Sporadic failure of P0645R10_text_formatting_formatting

[AlexGuteniev]

This run:
https://dev.azure.com/vclibs/STL/_build/results?buildId=9974&view=logs&j=e5928d94-6045-5a4b-1479-6369daf1c6e1&t=e774f207-3822-5f8e-5e8b-2cdb9d7ebe0b

The subsequent run succeeded, though the change between the runs was in unrelated area, and except of one variable un-shadowing that does not affect anything, it was whitespace/comments change: 2b43c89

[StephanTLavavej]

The error was:

Assertion failed: format("{:.2000a}", 1.0) == expected, file C:\a\1\s\tests\std\tests\P0645R10_text_formatting_formatting\test.cpp, line 836

I don't understand what happened here. Perhaps we should update the test with additional logging, to print the result of format() when it's unexpected here?

[StephanTLavavej]

Happened again in: https://dev.azure.com/vclibs/STL/_build/results?buildId=10119&view=ms.vss-test-web.build-test-results-tab&runId=1184952&resultId=102554&paneView=debug

Slightly different line/expression this time (note line 820 instead of line 836):

Assertion failed: format("{:.2000a}", value) == expected, file C:\a\1\s\tests\std\tests\P0645R10_text_formatting_formatting\test.cpp, line 820

[StephanTLavavej]

#2494 fortunately printed the actual string before the expected string (the latter was lost due to log truncation, #2557), statementreply in #2499 (comment) noticed what was wrong with that string indicating a bug in the "huge precision, clamp it and then append zeros later" codepath, I was able to repro it locally and find the root cause in

STL/stl/inc/format

Lines 2510 to 2546 in afe0800

	auto _Exponent_start = _Result.ptr;
	auto _Radix_point = _Result.ptr;
	auto _Integral_end = _Result.ptr;
	auto _Zeroes_to_append = 0;
	auto _Separators = 0;
	string _Groups;
	if (_Is_finite) {
	if (_Specs._Alt || _Specs._Localized) {
	for (auto _It = _Buffer_start; _It < _Result.ptr; ++_It) {
	if (*_It == '.') {
	_Radix_point = _It;
	} else if (*_It == _Exponent) {
	_Exponent_start = _It;
	}
	}
	_Integral_end = (_STD min)(_Radix_point, _Exponent_start);
	if (_Specs._Alt && _Radix_point == _Result.ptr) {
	// TRANSITION, decimal point may be wider
	++_Width;
	_Append_decimal = true;
	}
	if (_Specs._Localized) {
	_Groups = _STD use_facet<numpunct<_CharT>>(_Locale._Get()).grouping();
	_Separators = _Count_separators(static_cast<size_t>(_Integral_end - _Buffer_start), _Groups);
	}
	}
	switch (_Format) {
	case chars_format::hex:
	case chars_format::scientific:
	if (_Extra_precision != 0) {
	while (*_Exponent_start != _Exponent) { // Trailing zeroes are in front of the exponent
	--_Exponent_start;
	}

where while (*_Exponent_start != _Exponent) initially reads _Result.ptr (this is past-the-end of what to_chars() wrote), and Charlie Barto (@barcharcraz) has a fix. 🎉

This sporadically failed because the garbage character being read is almost always not the exponent character ('p' for hex, 'e' for scientific), but if it is, the code gets confused about where the exponent starts, leading to catastrophe shortly afterwards.