xjb : float/double to string algorithm

[xjb714]

src : https://github.com/xjb714/xjb
benchmark : https://github.com/xjb714/f2dec_bench

Thank you. This algorithm was inspired by my reading of your code and there is a mathematical proof in the paper. I tested the performance of the relevant algorithms (dragonbox, yy, ryu, teju, schubfach), but only included the binary to decimal part. After completing the decimal to string, I measured again, and the results might be inconsistent. The above contains source code links and test links.

[ibireme]

I took a quick look at your code.
From what I can see, xjb appears almost identical to my yy_double implementation, with only a few minor modifications.
I didn't notice any meaningful performance improvement.

[xjb714]

@ibireme I have implemented the complete floating-point number printing algorithm. The source code is at https://github.com/xjb714/xjb, in the bench directory. Part of the source code comes from your code. Thank you. You can execute the make command in the bench directory to obtain the result. The code also contains example.cpp. Currently, there is no fast path implementation for small integers.

[ibireme]

I noticed that your optimizations are mostly in the decimal-to-string part,
with heavy use of SIMD, and that the benchmark is built with -march=native.

In yyjson, SIMD-specific code and -march=native are intentionally avoided
to keep the library portable across different platforms and compilers.

These optimizations may be interesting to other projects.

[xjb714]

@ibireme In fact, even if -march=native is removed, the code can still compile normally and the program can still run normally. The gcc compiler provides default support for sse2 of x86-64. You can run gcc -dM -E - < /dev/null | grep "SSE" on Linux to see the result，or run gcc -dM -E - < /dev/null | grep "NEON" on MacOS with apple M1 CPU.