Using the public PyUnicodeWriter C API made the json module slower

[vstinner]

I modified the json module to replace the private _PyUnicodeWriter C API with the public PyUnicodeWriter C API:

• 6e63c47
• c914212

Problem: it made the json module slower. Let's investigate what's going on.

Linked PRs

• gh-133968: Use private unicode writer for json #133832
• gh-133968: Add fast path to PyUnicodeWriter_WriteStr() #133969
• [3.14] gh-133968: Add fast path to PyUnicodeWriter_WriteStr() (GH-133969) #133971
• gh-133968: Add PyUnicodeWriter_WriteASCII() function #133973

[vstinner]

See also #133186.

[encukou]

Could you try overoptimizing PyUnicodeWriter_WriteUTF8 and benchmarking that, to see how fast the existing function can “theoretically” be?
I'm thinking something like:

• in the JSON module, force the constants ("null" etc) to be size_t-aligned
• add private (ABI-only) function _PyUnicodeWriter_WriteUTF8_SmallAligned, which requires size_t-aligned input and size <= sizeof(size_t)
• make PyUnicodeWriter_WriteUTF8 a macro that calls _PyUnicodeWriter_WriteUTF8_SmallAligned if the need are met, and PyUnicodeWriter_WriteUTF8 otherwise

[vstinner]

in the JSON module, force the constants ("null" etc) to be size_t-aligned

I don't know how to guarantee that.

add private (ABI-only) function _PyUnicodeWriter_WriteUTF8_SmallAligned, which requires size_t-aligned input and size <= sizeof(size_t)

Why do you want to limit the size to sizeof(size_t)? Which kind of optimization are you thinking of? find_first_nonascii() already takes care of the different cases with different optimizations.

ascii_decode() is already highly optimized.