Francis Glassborow <francis.glassborow@btinternet.com> writes:

> On 01/12/2020 11:39, Tim Rentsch wrote:
>
>> Keith Thompson <Keith.S.Thompson+u@gmail.com> writes:
>>
>>> Ian Abbott <ijabbott63@gmail.com> writes:
>>>
>>>> A question (not posted by myself) from
>>>> https://stackoverflow.com/questions/65077630 :
>>>>
>>>> Consider the following:
>>>>
>>>> union { int i; char c; } x = {0};
>>>> x.c = x.i;
>>>>
>>>> Does the assignment x.c = x.i result in undefined behavior?
>>>>
>>>> C18 6.15.16.1/3 says:
>>>>
>>>> | If the value being stored in an object is read from another
>>>> | object that overlaps in any way the storage of the first object,
>>>> | then the overlap shall be exact and the two objects shall have
>>>> | qualified or unqualified versions of a compatible type;
>>>> | otherwise, the behavior is undefined.
>>>>
>>>> The objects x.c and x.i overlap, but have incompatible types, so on
>>>> first glance it appears to be UB.
>>>
>>> On first glance, yes, but I think this passage needs to be updated
>>> to reflect its intent.
>>>
>>> The RHS of an assignment is not an lvalue. If it starts out as an
>>> lvalue, then lvalue conversion is applied. Logically the value is
>>> retrieved *and then* copied into the destination object.
>>>
>>> A simple case like this is not likely to cause problems (in the
>>> absence of agressive optimization), but we can construct more
>>> problematic cases where the object being copied is arbitrarily large
>>> and the overlap might not be detectable at compile time.
>>>
>>> I've written an answer:
>>> https://stackoverflow.com/a/65080498/827263
>>
>> Let me start with where we agree. I agree that 6.5.16.1 p3 deserves
>> some clarification. After that however my reading and yours reach
>> different conclusions.
>>
>> First I think the passage as written clearly conveys the meaning
>> intended in this case, that this assignment is undefined behavior.
>> The value being stored was read out of x.i, and is being stored
>> into x.c. The types of those two objects don't mesh, and so the
>> assignment is undefined behavior. This assignment is about the
>> clearest possible case that would violate 6.5.16.1 p3, and the
>> passage as written conveys that, IMO with no room for argument.
>>
>> Second I think whether the RHS is an lvalue has no bearing on the
>> question. The cited paragraph does not mention anything about
>> lvalues; it speaks only of "the value being stored". Consider a
>> slight variation:
>>
>> x.c = (printf( "hello world\n" ), x.i);
>>
>> The RHS of this assignment is not an lvalue. But the /value/
>> being stored was read from x.i. As I read the Standard this
>> assignment too is undefined behavior, and IMO the Standard does
>> convey that intention.
>>
>> Here is another variation:
>>
>> x.c = 0 ? printf( "hello world\n" ) : x.i;
>>
>> Once again the value being stored was read from x.i, and again
>> as I read the Standard this assignment too is undefined behavior,
>> and IMO the Standard does convey that intention.

(see note given below)

>> Now let's look at an example you give in your stackoverflow answer.
>> Quoting a whole paragraph:
>>
>> The passage says that the value "is read from another
>> object". That's ambiguous. Must name of the object be the
>> entire RHS expression, or can it be just a subexpression?
>> If the latter, then x.c = x.i + 1; would have undefined
>> behavior, which in my opinion would be absurd.

My apologies for taking so long to respond to your comments.

> if
> x.c = x.i + 1;
>
> has defined behaviour then
>
> x.c = x.i + 0;
>
> also has defined behaviour.

Yes. It does, and it should.

> So IMO, consistency requires that we tighten the requirement so that
>
> x.c = any expression that uses x.i;
>
> has undefined behaviour.

Surely that is not the intent. It is only when the value being
stored is _read directly_ from an overlapping object, and is
not instead _formed as the result of a computation using_ an
overlapping object, that undefined behavior occurs.

> However I still have reservations because:
>
> {int const i = x.i; x.c = i;}
>
> is surely OK. However any optimising compiler will concatenate that to
>
> x.c = x.i;
>
> Note the braces limit the scope of i.

What an optimizing compiler might do has no effect on the program's
semantics, which is defined in terms of an abstract machine where
no optimizations occur. Each compiler has a responsibility to
ensure its optimizer faithfully preserves the program semantics
that the Standard requires, which in this case has defined behavior
because the value being stored is read out of 'i' and not out of
'x.i'.

(Note for the "see below" remark: I'm willing to be convinced that
the example

x.c = 0 ? printf( "hello world\n" ) : x.i;

has defined behavior, because the value being stored is the result
of a ?: operator, and thus not just a direct read out of x.i. This
argument can be seen more clearly if we consider

x.c = 0 ? 0ULL : x.i;

because the value of the RHS clearly is not the same as what is
read out of x.i, which is an int. Similarly

x.c = (unsigned long long) x.i;

has defined behavior, because the value being stored is the result
of a conversion operation, not the value read from x.i.)