One of the reasons that a bunch of the reftests are failing on Windows is that Windows native theme code doesn't seem to draw correctly into canvas drawWindow (which is what the reftest harness uses to do the comparisons). The easiest testcase for this is running the reftest harness. The worst example is mozilla/layout/reftests/bugs/346774-1a.html and mozilla/layout/reftests/bugs/346774-1b.html, where the combobox just shows up as mostly white with occasional pixels of aqua. Scrollbars seem to do a good bit better than that, but still occasionally have pieces missing, sometimes replaced by white with occasional pixels of aqua. See mozilla/layout/tools/reftest/README.txt for instructions on running reftest; you can trim down the manifest to include only the tests mentioned above if you want. Doing this will print PNG images as data: URLs for the test failures to stdout.

There's some wacky stateful stuff going on here. It depends what order the tests are run in. If I run 28811-1, then it messes up only the next test that I run (of 346774-1a or 142333-1), but later tests are ok.

I debugged various pieces of state in nsNativeThemeWin::DrawWidgetBackground and I couldn't find anything different between the broken and working states.

(One clarification on the stateful stuff -- it seems like the first test run with native theme drawing is always fine.)

Should be fixed by 339598; at least, the tests seem to give the correct results now. Reopen if that's not the case...