(j3.2006) Some colleagues wish they could count on these working

[Malcolm Cohen]

>Does anybody not effectively give constants the SAVE attribute?

Yes.

>Does anybody actually kill pointers according to 12.5.2.4p11 if the
>target is a constant or has the SAVE attribute, at least not in "check
>for undefined pointers" mode?

Yes.

Whether the value of a constant expression is materialised at compile time, link 
time, or run time, for us depends on the characteristics of the expression.  For 
example, scalar real/integer/logical are highly likely to be materialised on the 
stack, even with no optimisation or checking enabled, and this is likely to end 
up with the pointer pointing to random junk later on.

>It would be helpful if literal and named constants effectively had the
>SAVE attribute,

This is more than adequately catered for by using a variable.  Variables are 
also uniquely determined (one per name) whereas whether different constant 
expressions share some or all of their storage is an implementation detail.

> and pointers whose targets ultimately are effective
>arguments with the SAVE attribute do not become undefined as specified
>in 12.5.2.4p11.

Actually pointers are never pointer-associated with something that does not have 
the TARGET attribute, though they can be pointer-associated with a TARGET that 
is argument-associated with something that does not have the TARGET attribute. 
It is impossible to tell the difference in a standard-conforming program, but 
obviously different argument-passing technologies will lead to different 
behaviours of non-conforming programs - and if you are lucky, an explanatory 
error message.

I really do not think you want to give everything the TARGET attribute.  Those 
who "wish they could count on these working" would seem to be looking at the 
world through C-coloured spectacles.

Cheers,
-- 
................................Malcolm Cohen, Nihon NAG, Tokyo.