feat: 新增采样器按指令周期记录样本并输出热点直方图

Cargo.lock

@@ -1095,7 +1095,7 @@ dependencies = [
 
 [[package]]
 name = "oxide_builtins"
-version = "0.6.5"
+version = "0.6.6"
 dependencies = [
  "chrono",
  "fast-float",
@@ -1120,7 +1120,7 @@ dependencies = [
 
 [[package]]
 name = "oxide_bytecode"
-version = "0.6.5"
+version = "0.6.6"
 dependencies = [
  "num-bigint",
  "oxide_log",
@@ -1128,7 +1128,7 @@ dependencies = [
 
 [[package]]
 name = "oxide_cfg"
-version = "0.6.5"
+version = "0.6.6"
 dependencies = [
  "oxide_bytecode",
  "oxide_emit",
@@ -1139,7 +1139,7 @@ dependencies = [
 
 [[package]]
 name = "oxide_cli"
-version = "0.6.5"
+version = "0.6.6"
 dependencies = [
  "ansi_term",
  "bumpalo",
@@ -1159,7 +1159,7 @@ dependencies = [
 
 [[package]]
 name = "oxide_code_cache"
-version = "0.6.5"
+version = "0.6.6"
 dependencies = [
  "lru",
  "oxide_bytecode",
@@ -1168,7 +1168,7 @@ dependencies = [
 
 [[package]]
 name = "oxide_compiler"
-version = "0.6.5"
+version = "0.6.6"
 dependencies = [
  "oxide_bytecode",
  "oxide_cfg",
@@ -1187,7 +1187,7 @@ dependencies = [
 
 [[package]]
 name = "oxide_dce"
-version = "0.6.5"
+version = "0.6.6"
 dependencies = [
  "oxide_bytecode",
  "oxide_cfg",
@@ -1201,7 +1201,7 @@ dependencies = [
 
 [[package]]
 name = "oxide_emit"
-version = "0.6.5"
+version = "0.6.6"
 dependencies = [
  "num-bigint",
  "oxc_allocator",
@@ -1216,7 +1216,7 @@ dependencies = [
 
 [[package]]
 name = "oxide_ir"
-version = "0.6.5"
+version = "0.6.6"
 dependencies = [
  "oxide_bytecode",
  "oxide_log",
@@ -1225,7 +1225,7 @@ dependencies = [
 
 [[package]]
 name = "oxide_kernel"
-version = "0.6.5"
+version = "0.6.6"
 dependencies = [
  "bumpalo",
  "dashmap",
@@ -1240,7 +1240,7 @@ dependencies = [
 
 [[package]]
 name = "oxide_liveness"
-version = "0.6.5"
+version = "0.6.6"
 dependencies = [
  "oxide_bytecode",
  "oxide_cfg",
@@ -1252,7 +1252,7 @@ dependencies = [
 
 [[package]]
 name = "oxide_log"
-version = "0.6.5"
+version = "0.6.6"
 dependencies = [
  "tracing",
  "tracing-appender",
@@ -1261,7 +1261,7 @@ dependencies = [
 
 [[package]]
 name = "oxide_parser"
-version = "0.6.5"
+version = "0.6.6"
 dependencies = [
  "oxc_allocator",
  "oxc_ast",
@@ -1276,7 +1276,7 @@ dependencies = [
 
 [[package]]
 name = "oxide_regalloc"
-version = "0.6.5"
+version = "0.6.6"
 dependencies = [
  "oxide_bytecode",
  "oxide_cfg",
@@ -1292,7 +1292,7 @@ dependencies = [
 
 [[package]]
 name = "oxide_runtime_api"
-version = "0.6.5"
+version = "0.6.6"
 dependencies = [
  "num-bigint",
  "num-traits",
@@ -1304,7 +1304,7 @@ dependencies = [
 
 [[package]]
 name = "oxide_test262"
-version = "0.6.5"
+version = "0.6.6"
 dependencies = [
  "oxide_bytecode",
  "oxide_compiler",
@@ -1321,7 +1321,7 @@ dependencies = [
 
 [[package]]
 name = "oxide_types"
-version = "0.6.5"
+version = "0.6.6"
 dependencies = [
  "bumpalo",
  "hashbrown 0.14.5",
@@ -1333,7 +1333,7 @@ dependencies = [
 
 [[package]]
 name = "oxide_vm"
-version = "0.6.5"
+version = "0.6.6"
 dependencies = [
  "bumpalo",
  "chrono",

Cargo.toml

@@ -22,7 +22,7 @@ members = [
 ]
 
 [workspace.package]
-version = "0.6.5"
+version = "0.6.6"
 edition = "2021"
 license = "MIT"
 rust-version = "1.80"

crates/oxide_cli/src/bench/js_stress.rs

@@ -86,6 +86,11 @@ pub fn run_js_stress_bench(config: &BenchConfig, kernel: &Arc<KernelCore>, pool:
             let mut guard = pool.spawn();
             let vm = guard.vm_mut();
 
+            // 指令周期采样只对测量迭代开启：预热不采样，直方图不混入
+            // 预热样本，stderr 输出也不随预热迭代重复。
+            vm.set_sample_period(config.sample_period);
+            vm.set_sample_top_k(config.sample_top);
+
             let pre_session = vm.session_object_count();
             let pre_epoch = vm.epoch_object_count();
 

crates/oxide_cli/src/bench/mod.rs

@@ -32,6 +32,11 @@ pub struct BenchConfig {
     pub iterations: u32,
     pub update_baseline: bool,
     pub leak_check_interval: usize,
+    /// 指令周期采样周期（2 的幂，0 关闭）：开启时测量迭代每 period 条指令
+    /// 记一条样本，run 末按 flat_id 输出 top-K 直方图。
+    pub sample_period: u64,
+    /// 采样直方图 top-K 大小（默认 10）。
+    pub sample_top: usize,
 }
 
 /// 按 mode 分派到具体 benchmark 实现（js/rust/leak）。

crates/oxide_cli/src/main.rs

@@ -87,6 +87,13 @@ enum Commands {
         update_baseline: bool,
         #[arg(long, default_value = "1000")]
         leak_check_interval: usize,
+        /// 指令周期采样周期（2 的幂，0 关闭，默认关闭）。开启时测量迭代每
+        /// period 条指令记一条样本，run 末按 flat_id 输出 top-K 直方图。
+        #[arg(long, default_value = "0")]
+        sample: u64,
+        /// 采样直方图 top-K 大小（默认 10）。
+        #[arg(long, default_value = "10")]
+        sample_top: usize,
     },
     Test {
         suite: Option<String>,
@@ -130,9 +137,17 @@ fn main() -> ExitCode {
             iterations,
             update_baseline,
             leak_check_interval,
+            sample,
+            sample_top,
         }) => {
             let kernel = make_kernel(false, false);
             let pool = make_pool(&kernel);
+            // 采样周期边界检查：须为 0（关闭）或 2 的幂，否则指令边界的
+            // `steps & (period - 1) == 0` 判定退化（周期不整除步数序列）。
+            if sample != 0 && sample & (sample - 1) != 0 {
+                eprintln!("--sample period must be 0 (disabled) or a power of two, got {sample}");
+                return ExitCode::FAILURE;
+            }
             let config = bench::BenchConfig {
                 mode: mode.unwrap_or_else(|| "js".to_string()),
                 filter,
@@ -140,6 +155,8 @@ fn main() -> ExitCode {
                 iterations,
                 update_baseline,
                 leak_check_interval,
+                sample_period: sample,
+                sample_top,
             };
             bench::run_benchmarks(config, kernel, pool)
         }

crates/oxide_vm/src/vm/dispatch_loop.rs

@@ -234,6 +234,15 @@ impl Vm {
                 }
             }
 
+            // 指令周期采样：开启时（周期为 2 的幂）每 period 条指令记一条样本
+            // (flat_id, pc, opcode, frames 深度)，run 末聚合成 top-K 直方图；
+            // 关闭时仅这一条可预测分支，零写零分配。
+            if self.sampling.period != 0 && (steps & (self.sampling.period - 1)) == 0 {
+                self.sampling
+                    .records
+                    .push((self.active_flat_id, pc as u32, op as u8, self.frames.len() as u32));
+            }
+
             match op {
                 OpCode::NOP => {}
 

crates/oxide_vm/src/vm/mod.rs

@@ -18,7 +18,7 @@ use smallvec::SmallVec;
 pub use crate::bindings::init_kernel_builtins;
 use crate::native::NativeFn;
 use crate::vm_debug;
-use crate::vm_state::{GcState, IterState, ProfilingState, SymbolState};
+use crate::vm_state::{GcState, IterState, ProfilingState, SampleState, SymbolState};
 use oxide_kernel::kernel::{KernelCore, KernelSession};
 use oxide_types::error::JsErrorKind;
 use oxide_types::mem::P;
@@ -354,6 +354,8 @@ pub struct Vm {
     pub(crate) iters: IterState,
     /// 分组保存 inline cache 与指令计数器。
     pub(crate) profiling: ProfilingState,
+    /// 分组保存指令周期采样状态（周期、top-K、样本记录）。
+    pub(crate) sampling: SampleState,
     pub(crate) cell_stack: Vec<Vec<*mut Cell>>,
     /// 标签模板对象缓存（GetTemplateObject）：键 = (表代际, 模块 flat_id, site 序号)。
     /// 同一代际同 flat_id 同 site 恒返回同一对象；每次 `run()` 清空（缓存只留
@@ -455,6 +457,19 @@ impl Vm {
         self.pc_watch = path;
     }
 
+    /// 设置指令周期采样周期（2 的幂，0 关闭）。开启后 dispatch 主循环每
+    /// `period` 条指令记一条样本（flat_id、pc、opcode、frames 深度），run 末
+    /// 按 flat_id 聚合并向 stderr 输出 top-K 直方图；关闭时热路径仅一次
+    /// 可预测分支，零写零分配。
+    pub fn set_sample_period(&mut self, period: u64) {
+        self.sampling.period = period;
+    }
+
+    /// 设置采样直方图的 top-K 大小（默认 10）。
+    pub fn set_sample_top_k(&mut self, k: usize) {
+        self.sampling.top_k = k;
+    }
+
     /// 若 `val` 是字符串，返回其内容的 `String` 副本；否则返回 `None`。
     pub fn lookup_str(&self, val: JsValue) -> Option<String> {
         if !val.is_string() {

crates/oxide_vm/src/vm_runtime.rs

@@ -1,6 +1,8 @@
+use std::collections::HashMap;
 use std::sync::{Arc, OnceLock};
 
 use oxide_bytecode::module::CompiledModule;
+use oxide_bytecode::opcode::OpCode;
 
 use crate::vm::{CallFrame, FrameArgs, FrameContinuation, InlineSyncState, TableGen, Vm};
 use crate::{vm_debug, vm_info, vm_trace, vm_warn};
@@ -702,9 +704,81 @@ impl Vm {
         let result = self.dispatch();
         // 顶层执行结束后 drain 微任务队列：Promise reactions 与 thenable 委托在此执行。
         self.drain_job_queue();
+        // 指令周期采样：run 末聚合本 run 的样本并输出 top-K 直方图（关闭时零开销
+        // 短路）。放 dispatch 返回后而非循环内：嵌套 dispatch 会多次返回，直方图
+        // 只应输出一次。
+        self.print_sample_histogram();
         result
     }
 
+    /// 按 flat_id 聚合采样记录并向 stderr 输出 top-K 热点直方图。
+    ///
+    /// 聚合键取 (flat_id, pc)：样本数按 flat_id 汇总，top pc 取该 flat_id 下
+    /// 样本数最多的字节码偏移（与反汇编 offset 列同单位）。函数名经当前表代际
+    /// 平表解析（顶层为 0），未知 flat_id 输出 `<flat N>`。
+    ///
+    /// # 边界
+    /// - 采样关闭（`period == 0`）或无样本时直接返回，零输出零分配。
+    /// - 只读采样记录与平表，不改写任何执行状态。
+    pub(crate) fn print_sample_histogram(&self) {
+        if self.sampling.period == 0 || self.sampling.records.is_empty() {
+            return;
+        }
+        // 聚合：(flat_id, pc) → 样本数与 opcode；flat_id → 最大 frames 深度。
+        let mut pc_hist: HashMap<(u32, u32), (u64, u8)> = HashMap::new();
+        let mut flat_max_frames: HashMap<u32, u32> = HashMap::new();
+        for &(flat, pc, op, frames) in &self.sampling.records {
+            let entry = pc_hist.entry((flat, pc)).or_insert((0, op));
+            entry.0 += 1;
+            let top = flat_max_frames.entry(flat).or_insert(frames);
+            if frames > *top {
+                *top = frames;
+            }
+        }
+        // 按 flat_id 汇总样本数，降序排列后取 top-K。
+        let mut flat_total: HashMap<u32, u64> = HashMap::new();
+        for (&(flat, _), &(count, _)) in &pc_hist {
+            *flat_total.entry(flat).or_insert(0) += count;
+        }
+        let mut ranked: Vec<(u32, u64)> = flat_total.into_iter().collect();
+        ranked.sort_by(|a, b| b.1.cmp(&a.1).then(a.0.cmp(&b.0)));
+        let k = self.sampling.top_k.min(ranked.len());
+        eprintln!(
+            "sample top {k} (period {}, {} samples, {} instructions):",
+            self.sampling.period,
+            self.sampling.records.len(),
+            self.profiling.instruction_count
+        );
+        for (rank, &(flat, count)) in ranked.iter().take(k).enumerate() {
+            // top pc：该 flat_id 下样本数最多的字节码偏移（并列取小 pc）。
+            let mut top_pc: Option<(u64, u32, u8)> = None;
+            for (&(f, pc), &(c, op)) in &pc_hist {
+                if f != flat {
+                    continue;
+                }
+                match top_pc {
+                    None => top_pc = Some((c, pc, op)),
+                    Some((tc, _, _)) if c > tc => top_pc = Some((c, pc, op)),
+                    Some(_) => {}
+                }
+            }
+            let (top_count, top_pc, top_op) = top_pc.unwrap_or((0, 0, 0));
+            let op_name = OpCode::try_from(top_op)
+                .map(|o| o.to_string())
+                .unwrap_or_else(|_| format!("op{top_op}"));
+            let name = self
+                .tables
+                .get(&self.current_gen)
+                .and_then(|t| t.modules.get(flat as usize))
+                .and_then(|m| m.function_name.clone())
+                .unwrap_or_else(|| format!("<flat {flat}>"));
+            let frames = flat_max_frames.get(&flat).copied().unwrap_or(0);
+            eprintln!(
+                "  {rank:>2}. {name:<24} (flat {flat}) {count:>8} samples  top pc={top_pc} {op_name} ({top_count})  frames={frames}"
+            );
+        }
+    }
+
     /// 异常展开：沿 `try_stack` 查找最近可接管的 catch/finally，无则逐帧回退并返回未捕获错误。
     ///
     /// 与 `TRY_FINALLY_END` 的完成穿越逻辑对应：finally 体在途异常经 `pending_exception`

crates/oxide_vm/src/vm_state.rs

@@ -235,6 +235,28 @@ impl IterState {
     }
 }
 
+/// 指令周期采样状态：采样周期（2 的幂）与样本记录表。
+///
+/// 开启时 dispatch 主循环每 `period` 条指令记一条样本（flat_id、pc、opcode、
+/// frames 深度），run 末按 flat_id 聚合并输出 top-K 直方图（stderr）。
+/// 关闭（`period == 0`）时热路径仅一次可预测分支，零写零分配。
+pub(crate) struct SampleState {
+    /// 采样周期（2 的幂，0 = 关闭）。
+    pub(crate) period: u64,
+    /// 直方图 top-K 大小（默认 10）。
+    pub(crate) top_k: usize,
+    /// 样本记录：(flat_id, pc, opcode, frames 深度)，run 边界清空。
+    pub(crate) records: Vec<(u32, u32, u8, u32)>,
+}
+
+impl SampleState {
+    /// 清空样本记录（run 边界）。只清记录：period 与 top_k 是调用方
+    /// 设置的配置，不随 run 边界重置。
+    pub(crate) fn clear_records(&mut self) {
+        self.records.clear();
+    }
+}
+
 /// inline cache 命中/未命中计数与指令计数。
 pub(crate) struct ProfilingState {
     pub(crate) ic_hits: Cell<u64>,

crates/oxide_vm/src/vm_support.rs

@@ -8,7 +8,7 @@ use oxide_bytecode::module::Constant;
 use crate::bindings;
 use crate::vm::{TableGen, Vm};
 use crate::vm_info;
-use crate::vm_state::{GcState, IterState, ProfilingState, SymbolState};
+use crate::vm_state::{GcState, IterState, ProfilingState, SampleState, SymbolState};
 use oxide_kernel::kernel::{KernelConfig, KernelCore, KernelSession};
 use oxide_kernel::shape_forge::EMPTY_SHAPE_ID;
 use oxide_types::mem::P;
@@ -149,6 +149,11 @@ impl Vm {
                 ic_misses: std::cell::Cell::new(0),
                 instruction_count: 0,
             },
+            sampling: SampleState {
+                period: 0,
+                top_k: 10,
+                records: Vec::new(),
+            },
             cell_stack: Vec::new(),
             template_objects: std::collections::HashMap::new(),
             active_flat_id: 0,
@@ -283,6 +288,11 @@ impl Vm {
                 ic_misses: std::cell::Cell::new(0),
                 instruction_count: 0,
             },
+            sampling: SampleState {
+                period: 0,
+                top_k: 10,
+                records: Vec::new(),
+            },
             cell_stack: Vec::new(),
             template_objects: std::collections::HashMap::new(),
             active_flat_id: 0,
@@ -522,6 +532,9 @@ impl Vm {
         // per-run 指标（ic_hit_rate）跨文件累积，报告失真。
         self.profiling.ic_hits.set(0);
         self.profiling.ic_misses.set(0);
+        // 采样记录同属执行期状态：池化 Vm 跨 run 复用，不清空则上一 run 的
+        // 样本混入本 run 的直方图，热点分布失真。
+        self.sampling.clear_records();
         // GC 统计同属执行期状态：池化 Vm 跨 run 复用，不清零则 per-run
         // 指标（gc_trigger_count）跨文件累积，报告失真。
         let gc = &mut self.gc_state.session_gc;