[libcxx-commits] [libcxx] [libc++] Optimize make_heap() and sift_down() (PR #121480)

[Yang Kun via libcxx-commits]

omikrun wrote:

> Please provide the benchmark you ran. You can see examples in `libcxx/test/benchmarks`.

Oh, these results are the time taken to call `std::make_heap` on array of 10 million random elements. I applied [these changes](https://github.com/llvm/llvm-project/pull/121480.diff) manually to the headers in my toolchain (MSYS2).

The source code of benchmarks just like this.

```cpp
#include <algorithm>
#include <chrono>
#include <numeric>
#include <print>
#include <random>
#include <vector>

void make_heap_before(std::vector<int>::iterator first, std::vector<int>::iterator last);
void make_heap_after(std::vector<int>::iterator first, std::vector<int>::iterator last);

class timer {
    std::string_view m_msg;
    std::chrono::high_resolution_clock::time_point m_start = std::chrono::high_resolution_clock::now();

public:
    explicit timer(std::string_view msg) :
        m_msg(msg) {}
    ~timer() {
        std::println("{}: {}", m_msg, std::chrono::high_resolution_clock::now() - m_start);
    }
};

int main() {
    std::vector<int> vec1(10000000), vec2;
    std::iota(vec1.begin(), vec1.end(), 0);
    std::shuffle(vec1.begin(), vec1.end(), std::random_device{});
    vec2 = vec1;
    {
        timer t{"Before"};
        make_heap_before(vec1.begin(), vec1.end());
    }
    {
        timer t{"After"};
        make_heap_after(vec2.begin(), vec2.end());
    }
    return 0;
}
```

```cpp
// make_heap_before.cc
#include <algorithm>
#include <vector>

void make_heap_before(std::vector<int>::iterator first, std::vector<int>::iterator last) {
    std::make_heap(first, last);
}
```

```cpp
// make_heap_after.cc
#include <algorithm>
#include <vector>

void make_heap_before(std::vector<int>::iterator first, std::vector<int>::iterator last) {
    std::make_heap(first, last);
}
```

`make_heap_before.cc` was being compiled before I applied these changes and `make_heap_after.cc` was being compiled after I applied these changes.

I've also ran the benchmark program in libcxx/test/benchmarks/algorithms/make_heap.bench.cpp, here are partial results taken from the output.

```text
Running make_heap.bench.before.exe
Run on (16 X 3194 MHz CPU s)
CPU Caches:
  L1 Data 32 KiB (x8)
  L1 Instruction 32 KiB (x8)
  L2 Unified 512 KiB (x8)
  L3 Unified 16384 KiB (x1)
--------------------------------------------------------------------------------------------------------------
Benchmark                                                                    Time             CPU   Iterations
--------------------------------------------------------------------------------------------------------------
BM_MakeHeap_uint32_Random_1                                               1.02 ns        0.958 ns    783024128
BM_MakeHeap_uint32_Random_4                                               3.57 ns         3.66 ns    234881024
BM_MakeHeap_uint32_Random_16                                              3.85 ns         3.67 ns    195821568
BM_MakeHeap_uint32_Random_64                                              3.91 ns         3.91 ns    167772160
BM_MakeHeap_uint32_Random_256                                             3.95 ns         3.75 ns    195821568
BM_MakeHeap_uint32_Random_1024                                            3.95 ns         3.91 ns    195821568
BM_MakeHeap_uint32_Random_16384                                           4.07 ns         4.07 ns    195821568
BM_MakeHeap_uint32_Random_262144                                          4.52 ns         4.47 ns    146800640
```

```text
Running make_heap.bench.after.exe
Run on (16 X 3194 MHz CPU s)
CPU Caches:
  L1 Data 32 KiB (x8)
  L1 Instruction 32 KiB (x8)
  L2 Unified 512 KiB (x8)
  L3 Unified 16384 KiB (x1)
--------------------------------------------------------------------------------------------------------------
Benchmark                                                                    Time             CPU   Iterations
--------------------------------------------------------------------------------------------------------------
BM_MakeHeap_uint32_Random_1                                               1.00 ns         1.16 ns    618135552
BM_MakeHeap_uint32_Random_4                                               3.36 ns         3.45 ns    167772160
BM_MakeHeap_uint32_Random_16                                              3.12 ns         3.06 ns    234881024
BM_MakeHeap_uint32_Random_64                                              2.80 ns         2.93 ns    293601280
BM_MakeHeap_uint32_Random_256                                             2.56 ns         2.55 ns    293601280
BM_MakeHeap_uint32_Random_1024                                            2.50 ns         2.55 ns    293601280
BM_MakeHeap_uint32_Random_16384                                           2.57 ns         2.50 ns    262144000
BM_MakeHeap_uint32_Random_262144                                          3.18 ns         3.19 ns    293601280
```

I think it is better than the original ones.

https://github.com/llvm/llvm-project/pull/121480