REGRESSION(r126189): Reftest mismatches are (again) run through ImageDiff with 0.1 tolerance
Created attachment 159923 [details] Patch
Perhaps a reincarnation of bug #92847, this caused some tests to pop up as passing. Too bad.
Oh, that's my fault. Will look into it in a bit ...
How do we test this so it doesn't happen again?
Actually, to clarify ... this is not a regression of bug #92847. This was caused by the fix I landed for bug 94540 (r126189). At some point, we stopped calling diff_image() for reftests and were just comparing checksums. This seemed to be bad (at least for the Apple Mac port) and so I added the diff_image as a sanity check. It wasn't clear to me at the time (and still isn't) whether we should use the default tolerance for the port, or use 0 (and thus ref tests and pixel tests would potentially be compared with different tolerances). I can see arguments for either way, and perhaps this should be a separate default setting. In the absence of agreeing over what we want to do here, it's not obvious what the test should be testing ;). That said, I'll R+ this as I *think* the Apple Mac port doesn't actually need fuzzy diffs and so this shouldn't break things. I guess we can find out.
Comment on attachment 159923 [details] Patch Also, we should add some sort of a test for this (per eric's comment), but I don't want to hold up this fix for that. I'll file a separate bug.
Comment on attachment 159923 [details] Patch Clearing flags on attachment: 159923 Committed r126353: <http://trac.webkit.org/changeset/126353>
All reviewed patches have been landed. Closing bug.