From mboxrd@z Thu Jan  1 00:00:00 1970
From: hjl@lucon.org (H.J. Lu)
To: jason@cygnus.com (Jason Merrill)
Cc: wilson@cygnus.com, egcs@cygnus.com, g++@cygnus.com
Subject: Re: A new alpha bug
Date: Tue, 16 Sep 1997 11:09:00 -0000
Message-id: <m0xB1Uq-0004ecC@ocean.lucon.org>
References: <u9k9ghfaa8.fsf@yorick.cygnus.com>
X-SW-Source: 1997-09/msg00628.html

> > BTW, both yours and Jim's patches sound like a kludge to me.
> > I don't understand why it takes TImode to expose this bug.
> > Is there a more subtle bug somehwhere else? Also how many
> > other places do we have to change to make it work right
> > for all cases? Is there a cleaner way to fix it?
> 
> I don't imagine anyplace else needs to be changed.  Bashing the RTL for a
> variable is not a common operation.
> 

I have seen places with

	TREE_ADDRESSABLE (init) = 1;
	DECL_RTL (init) = return_target;

or something like it. Should they be replaces by


	DECL_RTL (init) = return_target;
	TREE_ADDRESSABLE (init) = 0;
	mark_addressable (init);

Maybe we should make a function/macro out of it and use it:

#define MARK_TREE_ADDRESSABLE(node) \
	if (TREE_ADDRESSABLE ((node)) \
	  { \
	    TREE_ADDRESSABLE ((node)) = 0;
	    mark_addressable ((node));
	  }

	TREE_ADDRESSABLE (init) = 1;
	DECL_RTL (init) = return_target;
	MARK_TREE_ADDRESSABLE (init);

	
-- 
H.J. Lu (hjl@gnu.ai.mit.edu)