[StableAgOH]

Fix #1845.

[StableAgOH]

@an-tao CI was passed in my repository. Request for re-run.

[marty1885]

@StableAgOH Can you benchmark the performance difference? The reason for thread_local was to buffer some randomness generated from the random API. Random can have a high overhead per call. thread_local allows extra randomness be reused. (static does not work as the random function must be thread safe)

[Nahida1O27]

@an-tao @marty1885 我做了benchmark。
Hello, recently I’ve finished packaging Drogon for MSYS2. Today I do benchmark on my Win11 PC.
Below info to refer:

• CPU: i5-1155G7
• OS: Windows 10.0.26300.9711
• compiler: clang++ 23.1.3 (aka也就是mingw后端的llvm)
• target: x86_64-w64-windows-gnu
• thread model: posix
• Trantor version: 1.5.29

测试源码：

// test.cc
//静态链接trantor，动态链接benchmark
//static link against libtrantor.a, dynamic link against libbenchmark.dll
#define TRANTOR_STATIC_DEFINE

#include <cassert>
#include <trantor/utils/Utilities.h>
#include <benchmark/benchmark.h>

namespace A {
namespace B {
std::string former(size_t size) { /* thread_local */ }

std::string later(size_t size) {/* this PR change */}
}   // namespace B
}   // namespace A

static void BM_former(benchmark::State& state) {
    for(auto _: state) {
        auto s = A::B::former(static_cast<size_t>(state.range(0)));
        benchmark::DoNotOptimize(s);
    }
}
static void BM_later(benchmark::State& state) {
    for(auto _: state) {
        auto s = A::B::later(static_cast<size_t>(state.range(0)));
        benchmark::DoNotOptimize(s);
    }
}


BENCHMARK(BM_former)->Arg(16)->Arg(64)->Arg(256)->Arg(1024);
BENCHMARK(BM_later)->Arg(16)->Arg(64)->Arg(256)->Arg(1024);
BENCHMARK_MAIN(); 

输出结果result：

Glarboy@LAPTOP-KMLVGRGJ CLANG64 D:/Glarboy/test/build
$ ./test.exe --benchmark_min_time=1.0
2026-10-29T22:04:47+08:00
Running D:\Glarboy\test\build\test.exe
Run on (8 X 2496 MHz CPU s)
CPU Caches:
  L1 Data 48 KiB (x4)
  L1 Instruction 32 KiB (x4)
  L2 Unified 1280 KiB (x4)
  L3 Unified 8192 KiB (x1)
-- LOG(0): Value passed to --benchmark_min_time should have a suffix. Eg., `30s` for 30-seconds.
---------------------------------------------------------
Benchmark               Time             CPU   Iterations
---------------------------------------------------------
BM_former/16          167 ns          165 ns      8145455
BM_former/64          718 ns          722 ns      1991111
BM_former/256        2805 ns         2668 ns       527059
BM_former/1024      10654 ns        10781 ns       100000
BM_later/16           301 ns          302 ns      4715789
BM_later/64           739 ns          738 ns      1991111
BM_later/256         2607 ns         2579 ns       527059
BM_later/1024       11224 ns        11108 ns       128000

Glarboy@LAPTOP-KMLVGRGJ CLANG64 D:/Glarboy/test/build
$ ./test.exe --benchmark_min_time=10.0
2026-10-29T22:09:15+08:00
Running D:\Glarboy\test\build\test.exe
Run on (8 X 2496 MHz CPU s)
CPU Caches:
  L1 Data 48 KiB (x4)
  L1 Instruction 32 KiB (x4)
  L2 Unified 1280 KiB (x4)
  L3 Unified 8192 KiB (x1)
-- LOG(0): Value passed to --benchmark_min_time should have a suffix. Eg., `30s` for 30-seconds.
---------------------------------------------------------
Benchmark               Time             CPU   Iterations
---------------------------------------------------------
BM_former/16          170 ns          167 ns     81454545
BM_former/64          739 ns          737 ns     19352052
BM_former/256        2836 ns         2827 ns      5239766
BM_former/1024      12931 ns        12610 ns      1280000
BM_later/16           320 ns          319 ns     44577114
BM_later/64           764 ns          764 ns     17164751
BM_later/256         2857 ns         2830 ns      5090909
BM_later/1024       10566 ns        10535 ns      1280000

[Nahida1O27]

看起来生成短字符更慢了

[StableAgOH]

看起来生成短字符更慢了

Thanks for the benchmark. I've updated this PR with a smaller fix: moving the TLS declarations outside the lambda while preserving cross-call caching.

I compared all three implementations on Windows 11, Ryzen 7 9800X3D, Clang 23.1.3 (MSYS2 UCRT64), using Trantor 1.5.29 with OpenSSL 3.6.5 and Google Benchmark 1.10.0 (-O3 -DNDEBUG, no LTO).

Single-thread median wall-clock time over 7 randomly interleaved repetitions:

Length	Original	Previous PR	Updated PR
16	164 ns	285 ns	157 ns
32	306 ns	293 ns	288 ns
33	316 ns	539 ns	310 ns
64	573 ns	558 ns	559 ns
256	2118 ns	1992 ns	2063 ns
1024	7971 ns	7810 ns	7628 ns

The updated version avoids the short-string regression seen in the previous PR version.