> I suggest apply patch A to 3.3 as it fixes performance > regression (2x) and is very simple. ASCII and UTF-8 are the two most common codecs in the world, so it's justified to have heavily optimized encoders and decoders. I don't know any application using UTF-32-LE or UTF-32-BE. So I don't want to waste Python memory/code size with a heavily optimized decoder. The patch A looks to be enough. -- 32 bit units is commonly used with wchar_t, but this format already has a fast decoder, PyUnicode_FromWideChar(), which uses memcpy() or _PyUnicode_CONVERT_BYTES().