finit-local-zero and default initialization

[Brooks Moses]

Asher Langton wrote:
> I'm planning to add support for -finit-local-zero (PR20441), as well
> as for arbitrary default initializations specified at compile-time.
> One request from within LLNL has been for a way to control the default
> initialization values at runtime, via some simple library routines
> (set_default_integer, get_default_integer, etc.).  This would be very
> useful for debugging large codes without recompiling an entire
> project.  Any comments, suggestions, or objections?

My off-the-cuff reaction is that the interface for this would be a very 
good place to start using intrinsic modules for encapsulating 
non-standard extensions.

Also, making this thread-safe is likely to be a pleasant little 
nightmare, but that's your fun, not mine.  :)

I do think it would be good to have the runtime stuff _only_ enabled if 
it's explictly enabled at runtime; a program compiled to initialize 
everything to a constant value such as zero or NaN should not include 
all the machinery necessary to check (in a thread-safe way) for the 
current initialization value every time a variable is initialized.

- Brooks