------- Comment #13 from ubizjak at gmail dot com 2010-06-10 14:59 -------
(In reply to comment #12)
> (In reply to comment #11)
>
> > > > I am not sure this is correct. The code prior to revision 160394 was
> > > > written in such a way to support X86_TUNE_OPT_AGU. We may have missed
> > > > some cases. But it is mostly correct. I'd like to know what specific
> > > > cases revision 160394 tried to fix. Thanks.
> > >
> > > The patch fixes general problem, where we simply output lea from
> > > add{si,di}
> > > instruction. This is not correct, since lea does not clobber flags, so we
> > > should split add RTX to lea RTX to model this fact.
> > >
> >
> > ADD is always faster than LEA for adding a register. However
> > there is a special case on Atom where ADD should be avoided.
> > It is true that LEA doesn't touch flags and we used it instead
> > of ADD which touches flags. It is an optimization specific for
> > Atom. You can think it as a special ADD for Atom and pretend
> > it clobbers flags. We shouldn't model it as a real LEA since
> > it is used to implement a faster ADD in this special case.
> >
> > Uros, I'd like to revert revision 160394 to fix bootstrap on
> > Atom. I will work with you to resolve issues you discovered
> > without affecting Atom performance. Thanks.
>
> OK, please go ahead. The correct fix is not as simple as I have thought.
Please revert only this part:
* config/i386/i386.md (*add<mode>_1): Remove alternative 2.
<TYPE_LEA>: Split instruction.
<default>: Remove alternative 2 handling.
(*addsi_1_zext) <TYPE_LEA>: Split instruction.
(add lea splitter): Generate SImode lea for mode sizes <= SImode.
(add_zext lea splitter): Use ix86_lea_for_add_ok in insn predicate.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44470
