All graphs for this bug:
  https://chromeperf.appspot.com/group_report?bug_id=798335

(For debugging:) Original alerts at time of bug-filing:
  https://chromeperf.appspot.com/group_report?sid=54664dce795bbba5b900a13ab26e691cd1563395935addccd1572d7c72117523


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

android-nexus5X
chromium-rel-win10

📍 Pinpoint job started.
https://pinpoint-dot-chromeperf.appspot.com/job/14e41ced040000

📍 Found a significant difference after 1 commit.
https://pinpoint-dot-chromeperf.appspot.com/job/14e41ced040000

Do not perform per-character font fallback for PUA codepoints
By drott@chromium.org · Thu Dec 21 12:07:14 2017
chromium @ 725d3699e90b2ad35a7b9d139b890a55457d9a80

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

CC bungeman@, I can't explain why this change would lead to Skia reporting a higher "Effective Size" memory usage. We're not doing system fallback for PUA area characters, but instead use the last resort font. This change is intentional to avoid looking up semantically not suitable system fonts for PUA codepoints as used for example in icon fonts.

I think this just means we're using the cache better (through fewer evictions). There is quite a spread of values in the effective_size both before and after this change. The top end values don't seem too different (before vs after), but the low end went up, which skews the average up.

I don't know enough about the 'effective_size' number or what it represents. If it's just a number measured right after the test is run then it's probably dominated by the largest thing most recently evicted from the font caches. If we're not putting large things in there in the first place we'll end up with better cache use.

In other words, if this is taking into account cache usage, this may actually represent an improvement. However, I'm not entirely sure what this is measuring at the moment.

Thanks, Ben. I'll mark this as WontFix, the change of not doing fallback for Unicode Private Use Area codepoints is intentional, side effects of different Skia cache utilization are outside of what I can control with CL.