[Bug target/59644] [4.9 Regression] r206243 miscompiles Linux kernel

["jakub at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59644

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |ASSIGNED
   Last reconfirmed|                            |2014-01-06
           Assignee|unassigned at gcc dot gnu.org      |jakub at gcc dot gnu.org
     Ever confirmed|0                           |1

--- Comment #23 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 31755
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=31755&action=edit
gcc49-pr59644.patch

Untested fix.  For the code quality cases this patch regresses, perhaps it
would be nice to have some LRA target hook where we could finalize the stack
realign flags earlier, still during LRA, but late enough that nothing would
need to be spilled in the various vector modes, if LRA would be able to cope
with just reshuffling the elimination offsets because of that (and even better
if it could deal with frame_pointer_needed -> !frame_pointer_needed change at
that point).
Vlad, do you think there could be such place in LRA (surely it would need to be
done before the final lra_eliminate).  If we could even more accurately find
out if we really need aligned stack or not then, I'd hope it could improve code
quality.