[diff.mods.to.headers] Cover type-generic macro suppression

[hubert-reinterpretcast]

No description provided.

[eisenwave]

This seems tutorial-like to be honest. The wording is clear that the tgmath.h macros don't exist in C++, and that's all that really needs to be said about compatibility differences.

Even if we want to have this new sentence, it seems like something that goes into a \begin{note} context.

I'm also not sure what you mean by "casts may be used instead". Are you casting the function itself to double(&)(double)? That isn't valid because of addressability restrictions. You could cast all the function arguments like sqrt((double)x), but that's possible in both languages, so it doesn't seem like it needs to be called out in [diff].

[jensmaurer]

I agree with the statement about casts; @hubert-reinterpretcast , please tell us what you mean by those "casts".

I do want to point out that Annex C is tutorial-like in nature (nothing new in that Annex, hopefully), so I'm not objecting to adding a clarifying sentence. However, I wonder whether an actual code example would convey the consequences more clearly.

#include <tgmath.h>

double x = (sqrt)(2.0f);   // calls sqrt(double) in C, but sqrt(float) in C++

Maybe just striking "casts may be used instead" is a way forward?

[jensmaurer]

@eisenwave , we didn't get a priority label here :-)

[eisenwave]

How come all major compilers accept the following:

#include <cmath>
void f() {
    (std::sqrt)(0.0);
}

Wouldn't this need to be an error because std::sqrt is an lvalue that refers to an overload set, and parenthesized expressions presumably need to have a type?

[eisenwave]

Hold on, I do think there is actually an incompatibility in some sense. I believe the wording currently mandates that if you include <math.h> you get the same stuff as <cmath>, minus things like std::lerp.

This means that the type of the expression sqrt(0.f) after including <math.h> in C++ is float, but in C, the type is double. In other words, the C++ behavior is always as if you had tgmath.h included, and there is no way to suppress it. I think that's what Hubert is aiming to address, even though the wording doesn't make it so obvious.

#include <math.h>
#include <type_traits>
// passes, but doesn't match the C behavior:
static_assert(std::is_same_v<decltype(sqrt(0.f)), float>);
// same:
static_assert(std::is_same_v<decltype((sqrt)(0.f)), float>);

So P2 because there is a missing incompatibility from [diff].