Performance and Benchmarks

性能与基准

Performance and Benchmarks

适用版本：4.0.0。

Applies to 4.0.0.

首页 · 语言指南

Home · Language Guide

性能数据用于比较版本和发现回归，不是跨机器的承诺。部署前请在目标 CPU、运行时、宿主配置和真实脚本上重新测量。

Performance data is for comparing versions and detecting regressions, not a cross-machine guarantee. Remeasure on the target CPU, runtime, host configuration, and real scripts before deployment.

运行基准

Running Benchmarks

统一 Benchmark 项目位于 benchmark/，同时覆盖运行时热路径与编译 pipeline。

The unified benchmark project lives in benchmark/ and covers both runtime hot paths and the compiler pipeline.

# 快速冒烟检查
dotnet run --project benchmark/Benchmark.csproj -c Release -- --smoke

# 输出简易 CSV 对比
dotnet run --project benchmark/Benchmark.csproj -c Release -- --compare

# 运行完整 BenchmarkDotNet
dotnet run --project benchmark/Benchmark.csproj -c Release

Run the smoke check, simple CSV comparison, or the complete BenchmarkDotNet suite with the commands above.

重点覆盖 Domain 创建、空调用、函数调用、模块调用、闭包调用、对象、数组、HashMap、字符串、JSON、TDoc、Regex、CLR 互操作，以及 Lexer、Parser、Emitter、单模块/多模块编译与 CompileBlock。

Coverage includes domain creation, empty calls, function calls, module calls, closure calls, objects, arrays, HashMap, strings, JSON, TDoc, Regex, CLR interop, and the lexer, parser, emitter, single-/multi-module compilation, and CompileBlock.

TDoc 基准包含原始值读写、普通嵌套对象读写、一般 Array、Int32Array 与宽对象读取，用于跟踪解析热路径、普通对象写入和 Packed Array 的分配回归。

TDoc benchmarks cover primitive reads/writes, ordinary nested-object reads/writes, general Array, Int32Array, and wide-object reads to track parser hot paths, ordinary-object writing, and packed-array allocation regressions.

dotnet run --project benchmark/Benchmark.csproj -c Release -- --filter "*TypedDocument*"

TDoc 热路径基线

TDoc Hot-path Baseline

下列结果来自 2026-08-20 的 Release/net10.0 BenchmarkDotNet ShortRun。环境为 Windows 11、Intel Xeon W-2235、.NET SDK 10.0.400 与 Runtime .NET 10.0.11。ShortRun 用于回归对比，不应视为跨机器的绝对服务等级承诺。

The results below are from a Release/net10.0 BenchmarkDotNet ShortRun on 2026-08-20: Windows 11, Intel Xeon W-2235, .NET SDK 10.0.400, and Runtime .NET 10.0.11. ShortRun is for regression comparison, not an absolute cross-machine service-level guarantee.

指标	Mean	Allocated	说明
SerializePrimitive	226.0 ns	32 B	结果字符串分配；无中间树
DeserializePrimitive	246.7 ns	0 B	流式扫描与直接绑定
SerializeNestedObject	1.289 µs	296 B	规范化文本与结果字符串
DeserializeNestedObject	2.410 µs	640 B	构造目标脚本对象
SerializePackedInt32Array1024	144.817 µs	20,360 B	文本结果为主要分配
DeserializePackedInt32Array1024	168.155 µs	4,192 B	目标 int[]/Packed Array
DeserializeGenericArray1024	178.815 µs	32,992 B	每个元素的通用 ScriptDatum 存储
DeserializeWideObject256	371.489 µs	16,608 B	256 个目标对象属性

The table records mean time and managed allocation. Primitive parsing stays allocation-free; object and array allocations shown here are the returned data or output text, not an intermediate syntax tree.

已记录的热路径结果

Recorded Hot-path Results

下列数据来自 2026-08-15 的 Release/net10.0 BenchmarkDotNet ShortRun。它们是文档基线；完整套件可用不带 --filter 的命令重建。

The following results are from a Release/net10.0 BenchmarkDotNet ShortRun on 2026-08-15. They are a documentation baseline; regenerate the full suite by running without --filter.

dotnet run --project benchmark/Benchmark.csproj -c Release -- \
  --filter "*FunctionCallLoop*" "*ModuleCallLoop*" "*NumericLoop*"

指标	规模	Mean	Allocated	观察
NumericLoop	1,000	1.477 µs	0 B	数值 local 和算术保持为原生 CIL
NumericLoop	10,000	13.483 µs	0 B	时间线性增长，无循环内分配
FunctionCallLoop	1,000	82.668 µs	0 B	同模块数值调用使用原生 specialization
FunctionCallLoop	10,000	766.803 µs	0 B	调用帧复用，无逐次上下文分配
ModuleCallLoop	1,000	122.591 µs	0 B	跨模块通用调用边界零分配
ModuleCallLoop	10,000	1.186 ms	0 B	动态分发成本可控且无逐次分配

The table records means and allocations for the documented hot paths. All six cases measured zero managed allocations on the execution path.

测试环境为 BenchmarkDotNet 0.15.8、Windows 11 10.0.28000.2704、Intel Core i7-13700KF、.NET SDK 10.0.400、Runtime .NET 10.0.11 和 ShortRun Job。Benchmark 复用预构造的宿主参数数组，因此 Allocated 只反映实际执行路径。

The environment was BenchmarkDotNet 0.15.8, Windows 11 10.0.28000.2704, Intel Core i7-13700KF, .NET SDK 10.0.400, Runtime .NET 10.0.11, and the ShortRun job. The benchmark reuses prebuilt host argument arrays, so Allocated reflects the actual execution path only.

写法建议

Writing Recommendations

• 在热点循环中缓存 length、模块常量和重复属性读取。
• 数值密集且长度已知时使用 Packed Array，而不是通用 Array。
• 使用 StringBuffer 构造大量文本，避免循环中的字符串反复拼接。
• 只在真正需要模块持久化输出或调试符号时选择 Persistence；常规执行优先选择满足需求的较轻模式。
• 优化前先测量，优化后同时运行功能测试和 benchmark。

• Cache length, module constants, and repeated property reads inside hot loops.
• Use packed arrays rather than general Array for known-length numeric work.
• Use StringBuffer for substantial text construction instead of repeated string concatenation in a loop.
• Choose Persistence only when durable module output or debugging symbols are truly required; otherwise use the lightest mode that meets the need.
• Measure before optimizing, then run both functional tests and benchmarks afterward.