Bug 199505

Reported test failure was, --- /home/phil/WebKit/WebKitBuild/Debug/layout-test-results/media/video-volume-expected.txt +++ /home/phil/WebKit/WebKitBuild/Debug/layout-test-results/media/video-volume-actual.txt @@ -12,7 +12,7 @@ EVENT(canplaythrough) EXPECTED (video.volume == '0') OK RUN(video.volume = 0.5) -EXPECTED (video.volume == '0.5') OK +EXPECTED (video.volume == '0.5'), OBSERVED '0.5000011920928955' FAIL TEST(video.volume = 1.5) THROWS(DOMException.INDEX_SIZE_ERR) OK TEST(video.volume = -0.5) THROWS(DOMException.INDEX_SIZE_ERR) OK END OF TEST This failure was caused by r246730, although it's not a regression. I believe it's a bug in the PA project. PA does it's own linear<->cubic conversions, which was part of the motivation for r246730 in the first place. However, despite the APIs for this conversion promising a inverse operation, we get, pa_sw_volume_to_linear(pa_sw_volume_from_linear(0.5)) -> 0.500001 The reason this worked before was due to the cubic scaling we initially applied. When you cube 0.5, you get 0.125 and pa_sw_volume_to_linear(pa_sw_volume_from_linear(0.125)) -> 0.125, so the cbrt gives back 0.5. The test passes, but not in the way you would expect. I believe the bug in PA is due to their use of implicitly widening 32-bit integers to 64-bits in their interpolation code, and getting garbage (undefined memory) in the resulting value which when cast back to float, gives us a rounding issue. I am talking with the developers about it.