2]);
+      count_unroll[3] += bit_util::PopCount(u64_data[3]);
       u64_data += kCountUnrollFactor;
     }
     for (int64_t k = 0; k < kCountUnrollFactor; k++) {