Like CBOR before #5824, MessagePack integer arrays were no faster than in
3.12.0: the type switch (number_integer vs. number_unsigned), the sign
test, and the fixint/8/16/32/64-bit ladder mispredict on mixed data.
They now take the CBOR approach:
- number_integer and number_unsigned share one case of write_msgpack.
- The format comes from the size class of x, where x is the value itself
if it is not negative and 2 * (-1 - n) + 1 for a negative n: n fits
int 8/16/32 exactly if that fits uint 8/16/32, so one table lookup
serves both signs. The fixint bound (127, or -32) is selected through a
mask, and the marker is 0xCB + class for uint and 0xCF + class for int.
- Arrays are written by an out-of-line write_msgpack_array() with runs of
integers encoded straight into space reserved from the sink.
The CBOR integer code is factored into helpers both formats share
(integer_bits, integer_size_class, encode_first_byte_and_tail,
write_array_elements). The new path picks the format from the low 64
bits, so number types wider than that keep the previous range checks.
The output is unchanged; the integer-array test now also checks
MessagePack against an independent encoder.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Niels Lohmann
5010011656