Switch scheduler to iterative floating control placement.

Switch scheduler to iterative floating control placement.

R=jarin@chromium.org

Committed: https://code.google.com/p/v8/source/detail?r=24968

[Michael Starzinger, 2014-10-28 17:34:44]

Note that this CL contains several cheap hacks (should mostly be marked with
TODOs) but is small step towards a fully iterative scheduler. Let me know if you
think the amount of hackery in this CL is too damn high to be landed.

[Jarin, 2014-10-29 09:22:33]

lgtm. Very cool.

https://codereview.chromium.org/685773002/diff/1/src/compiler/scheduler.cc
File src/compiler/scheduler.cc (right):

https://codereview.chromium.org/685773002/diff/1/src/compiler/scheduler.cc#ne...
src/compiler/scheduler.cc:120: DCHECK_EQ(Scheduler::kCoupled, data->placement_);
We might want to DCHECK that the 'placement' parameter is what it should be
(kFixed in this case, I believe).

https://codereview.chromium.org/685773002/diff/1/src/compiler/scheduler.cc#ne...
src/compiler/scheduler.cc:127: ++(GetData(control)->unscheduled_count_);
Is this here to keep the unscheduled_count above zero? (It seems that once the
node is kFixed, we never decrement its unscheduled use count.)

https://codereview.chromium.org/685773002/diff/1/src/compiler/scheduler.cc#ne...
src/compiler/scheduler.cc:276: is_dom &=
!scheduler_->GetData(node->InputAt(i))->is_floating_control_;
I really do not like mixing bitwise and bool operators. Could you rewrite this
to use && or ?: operator?

[Michael Starzinger, 2014-10-29 10:02:40]

Addressed comments. Thanks for review. Landing.

https://codereview.chromium.org/685773002/diff/1/src/compiler/scheduler.cc
File src/compiler/scheduler.cc (right):

https://codereview.chromium.org/685773002/diff/1/src/compiler/scheduler.cc#ne...
src/compiler/scheduler.cc:120: DCHECK_EQ(Scheduler::kCoupled, data->placement_);
On 2014/10/29 09:22:32, jarin wrote:
> We might want to DCHECK that the 'placement' parameter is what it should be
> (kFixed in this case, I believe).

Done. Here and also for the default case below.

https://codereview.chromium.org/685773002/diff/1/src/compiler/scheduler.cc#ne...
src/compiler/scheduler.cc:127: ++(GetData(control)->unscheduled_count_);
On 2014/10/29 09:22:32, jarin wrote:
> Is this here to keep the unscheduled_count above zero? (It seems that once the
> node is kFixed, we never decrement its unscheduled use count.)

As discussed offline: Yes, this is to keep the unscheduled_count above zero.
Unfortunately the control node is not yet marked as kFixed when we fix the phis.
I will work on cleaning up these TODOs next.

https://codereview.chromium.org/685773002/diff/1/src/compiler/scheduler.cc#ne...
src/compiler/scheduler.cc:276: is_dom &=
!scheduler_->GetData(node->InputAt(i))->is_floating_control_;
On 2014/10/29 09:22:32, jarin wrote:
> I really do not like mixing bitwise and bool operators. Could you rewrite this
> to use && or ?: operator?

Done.

[Michael Starzinger, 2014-10-29 10:17:59]

Committed patchset #2 (id:20001) manually as 24968 (presubmit successful).