Re: Static variables in template functions

[Mark Mitchell <mark at codesourcery dot com>]

Jason Merrill wrote:

> I was only suggesting that we don't warn for const, non-addressed
> variables, which I believe was the case in question here.  For others I
> would leave the warning.

What I proposed was that on affected platforms we make all 
initializations (independently of const-ness/addressed-ness) dynamic, 
and thereby eliminate the warning, as the compiler would then be 
generating correct code.

We could in theory refine that to do dynamic initializations only for 
variables that are either non-const or addressed, and still do static 
initializations for the remainder.  However, I don't think we can easily 
implement that refinement, as we decide whether to do a dynamic 
initialization before we have information about addressed-ness.

I'm not sure if we're talking past each other or not, so I'll try to 
make a clearer statement.

I'm proposing that on platforms without weak symbols we perform all 
initializations of function-scoped static data dynamically, thereby 
implementing the semantics required by the standard, even though it is 
sub-optimal in terms of performance.

If I understand correctly, your position is that we should instead, 
continue to have the current behavior (generate incorrect code, with a 
warning), except that we should avoid issuing the warning in the case of 
a const, non-addressed variable, as in that case the code we generate is 
in fact correct.

Do I have it right?

Thanks,

--
Mark Mitchell
CodeSourcery, LLC
mark@codesourcery.com
(916) 791-8304