Support clip masks for text runs.

[glennw]

This switches our subpixel rendering over to a two pass method
that supports normal alpha blending.

There's some tradeoffs involved here:

• +++ We get clip masks on subpixel text.
• +++ CPU batching is much better (no batch breaks on text color change).
• --- We draw the subpixel text pixels twice.

On balance, this is likely to be a reasonable performance win, because:
(a) We are typically CPU limited rather than GPU limited.
(b) The time in text run FS is typically a small portion of GPU time.

More importantly, this gives us correct clip masking on text, and also
allows us to apply some optimizations in the future:

• Take advantage of dual source blending where available. This is
  available on most recent hardware, and allows clip masking plus
  subpixel rendering in a single pass.
• Consider using the old constant color blend method when there
  is no clip mask present, if the two pass method does show up
  as a significant cost.

---

This change is 

[glennw]

r? @kvark

(Passes servo/wrench tests OK, but I still need to do a gecko try with this).

[glennw]

Gecko try:

https://hg.mozilla.org/try/rev/2d2e7a960008f0724f2f2799233d21bac923763d
https://treeherder.mozilla.org/#/jobs?repo=try&revision=2d2e7a960008f0724f2f2799233d21bac923763d

[glennw]

Gecko try is green on all configs.

[kvark]

Code looks good. I haven't evaluated the performance impact yet though.

---

Reviewed 16 of 16 files at r1.
Review status: all files reviewed at latest revision, 3 unresolved discussions.

---

webrender/res/ps_text_run.glsl, line 88 at r1 (raw file):

    switch (uMode) {
        case MODE_ALPHA:
            oFragColor = alpha * color * vColor;

I vaguely recall driver issues with output variables assigned in a switch

---

webrender/src/device.rs, line 1466 at r1 (raw file):

            .uniform_1f(program.u_device_pixel_ratio, self.device_pixel_ratio);
        self.gl
            .uniform_1i(program.u_mode, mode);

I wonder if we can get any perf from caching those uniforms. Device pixel ratio and the mode don't change often.

---

webrender/src/renderer.rs, line 900 at r1 (raw file):

        mode: M,
        renderer_errors: &mut Vec<RendererError>,
    ) where M: Into<i32> {

would not harm to have a proper name for this type instead of i32, even if just a typedef

---

Comments from Reviewable

[kvark]

More importantly, this gives us correct clip masking on text

I guess there is no alternative atm :)
@bors-servo r+

---

Review status: all files reviewed at latest revision, 3 unresolved discussions.

---

Comments from Reviewable

[bors-servo]

📌 Commit dfaa0de has been approved by kvark

[kvark]

Actually, Glenn is here soon, so I'll wait to let him address the nits.
@bors-servo r-

[glennw]

Review status: 15 of 16 files reviewed at latest revision, 3 unresolved discussions.

---

webrender/res/ps_text_run.glsl, line 88 at r1 (raw file):

Previously, kvark (Dzmitry Malyshau) wrote…

I vaguely recall driver issues with output variables assigned in a switch

I think that was only in a very specific case, not in any switch. Maybe merge and see if we encounter any issues?

---

webrender/src/device.rs, line 1466 at r1 (raw file):

Previously, kvark (Dzmitry Malyshau) wrote…

I wonder if we can get any perf from caching those uniforms. Device pixel ratio and the mode don't change often.

#1864

---

webrender/src/renderer.rs, line 900 at r1 (raw file):

Previously, kvark (Dzmitry Malyshau) wrote…

would not harm to have a proper name for this type instead of i32, even if just a typedef

Done.

---

Comments from Reviewable

[glennw]

@kvark Thanks for the quick review!

Comments addressed.

[kvark]

I think that was only in a very specific case, not in any switch. Maybe merge and see if we encounter any issues?

I think we should make it an informal rule to have a single code path for all output writes. It makes drivers happy, and allows easier code navigation and reasoning.

Anyhow, it's fine to follow-up with. Thanks for addressing my nits!
@bors-servo r+

[bors-servo]

📌 Commit 2e9eb03 has been approved by kvark

[glennw]

Yea, that sounds like a good idea - perhaps we need a style guide for shader writing...

[bors-servo]

⌛ Testing commit 2e9eb03 with merge 4a0a2ce...

[bors-servo]

☀️ Test successful - status-appveyor, status-travis
Approved by: kvark
Pushing 4a0a2ce to master...

[staktrace]

@glennw I believe this change is causing shader compilation failures on Windows:

thread '' panicked at 'called Result::unwrap() on an Err value: Link("debug_font", "Precisions for uniform 'uMode' differ between vertex and fragment shaders\n")', src\libcore\result.rs:860

Full regression range is in https://bugzilla.mozilla.org/show_bug.cgi?id=1407213#c10 but this PR is the one that touches uMode.

[kvark]

@staktrace could you try with #1872 ?

[staktrace]

Thanks for the quick PR! Try push is in progress at https://treeherder.mozilla.org/#/jobs?repo=try&revision=911f826e93460538f7906168377958a3259b30af

[staktrace]

I screwed up the reftest.list files in that try push so it didn't run the reftests. Here's another one (this includes #1873 as well): https://treeherder.mozilla.org/#/jobs?repo=try&revision=1f38daa67ff395256f26bc280b099a930dc26c38

[staktrace]

Fix verified, thanks!