Re: Patch for regression PR target/28490 (IA64 ICE)

[Jim Wilson <wilson at specifix dot com>]

Steve Ellcey wrote:
> I was talking to someone locally and he wondered why we didn't generate:
> 	addl r35 = @ltoff(cl_symbols#+8), r1
> 	;;
> 	ld8 r35 = [r35]

Good question.  I was looking at the psabi yesterday, and I noticed it 
said "@ltoff(expr)" and I was kind of wondering why we were going to so 
much trouble to avoid using an expr here.  I didn't see expr defined 
anywhere, but presumably anything that qualifies as a CONST is a valid 
expr here.  We probably just screwed up in the beginning, and never 
noticed the mistake.

> I was looking at how to keep the addend attached to the symbol but I am
> confused by the code in ia64_expand_move.  It seems like, if addend_lo
> is zero, we ignore addend_hi.  Is it not possible for addend_lo to be
> zero and addend_hi to be non-zero?  If so do we do the right thing?  I
> don't understand how.

Note that op1 is not set in this case, and hence op1 continues to be a 
CONST that contains the addend.  It is only when we set op1 to sym that 
we have a separate addend that we need to handle.

Apparently large addends are OK to use in @ltoff, but not small ones. 
Maybe this was a misguided optimization attempt, i.e. we extract out all 
of the small offsets thinking that we can fold them into following code, 
either by folding them into an address add, or an addressing mode.  The 
assumption here is that we would have common base addresses (the large 
offsets), and then the small offsets would all be relative to the base 
address.  This would reduce the number of got entries we need.  However, 
this would work better if we have reg+offset addresses which we don't. 
This trick is used on a number of other targets, but may not make sense 
on IA-64.  Maybe it was accidentally copied in from another target 
without realizing that it isn't a good idea here.  Also, this would work 
better if we had a really good auto-increment optimization pass, which 
we don't.  Unfortunately, all I can do at the moment is speculate.  I 
can comment on this in more detail when I get back.
--
Jim Wilson, GNU Tools Support, http://www.specifix.com