detail::write in one more place relevant to printf with long argument…

[rimathia]

…s, requires moving the definition of detail::write up in the file

I agree that my contributions are licensed under the {fmt} license, and agree to future changes to the licensing.

The benchmark just looks at the scaling for a long string argument:

#include <benchmark/benchmark.h>
#include <fmt/printf.h

void long_printf_format_arg(benchmark::State& state) {
  std::string format_string = "%s";
  std::string message = std::string(state.range(0), 'A');
  for (auto _ : state) {
    benchmark::DoNotOptimize(fmt::sprintf(format_string, message));
  };
}
BENCHMARK(long_printf_format_arg)->Arg(10)->Arg(100)->Arg(1000)->Arg(10000);

void long_fmt_format_arg(benchmark::State& state) {
  std::string format_string = "{}";
  std::string message = std::string(state.range(0), 'A');
  for (auto _ : state) {
    benchmark::DoNotOptimize(fmt::format(format_string, message));
  };
}
BENCHMARK(long_fmt_format_arg)->Arg(10)->Arg(100)->Arg(1000)->Arg(10000);

BENCHMARK_MAIN();

Timings on main (986fa00):

-----------------------------------------------------------------------
Benchmark                             Time             CPU   Iterations
-----------------------------------------------------------------------
long_printf_format_arg/10          38.9 ns         38.9 ns     17873649
long_printf_format_arg/100          192 ns          192 ns      3640240
long_printf_format_arg/1000        1297 ns         1297 ns       526502
long_printf_format_arg/10000      11060 ns        11059 ns        62581
long_fmt_format_arg/10             15.7 ns         15.7 ns     44365010
long_fmt_format_arg/100            75.3 ns         75.3 ns      9133969
long_fmt_format_arg/1000            121 ns          121 ns      5455920
long_fmt_format_arg/10000           275 ns          275 ns      2542404

Timings on the branch (6684314):

-----------------------------------------------------------------------
Benchmark                             Time             CPU   Iterations
-----------------------------------------------------------------------
long_printf_format_arg/10          37.0 ns         37.0 ns     18809721
long_printf_format_arg/100         94.1 ns         94.1 ns      7307423
long_printf_format_arg/1000         250 ns          250 ns      2756264
long_printf_format_arg/10000        443 ns          443 ns      1569507
long_fmt_format_arg/10             15.8 ns         15.8 ns     43821209
long_fmt_format_arg/100            75.8 ns         75.8 ns      9125515
long_fmt_format_arg/1000            118 ns          118 ns      5865348
long_fmt_format_arg/10000           294 ns          294 ns      2473629

[vitaut]

Thanks for the PR. This is a nice optimization but instead of replacing copy_str with write, please move the buffer handling logic from https://github.com/fmtlib/fmt/blob/master/include/fmt/format.h#L2058-L2063 to copy_str.

[vitaut]

Please move these back.

[vitaut]

This overload can be removed now.

[rimathia]

According to git diff origin/master...printf_long_format_arg the shifts of detail::write aren't part of the updated pull request anymore, please let me know if I'm confused here, then I'll just make a new branch for a pull request with a clean history

[rimathia]

The present specialization misses the opportunity to use buffer::append when converting between char_8 and char, but that seemed a bit tricky to get to work. Am I correct in assuming that this isn't a particularly relevant use case?

[vitaut]

Am I correct in assuming that this isn't a particularly relevant use case?

Yes, char8_t should be avoided anyway.

[vitaut]

Thank you!