Convexity is determined with a two pass algorithm, first by a sign
measure and then by a winding measure. Convexity also is meant to not be
affected by leading moveTos (other than the last leading moveTo before
real verbs) and not affected by trailing moveTos (since no additional
contour has actually started).
The old code would incorrectly reduce pointCount when the last moveTo
index was greater than 0, so the BySign pass was skipped or calculated
on an incomplete set of points. When a path (as the one added in this
CL's new test) is convex by winding but not by sign, it would be
incorrectly identified as convex. This led to further cascading issues
during rasterization.
However, the old code also had the effect of correctly ignoring any
last trailing moveTo from being included in the BySign test. Without
the new loop decrementing pointCount, trailing moveTo locations
would incorrectly create concave paths (and would in fact be concave
if the verb was anything other than a move).
I also realized that if the last moveTo index is not at the end of the
initial leading block, or at the end of the path entirely, then it
means the path must have multiple contours, at which point the path
cannot be convex, so we take the early out.
TBR: reed@google.com, bsalomon@google.com
Bug: skia:1220754
Change-Id: I9bd38f2eaaa3dbee135c190ade46fce0bd20257a
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/420238
Reviewed-by: Michael Ludwig <michaelludwig@google.com>
Commit-Queue: Michael Ludwig <michaelludwig@google.com>
9f73b04b43