Support snapping on clip masks, and fast path rectangle + clip out mode.

[glennw]

This change is 

[glennw]

This supersedes #2608 (it includes that commit).

This PR achieves two things:

• Support snapping of axis-aligned clip masks (mostly reviewed in #2608).
• Native Rectangle + ClipOut mode (an optimization, and accuracy win).

This allows primitives that only contain clip rects to avoid generating
a clip mask, and also fixes a number of reftests.

Fixes #2595.
Fixes #2501.

Try run:
https://treeherder.mozilla.org/#/jobs?repo=try&revision=89310dbe1725653477a524a69804cd34dcb084e1

Of the relevant reftest differences:

R4 - 1 new PASS
R6 - 1 new PASS
R7 - 1 new PASS, 1 new FAIL (agreed to fuzz #2608 (comment))

r? @mrobinson or @kvark

[kvark]

Reviewed 13 of 13 files at r1.
Review status: all files reviewed at latest revision, all discussions resolved.

---

webrender/res/clip_shared.glsl, line 85 at r1 (raw file):

    // compute the point position inside the scroll node, in CSS space
    vec2 vertex_pos = device_pos +

I don't understand this part. When we snap, we want the final VS results to be snapped. But this code uses the pre-snapped device pos

---

webrender/src/prim_store.rs, line 1576 at r1 (raw file):

        }

        if is_large || rect_clips_only {

why are we segmenting a small but rect-only primitive?

---

Comments from Reviewable

[glennw]

Review status: all files reviewed at latest revision, 2 unresolved discussions.

---

webrender/res/clip_shared.glsl, line 85 at r1 (raw file):

Previously, kvark (Dzmitry Malyshau) wrote…

I don't understand this part. When we snap, we want the final VS results to be snapped. But this code uses the pre-snapped device pos

Because in the clip shaders, the vertices are the flat coordinates of the corners of the 2d bounding rect of the clip task. So I don't see how / why we'd be able to apply snapping here? What this PR does is keep the original vertex positions the same as the task rect, but adjust the interpolants for the local position at each corner vertex such that the snapping is accounted for. At least, that's the intent :) Does that sound reasonable?

---

webrender/src/prim_store.rs, line 1576 at r1 (raw file):

Previously, kvark (Dzmitry Malyshau) wrote…

why are we segmenting a small but rect-only primitive?

Because rect only primitives can be efficiently segmented into CPU regions and thus completely avoid any clip-mask at all. For example, I intend to draw simple borders as a rectangle with a clip + clip-out rectangle (got a local patch nearly ready for this). This means that for the most common border types we don't invoke the border shader at all, and don't allocate / render any clip masks in the case of no rounded corners.

---

Comments from Reviewable

[bors-servo]

☔ The latest upstream changes (presumably #2614) made this pull request unmergeable. Please resolve the merge conflicts.

[glennw]

Rebased.

[kvark]

---

Reviewed 4 of 4 files at r2.
Review status: all files reviewed at latest revision, all discussions resolved.

---

Comments from Reviewable

[kvark]

Thank you!
@bors-servo r+

[bors-servo]

📌 Commit 076b443 has been approved by kvark

[bors-servo]

⌛ Testing commit 076b443 with merge fe24137...

[bors-servo]

☀️ Test successful - status-appveyor, status-taskcluster
Approved by: kvark
Pushing fe24137 to master...