Subzero: Recognize single-block loops during loop depth analysis.

Subzero: Recognize single-block loops during loop depth analysis.

BUG= none
R=jpp@chromium.org

Committed: https://gerrit.chromium.org/gerrit/gitweb?p=native_client/pnacl-subzero.git;a=commit;h=f49b2396547a7cf3b8a9b5198f62cbc545c2572c

[Jim Stichnoth, 2015-11-09 17:20:58]

Description was changed from

==========
Subzero: Recognize single-block loops during loop depth analysis.

BUG= none
==========

to

==========
Subzero: Recognize single-block loops during loop depth analysis.

BUG= none
==========

[Jim Stichnoth, 2015-11-09 17:20:59]

stichnot@chromium.org changed reviewers:
+ jpp@chromium.org, kschimpf@google.com, sehr@chromium.org

[Jim Stichnoth, 2015-11-09 17:23:35]

Shockingly, this appears to give a 2.5% geomean improvement, including a 20+%
improvement in bzip2.  (I don't know how this bzip2 gain interacts with the gain
from the CL that allows use of high 8-bit registers.)

[John, 2015-11-09 18:04:16]

lgtm

If I understand this correctly, it is effectively finding teenie tiny loops,
which will then increase the likelyhood of the variables used in them to be
register allocated, right?

[Jim Stichnoth, 2015-11-09 18:10:56]

On 2015/11/09 18:04:16, John wrote:
> lgtm
> 
> If I understand this correctly, it is effectively finding teenie tiny loops,
> which will then increase the likelyhood of the variables used in them to be
> register allocated, right?

That's right, it ultimately gives a higher weight delta to Variable references
inside that tiny loop.

[Jim Stichnoth, 2015-11-09 18:46:17]

Committed patchset #1 (id:1) manually as
f49b2396547a7cf3b8a9b5198f62cbc545c2572c (presubmit successful).