Write doubles with the shortest digits (Żmij), digits in registers

Write doubles with the conversion of Zmij by Victor Zverovich (MIT),
ported to C++11 (detail/conversions/zmij.hpp). It finds the
shortest decimal that reads back as the same double, and the
closest one if there are several. Grisu2, used until now, is fast
but not always shortest: it sometimes writes a 17th digit where 16
suffice, or a last digit that is not the closest. The layout is
unchanged (1.5, 100.0, 1e+100, -0.0); float keeps Grisu2.

Digits are converted eight at a time with the BCD conversion of
Xiang JunBo, as in Zmij, and written with one byte swap per eight
digits and fixed-size moves instead of per-digit loops. Leading and
trailing zeros are counted from those bytes. to_chars() uses a
local buffer when the caller's is shorter than the 41 bytes this
may write. The powers of ten come from the number-parsing table,
adjusted where it holds values rounded up, and extended with Zmij's
compressed tables beyond 10^308.

write_shortest() converts its 16 digits in one vector register
(SSE2 on x86-64, NEON on 64-bit Arm, both baseline) and inserts the
decimal point inside the register, avoiding a store-forwarding
stall that cost about 25% of the time to write a double. dump()
writes floats and integers straight into the serializer's write
buffer instead of copying them from a member buffer, and small
integers eight digits at a time. read_eight_bytes() and
parse_eight_digits() are marked always-inline, which GCC had been
calling out of line in the number-parsing loops.

Of one million random doubles, about 0.14% are now written with
different digits, always to a value that still reads back as the
same double.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

include/nlohmann/detail/bit_ops.hpp

@@ -86,8 +86,9 @@ inline uint128_parts full_multiplication(std::uint64_t a, std::uint64_t b) noexc
 }
 
 /// eight bytes as a little-endian word (compilers fold this into one load on
-/// little-endian targets)
-inline std::uint64_t read_eight_bytes(const unsigned char* b) noexcept
+/// little-endian targets; always inlined, as GCC otherwise calls it in the
+/// number loops)
+JSON_HEDLEY_ALWAYS_INLINE std::uint64_t read_eight_bytes(const unsigned char* b) noexcept
 {
     return static_cast<std::uint64_t>(b[0]) | (static_cast<std::uint64_t>(b[1]) << 8u)
            | (static_cast<std::uint64_t>(b[2]) << 16u) | (static_cast<std::uint64_t>(b[3]) << 24u)
@@ -96,7 +97,7 @@ inline std::uint64_t read_eight_bytes(const unsigned char* b) noexcept
 }
 
 /// eight bytes as a little-endian word
-inline std::uint64_t read_eight_bytes(const char* p) noexcept
+JSON_HEDLEY_ALWAYS_INLINE std::uint64_t read_eight_bytes(const char* p) noexcept
 {
     return read_eight_bytes(reinterpret_cast<const unsigned char*>(p)); // NOLINT(cppcoreguidelines-pro-type-reinterpret-cast)
 }