We've already beaten this topic to death, so let's put a final nail in
the coffin:


__to_chars_10_impl is quite fast. According to the IACA the main loop
takes only 6.0 cycles, the whole function with one iteration takes
10.0 cycles. Replacing the __first[pos] and __first[pos - 1] with
__first[0] and __first[1] drops the function time to 7.53 cycles.

Changelog:

2019-09-08  Antony Polukhin  <antoshkka@gmail.com>

    * include/bits/charconv.h (__detail::__to_chars_10_impl): Replace
    final offsets with constants.


And that's the only optimization that improves all the usecases and
reduces code size on 3 instructions.

Different approaches for optimizing the loop were showing different
results depending on the workload. The most interesting result gives
the compressed table of binary coded decimals:

static constexpr unsigned char __binary_coded_decimals[50] =
{ 0x00, 0x02, 0x04, 0x06, 0x08,  0x10... 0x98 };
unsigned __pos = __len - 1;
while (__val >= 100)
{
  auto const addition = __val & 1;
  auto const __num = (__val % 100) >> 1;
  __val /= 100;
  auto const __bcd = __binary_coded_decimals[__num];
  __first[__pos] = '0' + (__bcd & 0xf) + addition;
  __first[__pos - 1] = '0' + (__bcd >> 4);
  __pos -= 2;
}

That approach shows the same results or even outperforms the existing
approach with __digits[201] = "0001020304..." in case of cold cache.
It also produces slightly smaller binaries. Unfortunately on a warmed
up cache it's slower than the existing approach. I don't think that
it's a worth change.

Attaching some of the benchmarks as a separate file (not for merge,
just something to experiment with).


-- 
Best regards,
Antony Polukhin