Re: Fix latent bug in add_referenced_vars

[Diego Novillo <dnovillo at redhat dot com>]

Paul Schlie wrote:

> Therefore the following comment may not be correct?:
>
>   /* Initializers of external variables are used solely by frontends
>      and serve no usefull value for us.  */
>
> (although I may be misunderstanding the patch's effect/intent?)
Hmm, no, that is not what the patch does.  Every symbol referenced by 
the program has to be registered with the optimizers so that we can put 
the symbol in SSA or FUD-chain form.  So, we do an initial scan of the 
program looking and registering symbols.  If you had a global variable

	int V;
	const int *GV = &V;

and then you had code like this:

	*GV = ...;

Constant propagation would propagate &V into *GV, yielding *&V, which 
ends up looking like:

	V = ...;

And now we have a completely different symbol on the LHS of the 
expression.  That symbol V needs to be registered with the optimizers, 
otherwise we ICE sooner or later.

The patch simply ignores non-const DECL_EXTERNALs because the optimizers 
cannot assume that the initial value of a DECL_EXTERNAL stays the same 
all the time.  So, by ignoring non-const DECL_EXTERNALs, we don't get 
into walking initializers for things that cgraph doesn't know about.


Diego.