//! 基准测试模块。
use super::{
    event::CompletedTest,
    options::BenchMode,
    test_result::TestResult,
    types::{TestDesc, TestId},
    Sender,
};

use crate::stats;
use std::cmp;
use std::io;
use std::panic::{catch_unwind, AssertUnwindSafe};
use std::sync::{Arc, Mutex};
use std::time::{Duration, Instant};

/// 一个标识函数，*__hints__* 编译器对 `black_box` 能做的事情保持最大限度的悲观。
///
///
/// 有关详细信息，请参见 [`std::hint::black_box`]。
#[inline(always)]
pub fn black_box<T>(dummy: T) -> T {
    std::hint::black_box(dummy)
}

/// 基准测试管理。
///
/// 这被馈入标有 `#[bench]` 的函数中，以便在通过调用 `iter` 重复运行一段代码之前进行设置和拆除。
///
///
#[derive(Clone)]
pub struct Bencher {
    mode: BenchMode,
    summary: Option<stats::Summary>,
    pub bytes: u64,
}

impl Bencher {
    /// 基准函数在其体内运行的回调。
    pub fn iter<T, F>(&mut self, mut inner: F)
    where
        F: FnMut() -> T,
    {
        if self.mode == BenchMode::Single {
            ns_iter_inner(&mut inner, 1);
            return;
        }

        self.summary = Some(iter(&mut inner));
    }

    pub fn bench<F>(&mut self, mut f: F) -> Result<Option<stats::Summary>, String>
    where
        F: FnMut(&mut Bencher) -> Result<(), String>,
    {
        let result = f(self);
        result.map(|_| self.summary)
    }
}

#[derive(Debug, Clone, PartialEq)]
pub struct BenchSamples {
    pub ns_iter_summ: stats::Summary,
    pub mb_s: usize,
}

pub fn fmt_bench_samples(bs: &BenchSamples) -> String {
    use std::fmt::Write;
    let mut output = String::new();

    let median = bs.ns_iter_summ.median as usize;
    let deviation = (bs.ns_iter_summ.max - bs.ns_iter_summ.min) as usize;

    write!(
        output,
        "{:>11} ns/iter (+/- {})",
        fmt_thousands_sep(median, ','),
        fmt_thousands_sep(deviation, ',')
    )
    .unwrap();
    if bs.mb_s != 0 {
        write!(output, " = {} MB/s", bs.mb_s).unwrap();
    }
    output
}

// 用数千个分隔符设置数字格式
fn fmt_thousands_sep(mut n: usize, sep: char) -> String {
    use std::fmt::Write;
    let mut output = String::new();
    let mut trailing = false;
    for &pow in &[9, 6, 3, 0] {
        let base = 10_usize.pow(pow);
        if pow == 0 || trailing || n / base != 0 {
            if !trailing {
                write!(output, "{}", n / base).unwrap();
            } else {
                write!(output, "{:03}", n / base).unwrap();
            }
            if pow != 0 {
                output.push(sep);
            }
            trailing = true;
        }
        n %= base;
    }

    output
}

fn ns_iter_inner<T, F>(inner: &mut F, k: u64) -> u64
where
    F: FnMut() -> T,
{
    let start = Instant::now();
    for _ in 0..k {
        black_box(inner());
    }
    start.elapsed().as_nanos() as u64
}

pub fn iter<T, F>(inner: &mut F) -> stats::Summary
where
    F: FnMut() -> T,
{
    // 最初的替补席上跑来获得大致数字。
    let ns_single = ns_iter_inner(inner, 1);

    // 如果第一次运行的时间小于 1ns，请尝试估计 1ms 的迭代次数，然后减少到 1m 迭代次数。
    //
    let ns_target_total = 1_000_000; // 1ms
    let mut n = ns_target_total / cmp::max(1, ns_single);

    // 如果第一次运行花费的时间超过 1 毫秒，我们不想在每个循环中仅进行 0 次迭代。
    // 下面的统计分析自动处理了不能做那么多运行的不幸的副作用 (即较大的错误条)。
    //
    //
    //
    n = cmp::max(1, n);

    let mut total_run = Duration::new(0, 0);
    let samples: &mut [f64] = &mut [0.0_f64; 50];
    loop {
        let loop_start = Instant::now();

        for p in &mut *samples {
            *p = ns_iter_inner(inner, n) as f64 / n as f64;
        }

        stats::winsorize(samples, 5.0);
        let summ = stats::Summary::new(samples);

        for p in &mut *samples {
            let ns = ns_iter_inner(inner, 5 * n);
            *p = ns as f64 / (5 * n) as f64;
        }

        stats::winsorize(samples, 5.0);
        let summ5 = stats::Summary::new(samples);

        let loop_run = loop_start.elapsed();

        // 如果我们已经运行了 100 毫秒，并且似乎已经收集到一个稳定的中间值。
        //
        if loop_run > Duration::from_millis(100)
            && summ.median_abs_dev_pct < 1.0
            && summ.median - summ5.median < summ5.median_abs_dev
        {
            return summ5;
        }

        total_run += loop_run;
        // 我们运行的最长时间是 3 秒。
        if total_run > Duration::from_secs(3) {
            return summ5;
        }

        // 如果我们在这里溢出，请返回到目前为止的结果。
        // 我们将乘数检查为 10，因为我们将乘以 2，并且循环的下一次迭代也将乘以 5 (以计算 summ5 结果)
        //
        //
        n = match n.checked_mul(10) {
            Some(_) => n * 2,
            None => {
                return summ5;
            }
        };
    }
}

pub fn benchmark<F>(
    id: TestId,
    desc: TestDesc,
    monitor_ch: Sender<CompletedTest>,
    nocapture: bool,
    f: F,
) where
    F: FnMut(&mut Bencher) -> Result<(), String>,
{
    let mut bs = Bencher { mode: BenchMode::Auto, summary: None, bytes: 0 };

    let data = Arc::new(Mutex::new(Vec::new()));

    if !nocapture {
        io::set_output_capture(Some(data.clone()));
    }

    let result = catch_unwind(AssertUnwindSafe(|| bs.bench(f)));

    io::set_output_capture(None);

    let test_result = match result {
        // bs.bench(f) {
        Ok(Ok(Some(ns_iter_summ))) => {
            let ns_iter = cmp::max(ns_iter_summ.median as u64, 1);
            let mb_s = bs.bytes * 1000 / ns_iter;

            let bs = BenchSamples { ns_iter_summ, mb_s: mb_s as usize };
            TestResult::TrBench(bs)
        }
        Ok(Ok(None)) => {
            // 它没有被调用，所以没有数据。
            // FIXME: 在这种情况下的错误？
            let samples: &mut [f64] = &mut [0.0_f64; 1];
            let bs = BenchSamples { ns_iter_summ: stats::Summary::new(samples), mb_s: 0 };
            TestResult::TrBench(bs)
        }
        Err(_) => TestResult::TrFailed,
        Ok(Err(_)) => TestResult::TrFailed,
    };

    let stdout = data.lock().unwrap().to_vec();
    let message = CompletedTest::new(id, desc, test_result, None, stdout);
    monitor_ch.send(message).unwrap();
}

pub fn run_once<F>(f: F) -> Result<(), String>
where
    F: FnMut(&mut Bencher) -> Result<(), String>,
{
    let mut bs = Bencher { mode: BenchMode::Single, summary: None, bytes: 0 };
    bs.bench(f).map(|_| ())
}