to_chars.cpp, 2 digits at a time and avoiding reversal loop

[fior512]

Summary

This PR propose an improvement of src/to_chars::dragonbox() by using 2-at-a-time method (Alexandrescu, "Three Optimization Tips for C++", 2012) and back-to-front+memcpy write in buffer to avoid reversal loop. Reducing by 3x the number of loop iterations in this function.

Context

This PR continue:

• grisu2 to dragonbox as underlying logic for to_chars #2777

Issues related to this PR:

• Faster double-to-string operation when writing out JSON #1692
• Optimizations opportunities in to_chars (grisu2) #1450

100% on tests with:

cmake -B build -D SIMDJSON_DEVELOPER_MODE=ON
cmake --build build
ctest --test-dir build

benchmarked with benchmark/car_builder/benchmark_car_builder.cpp:

cmake -B build -D SIMDJSON_DEVELOPER_MODE=ON -D CMAKE_BUILD_TYPE=Release
cmake --build build --target benchmark_car_builder -j
./build/benchmark/car_builder/benchmark_car_builder

Bench

master:

Trial 1: string_builder:     1.21 ns  0.82 GB/s  5.24 GHz  6.35 cycles/char  23.39 ins./char  3.68 i/c
Trial 2: string_builder:     1.21 ns  0.82 GB/s  5.23 GHz  6.36 cycles/char  23.39 ins./char  3.68 i/c
Trial 3: string_builder:     1.21 ns  0.82 GB/s  5.25 GHz  6.37 cycles/char  23.38 ins./char  3.67 i/c

this PR:

Trial 1: string_builder:     0.95 ns  1.05 GB/s  5.23 GHz  4.98 cycles/char  17.03 ins./char  3.42 i/c
Trial 2: string_builder:     0.95 ns  1.05 GB/s  5.25 GHz  4.98 cycles/char  17.03 ins./char  3.42 i/c
Trial 3: string_builder:     0.95 ns  1.05 GB/s  5.23 GHz  4.97 cycles/char  17.03 ins./char  3.42 i/c

computed with:

• AMD Ryzen 5 7600X (6cores)
• CPU governor pinned to Performance

Notes

• Latency of to_chars.cpp of this PR, over benchmark/car_builder/benchmark_car_builder.cpp, drop from 0.95 ns/char to 0.85 ns/char if merged with to_chars.cpp, fast path to avoid i128 emulation when possible #2780
• No license needed for 2-at-a-time method; credit given inside to_chars.cpp::dragonbox() doc as comments.
• tried jeaiii algorithm inside dragonbox() to avoid completly modulo and division but it become 0.02ns/char slower than this PR.

[lemire]

Running tests.