62349 2011-06-08 20:24:47 -0700 REGRESSION(88260): 10-50% performance regression across many page cyclers 2011-06-13 12:29:47 -0700 1 1 1 Unclassified WebKit New Bugs 528+ (Nightly build) Unspecified Unspecified RESOLVED FIXED P2 Normal --- 1 jamesr jamesr aroben fishd ggaren hyatt jamesr mitz msaboff tonyg webkit.review.bot oldest_to_newest 417774 0 jamesr 2011-06-08 20:24:47 -0700 REGRESSION(88260): 10-50% performance regression across many page cyclers 417775 1 96536 jamesr 2011-06-08 20:27:38 -0700 Created attachment 96536 Patch 417777 2 jamesr 2011-06-08 20:28:46 -0700 Patch for consideration. Please let me know if you would prefer to retain the old thresholds for some platforms. We definitely want to change them for Chromium. 417814 3 96536 fishd 2011-06-08 21:59:30 -0700 Comment on attachment 96536 Patch R=me 417855 4 96536 webkit.review.bot 2011-06-08 23:38:21 -0700 Comment on attachment 96536 Patch Clearing flags on attachment: 96536 Committed r88428: <http://trac.webkit.org/changeset/88428> 417856 5 webkit.review.bot 2011-06-08 23:38:26 -0700 All reviewed patches have been landed. Closing bug. 418110 6 msaboff 2011-06-09 11:00:57 -0700 Please make the larger values platform specific. We do not see a need for these thresholds to be this large. 418116 7 jamesr 2011-06-09 11:13:26 -0700 I'm happy to #ifdef values, what platforms and what values do you want? Given that this regresses performance across the board I imagine most ports will want to keep the old values. 418168 8 ggaren 2011-06-09 12:05:13 -0700 James, Darin, I don't think "svn revert" was a very polite way to raise this technical issue. But, now that you have raised the issue, let's try to find a solution that preserves rendering speed without a memory footprint regression. James, you mentioned that the Page Cycler uses up to 105 fonts. Do you think that's representative of typical web browsing, or is your goal here just to keep the whole Page Cycler in cache? How quickly does the Page Cycler cycle? Would a time-based approach to purging inactive fonts keep the whole Page Cycler in cache? 418179 9 jamesr 2011-06-09 12:18:30 -0700 I did not svn revert anything, I restored the previous constants and left the rest of the patch untouched. I do believe that our page cyclers are representative of user's browsing habits - that's why we have them in the first place - and as I mentioned on the other bug plan to gather some data from users in the field to get some more insight into what the font use looks like more broadly. 418188 10 ggaren 2011-06-09 12:29:58 -0700 > I did not svn revert anything, I restored the previous constants and left the rest of the patch untouched. OK. I don't think manually reverting a portion of a patch while leaving the rest untouched was a very polite way to raise this technical issue. > and as I mentioned on the other bug plan to gather some data from users in the field to get some more insight into what the font use looks like more broadly. Great. How soon do you think you'll have data to evaluate? In the meantime, do you think a time-based approach would succeed at keeping the whole Page Cycler in cache? How quickly does it cycle between its 105 fonts? 418330 11 msaboff 2011-06-09 14:56:45 -0700 Let's keep in mind we are not talking about total fonts, we are talking about inactive fonts. The original changes in <http://trac.webkit.org/changeset/88260> made the previously abandon system fallback fonts be inactive immediately after they are accessed. This could increase the number of inactive fonts. The change set also moved the deletion check to be after we are done using fonts for layout and painting operations instead of when we request a font from the cache. This change set did not place or modify any limit on the number of active fonts. From the testing I did, it appears that there are more active fonts than inactive fonts. In many cases the only inactive fonts that result from a page are when it is closed. If a page being accessed / rendered is Latin-1 based there may be little or no inactive fonts due to system fallback font usage as few if any glyphs are needed outside those provided by a specified font. From my tests the pages with the greatest number of inactive fonts due to system fallback fonts are asian or other non Latin-1 based. Although we have and use tests with a wide variety of pages with diverse languages, we don't believe that users in the real world actively access pages with more than two or three languages and therefore more than a handful of alphabets. There may be some inadvertent access to a page foreign to a user, but that is not the general case. This info coupled with various page load tests results led us to the smaller inactive font constants. I would prefer that the higher limits be for the chromium platform. Other platform developers can make a choice of the two sets or add their own values. 418427 12 jamesr 2011-06-09 17:37:12 -0700 Performance is still ~50% worse than pre-88260 levels on some of our tests, even with these font cache limits: http://build.chromium.org/f/chromium/perf/mac-release-10.5/intl1/report.html?history=100&rev=88610 The initial bump is r88620, the small drop is r88428. Given WebKit's no-regression policy this seems like a strong reason to revert the whole thing until the regression is understood and fixed, otherwise we are likely to introduce even more regressions on top of this one that will be more difficult to identify and fix. 418688 13 msaboff 2011-06-10 06:18:21 -0700 As I indicated in a prior comment, the change in r88260 fixed reference count leakage for system fallback fonts. Your tests could be creating far more inactive fonts so that 120 is now too small. Before reverting all of r88260, I suggest that you try the following test. - Change cMaxInactiveFontData to INT_MAX. - Rerun your tests to see what that does to performance. - At the end of the tests, display how many fonts are in the cache with fontDataCount() and inactiveFontDataCount(). If you do this after windows are closed, all or almost all fonts should be inactive. Better to display these values while some or all windows / tabs are still open. Or, add code to monitor the inactive count, keeping the highest value before all windows / tabs are closed. This should help you determine the new value for cMaxInactiveFontData to keep all fonts in the cache. Reply back with the results. 419069 14 96827 msaboff 2011-06-10 17:32:55 -0700 Created attachment 96827 Patch with platform specific values for FontCache inactive constants Proposed patch making Chromium specific values of cMaxInactiveFontData and cTargetInactiveFontData. Note that set cMaxInactiveFontData to INT_MAX with the expectation that someone from the Chromium team will investigate a better value. 419074 15 96827 jamesr 2011-06-10 17:39:00 -0700 Comment on attachment 96827 Patch with platform specific values for FontCache inactive constants INT_MAX is a terrible value as well, I'd be much happier with the previous values of 120/100. 419089 16 ggaren 2011-06-10 17:54:14 -0700 > INT_MAX is a terrible value as well, I'd be much happier with the previous values of 120/100. There's no such thing as "the previous values". Previously, WebCore leaked most fonts. A 120/100 limit in a world where most fonts are leaked doesn't tell us what the limit should be once the leak is fixed. I think you confirmed this finding when you measured that restoring the 120/100 constants doesn't fully restore performance to the Page Cycler. Are you saying that you're OK with not immediately restoring Chromium's score on the Page Cycler? As an aside, I'm not super interested in deciding what value goes into the #IF PLATFORM(CHROMIUM). You guys can pick any value you want. My goal is to bring memory use under control for the other ports. 419105 17 jamesr 2011-06-10 18:20:15 -0700 "leaking most" sounds different from setting it to INT_MAX, although perhaps it isn't - I'm still not fully familiar with how this code works. it seems like for Chromium-Mac, a value of 250 will be high enough to keep all inactive fonts in the cache for the page cycler data we have. This value is twice as high as what's needed for Chromium-linux, which is a bit surprising. Let's try 250 and see what happens. Longer term we'll have to do more detailed research on what values make sense for our user's browsing habits. Just doing a simple test of navigating around yahoo.jp, though, shows that the inactive font cache is over 80 after visiting 4 randomly selected pages within a domain. I'm curious what data led to a value of 50 being chosen. 419256 18 96868 msaboff 2011-06-11 17:27:35 -0700 Created attachment 96868 Change with chromium recommended value Set cMaxInactiveFontData to 250 per comment 17. Set cTargetInactiveFontData to 200 so that 50 inactive fonts will be cleaned up at a time. 419259 19 96868 jamesr 2011-06-11 18:37:27 -0700 Comment on attachment 96868 Change with chromium recommended value This looks good for now, thanks. 419267 20 msaboff 2011-06-11 23:01:49 -0700 Committed r88611: <http://trac.webkit.org/changeset/88611> 96536 2011-06-08 20:27:38 -0700 2011-06-10 17:32:55 -0700 Patch bug-62349-20110608202737.patch text/plain 1997 jamesr U3VidmVyc2lvbiBSZXZpc2lvbjogODgzOTAKZGlmZiAtLWdpdCBhL1NvdXJjZS9XZWJDb3JlL0No YW5nZUxvZyBiL1NvdXJjZS9XZWJDb3JlL0NoYW5nZUxvZwppbmRleCAwYjBmZmVjNzVmY2E4NWZi MWJjZWY1Zjk4Yjc1OWYwMjc4YzM5MWEwLi4yN2Q4YjkxMDhkOTJkYmM1YmVjY2UwMjg0NGYxOGMx NmUwYzdlNjJlIDEwMDY0NAotLS0gYS9Tb3VyY2UvV2ViQ29yZS9DaGFuZ2VMb2cKKysrIGIvU291 cmNlL1dlYkNvcmUvQ2hhbmdlTG9nCkBAIC0xLDMgKzEsMTkgQEAKKzIwMTEtMDYtMDggIEphbWVz IFJvYmluc29uICA8amFtZXNyQGNocm9taXVtLm9yZz4KKworICAgICAgICBSZXZpZXdlZCBieSBO T0JPRFkgKE9PUFMhKS4KKworICAgICAgICBSRUdSRVNTSU9OKDg4MjYwKTogMTAtNTAlIHBlcmZv cm1hbmNlIHJlZ3Jlc3Npb24gYWNyb3NzIG1hbnkgcGFnZSBjeWNsZXJzCisgICAgICAgIGh0dHBz Oi8vYnVncy53ZWJraXQub3JnL3Nob3dfYnVnLmNnaT9pZD02MjM0OQorCisgICAgICAgIHI4ODI2 MCBmaXhlZCBhIGZvbnQgY2FjaGUgcmVzb3VyY2UgbGVhayBhbmQgbG93ZXJlZCB0aGUgaW5hY3Rp dmUgZm9udCBjYWNoZSB0aHJlc2hvbGQuICBUaGUgbGF0dGVyIGNhdXNlZCBhCisgICAgICAgIHNp Z25pZmljYW50IHBlcmZvcm1hbmNlIHJlZ3Jlc3Npb24gYWNyb3NzIG1hbnkgY2hyb21pdW0gcGFn ZSBjeWNsZXJzLCBmb3IgZXhhbXBsZQorICAgICAgICBodHRwOi8vYnVpbGQuY2hyb21pdW0ub3Jn L2YvY2hyb21pdW0vcGVyZi9saW51eC1yZWxlYXNlLXdlYmtpdC1sYXRlc3QvbW96L3JlcG9ydC5o dG1sP2hpc3Rvcnk9NTAmcmV2PTg4Mjc5LgorCisgICAgICAgIFRoaXMgcmVzdG9yZXMgdGhlIHBy ZXZpb3VzIGluYWN0aXZlIGZvbnQgc2l6ZSB0aHJlc2hvbGRzIHRvIHRoZWlyIHByZXZpb3VzIHZh bHVlcywgYnV0IHJldGFpbnMgdGhlIGZvbnQgY2xlYW51cAorICAgICAgICBsb2dpYy4KKworICAg ICAgICAqIHBsYXRmb3JtL2dyYXBoaWNzL0ZvbnRDYWNoZS5jcHA6CisKIDIwMTEtMDYtMDggIE1p a2UgUmVlZCAgPHJlZWRAZ29vZ2xlLmNvbT4KIAogICAgICAgICBSZXZpZXdlZCBieSBKYW1lcyBS b2JpbnNvbi4KZGlmZiAtLWdpdCBhL1NvdXJjZS9XZWJDb3JlL3BsYXRmb3JtL2dyYXBoaWNzL0Zv bnRDYWNoZS5jcHAgYi9Tb3VyY2UvV2ViQ29yZS9wbGF0Zm9ybS9ncmFwaGljcy9Gb250Q2FjaGUu Y3BwCmluZGV4IDZmNzhmNjJhNWVjOTRmOTlkNDVmYzg1OTVjMmE5MWM0YzBjMjE3ODIuLjQyZTJi ZjkxODk4ZjI2MzM3MjgxMzQ4ZDg5NWRmYWEyZGJkYzNmY2MgMTAwNjQ0Ci0tLSBhL1NvdXJjZS9X ZWJDb3JlL3BsYXRmb3JtL2dyYXBoaWNzL0ZvbnRDYWNoZS5jcHAKKysrIGIvU291cmNlL1dlYkNv cmUvcGxhdGZvcm0vZ3JhcGhpY3MvRm9udENhY2hlLmNwcApAQCAtMjUxLDggKzI1MSw4IEBAIHR5 cGVkZWYgSGFzaE1hcDxGb250UGxhdGZvcm1EYXRhLCBwYWlyPFNpbXBsZUZvbnREYXRhKiwgdW5z aWduZWQ+LCBGb250RGF0YUNhY2hlCiAKIHN0YXRpYyBGb250RGF0YUNhY2hlKiBnRm9udERhdGFD YWNoZSA9IDA7CiAKLWNvbnN0IGludCBjTWF4SW5hY3RpdmVGb250RGF0YSA9IDUwOyAvLyBQcmV0 dHkgTG93IFRocmVzaG9sZAotY29uc3QgaW50IGNUYXJnZXRJbmFjdGl2ZUZvbnREYXRhID0gMzA7 Citjb25zdCBpbnQgY01heEluYWN0aXZlRm9udERhdGEgPSAxMjA7IC8vIFByZXR0eSBMb3cgVGhy ZXNob2xkCitjb25zdCBpbnQgY1RhcmdldEluYWN0aXZlRm9udERhdGEgPSAxMDA7CiBzdGF0aWMg TGlzdEhhc2hTZXQ8Y29uc3QgU2ltcGxlRm9udERhdGEqPiogZ0luYWN0aXZlRm9udERhdGEgPSAw OwogCiBTaW1wbGVGb250RGF0YSogRm9udENhY2hlOjpnZXRDYWNoZWRGb250RGF0YShjb25zdCBG b250RGVzY3JpcHRpb24mIGZvbnREZXNjcmlwdGlvbiwgY29uc3QgQXRvbWljU3RyaW5nJiBmYW1p bHksIGJvb2wgY2hlY2tpbmdBbHRlcm5hdGVOYW1lLCBTaG91bGRSZXRhaW4gc2hvdWxkUmV0YWlu KQo= 96827 2011-06-10 17:32:55 -0700 2011-06-11 17:27:35 -0700 Patch with platform specific values for FontCache inactive constants 62349-2.patch text/plain 1908 msaboff SW5kZXg6IFNvdXJjZS9XZWJDb3JlL0NoYW5nZUxvZwo9PT09PT09PT09PT09PT09PT09PT09PT09 PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09Ci0tLSBTb3VyY2UvV2Vi Q29yZS9DaGFuZ2VMb2cJKHJldmlzaW9uIDg4NTc4KQorKysgU291cmNlL1dlYkNvcmUvQ2hhbmdl TG9nCSh3b3JraW5nIGNvcHkpCkBAIC0xLDMgKzEsMjAgQEAKKzIwMTEtMDYtMTAgIE1pY2hhZWwg U2Fib2ZmICA8bXNhYm9mZkBhcHBsZS5jb20+CisKKyAgICAgICAgUmV2aWV3ZWQgYnkgTk9CT0RZ IChPT1BTISkuCisKKyAgICAgICAgUkVHUkVTU0lPTig4ODI2MCk6IDEwLTUwJSBwZXJmb3JtYW5j ZSByZWdyZXNzaW9uIGFjcm9zcyBtYW55IHBhZ2UgY3ljbGVycworICAgICAgICBodHRwczovL2J1 Z3Mud2Via2l0Lm9yZy9zaG93X2J1Zy5jZ2k/aWQ9NjIzNDkKKworICAgICAgICBNYWRlIENIUk9N SVVNIHBsYXRmb3JtIHNwZWNpZmljIHZhbHVlcyBmb3IgY01heEluYWN0aXZlRm9udERhdGEgYW5k CisgICAgICAgIGNUYXJnZXRJbmFjdGl2ZUZvbnREYXRhIHRvIG1pdGlnYXRlIHRoZSBjdXJyZW50 IHBlcmZvcm1hbmNlIGlzc3VlIG9uIENocm9taXVtCisgICAgICAgIHBhZ2UgY3ljbGVyIHRlc3Rz LiAgTWFkZSB0aGUgY2hyb21pdW0gdmFsdWUgZm9yIGNNYXhJbmFjdGl2ZUZvbnREYXRhIElOVF9N QVggdG8KKyAgICAgICAgZWZmZWN0aXZlbHkgdHVybiBvZmYgcHVyZ2luZyBvZiBpbmFjdGl2ZSBm b250cyBmb3IgdGhhdCBwbGF0Zm9ybS4gIFNvbWVvbmUgb24gdGhlCisgICAgICAgIENocm9taXVt IHRlYW0gY2FuIGNoYW5nZSB0byBhIG1vcmUgc3VpdGFibGUgdmFsdWUgd2hlbiB0aGF0IGlzIGRl dGVybWluZWQuCisKKyAgICAgICAgQ2hhbmdlIG1hZGUgdG8gYWRkcmVzcyBleGlzdGluZyBDaHJv bWl1bSBwZXJmb3JtYW5jZSB0ZXN0IHJlZ3Jlc3Npb24uCisKKyAgICAgICAgKiBwbGF0Zm9ybS9n cmFwaGljcy9Gb250Q2FjaGUuY3BwOgorCiAyMDExLTA2LTEwICBEYXZpZCBMZXZpbiAgPGxldmlu QGNocm9taXVtLm9yZz4KIAogICAgICAgICBSZXZpZXdlZCBieSBEbWl0cnkgVGl0b3YuCkluZGV4 OiBTb3VyY2UvV2ViQ29yZS9wbGF0Zm9ybS9ncmFwaGljcy9Gb250Q2FjaGUuY3BwCj09PT09PT09 PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09 PT0KLS0tIFNvdXJjZS9XZWJDb3JlL3BsYXRmb3JtL2dyYXBoaWNzL0ZvbnRDYWNoZS5jcHAJKHJl dmlzaW9uIDg4MjczKQorKysgU291cmNlL1dlYkNvcmUvcGxhdGZvcm0vZ3JhcGhpY3MvRm9udENh Y2hlLmNwcAkod29ya2luZyBjb3B5KQpAQCAtMjUxLDggKzI1MSwxMyBAQCB0eXBlZGVmIEhhc2hN YXA8Rm9udFBsYXRmb3JtRGF0YSwgcGFpcjxTCiAKIHN0YXRpYyBGb250RGF0YUNhY2hlKiBnRm9u dERhdGFDYWNoZSA9IDA7CiAKKyNpZiBQTEFURk9STShDSFJPTUlVTSkKK2NvbnN0IGludCBjTWF4 SW5hY3RpdmVGb250RGF0YSA9IElOVF9NQVg7Citjb25zdCBpbnQgY1RhcmdldEluYWN0aXZlRm9u dERhdGEgPSAxMDA7CisjZWxzZQogY29uc3QgaW50IGNNYXhJbmFjdGl2ZUZvbnREYXRhID0gNTA7 IC8vIFByZXR0eSBMb3cgVGhyZXNob2xkCiBjb25zdCBpbnQgY1RhcmdldEluYWN0aXZlRm9udERh dGEgPSAzMDsKKyNlbmRpZgogc3RhdGljIExpc3RIYXNoU2V0PGNvbnN0IFNpbXBsZUZvbnREYXRh Kj4qIGdJbmFjdGl2ZUZvbnREYXRhID0gMDsKIAogU2ltcGxlRm9udERhdGEqIEZvbnRDYWNoZTo6 Z2V0Q2FjaGVkRm9udERhdGEoY29uc3QgRm9udERlc2NyaXB0aW9uJiBmb250RGVzY3JpcHRpb24s IGNvbnN0IEF0b21pY1N0cmluZyYgZmFtaWx5LCBib29sIGNoZWNraW5nQWx0ZXJuYXRlTmFtZSwg U2hvdWxkUmV0YWluIHNob3VsZFJldGFpbikK 96868 2011-06-11 17:27:35 -0700 2011-06-11 18:37:27 -0700 Change with chromium recommended value 62349-3.patch text/plain 1881 msaboff SW5kZXg6IFNvdXJjZS9XZWJDb3JlL0NoYW5nZUxvZwo9PT09PT09PT09PT09PT09PT09PT09PT09 PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09Ci0tLSBTb3VyY2UvV2Vi Q29yZS9DaGFuZ2VMb2cJKHJldmlzaW9uIDg4NTc4KQorKysgU291cmNlL1dlYkNvcmUvQ2hhbmdl TG9nCSh3b3JraW5nIGNvcHkpCkBAIC0xLDMgKzEsMjAgQEAKKzIwMTEtMDYtMTAgIE1pY2hhZWwg U2Fib2ZmICA8bXNhYm9mZkBhcHBsZS5jb20+CisKKyAgICAgICAgUmV2aWV3ZWQgYnkgTk9CT0RZ IChPT1BTISkuCisKKyAgICAgICAgUkVHUkVTU0lPTig4ODI2MCk6IDEwLTUwJSBwZXJmb3JtYW5j ZSByZWdyZXNzaW9uIGFjcm9zcyBtYW55IHBhZ2UgY3ljbGVycworICAgICAgICBodHRwczovL2J1 Z3Mud2Via2l0Lm9yZy9zaG93X2J1Zy5jZ2k/aWQ9NjIzNDkKKworICAgICAgICBNYWRlIENIUk9N SVVNIHBsYXRmb3JtIHNwZWNpZmljIHZhbHVlcyBmb3IgY01heEluYWN0aXZlRm9udERhdGEgYW5k CisgICAgICAgIGNUYXJnZXRJbmFjdGl2ZUZvbnREYXRhIHRvIG1pdGlnYXRlIHRoZSBjdXJyZW50 IHBlcmZvcm1hbmNlIGlzc3VlIG9uIENocm9taXVtCisgICAgICAgIHBhZ2UgY3ljbGVyIHRlc3Rz LiAgTWFkZSB0aGUgY2hyb21pdW0gdmFsdWUgZm9yIGNNYXhJbmFjdGl2ZUZvbnREYXRhIDI1MCBi YXNlZAorICAgICAgICBvbiByZWNvbW1lbmRhdGlvbiBmcm9tIEphbWVzIFJvYmluc29uLiAgQ2hh bmdlZCBjVGFyZ2V0SW5hY3RpdmVGb250RGF0YSB0byAyMDAKKyAgICAgICAgc28gdGhhdCA1MCBp bmFjdGl2ZSBmb250cyB3aWxsIGJlIGNsZWFuZWQgdXAgYXQgYSB0aW1lLgorCisgICAgICAgIENo YW5nZSBtYWRlIHRvIGFkZHJlc3MgZXhpc3RpbmcgQ2hyb21pdW0gcGVyZm9ybWFuY2UgdGVzdCBy ZWdyZXNzaW9uLgorCisgICAgICAgICogcGxhdGZvcm0vZ3JhcGhpY3MvRm9udENhY2hlLmNwcDoK KwogMjAxMS0wNi0xMCAgRGF2aWQgTGV2aW4gIDxsZXZpbkBjaHJvbWl1bS5vcmc+CiAKICAgICAg ICAgUmV2aWV3ZWQgYnkgRG1pdHJ5IFRpdG92LgpJbmRleDogU291cmNlL1dlYkNvcmUvcGxhdGZv cm0vZ3JhcGhpY3MvRm9udENhY2hlLmNwcAo9PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09 PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09Ci0tLSBTb3VyY2UvV2ViQ29yZS9w bGF0Zm9ybS9ncmFwaGljcy9Gb250Q2FjaGUuY3BwCShyZXZpc2lvbiA4ODI3MykKKysrIFNvdXJj ZS9XZWJDb3JlL3BsYXRmb3JtL2dyYXBoaWNzL0ZvbnRDYWNoZS5jcHAJKHdvcmtpbmcgY29weSkK QEAgLTI1MSw4ICsyNTEsMTMgQEAgdHlwZWRlZiBIYXNoTWFwPEZvbnRQbGF0Zm9ybURhdGEsIHBh aXI8UwogCiBzdGF0aWMgRm9udERhdGFDYWNoZSogZ0ZvbnREYXRhQ2FjaGUgPSAwOwogCisjaWYg UExBVEZPUk0oQ0hST01JVU0pCitjb25zdCBpbnQgY01heEluYWN0aXZlRm9udERhdGEgPSAyNTA7 Citjb25zdCBpbnQgY1RhcmdldEluYWN0aXZlRm9udERhdGEgPSAyMDA7CisjZWxzZQogY29uc3Qg aW50IGNNYXhJbmFjdGl2ZUZvbnREYXRhID0gNTA7IC8vIFByZXR0eSBMb3cgVGhyZXNob2xkCiBj b25zdCBpbnQgY1RhcmdldEluYWN0aXZlRm9udERhdGEgPSAzMDsKKyNlbmRpZgogc3RhdGljIExp c3RIYXNoU2V0PGNvbnN0IFNpbXBsZUZvbnREYXRhKj4qIGdJbmFjdGl2ZUZvbnREYXRhID0gMDsK IAogU2ltcGxlRm9udERhdGEqIEZvbnRDYWNoZTo6Z2V0Q2FjaGVkRm9udERhdGEoY29uc3QgRm9u dERlc2NyaXB0aW9uJiBmb250RGVzY3JpcHRpb24sIGNvbnN0IEF0b21pY1N0cmluZyYgZmFtaWx5 LCBib29sIGNoZWNraW5nQWx0ZXJuYXRlTmFtZSwgU2hvdWxkUmV0YWluIHNob3VsZFJldGFpbikK