`<format>`: Alternate form general floating-point can mishandle width

[StephanTLavavej]

Reported to me by Attila Feher. https://godbolt.org/z/djh5KxjW6

#include <format>
#include <iostream>
using namespace std;

int main() {
    std::cout << "Expected: [1.e-37]\n";
    std::cout << format("  Actual: [{:#6.0g}]\n", 1.234e-37);
}

libstdc++ and libc++ behave as expected, but MSVC prints:

Expected: [1.e-37]
  Actual: [ 1.e-37]

🕵️ Analysis

I debugged into this, and although I'm not absolutely certain of the root cause yet, this line appears to be involved:

STL/stl/inc/format

Line 2278 in 926d458

_Zeroes_to_append = _Extra_precision + _Precision - _Digits;

_Zeroes_to_append is computed as -1 because _Extra_precision and _Precision are both 0, and _Digits is 1.

We've correctly computed _Width to be 6, but the negative _Zeroes_to_append damages it, which appears to be the proximate cause of the bug:

STL/stl/inc/format

Line 2297 in 926d458

_Width += _Zeroes_to_append;

Later, we use _Zeroes_to_append in a clamped manner (i.e. negative values are treated as zero):

STL/stl/inc/format

Lines 2328 to 2330 in 926d458

	for (; _Zeroes_to_append > 0; --_Zeroes_to_append) {
	*_Out++ = '0';
	}

However, we need to completely understand the problem and all possible codepaths before developing a fix (i.e. we can't just clamp it to zero and call it a day). Note that the chars_format::general codepath further adjusts _Zeroes_to_append:

STL/stl/inc/format

Lines 2278 to 2289 in 926d458

	_Zeroes_to_append = _Extra_precision + _Precision - _Digits;
	// Leading zeroes are not significant if we used fixed point notation.
	if (_Exponent_start == _Result.ptr && _STD abs(_Value) < 1.0 && _Value != 0.0) {
	for (auto _It = _Buffer_start; _It < _Result.ptr; ++_It) {
	if (*_It == '0') {
	++_Zeroes_to_append;
	} else if (*_It != '.') {
	break;
	}
	}
	}

[cpplearner]

For the g specifier (which corresponds to chars_format::general, which corresponds to printf's %g), a precision of 0 should behave as if it were 1. It seems that we are missing this adjustment.

[StephanTLavavej]

Ah, great catch! We correctly handle that in <charconv>, so it seems that <format> is missing the same logic:

STL/stl/inc/charconv

Lines 2346 to 2356 in 926d458

	// C11 7.21.6.1 "The fprintf function"/5:
	// "A negative precision argument is taken as if the precision were omitted."
	// /8: "g,G [...] Let P equal the precision if nonzero, 6 if the precision is omitted,
	// or 1 if the precision is zero."
	// Performance note: It's possible to rewrite this for branchless codegen,
	// but profiling will be necessary to determine whether that's faster.
	if (_Precision < 0) {
	_Precision = 6;
	} else if (_Precision == 0) {
	_Precision = 1;