[webkit-dev] Rebaselining render tree dumps

[David Hyatt]

On Dec 6, 2010, at 3:20 PM, Dirk Pranke wrote:

> I think this is a great idea. I have one suggestion and one
> potentially large, derailing comment :)
> 
> Suggestion: in theory, fixing the render tree output shouldn't change
> the pixel output. So, I would suggest that as many ports as possible
> should be running (up-to-date) pixel tests when do have this flag day.
> 

Yeah, that's a really good observation.  The pixel tests will help us know that the changes to the dump didn't regress anything.

> Large, potentially derailing comment -- would this be an opportune
> time to try and codify the intent of the rendertree output at all? We
> are geting close to having support for reftests (finally!), and we
> have dumpAsMarkup() now as well. It seems to me that a lot of the
> original intent for render tree output may have been replaced by
> better alternatives now (not that I'm suggesting we rewrite all of the
> tests prior to this flag day).
> 

We definitely should be reducing the use of the render tree dump.  There are a number of areas where reftests are extremely difficult though.  Vertical text and multi-column layout are both examples where reftests are virtually impossible.  I think reftests for certain kinds of transforms and compositing could be difficult as well.

The biggest benefit of render tree dumps is that they tell you about the geometry and style of scrolled or clipped content.  Therefore they catch bugs that pixel tests don't see (and if reftests compare based off visible pixels then they would not see the error either).  This is the case with any tests that have scrolled or clipped content.

dave
(hyatt at apple.com)