PATCH for global register error message

[Andrew Haley]

> Date: Tue, 26 Oct 1999 02:36:46 -0600
> From: Jeffrey A Law <law@cygnus.com>
> 
>   In message < 19991022133445.4392.qmail@pasanda.cygnus.co.uk >you write:
>   > This patch moves the error checking for global register variables from
>   > varasm to regclass.  The reason for this is that the C++ front end may
>   > define some functions before parsing any source, thus making it
>   > impossible to declare any global register variables; better to defer
>   > the checking until registers are allocated.
> Is that still the case after all the work to move towards functions as
> trees for the C++ front end?   

I think so, yes.

> And I'd be curious about what function gets defined before we parse
> any source -- can you give a few details about that?

Sure.  C++'s init_decl_processing calls c_common_nodes_and_builtins,
which in turn defines a bunch of functions such as __builtin_constant_p.
A file which only contains a global register declaration is rejected
by the C++ compiler.

In any case, I would have thought that we should leave the prohibition
of global register variables until as late as possible.

Andrew.