[Bug c/114526] ISO C does not prohibit extensions: fix misconception.

["harald at gigawatt dot nl" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114526

--- Comment #16 from Harald van Dijk <harald at gigawatt dot nl> ---
(In reply to Joseph S. Myers from comment #15)
> In the cases where there is no statement either way, the behavior is
> undefined as a property of the translation unit (not just of the execution):
> it is not defined whether such a conversion may occur in a translation unit,

This is still not stated anywhere in the standard though.

> Being undefined
> through omission of definition has, as per clause 4, not difference in
> meaning or emphasis from being explicitly undefined.

Of course, but if the standard had explicitly stated that conversion between
function pointers and object pointers was undefined, it might be phrased in a
way that applies even to dead code. If you are relying on being undefined by
omission, you have to be really sure the behaviour is not defined *anywhere*,
including by general rules about dead code.

I will grant that the standard never explicitly says dead code is not executed
and has no effect, but if this is in dispute, we have a bigger problem.

> I'd suggest working with the Undefined Behavior Study Group on making it
> more explicit for each instance of undefined behavior whether it is a
> property of the program or of an execution thereof, but if any case seems
> particularly unclear, filing an issue once the new C standard issue tracker
> is up and running would probably be reasonable (but it seems likely that
> such issues would be referred to the UB study group to recommend a
> resolution just as floating-point issues would likely be referred to the CFP
> group).

Considering my stance is that WG14 have repeatedly and consistently stated what
the rules are, I see this as a waste of their time.

> It's *not* the case that the same rules apply everywhere, because there are
> two different kinds of UB depending on whether what's undefined is a
> property of the program or an execution thereof. Division by zero is
> obviously UB as a property of an execution, because whether a value is zero
> is a property of the execution.

Considering this example of 1/0 has been the subject of two separate DRs that I
referenced, I have to say it is not obvious from the standard itself. Keeping
in mind that the operands are constants and implementations are required to be
capable of constant expression evaluation in some contexts, a hypothetical
standard that permitted, or even required, this to be evaluated at translation
time (with undefined behaviour) even in otherwise dead code would make perfect
sense. But that is not the C standard we have, at least not the official
interpretation of it.

> Different types for the same identifier with
> external linkage in different translation units is obviously UB as a
> property of the program (and not widely diagnosed without LTO), as the whole
> concept of an identifier corresponding to an object with a particular value
> depends on a globally consistent notion of its type and the UB is about
> presence of declarations rather than a particular path of execution.

Yes, because a program that does not reference these identifiers still violates
the rule that specifies they must have compatible type. This means that there
is no execution of the program that avoids UB.

But in my program, there is no rule that is violated. Perhaps the rule that you
describe in your comment, that no program may contain any unsupported
conversion anywhere, regardless of whether the conversion is ever performed,
should exist, but it is simply not the case that there is such a rule to be
found anywhere in the standard.

One additional comment, though:

The fact that conversions between function pointers and object pointers are
rejected under -pedantic-errors mean that 'gcc -std=c99 -pedantic-errors'
cannot  be used as the implementation for POSIX's c99 utility, as POSIX's c99
utility is required to conform to the C99 standard, and simultaneously, permit
conversions between function pointers and object pointers (at least in some
cases). (Adjust for later versions as needed.) This is unfortunate, and
regardless of whether the C standard allows such programs to be rejected, can
we agree that the C standard also allows them to be accepted, and POSIX
requires them to be accepted? Is that not already sufficient reason to
reconsider?