<xlocnum>: "multiply by power of 10" logic is imprecise

[statementreply]

Describe the bug

_Stodx_v2 and _Stofx_v2 contains:

STL/stl/inc/xlocnum

Lines 58 to 60 in b1fbbdc

	if (_Pten != 0) {
	_Val *= _CSTD pow(10.0, static_cast<double>(_Pten));
	}

Like almost all of our pre-charconv floating-point formatting, I believe that this is incorrect - dividing by a power of 10 isn't an exact operation.

I believe that this code path is only activated when the integral part of the significand is longer than 768 digits (which overflows without a negative exponent).

Command-line test case

D:\Temp>type long_num_get.cpp

#include <cassert>
#include <sstream>
#include <string>

using namespace std;

int main() {
    istringstream iss{"1" + string(2000, '0') + "e-2000"};
    double x;
    assert(iss >> x);
    assert(x == 1.0);
    return 0;
}

D:\Temp>cl /EHsc /W4 /WX long_num_get.cpp
Microsoft (R) C/C++ Optimizing Compiler Version 19.28.29812 for x64
Copyright (C) Microsoft Corporation.  All rights reserved.

long_num_get.cpp
Microsoft (R) Incremental Linker Version 14.28.29812.0
Copyright (C) Microsoft Corporation.  All rights reserved.

/out:long_num_get.exe
long_num_get.obj

D:\Temp>.\long_num_get.exe
Assertion failed: iss >> x, file long_num_get.cpp, line 10

STL version

Microsoft Visual Studio Community 2019 Preview
版本 16.9.0 Preview 3.0

[CaseyCarter]

I suspect we could grep the sources for occurrences of \b10\b and every one would be a floating-point bug.

[StephanTLavavej]

Thanks for the example - I had noted this line in #264 but thought it was dead.

[frederick-vs-ja]

Perhaps #1792 is quite similar to this.