CWG3129 [lex.fcon] Clarify which floating-point-literals are valid

[eisenwave]

Reference (section label): [lex.fcon]

Issue description

Implementations diverge on whether the following declaration is well-formed (https://godbolt.org/z/hT6nza5hr):

float x = 1e100000000000000000000000000000000000000000000000.f;

[lex.fcon] paragraph 3 states:

If the scaled value is not in the range of representable values for its type, the program is ill-formed. Otherwise, the value of a floating-point-literal is the scaled value if representable, else the larger or smaller representable value nearest the scaled value, chosen in an implementation-defined manner.

Following the changes made in CWG2723, the range of representable values includes all real numbers if the infinities can be represented in float, meaning that x is infinity, and the program is well-formed. GCC and Clang are correct; MSVC and EDG are not. This should be clarified.

Note that GCC and Clang have been using the definition of "range of representable values" that dates back to at least C11, even though it did not exist in C++ until recently.

Suggested resolution

Add an example below [lex.fcon] paragraph 3:

std::float32_t x = 0.0f32;          // value 0 is exactly representable
std::float32_t y = 0.1f32;          // rounded to one of two values nearest to 0.1
std::float32_t z = 1e1000000000f32; // either greatest finite value or positive infinity

[eisenwave]

Vaguely related to CWG2752 (cplusplus/papers#1584) and cplusplus/papers#2136 (various excess-precision-related topics).

However, the problem in this issue is not about excess precision, but about whether the floating-point-literal is well-formed in the first place. It's actually irrelevant to the issue whether 0.1 is represented using excess range/precision; it will have to be rounded at some point when storing the value in y in the example.

[eisenwave]

It's worth noting that the C wording is of no help here because it is equally defective:

Each constant shall have a type and the value of a constant shall be in the range of representable
values for its type.

-- N3550 6.4.5 Constants

WG14 recently added an identical definition of "range of representable values" for floating-point types which breaks what C every compiler has been doing, just like C++. We basically copied and pasted C's defects.

Further semantics are specified in N3550 6.4.5.3 Floating literals, and are somewhat similar to C++.

[eisenwave]

I've talked this issue through with Matthias Kretz and am now of the opinion that GCC and Clang are correct and have been correct for a very long time. The C11 definition which CWG2723 imported is ancient, and the current wording is unambiguous.

However, it would be beneficial if we clarified that we really, actually, for real, like, totally, mean that 1e1000000000f32 is well-formed and equal to infinity.

[jensmaurer]

CWG3129