All graphs for this bug:
  https://chromeperf.appspot.com/group_report?bug_id=820723

(For debugging:) Original alerts at time of bug-filing:
  https://chromeperf.appspot.com/group_report?sid=7e853dfc47021552e21080ffe50876d9af0c0c186e3f1458818d66cf06782d01


Bot(s) for this bug's original alert(s):

chromium-rel-win7-gpu-nvidia

📍 Pinpoint job started.
https://pinpoint-dot-chromeperf.appspot.com/job/1685c6da440000

📍 Found a significant difference after 1 commit.
https://pinpoint-dot-chromeperf.appspot.com/job/1685c6da440000

Roll src/third_party/angle/ 19fa1c6f8..e8a93c6ed (3 commits) by angle-chromium-autoroll@skia-buildbots.google.com.iam.gserviceaccount.com
https://chromium.googlesource.com/chromium/src/+/bdb3104733f57f58963321f20e079bea4eab8cb4

Understanding performance regressions:
  http://g.co/ChromePerformanceRegressions

James, I think this is probably related to your buffer binding changes. Can you investigate?

Also filed  issue 820977  about the pinpoint job not recursing into third_party/angle.

Looks like the major reason for the slowdown in Bindings_gl_100_objects_allocated_every_iteration is that I added the three loops here: https://cs.chromium.org/chromium/src/third_party/angle/src/libANGLE/State.cpp?l=1448&rcl=64e5ed2abd8a4cc77d3647e4b245db57a8c42ae5

Before, we weren't checking those binding points when detaching buffers.

For the other tests, it's simply the overhead of doing more refcounting, which these microbenchmarks exercise a lot. For WebGL it's necessary for spec compliance. If you think it's necessary I can add some checks to see whether it's a WebGL context or not before doing the refcounting.

Maybe you could add an early out if the ref count == 0? It should be this in many cases where the app deletes an unbound buffer.

The following revision refers to this bug:
  https://chromium.googlesource.com/angle/angle/+/4d9d483936e6e845f24e851d1cf288bb37bfaf1c

commit 4d9d483936e6e845f24e851d1cf288bb37bfaf1c
Author: James Darpinian <jdarpinian@chromium.org>
Date: Tue Mar 13 22:34:45 2018

Optimize buffer deletion by returning early if buffer is not bound.

This improves angle_perftests
bindings_gl_100_objects_allocated_every_iteration by over 30%.

Bug:  820723 
Change-Id: Icfaa0f54f7c61db6475948e3a95296f37ab4f7dc
Reviewed-on: https://chromium-review.googlesource.com/961302
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Commit-Queue: Jamie Madill <jmadill@chromium.org>

[modify] https://crrev.com/4d9d483936e6e845f24e851d1cf288bb37bfaf1c/src/libANGLE/Buffer.cpp
[modify] https://crrev.com/4d9d483936e6e845f24e851d1cf288bb37bfaf1c/src/libANGLE/State.cpp
[modify] https://crrev.com/4d9d483936e6e845f24e851d1cf288bb37bfaf1c/src/libANGLE/Context.cpp
[modify] https://crrev.com/4d9d483936e6e845f24e851d1cf288bb37bfaf1c/src/libANGLE/State.h
[modify] https://crrev.com/4d9d483936e6e845f24e851d1cf288bb37bfaf1c/src/libANGLE/Buffer.h
[modify] https://crrev.com/4d9d483936e6e845f24e851d1cf288bb37bfaf1c/src/libANGLE/Context.h

Can confirm graph has recovered: https://chromeperf.appspot.com/group_report?keys=agxzfmNocm9tZXBlcmZyFAsSB0Fub21hbHkYgICQ0uOw5AgM