PR middle-end/18160

[Adam Nemet]

Richard Henderson writes:
> On Tue, Oct 26, 2004 at 12:51:33PM -0700, Adam Nemet wrote:
>> The question is should the front-end mark the address expression of a
>> (local) register variable with an error_mark_node or should the
>> expander deal with it and assign it to memory like it used to.
> 
> The former.  We will have generated a diagnostic, after all.

Well, it is a warning.  So GCC can still generate code in terms of
ignoring the user's request to assign the variable to a register.
This is how 3.4 used to work.  I think 3.4's behavior is more
appropriate with a warning than to remove the whole function call
because of error_mark_node.

I agree that the former is cleaner (maybe with an error) but can be
considered a regression.  I actually had a patch for the former too
but then I checked 3.4 which suggested the latter approach.

Adam