All graphs for this bug:
  https://chromeperf.appspot.com/group_report?bug_id=839564

(For debugging:) Original alerts at time of bug-filing:
  https://chromeperf.appspot.com/group_report?sid=7010e6101da5a3fe4bb094f2a1293cf53613454574b7cbdcb1121ea513be78af


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

chromium-rel-mac11-air
chromium-rel-win7-dual

📍 Pinpoint job started.
https://pinpoint-dot-chromeperf.appspot.com/job/128d4573c40000

📍 Found significant differences after each of 2 commits.
https://pinpoint-dot-chromeperf.appspot.com/job/128d4573c40000

blink: Set up a cc::SoftwareImageDecodeCache for canvas image decodes. by khushalsagar@chromium.org
https://chromium.googlesource.com/chromium/src/+/49fc8a9c4a9eb46faa2258daea5bdbee48da7dcb

[LayoutNG] Stop removing trailing spaces in CollectInlines by kojii@chromium.org
https://chromium.googlesource.com/chromium/src/+/498ff64dc11978dcfdd294d466454ddb8cb0dd5c

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

Not mine, mine affects only when LayoutNG runtime flag is enabled.

khushalsagar@, can you please take a look?

Yup, its my patch. The regression in canvas_bench_many_images is a dupe of  issue 839763 . I'm looking at the tough_canvas_cases_rendering_throughput_put_get_image_data.html case.

Looks like the case has no images on the main thread in blink, its not even hitting the sw image cache that was added in the patch at all. There is a call to FlushRecording for canvas during each main frame ~2000 times though, so my guess is its just the overhead of hooking up the image provider during canvas raster that wasn't there earlier? In my local run that regressed the main frame time from ~153 to 157ms. I think we can ignore this much.

Just making sure I understand what's going on here.

There's a method which can be called 2000 times a frame which we made non-trivially slower, in cases where the new logic isn't being executed?

Extremely naively, this sounds like we should be able to get the overhead down.

Is my understanding incorrect? Is there any way we could lazily hook up the image provider?