Unreviewed patch

[Ulrich Weigand]

Zack Weinberg wrote:

> Could you provide a little more context?  The code that causes the
> ICE, for instance, and an analysis of why this is the right fix.

Unfortunately I don't have a test case that triggers the ICE on
either mainline or 'vanilla' 3.3; we saw it when using a compiler
based on 3.3 with this mainline patch backported:
http://gcc.gnu.org/ml/gcc-patches/2003-07/msg01701.html
(With that compiler, the ICE showed in one of the SPECint apps.)

On vanilla 3.3 the problem doesn't show because in that particular
case the ext_dependant giv isn't recognized (without the patch above),
on CVS head it doesn't show because the old unroller isn't really
used any more.

I just thought it would be best to fix it as the fix appears
to be rather obvious, and the problem can conceivably show up 
(in other cases) even in a vanilla 3.3 or else in CVS head when 
using -fold-unroll-loops.

As to why this is the right fix, if you look at all the other
places in unroll.c (or loop.c) where the BIV initial value
is used in some computation involving a GIV, it is *always*
passed through extend_value_for_giv.  This was initially
introduced by rth's patch
http://gcc.gnu.org/ml/gcc-patches/2000-09/msg00257.html

It looks like this was just forgotten at that one location,
and simply unnoticed until now because it's hard to trigger.

Bye,
Ulrich

-- 
  Dr. Ulrich Weigand
  weigand@informatik.uni-erlangen.de