See test results at http://test-results.appspot.com/dashboards/flakiness_dashboard.html#tests=compositing%2Fdirect-image-compositing.html&showExpectations=true&group=%40ToT%20GPU%20-%20chromium.org . Notice the difference in how the blurs on the left side of the result page look.
It looks as if the following failures are also caused by this change. Note that in these cases, the difference generally looks like something I'd rebaseline for, but since the test in comment 0 is clearly a significant change I'm going to just mark all these as failing until you can look. compositing/geometry/fixed-position.html compositing/geometry/horizontal-scroll-composited.html compositing/geometry/vertical-scroll-composited.html compositing/masks/masked-ancestor.html compositing/masks/simple-composited-mask.html compositing/overflow/fixed-position-ancestor-clip.html fast/canvas/canvas-text-alignment.html fast/canvas/image-object-in-canvas.html http://test-results.appspot.com/dashboards/flakiness_dashboard.html#tests=compositing%2Fgeometry%2Ffixed-position.html%2Ccompositing%2Fgeometry%2Fhorizontal-scroll-composited.html%2Ccompositing%2Fgeometry%2Fvertical-scroll-composited.html%2Ccompositing%2Fmasks%2Fmasked-ancestor.html%2Ccompositing%2Fmasks%2Fsimple-composited-mask.html%2Ccompositing%2Foverflow%2Ffixed-position-ancestor-clip.html%2Cfast%2Fcanvas%2Fcanvas-text-alignment.html%2Cfast%2Fcanvas%2Fimage-object-in-canvas.html&showExpectations=true&group=%40ToT%20GPU%20-%20chromium.org
Thanks for filing this with all the additional cases. It looks like it's an extremely small change and not anything as large as an off-by-one texel issue. The scrollbars on the fixed-position test go from RGB a4a4a4 to a3a3a3. My intuition is that this is floating point error in the texture transform going to the shader causing a minute amount of filtering.
Created attachment 82979 [details] Patch
This patch fixes some issues on Linux (but breaks tests because of the recent rebaselining). It fixes some issues on Windows, but I'll point out that it breaks a few issues from the pre-patch rebaselining. However, I think nearest sampling will give more consistent results in the future, so I will rebaseline all of the tests affected by this patch after this lands. Obviously once we have tiled content layers, we can add an option for bilinear filtering on those layers.
Comment on attachment 82979 [details] Patch Looks okay to me.
Comment on attachment 82979 [details] Patch Clearing flags on attachment: 82979 Committed r79084: <http://trac.webkit.org/changeset/79084>
All reviewed patches have been landed. Closing bug.