quickwin 3.0.1 → 3.0.3

package/test/bench_ffi_callraw.ts

@@ -1,235 +0,0 @@
-// 一次性性能 bench（归因场）：绕过 callPacked 直调 ffi.ffiCall，把 bind() 调用链拆层。
-// 不入 CI。A 臂 = bind() 同场对照（对齐 bench_ffi_overhead.ts 的 S0-S3 量级）。
-// 用法：make js 后 win11 POST /exec {"cmd":"qwin.exe test\\bench_ffi_callraw.js","timeout":600000}
-// 臂（同 proc、同帧布局，唯一变量是「每次调用经过哪些层」）：
-//   A bind      rest 参数 + callPacked 全链（每调 argFrame/DataView/retBuf/readRet-dv
-//               分配、getWidth 闭包、writeSlot 对象字面量、ffiCall、readRet）
-//   B raw       帧复用 + 每调最小写槽/读返 + ffiCall → native 桩 + QuickJS C 边界
-//   C raw-alloc 每调 new ArrayBuffer(32)+DataView+ArrayBuffer(8)+读返 DataView
-//               （复刻 callPacked 分配面）+ 写槽 + ffiCall（仅 S0/S1）
-// 差值解读：B = trampoline 地板；C−B = 每调分配；A−C = callPacked JS 层其余。
-// x64 帧约定：argFrame 恒 32B（4×8 槽，桩无条件读 4 槽），retBuf 8B（ffiCall ≥8 校验）。
-// S1/S2/S3 各臂消费返回值到 sink（A/B/C 同口径，防 DCE 疑虑 + 收尾打印证活）。
-// 结束必须 DestroyWindow + std.exit。
-import * as std from 'std'
-import * as gui from 'gui'
-import * as ffi from 'ffi'
-import { bind } from '../lib/ffi/bind.js'
-
-const WM_NULL = 0x0000
-const FRAME_BYTES = 32   // x64：4 槽 × 8B，桩无条件读 slots[0..3]
-const ROUNDS = 15
-const TARGET_MS = 400
-
-const med = (t: number[]): number => {
-    const s = [...t].sort((x, y) => x - y)
-    const m = s.length >> 1
-    return s.length % 2 ? s[m]! : (s[m - 1]! + s[m]!) / 2
-}
-
-// 文件日志（cwd=Z:\quickwin → 宿主 _build/bench_ffi_callraw_log.txt）
-const log = std.open('bench_ffi_callraw_log.txt', 'w')
-function step(msg: string): void {
-    std.printf('%s\n', msg)
-    if (log) { log.puts(msg + '\n'); log.flush() }
-}
-
-step('stage: imports ok')
-
-// ---- proc 地址（B/C 臂绕过 bind 直驱 ffiCall）----
-function procOf(dll: string, name: string): number {
-    const h = ffi.LoadLibrary(dll)
-    if (!h) throw new Error('LoadLibrary(' + dll + ') failed')
-    const p = ffi.GetProcAddress(h, name)
-    if (!p) throw new Error('GetProcAddress(' + dll + '!' + name + ') failed')
-    return p
-}
-const pNop0 = procOf('test/nop.dll', 'nop0')
-const pNop1 = procOf('test/nop.dll', 'nop1')
-const pPid = procOf('kernel32.dll', 'GetCurrentProcessId')
-const pSm = procOf('user32.dll', 'SendMessageW')
-
-// B 臂复用帧 + 视图（一次分配，跨轮复用）
-const frame = new ArrayBuffer(FRAME_BYTES)
-const retBuf = new ArrayBuffer(8)
-const fv = new DataView(frame)
-const rv = new DataView(retBuf)
-let sink = 0
-
-const fNop0 = bind('test/nop.dll', 'nop0', ' -> void')
-const fNop1 = bind('test/nop.dll', 'nop1', 'i32 -> i32')
-const fPid = bind('kernel32.dll', 'GetCurrentProcessId', ' -> u32')
-
-const winRaw = gui.CreateWindow('STATIC', 'bench-title', 0, 0, 0, 20, 20, null, null)
-if (!winRaw) throw new Error('CreateWindow(STATIC) failed')
-const hwnd: gui.HWND = winRaw
-const fSm = bind('user32.dll', 'SendMessageW', '<HWND>ptr u32 <>ptr <>ptr -> <>ptr')
-step('stage: window created')
-
-// ---- 可行性探针：任一臂签名/帧布局错位立刻炸 ----
-fNop0()
-if (fNop1(1) !== 2) throw new Error('sanity: bind nop1(1) !== 2')
-if (fPid() <= 0) throw new Error('sanity: bind pid <= 0')
-ffi.ffiCall(pNop0, frame, retBuf, 0)
-fv.setInt32(0, 1, true)
-ffi.ffiCall(pNop1, frame, retBuf, 0)
-if (rv.getInt32(0, true) !== 2) throw new Error('sanity: raw nop1 !== 2, got ' + rv.getInt32(0, true))
-ffi.ffiCall(pPid, frame, retBuf, 0)
-if (rv.getUint32(0, true) <= 0) throw new Error('sanity: raw pid <= 0')
-// S3 帧布局预演（hwnd 槽 + msg + wParam + lParam）
-fv.setBigUint64(0, BigInt(Number(hwnd)), true)
-fv.setUint32(8, WM_NULL, true)
-fv.setBigUint64(16, 0n, true)
-fv.setBigUint64(24, 0x1234n, true)
-ffi.ffiCall(pSm, frame, retBuf, 0)
-const smRet = rv.getBigUint64(0, true)
-step('stage: probes ok, raw SendMessage ret=' + String(smRet) + ' (expect 0n)')
-
-// ---- 臂循环：函数取到局部 const，两侧/各臂同形态 ----
-const runA0 = (n: number): number => {
-    const f = fNop0
-    const t0 = Date.now()
-    for (let i = 0; i < n; i++) f()
-    return Date.now() - t0
-}
-const runB0 = (n: number): number => {
-    const p = pNop0, fr = frame, rb = retBuf
-    const t0 = Date.now()
-    for (let i = 0; i < n; i++) ffi.ffiCall(p, fr, rb, 0)
-    return Date.now() - t0
-}
-const runC0 = (n: number): number => {
-    const p = pNop0
-    const t0 = Date.now()
-    for (let i = 0; i < n; i++) {
-        const f = new ArrayBuffer(FRAME_BYTES)
-        const d = new DataView(f)   // callPacked 无条件建 dv，照抄
-        void d
-        ffi.ffiCall(p, f, new ArrayBuffer(8), 0)
-    }
-    return Date.now() - t0
-}
-const runA1 = (n: number): number => {
-    const f = fNop1
-    const t0 = Date.now()
-    for (let i = 0; i < n; i++) sink += f(1)
-    return Date.now() - t0
-}
-const runB1 = (n: number): number => {
-    const p = pNop1, fr = frame, rb = retBuf, d = fv, rd = rv
-    const t0 = Date.now()
-    for (let i = 0; i < n; i++) {
-        d.setInt32(0, 1, true)
-        ffi.ffiCall(p, fr, rb, 0)
-        sink += rd.getInt32(0, true)
-    }
-    return Date.now() - t0
-}
-const runC1 = (n: number): number => {
-    const p = pNop1
-    const t0 = Date.now()
-    for (let i = 0; i < n; i++) {
-        const f = new ArrayBuffer(FRAME_BYTES)
-        const d = new DataView(f)
-        d.setInt32(0, 1, true)
-        const r = new ArrayBuffer(8)
-        ffi.ffiCall(p, f, r, 0)
-        sink += new DataView(r).getInt32(0, true)   // readRet 同构：每次 new DataView
-    }
-    return Date.now() - t0
-}
-const runA2 = (n: number): number => {
-    const f = fPid
-    const t0 = Date.now()
-    for (let i = 0; i < n; i++) sink += f()
-    return Date.now() - t0
-}
-const runB2 = (n: number): number => {
-    const p = pPid, fr = frame, rb = retBuf, rd = rv
-    const t0 = Date.now()
-    for (let i = 0; i < n; i++) {
-        ffi.ffiCall(p, fr, rb, 0)
-        sink += rd.getUint32(0, true)
-    }
-    return Date.now() - t0
-}
-const runA3 = (n: number): number => {
-    const f = fSm
-    const t0 = Date.now()
-    for (let i = 0; i < n; i++) sink += f(hwnd, WM_NULL, 0, 0x1234)
-    return Date.now() - t0
-}
-const runB3 = (n: number): number => {
-    const p = pSm, fr = frame, rb = retBuf, d = fv, rd = rv
-    const t0 = Date.now()
-    for (let i = 0; i < n; i++) {
-        d.setBigUint64(0, BigInt(Number(hwnd)), true)
-        d.setUint32(8, WM_NULL, true)
-        d.setBigUint64(16, 0n, true)
-        d.setBigUint64(24, 0x1234n, true)
-        ffi.ffiCall(p, fr, rb, 0)
-        sink += Number(rd.getBigUint64(0, true))
-    }
-    return Date.now() - t0
-}
-
-type Arm = { label: string; n: number; fn: (n: number) => number }
-
-// 自校准：探针逐级放大 n0 至 dt≥20ms（1ms 量化步长 ≤5%），按 TARGET_MS 定批量。
-// B 臂可能比 A 快两个量级，固定批量必有一侧撞量化墙——与 overhead 场同法。
-function calibrate(label: string, fn: (n: number) => number): number {
-    let n0 = 10_000
-    let dt = 0
-    for (;;) {
-        dt = fn(n0)
-        if (dt >= 20 || n0 >= 10_000_000) break
-        n0 *= 10
-    }
-    const per = dt / n0
-    const n = Math.max(10_000, Math.min(10_000_000, Math.round(TARGET_MS / per)))
-    step('  calib ' + label + ': probe n=' + String(n0) + ' dt=' + String(dt) +
-        'ms → batch=' + String(n))
-    return n
-}
-
-const mkArm = (label: string, fn: (n: number) => number): Arm =>
-    ({ label, n: calibrate(label, fn), fn })
-
-function suite(name: string, arms: Arm[]): void {
-    step('suite: ' + name + ' warmup')
-    for (const a of arms) a.fn(a.n)
-    const times: number[][] = arms.map(() => [])
-    for (let r = 0; r < ROUNDS; r++) {
-        for (let k = 0; k < arms.length; k++) {
-            const i = (r + k) % arms.length
-            times[i]!.push(arms[i]!.fn(arms[i]!.n))
-        }
-    }
-    for (let r = 0; r < Math.min(ROUNDS, 3); r++) {
-        step('  round ' + String(r) + ': ' +
-            arms.map((a, i) => a.label + '=' + String(times[i]![r]) + 'ms').join('  '))
-    }
-    const ns = (i: number, ms: number) => (ms * 1e6 / arms[i]!.n).toFixed(1)
-    const meds = times.map(med)
-    const mins = times.map(t => Math.min(...t))
-    step('  ** median: ' + arms.map((a, i) => a.label + '=' + ns(i, meds[i]!) + 'ns').join(' | '))
-    step('     min:    ' + arms.map((a, i) => a.label + '=' + ns(i, mins[i]!) + 'ns').join(' | '))
-    if (arms.length >= 3) {
-        const a = Number(ns(0, meds[0]!)), b = Number(ns(1, meds[1]!)), c = Number(ns(2, meds[2]!))
-        step('      split:  C-B=' + (c - b).toFixed(1) + 'ns (每调分配)  A-C=' +
-            (a - c).toFixed(1) + 'ns (callPacked JS 层其余)')
-    }
-}
-
-step('bench_ffi_callraw (attribution): A=bind / B=raw ffiCall / C=raw+per-call alloc; ' +
-    'calib ~' + String(TARGET_MS) + 'ms, ' + String(ROUNDS) + 'r (median)')
-suite('S0 nop0  0参 void', [mkArm('A bind', runA0), mkArm('B raw ', runB0), mkArm('C allc', runC0)])
-suite('S1 nop1  1×i32', [mkArm('A bind', runA1), mkArm('B raw ', runB1), mkArm('C allc', runC1)])
-suite('S2 GetCurrentProcessId', [mkArm('A bind', runA2), mkArm('B raw ', runB2)])
-suite('S3 SendMessageW WM_NULL', [mkArm('A bind', runA3), mkArm('B raw ', runB3)])
-
-step('sink=' + String(sink) + ' (liveness: >0 证明各臂 native 调用真实发生)')
-gui.DestroyWindow(hwnd)
-step('DONE')
-if (log) log.close()
-std.exit(0)

package/test/bench_ffi_overhead.js

@@ -1,139 +0,0 @@
-// 一次性性能 bench：我方 bind() FFI vs node:ffi(Node v26.10) 同机同场景对比（win11 VM）
-// 不入 CI。对照脚本 test/bench_nodeffi.mjs（node 侧同场景）。结论见对话，未落文档。
-// 用法：make js 后 POST /exec {"cmd":"qwin.exe test/bench_ffi_overhead.js"}
-// 场景（跨运行时按 ns/call 归一，每臂独立自校准到 ~400ms/样本）：
-//   S0 nop0  test/nop.dll 0 参 void —— 纯 JS→C 往返 trampoline（无封送无返回读取）
-//   S1 nop1  test/nop.dll 1×i32 —— 单标量封送往返
-//   S2 pid   kernel32 GetCurrentProcessId 0 参 —— 真实系统 DLL
-//   S3 sm    user32 SendMessageW WM_NULL 自有窗口 —— 完整链路 + gui C 直调地板参照
-// 方法学沿用 bench_sendmessage_nonnull：预热 1 + 臂序轮转 + median 主指标/min。
-// Date.now() 1ms 量化 → 批量自校准（探针 ≥20ms 才有 ≤5% 步长，目标 ~400ms/样本）。
-// 全部消息发自有窗口（直调 wndproc）。结束必须 DestroyWindow + std.exit。
-import * as std from 'std';
-import * as gui from 'gui';
-import { bind } from '../lib/ffi/bind.js';
-const WM_NULL = 0x0000;
-const ROUNDS = 15;
-const TARGET_MS = 400;
-// median（离群稳健）供 calibrate 与 suite 共用
-const med = (t) => {
-    const s = [...t].sort((x, y) => x - y);
-    const m = s.length >> 1;
-    return s.length % 2 ? s[m] : (s[m - 1] + s[m]) / 2;
-};
-// 文件日志（cwd=Z:\quickwin，宿主机 _build/bench_ffi_overhead_log.txt 可实时读）——
-// exec 只在进程退出后回包，挂了靠它定位卡点
-const log = std.open('bench_ffi_overhead_log.txt', 'w');
-function step(msg) {
-    std.printf('%s\n', msg);
-    if (log) {
-        log.puts(msg + '\n');
-        log.flush();
-    }
-}
-step('stage: imports ok');
-const winRaw = gui.CreateWindow('STATIC', 'bench-title', 0, 0, 0, 20, 20, null, null);
-if (!winRaw)
-    throw new Error('CreateWindow(STATIC) failed');
-const win = winRaw;
-step('stage: window created');
-const fNop0 = bind('test/nop.dll', 'nop0', ' -> void');
-const fNop1 = bind('test/nop.dll', 'nop1', 'i32 -> i32');
-const fPid = bind('kernel32.dll', 'GetCurrentProcessId', ' -> u32');
-const fSm = bind('user32.dll', 'SendMessageW', '<HWND>ptr u32 <>ptr <>ptr -> <>ptr');
-// 每臂循环体：函数先取到局部 const（与 node 侧同形态——模块槽访问 vs 局部槽访问的
-// 每迭代差在 ns 级，两边都取局部才对齐）；调用结果不使用（QuickJS 解释器不消除
-// 副作用调用，与原 bench 同口径）
-const runNop0 = (n) => {
-    const f = fNop0;
-    const t0 = Date.now();
-    for (let i = 0; i < n; i++)
-        f();
-    return Date.now() - t0;
-};
-const runNop1 = (n) => {
-    const f = fNop1;
-    const t0 = Date.now();
-    for (let i = 0; i < n; i++)
-        f(1);
-    return Date.now() - t0;
-};
-const runPid = (n) => {
-    const f = fPid;
-    const t0 = Date.now();
-    for (let i = 0; i < n; i++)
-        f();
-    return Date.now() - t0;
-};
-const runSm = (n) => {
-    const f = fSm;
-    const t0 = Date.now();
-    for (let i = 0; i < n; i++)
-        f(win, WM_NULL, 0, 0x1234);
-    return Date.now() - t0;
-};
-const runGuiSm = (n) => {
-    const f = gui.SendMessage;
-    const t0 = Date.now();
-    for (let i = 0; i < n; i++)
-        f(win, WM_NULL, 0, 0x1234);
-    return Date.now() - t0;
-};
-// 自校准：探针逐级放大 n0 直到 dt≥20ms（量化步长 1ms ≤ 5%），再按 TARGET_MS 定批量。
-// 不预设单次调用量级——我方与 gui 地板相差 100×，固定批量必有一侧撞量化墙。
-function calibrate(label, fn) {
-    let n0 = 10_000;
-    let dt = 0;
-    for (;;) {
-        dt = fn(n0);
-        if (dt >= 20 || n0 >= 10_000_000)
-            break;
-        n0 *= 10;
-    }
-    const per = dt / n0;
-    const n = Math.max(10_000, Math.min(10_000_000, Math.round(TARGET_MS / per)));
-    step('  calib ' + label + ': probe n=' + String(n0) + ' dt=' + String(dt) +
-        'ms → batch=' + String(n));
-    return n;
-}
-function mkArm(label, fn) {
-    return { label, n: calibrate(label, fn), fn };
-}
-function suite(name, arms) {
-    step('suite: ' + name + ' warmup');
-    for (const a of arms)
-        a.fn(a.n); // 预热各 1 轮
-    const times = arms.map(() => []);
-    for (let r = 0; r < ROUNDS; r++) {
-        // 臂序按 r 轮转起跑，抵消顺序性偏置（温度/调度漂移）
-        for (let k = 0; k < arms.length; k++) {
-            const i = (r + k) % arms.length;
-            times[i].push(arms[i].fn(arms[i].n));
-        }
-    }
-    // 每轮原始数据只 dump 前 3 轮（15 轮全打没意义，log 只留诊断线索）
-    for (let r = 0; r < Math.min(ROUNDS, 3); r++) {
-        step('  round ' + String(r) + ': ' +
-            arms.map((a, i) => a.label + '=' + String(times[i][r]) + 'ms').join('  '));
-    }
-    const ns = (i, ms) => (ms * 1e6 / arms[i].n).toFixed(1);
-    // median 是主指标（离群稳健）；min 留作参照
-    const meds = times.map(med);
-    const mins = times.map(t => Math.min(...t));
-    step('  ** median: ' + arms.map((a, i) => a.label + '=' + ns(i, meds[i]) + 'ns').join(' | '));
-    step('     min:    ' + arms.map((a, i) => a.label + '=' + ns(i, mins[i]) + 'ns').join(' | '));
-}
-step('bench_ffi_overhead (ours): per-arm calib ~' + String(TARGET_MS) + 'ms, ' +
-    String(ROUNDS) + 'r (median 主指标), 对照 test/bench_nodeffi.mjs');
-suite('S0 nop0  0参 void 纯往返', [mkArm('ours', runNop0)]);
-suite('S1 nop1  1×i32 封送往返', [mkArm('ours', runNop1)]);
-suite('S2 GetCurrentProcessId  0参真实API', [mkArm('ours', runPid)]);
-suite('S3 SendMessageW WM_NULL 自有窗口 (gui=直调地板)', [
-    mkArm('ours', runSm),
-    mkArm('gui ', runGuiSm),
-]);
-gui.DestroyWindow(win);
-step('DONE');
-if (log)
-    log.close();
-std.exit(0);

package/test/bench_ffi_overhead.ts

@@ -1,141 +0,0 @@
-// 一次性性能 bench：我方 bind() FFI vs node:ffi(Node v26.10) 同机同场景对比（win11 VM）
-// 不入 CI。对照脚本 test/bench_nodeffi.mjs（node 侧同场景）。结论见对话，未落文档。
-// 用法：make js 后 POST /exec {"cmd":"qwin.exe test/bench_ffi_overhead.js"}
-// 场景（跨运行时按 ns/call 归一，每臂独立自校准到 ~400ms/样本）：
-//   S0 nop0  test/nop.dll 0 参 void —— 纯 JS→C 往返 trampoline（无封送无返回读取）
-//   S1 nop1  test/nop.dll 1×i32 —— 单标量封送往返
-//   S2 pid   kernel32 GetCurrentProcessId 0 参 —— 真实系统 DLL
-//   S3 sm    user32 SendMessageW WM_NULL 自有窗口 —— 完整链路 + gui C 直调地板参照
-// 方法学沿用 bench_sendmessage_nonnull：预热 1 + 臂序轮转 + median 主指标/min。
-// Date.now() 1ms 量化 → 批量自校准（探针 ≥20ms 才有 ≤5% 步长，目标 ~400ms/样本）。
-// 全部消息发自有窗口（直调 wndproc）。结束必须 DestroyWindow + std.exit。
-import * as std from 'std'
-import * as gui from 'gui'
-import { bind } from '../lib/ffi/bind.js'
-
-const WM_NULL = 0x0000
-const ROUNDS = 15
-const TARGET_MS = 400
-
-// median（离群稳健）供 calibrate 与 suite 共用
-const med = (t: number[]): number => {
-    const s = [...t].sort((x, y) => x - y)
-    const m = s.length >> 1
-    return s.length % 2 ? s[m]! : (s[m - 1]! + s[m]!) / 2
-}
-
-// 文件日志（cwd=Z:\quickwin，宿主机 _build/bench_ffi_overhead_log.txt 可实时读）——
-// exec 只在进程退出后回包，挂了靠它定位卡点
-const log = std.open('bench_ffi_overhead_log.txt', 'w')
-function step(msg: string): void {
-    std.printf('%s\n', msg)
-    if (log) { log.puts(msg + '\n'); log.flush() }
-}
-
-step('stage: imports ok')
-
-const winRaw = gui.CreateWindow('STATIC', 'bench-title', 0, 0, 0, 20, 20, null, null)
-if (!winRaw) throw new Error('CreateWindow(STATIC) failed')
-const win: gui.HWND = winRaw
-step('stage: window created')
-
-const fNop0 = bind('test/nop.dll', 'nop0', ' -> void')
-const fNop1 = bind('test/nop.dll', 'nop1', 'i32 -> i32')
-const fPid = bind('kernel32.dll', 'GetCurrentProcessId', ' -> u32')
-const fSm = bind('user32.dll', 'SendMessageW', '<HWND>ptr u32 <>ptr <>ptr -> <>ptr')
-
-// 每臂循环体：函数先取到局部 const（与 node 侧同形态——模块槽访问 vs 局部槽访问的
-// 每迭代差在 ns 级，两边都取局部才对齐）；调用结果不使用（QuickJS 解释器不消除
-// 副作用调用，与原 bench 同口径）
-const runNop0 = (n: number): number => {
-    const f = fNop0
-    const t0 = Date.now()
-    for (let i = 0; i < n; i++) f()
-    return Date.now() - t0
-}
-const runNop1 = (n: number): number => {
-    const f = fNop1
-    const t0 = Date.now()
-    for (let i = 0; i < n; i++) f(1)
-    return Date.now() - t0
-}
-const runPid = (n: number): number => {
-    const f = fPid
-    const t0 = Date.now()
-    for (let i = 0; i < n; i++) f()
-    return Date.now() - t0
-}
-const runSm = (n: number): number => {
-    const f = fSm
-    const t0 = Date.now()
-    for (let i = 0; i < n; i++) f(win, WM_NULL, 0, 0x1234)
-    return Date.now() - t0
-}
-const runGuiSm = (n: number): number => {
-    const f = gui.SendMessage
-    const t0 = Date.now()
-    for (let i = 0; i < n; i++) f(win, WM_NULL, 0, 0x1234)
-    return Date.now() - t0
-}
-
-type Arm = { label: string; n: number; fn: (n: number) => number }
-
-// 自校准：探针逐级放大 n0 直到 dt≥20ms（量化步长 1ms ≤ 5%），再按 TARGET_MS 定批量。
-// 不预设单次调用量级——我方与 gui 地板相差 100×，固定批量必有一侧撞量化墙。
-function calibrate(label: string, fn: (n: number) => number): number {
-    let n0 = 10_000
-    let dt = 0
-    for (;;) {
-        dt = fn(n0)
-        if (dt >= 20 || n0 >= 10_000_000) break
-        n0 *= 10
-    }
-    const per = dt / n0
-    const n = Math.max(10_000, Math.min(10_000_000, Math.round(TARGET_MS / per)))
-    step('  calib ' + label + ': probe n=' + String(n0) + ' dt=' + String(dt) +
-        'ms → batch=' + String(n))
-    return n
-}
-
-function mkArm(label: string, fn: (n: number) => number): Arm {
-    return { label, n: calibrate(label, fn), fn }
-}
-
-function suite(name: string, arms: Arm[]): void {
-    step('suite: ' + name + ' warmup')
-    for (const a of arms) a.fn(a.n)   // 预热各 1 轮
-    const times: number[][] = arms.map(() => [])
-    for (let r = 0; r < ROUNDS; r++) {
-        // 臂序按 r 轮转起跑，抵消顺序性偏置（温度/调度漂移）
-        for (let k = 0; k < arms.length; k++) {
-            const i = (r + k) % arms.length
-            times[i]!.push(arms[i]!.fn(arms[i]!.n))
-        }
-    }
-    // 每轮原始数据只 dump 前 3 轮（15 轮全打没意义，log 只留诊断线索）
-    for (let r = 0; r < Math.min(ROUNDS, 3); r++) {
-        step('  round ' + String(r) + ': ' +
-            arms.map((a, i) => a.label + '=' + String(times[i]![r]) + 'ms').join('  '))
-    }
-    const ns = (i: number, ms: number) => (ms * 1e6 / arms[i]!.n).toFixed(1)
-    // median 是主指标（离群稳健）；min 留作参照
-    const meds = times.map(med)
-    const mins = times.map(t => Math.min(...t))
-    step('  ** median: ' + arms.map((a, i) => a.label + '=' + ns(i, meds[i]!) + 'ns').join(' | '))
-    step('     min:    ' + arms.map((a, i) => a.label + '=' + ns(i, mins[i]!) + 'ns').join(' | '))
-}
-
-step('bench_ffi_overhead (ours): per-arm calib ~' + String(TARGET_MS) + 'ms, ' +
-    String(ROUNDS) + 'r (median 主指标), 对照 test/bench_nodeffi.mjs')
-suite('S0 nop0  0参 void 纯往返', [mkArm('ours', runNop0)])
-suite('S1 nop1  1×i32 封送往返', [mkArm('ours', runNop1)])
-suite('S2 GetCurrentProcessId  0参真实API', [mkArm('ours', runPid)])
-suite('S3 SendMessageW WM_NULL 自有窗口 (gui=直调地板)', [
-    mkArm('ours', runSm),
-    mkArm('gui ', runGuiSm),
-])
-
-gui.DestroyWindow(win)
-step('DONE')
-if (log) log.close()
-std.exit(0)

package/test/bench_sendmessage.js

@@ -1,104 +0,0 @@
-// 一次性性能 bench：gui.SendMessage（C 模块直调） vs FFI bind SendMessageW（libffi trampoline）
-// 不入 CI。用法：make js 后 POST /exec {"cmd":"qwin.exe test/bench_sendmessage.js"}
-// 每场景预热 1 轮 + 5 轮计数（奇偶交叉起跑顺序），取各自最小值降抖动。
-// 注意：全部消息发给自有窗口（直调 wndproc）——跨进程消息走消息经纪，几十 μs/次且
-// IPC 开销掩盖入口链差异，不作测点。结束必须 DestroyWindow + std.exit：留活窗会让
-// qwin 主流程驻留消息循环，exec 永不返回。
-import * as std from 'std';
-import * as gui from 'gui';
-import { SendMessage } from '../lib/windows/user32.js';
-import { WCHAR } from '../lib/ffi/bind.js';
-import { PtrArrayBuffer } from '../lib/ffi/ctype.js';
-const N = 500_000;
-// S2 含 WCHAR.encode 每次分配：500k 时 GC 抖动把单轮拖到 190s+（实测 ~380μs/迭代），
-// 降到 50k——统计上仍远超噪声需求
-const N_S2 = 50_000;
-const ROUNDS = 5;
-const ROUNDS_S2 = 3;
-const WM_SETTEXT = 0x000c;
-const WM_GETTEXT = 0x000d;
-// 文件日志（cwd=Z:\quickwin，宿主机 _build/bench_log.txt 可实时读）——exec 只在进程
-// 退出后回包，挂了靠它定位卡点
-const log = std.open('bench_log.txt', 'w');
-function step(msg) {
-    std.printf('%s\n', msg);
-    if (log) {
-        log.puts(msg + '\n');
-        log.flush();
-    }
-}
-step('stage: imports ok');
-const winRaw = gui.CreateWindow('STATIC', 'bench-title', 0, 0, 0, 20, 20, null, null);
-if (!winRaw)
-    throw new Error('CreateWindow(STATIC) failed');
-const win = winRaw;
-step('stage: window created');
-// 两侧入口各自提为常量（循环内属性查找开销对齐）
-const gSM = gui.SendMessage;
-const fSM = SendMessage;
-const TEXT = 'bench';
-const buf = new PtrArrayBuffer(64);
-// —— S1（WM_NULL 纯数字，最纯入口链开销）：实现已随首轮跑数移除，数据见 .agents/SENDMESSAGE_BENCH.md ——
-// —— S2：WM_SETTEXT 字符串 lParam（任务 3 多态点的真实成本形态）：
-//    gui 侧 C 内编码；FFI 侧调用点 WCHAR.encode 在循环内（迁移后的真实写法）——
-function s2g() {
-    const t0 = Date.now();
-    for (let i = 0; i < N_S2; i++)
-        gSM(win, WM_SETTEXT, 0, TEXT);
-    return Date.now() - t0;
-}
-function s2f() {
-    const t0 = Date.now();
-    for (let i = 0; i < N_S2; i++) {
-        const w = WCHAR.encode(TEXT);
-        fSM(win, WM_SETTEXT, 0, w.ptr);
-    }
-    return Date.now() - t0;
-}
-// —— S3：WM_GETTEXT 缓冲出参往返，两侧 JS 实参完全相同（纯入口差 + C 侧拷贝共模）——
-function s3g() {
-    const t0 = Date.now();
-    for (let i = 0; i < N; i++)
-        gSM(win, WM_GETTEXT, 64, buf.ptr);
-    return Date.now() - t0;
-}
-function s3f() {
-    const t0 = Date.now();
-    for (let i = 0; i < N; i++)
-        fSM(win, WM_GETTEXT, 64, buf.ptr);
-    return Date.now() - t0;
-}
-function suite(name, n, rounds, rg, rf) {
-    step('suite: ' + name + ' warmup');
-    rg();
-    rf(); // 预热各 1 轮
-    const gs = [];
-    const fs = [];
-    for (let r = 0; r < rounds; r++) {
-        // 奇偶交叉起跑顺序，抵消顺序性偏置（温度/调度漂移）
-        if (r % 2 === 0) {
-            gs.push(rg());
-            fs.push(rf());
-        }
-        else {
-            fs.push(rf());
-            gs.push(rg());
-        }
-        step('  round ' + String(r) + ': gui=' + String(gs[gs.length - 1]) +
-            'ms ffi=' + String(fs[fs.length - 1]) + 'ms');
-    }
-    const gmin = Math.min(...gs), fmin = Math.min(...fs);
-    const ns = (ms) => (ms * 1e6 / n).toFixed(0);
-    step('  ==> gui min=' + String(gmin) + 'ms (' + ns(gmin) + ' ns/call), ffi min=' +
-        String(fmin) + 'ms (' + ns(fmin) + ' ns/call), ffi/gui=' + (fmin / gmin).toFixed(2) +
-        'x, delta=' + ns(fmin - gmin) + 'ns');
-}
-step('bench_sendmessage: round2 (S1 done in round1), S2 n=' + String(N_S2) +
-    '/' + String(ROUNDS_S2) + 'r, S3 n=' + String(N) + '/' + String(ROUNDS) + 'r (min)');
-suite('S2 WM_SETTEXT  字符串 lParam  (FFI 侧含 WCHAR.encode)', N_S2, ROUNDS_S2, s2g, s2f);
-suite('S3 WM_GETTEXT  缓冲出参往返  (两侧实参相同)', N, ROUNDS, s3g, s3f);
-gui.DestroyWindow(win);
-step('DONE');
-if (log)
-    log.close();
-std.exit(0);