On Tue, Nov 15, 2016 at 6:48 AM, Segher Boessenkool
<seg...@kernel.crashing.org> wrote:
> If we use ABI_V4 and we have a big stack frame, we end the epilogue
> with a "mr 1,11" (or similar) instruction. This instruction however
> has no dependencies on the earlier restores from stack (done via r11),
> so sched2 can end up reordering the insns, which is bad because we
> have no red zone so that you then restore from stack that is already
> deallocated.
>
> This fixes it by making that restore depend on the memory accesses.
>
> Tested on powerpc64-linux {-m32,-m64}; is this okay for trunk? Do we
> want a testcase for this?
>
>
> Segher
>
>
> 2016-11-15 Segher Boessenkool <seg...@kernel.crashing.org>
>
> * config/rs6000/rs6000.c (rs6000_emit_stack_reset): Emit the
> stack_restore_tie insn instead of stack_tie, for the SVR4 and
> SPE ABIs.
> * config/rs6000/rs6000.md (stack_restore_tie): New define_insn.
Okay.
A similar change may be necessary for other uses of rs6000_stack_tie
in the prologue.
It would be good to comment somewhere that stack_restore_tie is a
superset of rs6000_stack_tie.
Thanks, David
