[ShawnNotFound]

For some 80-bit long double values just below a power of 10, formatting with a precision gives a leading zero or rounds at the wrong digit:

auto ld = std::ldexp(static_cast<long double>(0xe596b7b0c643c719ULL), 79);
fmt::format("{:.19e}", ld);  // "1.0000000000000000000e+43", expected "9.9999999999999999997e+42"
fmt::format("{:.21g}", ld);  // "0.99999999999999999997e+43"
fmt::format("{:.0f}", ld);   // "09999999999999999999741184793924429452148736"

format_dragon estimates the exponent and corrects it in the dragon::fixup step, but that check also includes the upper boundary when a precision is given, so the estimate isn't lowered and the digits are generated one exponent too high. The upper boundary only matters in the shortest mode, where the output can round up to the next power of 10. With a fixed number of digits, rounding happens after digit generation, so this compares only the value in that case. Shortest output is unchanged.

The same thing happens for double when it's formatted at compile time, e.g. 9.9999999999999992e22 with {:.20e} gave 0.99999999999999991611e+23.

Added the cases above to the precision test. I also compared 200k random long double values against printf with {:.Ne} and {:.Nf} and found no differences.

[vitaut]

Looks good but please apply clang-format.

[ShawnNotFound]

Thanks, applied clang-format.

[vitaut]

Merged, thanks!