UB when converting a float-point value to a type that can't represent its value

[shafik]

clang used to flag this as undefined behavior:

#include <limits>

void foo() {
 constexpr double d2 = std::numeric_limits<double>::max();         
 constexpr float f = d2;
}

back in clang-8 but it has not since: https://godbolt.org/z/5K1da874d

Based on conv.double p1 I think this is still UB.

[shafik]

CC @zygoloid

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: Shafik Yaghmour (shafik)

clang used to flag this as undefined behavior:

#include <limits>

void foo() {
 constexpr double d2 = std::numeric_limits<double>::max();         
 constexpr float f = d2;
}

back in clang-8 but it has not since: https://godbolt.org/z/5K1da874d

Based on [conv.double p1](https://timsong-cpp.github.io/cppwp/n4659/conv.double#1) I think this is still UB.

[frederick-vs-ja]

I reported CWG2723 for this. It seems clear that this isn't UB now as long as float can represent positive infinity.

[zygoloid]

The range of representable values for floating-point types in Clang includes +/-infinity, so there is no case where we convert a floating-point value to a floating-point type that can't represent that value. Hence there's no UB here.

As I recall, this was changed at around the same time the LLVM uitofp and sitofp instructions were changed to not be documented as producing poison on integer inputs that exceed the finite range of the destination floating point type (eg, converting (unsigned __int128)-1 to float or converting (unsigned long long)-1 to _Float16).