[PATCH] Fix PR69274, 435.gromacs performance regression due to RA

[Vladimir Makarov]

On 02/08/2016 12:38 PM, Michael Matz wrote:
>
> I think the patch makes perfect sense.  ira_setup_alts should have no
> observable behaviour from the outside, except the returned value of merged
> acceptable alternatives.  Certainly it has no business to fiddle with
> recog_data.  It only does the swapping to merge alternatives, and
> accidentaly left them swapped; the merging could have been implemented
> without swapping recog_data.operands, and then the whole issue wouldn't
> have occurred (and addr-sel-1.c wouldn't have been added because it still
> would be "broken").
>
Sorry, I was confused by Richard's message thinking that his patch 
actually exchanges the operands.  I think we have some expression 
shaping optimizations and exchanging operands probably rejects the 
optimization effect.  With this point of view ira-costs.c should not 
exchange operands.   So the patch is not bogus as Richard wrote but 
perfectly legitimate.