I have two experimental CLs for this bug:
1. https://chromium-review.googlesource.com/c/chromium/src/+/1412805: [CompositeAfterPaint] Create PaintOffsetTranslation for all stacked elements
2. https://chromium-review.googlesource.com/c/chromium/src/+/1410165 [CompositeAfterPaint] Create PaintOffsetTranslation for compositing

1 creates more PaintOffsetTranslations than necessary including many objects that won't be composited.

Cluster telemetry result:
https://ct.skia.org/results/cluster-telemetry/tasks/chromium_perf_runs/wangxianzhu-20190115192807/html/index.html
20% regression of number of paint operations
10% regression of paint operation memory usage
1% *progression* of raster time
2.4% *progression* of pixel recorded and pixel rastered
Please ignore record time results because of bug 922642.

2 only creates PaintOffsetTranslation for objects having known direct compositing reason but misses PaintOffsetTranslation for layers that we decide to composite later in CAP.

Cluster telemetry result: https://ct.skia.org/results/cluster-telemetry/tasks/chromium_perf_runs/wangxianzhu-20190116170658/html/index.html
No performance regressions.

I'm inclined to 1. We can follow-up to find some heuristics to reduce the unnecessary PaintOffsetTranslations or keep them but improve their performance.