feat: 添加基准测试基础设施

crates/oxide_cli/Cargo.toml

@@ -22,3 +22,5 @@ oxide_api = { path = "../oxide_api" }
 clap.workspace = true
 rustyline.workspace = true
 ansi_term.workspace = true
+serde_json.workspace = true
+serde = { workspace = true, features = ["derive"] }

crates/oxide_cli/src/bench/baseline.rs

@@ -0,0 +1,94 @@
+use crate::bench::metrics::MetricCollection;
+
+pub struct Baseline {
+    pub entries: Vec<MetricCollection>,
+}
+
+impl Baseline {
+    pub fn empty() -> Self {
+        Self { entries: Vec::new() }
+    }
+}
+
+pub struct Regression {
+    pub metric: String,
+    pub test_name: String,
+    pub baseline: f64,
+    pub current: f64,
+    pub ratio: f64,
+    pub tolerance: f64,
+}
+
+pub fn load_baseline() -> Result<Baseline, String> {
+    let json_path = "benchmark_baseline.json";
+    let data = match std::fs::read_to_string(json_path) {
+        Ok(d) => d,
+        Err(_) => {
+            eprintln!("No baseline found at {} — run --update-baseline first", json_path);
+            return Ok(Baseline::empty());
+        }
+    };
+    let entries: Vec<MetricCollection> =
+        serde_json::from_str(&data).map_err(|e| format!("Failed to parse baseline: {}", e))?;
+    Ok(Baseline { entries })
+}
+
+pub fn save_baseline(results: &[MetricCollection]) -> Result<(), String> {
+    let json = serde_json::to_string_pretty(results).map_err(|e| format!("Failed to serialize baseline: {}", e))?;
+    std::fs::write("benchmark_baseline.json", &json)
+        .map_err(|e| format!("Failed to write benchmark_baseline.json: {}", e))?;
+    let md = crate::bench::output::format_text_table(results);
+    std::fs::write("BENCHMARK_BASELINE.md", md).map_err(|e| format!("Failed to write BENCHMARK_BASELINE.md: {}", e))?;
+    Ok(())
+}
+
+pub fn compare_baseline(current: &[MetricCollection]) -> Vec<Regression> {
+    let baseline = match load_baseline() {
+        Ok(b) => b,
+        Err(_) => return Vec::new(),
+    };
+    if baseline.entries.is_empty() {
+        eprintln!("No baseline data — skipping regression check");
+        return Vec::new();
+    }
+    let mut regressions = Vec::new();
+    for cur in current {
+        for base in &baseline.entries {
+            if base.test_name != cur.test_name {
+                continue;
+            }
+            for (name, cur_val) in cur.iter_metrics() {
+                if cur_val == 0.0 {
+                    continue;
+                }
+                let base_val = base
+                    .iter_metrics()
+                    .iter()
+                    .find(|(n, _)| *n == name)
+                    .map(|(_, v)| *v)
+                    .unwrap_or(cur_val);
+                if base_val == 0.0 {
+                    continue;
+                }
+                let tolerance = match name {
+                    "wall_time_us" => 0.10,
+                    "gc_trigger_count" | "gc_bytes_freed" | "gc_objects_scanned" | "gc_collection_us" => 0.50,
+                    "session_bytes" => 0.20,
+                    _ => 0.20,
+                };
+                let ratio = cur_val / base_val;
+                if ratio > 1.0 + tolerance {
+                    regressions.push(Regression {
+                        metric: name.to_string(),
+                        test_name: cur.test_name.clone(),
+                        baseline: base_val,
+                        current: cur_val,
+                        ratio,
+                        tolerance,
+                    });
+                }
+            }
+        }
+    }
+    regressions
+}

crates/oxide_cli/src/bench/js_stress.rs

@@ -0,0 +1,208 @@
+use std::fs;
+use std::process::ExitCode;
+use std::sync::Arc;
+use std::time::Instant;
+
+use oxide_compiler::compiler::{compiled_module_hash, Compiler};
+use oxide_kernel::kernel::KernelCore;
+use oxide_parser::Allocator;
+use oxide_vm::vm_pool::VmPool;
+
+use crate::bench::baseline::{compare_baseline, save_baseline};
+use crate::bench::metrics::MetricCollection;
+use crate::bench::output::{format_json, format_text_table};
+use crate::bench::BenchConfig;
+
+pub fn run_js_stress_bench(config: &BenchConfig, kernel: &Arc<KernelCore>, pool: &Arc<VmPool>) -> ExitCode {
+    let mut results: Vec<MetricCollection> = Vec::new();
+    let compiler = Compiler::new();
+
+    let test_dir = "tests/stress";
+    let mut files: Vec<_> = match fs::read_dir(test_dir) {
+        Ok(entries) => entries
+            .filter_map(|e| e.ok())
+            .filter(|e| {
+                e.path().extension().is_some_and(|ext| ext == "js")
+                    && config
+                        .filter
+                        .as_ref()
+                        .map_or(true, |f| e.file_name().to_string_lossy().contains(f.as_str()))
+            })
+            .map(|e| e.path())
+            .collect(),
+        Err(e) => {
+            eprintln!("Could not read {}: {}", test_dir, e);
+            return ExitCode::FAILURE;
+        }
+    };
+    files.sort();
+
+    if files.is_empty() {
+        eprintln!("No matching stress tests found in {}", test_dir);
+        return ExitCode::FAILURE;
+    }
+
+    for path in &files {
+        let test_name = path.file_stem().unwrap_or_default().to_string_lossy().to_string();
+        let mut js = match fs::read_to_string(path) {
+            Ok(s) => s,
+            Err(e) => {
+                eprintln!("Failed to read {:?}: {}", path, e);
+                continue;
+            }
+        };
+
+        if config.iterations > 0 {
+            js = override_iterations(&js, config.iterations);
+        }
+
+        let allocator = Allocator::default();
+        let program = match oxide_parser::parse(&allocator, &js) {
+            Ok(p) => p,
+            Err(errors) => {
+                eprintln!("Parse error in {}: {:?}", test_name, errors.first());
+                continue;
+            }
+        };
+
+        let hash = compiled_module_hash(&program);
+        let module = match kernel.code_forge().get_or_insert_with(hash, || compiler.compile(&program)) {
+            Ok(m) => m,
+            Err(e) => {
+                eprintln!("Compile error in {}: {}", test_name, e);
+                continue;
+            }
+        };
+
+        for _ in 0..config.warmup {
+            let mut guard = pool.spawn();
+            let _ = guard.vm_mut().run(&module);
+        }
+
+        let mut metrics = Vec::new();
+        for _ in 0..config.iterations {
+            let compile_start = Instant::now();
+            let mut guard = pool.spawn();
+            let vm = guard.vm_mut();
+            let compile_time = compile_start.elapsed();
+
+            let pre_session = vm.session_object_count();
+            let pre_epoch = vm.epoch_object_count();
+
+            let exec_start = Instant::now();
+            let _result = vm.run(&module);
+            let exec_time = exec_start.elapsed();
+
+            let gc_stats = vm.session_gc_stats();
+
+            metrics.push(MetricCollection {
+                test_name: test_name.clone(),
+                wall_time_us: (compile_time + exec_time).as_micros() as u64,
+                session_objects: vm.session_object_count().saturating_sub(pre_session),
+                session_bytes: vm.session_bytes_allocated(),
+                epoch_objects: vm.epoch_object_count().saturating_sub(pre_epoch),
+                epoch_bytes: 0,
+                gc_trigger_count: gc_stats.total_collections,
+                gc_bytes_freed: gc_stats.last_collection_bytes_freed,
+                gc_objects_scanned: gc_stats.last_collection_objects_scanned,
+                gc_collection_us: gc_stats.last_collection_duration_us,
+                instruction_count: vm.instruction_count(),
+                compile_time_us: compile_time.as_micros() as u64,
+                exec_time_us: exec_time.as_micros() as u64,
+                ic_hit_rate: vm.ic_hit_rate(),
+                ic_hits: vm.ic_hit_count(),
+                ic_misses: vm.ic_miss_count(),
+            });
+        }
+
+        let avg = average_metrics(&metrics);
+        results.push(avg);
+    }
+
+    let json = format_json(&results);
+    let table = format_text_table(&results);
+    eprintln!("{}", table);
+    println!("{}", json);
+
+    if config.update_baseline {
+        if let Err(e) = save_baseline(&results) {
+            eprintln!("Failed to save baseline: {}", e);
+            return ExitCode::FAILURE;
+        }
+        eprintln!("Baseline saved to BENCHMARK_BASELINE.md + benchmark_baseline.json");
+    } else {
+        let regressions = compare_baseline(&results);
+        if !regressions.is_empty() {
+            eprintln!("\nRegression detected:");
+            for r in &regressions {
+                eprintln!(
+                    "  {}: {} baseline={} current={} ratio={:.2}% tolerance={:.0}%",
+                    r.test_name,
+                    r.metric,
+                    r.baseline,
+                    r.current,
+                    (r.ratio - 1.0) * 100.0,
+                    r.tolerance * 100.0,
+                );
+            }
+            return ExitCode::FAILURE;
+        }
+    }
+
+    ExitCode::SUCCESS
+}
+
+fn average_metrics(metrics: &[MetricCollection]) -> MetricCollection {
+    let n = metrics.len() as f64;
+    if n == 0.0 || metrics.is_empty() {
+        return metrics.first().cloned().unwrap_or(MetricCollection {
+            test_name: String::new(),
+            wall_time_us: 0,
+            session_objects: 0,
+            session_bytes: 0,
+            epoch_objects: 0,
+            epoch_bytes: 0,
+            gc_trigger_count: 0,
+            gc_bytes_freed: 0,
+            gc_objects_scanned: 0,
+            gc_collection_us: 0,
+            instruction_count: 0,
+            compile_time_us: 0,
+            exec_time_us: 0,
+            ic_hit_rate: 0.0,
+            ic_hits: 0,
+            ic_misses: 0,
+        });
+    }
+    MetricCollection {
+        test_name: metrics[0].test_name.clone(),
+        wall_time_us: (metrics.iter().map(|m| m.wall_time_us as f64).sum::<f64>() / n) as u64,
+        session_objects: (metrics.iter().map(|m| m.session_objects as f64).sum::<f64>() / n) as usize,
+        session_bytes: (metrics.iter().map(|m| m.session_bytes as f64).sum::<f64>() / n) as usize,
+        epoch_objects: (metrics.iter().map(|m| m.epoch_objects as f64).sum::<f64>() / n) as usize,
+        epoch_bytes: (metrics.iter().map(|m| m.epoch_bytes as f64).sum::<f64>() / n) as u64,
+        gc_trigger_count: (metrics.iter().map(|m| m.gc_trigger_count as f64).sum::<f64>() / n) as u64,
+        gc_bytes_freed: (metrics.iter().map(|m| m.gc_bytes_freed as f64).sum::<f64>() / n) as u64,
+        gc_objects_scanned: (metrics.iter().map(|m| m.gc_objects_scanned as f64).sum::<f64>() / n) as u64,
+        gc_collection_us: (metrics.iter().map(|m| m.gc_collection_us as f64).sum::<f64>() / n) as u64,
+        instruction_count: (metrics.iter().map(|m| m.instruction_count as f64).sum::<f64>() / n) as u64,
+        compile_time_us: (metrics.iter().map(|m| m.compile_time_us as f64).sum::<f64>() / n) as u64,
+        exec_time_us: (metrics.iter().map(|m| m.exec_time_us as f64).sum::<f64>() / n) as u64,
+        ic_hit_rate: metrics.iter().map(|m| m.ic_hit_rate).sum::<f64>() / n,
+        ic_hits: (metrics.iter().map(|m| m.ic_hits as f64).sum::<f64>() / n) as u64,
+        ic_misses: (metrics.iter().map(|m| m.ic_misses as f64).sum::<f64>() / n) as u64,
+    }
+}
+
+fn override_iterations(js: &str, iters: u32) -> String {
+    js.lines()
+        .map(|line| {
+            if line.trim_start().starts_with("var ITERATIONS ") {
+                format!("var ITERATIONS = {};", iters)
+            } else {
+                line.to_string()
+            }
+        })
+        .collect::<Vec<_>>()
+        .join("\n")
+}

crates/oxide_cli/src/bench/leak_detect.rs

@@ -0,0 +1,141 @@
+use std::collections::VecDeque;
+use std::process::ExitCode;
+use std::sync::Arc;
+use std::time::Instant;
+
+use oxide_compiler::compiler::{compiled_module_hash, Compiler};
+use oxide_kernel::kernel::KernelCore;
+use oxide_parser::Allocator;
+use oxide_vm::vm_pool::VmPool;
+
+use crate::bench::metrics::MetricCollection;
+use crate::bench::output::format_json;
+use crate::bench::BenchConfig;
+
+pub struct LeakSampler {
+    window: VecDeque<(usize, f64)>,
+    window_size: usize,
+}
+
+impl LeakSampler {
+    pub fn new(window_size: usize) -> Self {
+        Self {
+            window: VecDeque::new(),
+            window_size,
+        }
+    }
+
+    pub fn add_sample(&mut self, iteration: usize, value: f64) -> Option<LeakVerdict> {
+        self.window.push_back((iteration, value));
+        if self.window.len() >= self.window_size {
+            self.window.pop_front();
+        }
+        if self.window.len() < 10 {
+            return None;
+        }
+        let n = self.window.len() as f64;
+        let sum_x: f64 = self.window.iter().map(|(i, _)| *i as f64).sum();
+        let sum_y: f64 = self.window.iter().map(|(_, v)| v).sum();
+        let sum_xy: f64 = self.window.iter().map(|(i, v)| *i as f64 * v).sum();
+        let sum_x2: f64 = self.window.iter().map(|(i, _)| (*i as f64).powi(2)).sum();
+        let slope = (n * sum_xy - sum_x * sum_y) / (n * sum_x2 - sum_x * sum_x);
+        let intercept = (sum_y - slope * sum_x) / n;
+        let y_mean = sum_y / n;
+        let ss_res: f64 = self
+            .window
+            .iter()
+            .map(|(i, v)| (v - (slope * *i as f64 + intercept)).powi(2))
+            .sum();
+        let ss_tot: f64 = self.window.iter().map(|(_, v)| (v - y_mean).powi(2)).sum();
+        let r2 = if ss_tot == 0.0 { 1.0 } else { 1.0 - ss_res / ss_tot };
+        if r2 > 0.9 && slope > 0.0 {
+            Some(LeakVerdict { slope, r2 })
+        } else {
+            None
+        }
+    }
+}
+
+pub struct LeakVerdict {
+    pub slope: f64,
+    pub r2: f64,
+}
+
+pub fn run_leak_detect(config: &BenchConfig, kernel: &Arc<KernelCore>, pool: &Arc<VmPool>) -> ExitCode {
+    let js =
+        "var ITERATIONS = 1000; var obj = {}; for (var i = 0; i < ITERATIONS; i++) { obj['key' + i] = i; } obj['key0']";
+    let allocator = Allocator::default();
+    let program = match oxide_parser::parse(&allocator, js) {
+        Ok(p) => p,
+        Err(_) => return ExitCode::FAILURE,
+    };
+    let compiler = Compiler::new();
+    let hash = compiled_module_hash(&program);
+    let module = match kernel.code_forge().get_or_insert_with(hash, || compiler.compile(&program)) {
+        Ok(m) => m,
+        Err(_) => return ExitCode::FAILURE,
+    };
+
+    let mut samplers: Vec<(&str, LeakSampler)> = vec![
+        ("session_objects", LeakSampler::new(20)),
+        ("session_bytes", LeakSampler::new(20)),
+        ("code_forge_entries", LeakSampler::new(20)),
+        ("symbol_registry", LeakSampler::new(20)),
+    ];
+
+    let mut results: Vec<MetricCollection> = Vec::new();
+
+    for i in 0..config.iterations as usize {
+        let start = Instant::now();
+        let mut guard = pool.spawn();
+        let vm = guard.vm_mut();
+        let _ = vm.run(&module);
+        let elapsed = start.elapsed();
+
+        for (name, sampler) in &mut samplers {
+            let value: f64 = match *name {
+                "session_objects" => vm.session_object_count() as f64,
+                "session_bytes" => vm.session_bytes_allocated() as f64,
+                "code_forge_entries" => kernel.code_forge().len() as f64,
+                "symbol_registry" => vm.symbol_registry_len() as f64,
+                _ => continue,
+            };
+            if let Some(verdict) = sampler.add_sample(i, value) {
+                eprintln!(
+                    "[LEAK] {}: slope={:.6} R²={:.4} over {} samples",
+                    name,
+                    verdict.slope,
+                    verdict.r2,
+                    sampler.window.len(),
+                );
+            }
+        }
+
+        if i == config.iterations as usize - 1 {
+            let guard = pool.spawn();
+            let vm = guard.vm();
+            results.push(MetricCollection {
+                test_name: "leak_detect".to_string(),
+                wall_time_us: elapsed.as_micros() as u64,
+                session_objects: vm.session_object_count(),
+                session_bytes: vm.session_bytes_allocated(),
+                epoch_objects: vm.epoch_object_count(),
+                epoch_bytes: 0,
+                gc_trigger_count: vm.session_gc_stats().total_collections,
+                gc_bytes_freed: vm.session_gc_stats().last_collection_bytes_freed,
+                gc_objects_scanned: vm.session_gc_stats().last_collection_objects_scanned,
+                gc_collection_us: vm.session_gc_stats().last_collection_duration_us,
+                instruction_count: vm.instruction_count(),
+                compile_time_us: 0,
+                exec_time_us: 0,
+                ic_hit_rate: vm.ic_hit_rate(),
+                ic_hits: vm.ic_hit_count(),
+                ic_misses: vm.ic_miss_count(),
+            });
+        }
+    }
+
+    let json = format_json(&results);
+    println!("{}", json);
+    ExitCode::SUCCESS
+}

crates/oxide_cli/src/bench/metrics.rs

@@ -0,0 +1,43 @@
+use serde::{Deserialize, Serialize};
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct MetricCollection {
+    pub test_name: String,
+    pub wall_time_us: u64,
+    pub session_objects: usize,
+    pub session_bytes: usize,
+    pub epoch_objects: usize,
+    pub epoch_bytes: u64,
+    pub gc_trigger_count: u64,
+    pub gc_bytes_freed: u64,
+    pub gc_objects_scanned: u64,
+    pub gc_collection_us: u64,
+    pub instruction_count: u64,
+    pub compile_time_us: u64,
+    pub exec_time_us: u64,
+    pub ic_hit_rate: f64,
+    pub ic_hits: u64,
+    pub ic_misses: u64,
+}
+
+impl MetricCollection {
+    pub fn iter_metrics(&self) -> Vec<(&'static str, f64)> {
+        vec![
+            ("wall_time_us", self.wall_time_us as f64),
+            ("session_objects", self.session_objects as f64),
+            ("session_bytes", self.session_bytes as f64),
+            ("epoch_objects", self.epoch_objects as f64),
+            ("epoch_bytes", self.epoch_bytes as f64),
+            ("gc_trigger_count", self.gc_trigger_count as f64),
+            ("gc_bytes_freed", self.gc_bytes_freed as f64),
+            ("gc_objects_scanned", self.gc_objects_scanned as f64),
+            ("gc_collection_us", self.gc_collection_us as f64),
+            ("instruction_count", self.instruction_count as f64),
+            ("compile_time_us", self.compile_time_us as f64),
+            ("exec_time_us", self.exec_time_us as f64),
+            ("ic_hit_rate", self.ic_hit_rate),
+            ("ic_hits", self.ic_hits as f64),
+            ("ic_misses", self.ic_misses as f64),
+        ]
+    }
+}

crates/oxide_cli/src/bench/mod.rs

@@ -0,0 +1,34 @@
+pub mod baseline;
+pub mod js_stress;
+pub mod leak_detect;
+pub mod metrics;
+pub mod output;
+pub mod rust_bench;
+
+use std::process::ExitCode;
+use std::sync::Arc;
+
+use oxide_kernel::kernel::KernelCore;
+use oxide_vm::vm_pool::VmPool;
+
+pub struct BenchConfig {
+    pub mode: String,
+    pub filter: Option<String>,
+    pub warmup: u32,
+    pub iterations: u32,
+    pub process: bool,
+    pub update_baseline: bool,
+    pub leak_check_interval: usize,
+}
+
+pub fn run_benchmarks(config: BenchConfig, kernel: Arc<KernelCore>, pool: Arc<VmPool>) -> ExitCode {
+    match config.mode.as_str() {
+        "js" => js_stress::run_js_stress_bench(&config, &kernel, &pool),
+        "rust" => rust_bench::run_rust_bench(config.filter.as_deref()),
+        "leak" => leak_detect::run_leak_detect(&config, &kernel, &pool),
+        _ => {
+            eprintln!("Unknown bench mode: {}. Use: js, rust, or leak", config.mode);
+            ExitCode::FAILURE
+        }
+    }
+}

crates/oxide_cli/src/bench/output.rs

@@ -0,0 +1,40 @@
+use crate::bench::metrics::MetricCollection;
+
+pub fn format_json(results: &[MetricCollection]) -> String {
+    serde_json::to_string_pretty(results).unwrap_or_else(|e| format!("{{\"error\": \"{}\"}}", e))
+}
+
+pub fn format_text_table(results: &[MetricCollection]) -> String {
+    let mut out = String::new();
+    out.push_str(&format!(
+        "{:<35} {:>10} {:>10} {:>10} {:>12} {:>8} {:>8} {:>10} {:>10}\n",
+        "Test Name", "Wall(ms)", "Comp(ms)", "Exec(ms)", "Instrs", "GC Ct", "IC Rate", "Sess Objs", "Sess B",
+    ));
+    out.push_str(&"-".repeat(120));
+    out.push('\n');
+
+    for m in results {
+        out.push_str(&format!(
+            "{:<35} {:>10.2} {:>10.3} {:>10.3} {:>12} {:>8} {:>7.1}% {:>10} {:>10}",
+            truncate(&m.test_name, 35),
+            m.wall_time_us as f64 / 1000.0,
+            m.compile_time_us as f64 / 1000.0,
+            m.exec_time_us as f64 / 1000.0,
+            m.instruction_count,
+            m.gc_trigger_count,
+            m.ic_hit_rate * 100.0,
+            m.session_objects,
+            m.session_bytes,
+        ));
+        out.push('\n');
+    }
+    out
+}
+
+fn truncate(s: &str, max: usize) -> &str {
+    if s.len() <= max {
+        s
+    } else {
+        &s[..max]
+    }
+}

crates/oxide_cli/src/bench/rust_bench.rs

@@ -0,0 +1,20 @@
+pub fn run_rust_bench(filter: Option<&str>) -> std::process::ExitCode {
+    let mut cmd = std::process::Command::new("cargo");
+    cmd.args(["bench", "-p", "oxide_vm"]);
+    if let Some(name) = filter {
+        cmd.args(["--bench", name]);
+    }
+    match cmd.status() {
+        Ok(status) => {
+            if status.success() {
+                std::process::ExitCode::SUCCESS
+            } else {
+                std::process::ExitCode::from(1)
+            }
+        }
+        Err(e) => {
+            eprintln!("Failed to run cargo bench: {}", e);
+            std::process::ExitCode::from(1)
+        }
+    }
+}

crates/oxide_cli/src/main.rs

@@ -1,4 +1,5 @@
 #![allow(clippy::arc_with_non_send_sync)]
+#![allow(dead_code)]
 
 use std::fs;
 use std::process::ExitCode;
@@ -15,6 +16,8 @@ use oxide_types::object::JsObject;
 use oxide_vm::vm_pool::VmPool;
 use oxide_vm::JsValue;
 
+mod bench;
+
 #[derive(Parser)]
 #[command(name = "oxide", version, about = "OxideJS - Rust JavaScript engine")]
 struct Cli {
@@ -42,7 +45,20 @@ enum Commands {
         file: Option<String>,
     },
     Bench {
-        suite: Option<String>,
+        #[arg(default_value = "js")]
+        mode: Option<String>,
+        #[arg(short, long)]
+        filter: Option<String>,
+        #[arg(long, default_value = "2")]
+        warmup: u32,
+        #[arg(long, default_value = "10")]
+        iterations: u32,
+        #[arg(long)]
+        process: bool,
+        #[arg(long)]
+        update_baseline: bool,
+        #[arg(long, default_value = "1000")]
+        leak_check_interval: usize,
     },
     Test {
         suite: Option<String>,
@@ -64,7 +80,28 @@ fn main() -> ExitCode {
             run(&file, &kernel, &pool)
         }
         Some(Commands::Compile { expr, file }) => compile(expr, file),
-        Some(Commands::Bench { .. }) => not_implemented("bench"),
+        Some(Commands::Bench {
+            mode,
+            filter,
+            warmup,
+            iterations,
+            process,
+            update_baseline,
+            leak_check_interval,
+        }) => {
+            let kernel = make_kernel();
+            let pool = make_pool(&kernel);
+            let config = bench::BenchConfig {
+                mode: mode.unwrap_or_else(|| "js".to_string()),
+                filter,
+                warmup,
+                iterations,
+                process,
+                update_baseline,
+                leak_check_interval,
+            };
+            bench::run_benchmarks(config, kernel, pool)
+        }
         Some(Commands::Test { .. }) => not_implemented("test"),
         None => repl(),
     }

crates/oxide_cli/tests/cli_smoke.rs

@@ -38,12 +38,12 @@ fn run_file() {
 }
 
 #[test]
-fn bench_not_implemented() {
+fn bench_help() {
     let output = Command::new("cargo")
-        .args(["run", "--", "bench"])
+        .args(["run", "--", "bench", "--help"])
         .output()
-        .expect("failed to run oxide bench");
+        .expect("failed to run oxide bench --help");
 
-    let stderr = String::from_utf8_lossy(&output.stderr);
-    assert!(stderr.contains("not yet implemented"), "bench should print not-implemented message");
+    let stdout = String::from_utf8_lossy(&output.stdout);
+    assert!(stdout.contains("--iterations"), "bench --help should list --iterations flag");
 }