Make dashed-border rendering more similar to Gecko / other browsers.

[emilio]

This mostly focuses on non-cornered dashed borders.

The worst case of this algorithm is the one where the borders are rounded and the space between corners (the solid segment) is small but non-zero.

Thus I'm a bit on the fence on the fourth commit... I could preserve the current algorithm for the rounded corner cases.

The tl;dr is that this makes a dashed segment always start with half a dash, which matches other browsers (modulo a TODO where other browsers start with the whole segment, and border-radius when each browser does something completely different so I suspect we may be a bit more lenient).

I think we should do this for now if it doesn't cause problems, though it's trivial to extend it to do the same as Gecko for discontinued borders, etc. It's just a bit more annoying because instead of passing a flag I need to pass the offset around in a new float value.

I ended up shifting the brush in brush_picture.glsl, where the rest of the repeating code is handled. That being said, I've done the transformation in local space, but there may be a better way to do it.

There are some cleanups to the border code that I can do if this is a decent / acceptable approach.

---

This change is 

[gw3583]

It's probably not ideal that we need some border-specific code / flags in the common brush.glsl code, which I guess is what you were talking about on IRC?

What we actually do for dashes in corners now

webrender/webrender/src/border.rs

Line 1096 in 9a2fbef

instances.push(BorderInstance {

is generate an instance for each dash / dot. We do this for corners because the size of the dash / dot varies with each instance as traced around the corner.

I think what we probably want to do is something similar for dots / dashes in the border edge segments. This would give us complete control over specifying a local rect for each of the dashes, without affecting the common brush code.

Previously, we wanted to do it in the shader to avoid generating a lot of instances per frame, but we now cache border segments to the texture cache, and re-use them between frames, so this isn't really a concern now.

Having said all that, given our current deadlines and bug list - it's probably fine to merge a solution like this and do the work above as a follow up later on - thoughts?

[emilio]

Try is green and this of course fixes #2828 and such.

Need to cleanup and remove the now unused flags and such which I'll do tomorrow, but @gw3583 does this look reasonable to you?

[emilio]

A little bit of a lie: It's green modulo some panic which is fixed in a commit here already.

Previous try run: https://treeherder.mozilla.org/#/jobs?repo=try&revision=706f34643c938dbeafaf4db23d332b4bba64592c&selectedJob=198528498

New one: https://treeherder.mozilla.org/#/jobs?repo=try&revision=d28f4117d92954f19cf509d5aef4e2542fd71983

[kvark]

Reviewed 10 of 10 files at r1.
Reviewable status: all files reviewed, 6 unresolved discussions (waiting on @emilio)

---

webrender/res/brush_image.glsl, line 99 at r1 (raw file):

        //       works. That assumption may not hold if this
        //       is used for other purposes in the future.
        if ((brush_flags & BRUSH_FLAG_SEGMENT_REPEAT_X) != 0)

nit: I don't think this formatting changes are good. Better be consistent with both Rust and the rest of GLSL code and always use the brackets

---

webrender/res/cs_border_segment.glsl, line 65 at r1 (raw file):

#define CLIP_NONE        0
#define CLIP_DASH_CORNER 1

nit: can we put the DASH_* types together?

---

webrender/res/cs_border_segment.glsl, line 367 at r1 (raw file):

            // |xxx|   |   |xxx|
            // +---+---+---+---+
            bool in_dash = is_vertical

nit: float pos = is_vertical ? vPos.y : vPos.x;

---

webrender/src/border.rs, line 247 at r1 (raw file):

#[derive(Debug, Clone)]
struct BorderCornerClipSource {
    // XXX this makes no sense as a name for dashed borders now that it

what exactly makes no sense now? max_clip_count? BorderCornerClipSource?

---

webrender/src/border.rs, line 369 at r1 (raw file):

                let mut current_length = 0.;

                dot_dash_data.reserve(max_clip_count / 4);

this should have +1 like the loop does

---

webrender/src/border.rs, line 992 at r1 (raw file):

        BrushSegment::new(
            image_rect,
            /* may_need_clip_mask = */ true,

asks to be an enum :)

[emilio]

Thanks for the review @kvark!

Addressed comments and squashed the commits. I'll cleanup BorderCornerClipSource as a followup I think. Shouldn't be really hard.

[emilio]

@bors-servo r=kvark

[bors-servo]

📌 Commit 7717221 has been approved by kvark

[bors-servo]

⌛ Testing commit 7717221 with merge 02f14d0...

[bors-servo]

☀️ Test successful - status-appveyor, status-taskcluster
Approved by: kvark
Pushing 02f14d0 to master...