Which platform did you see this on?

Sorry, saw this on Linux.

flackr's change in compositing scrollers that form a stacking context, or whatever that was? Seems unlikely to affect lots of pages, however.

Re #7, there are no overflow scrollers on wikipedia so it would be unaffected by that change.

Will need to bisect - how do I force rAF aligned touch and mouse input off?

out/Release/chrome --disable-features=RafAlignedTouchInput,RafAlignedMouseInput

I can't reproduce on my Linux/NVidia machine at ed16c60bdef2a6fca694e312d5a77e5b93d538d2 (revision listed in #1)

Running chrome using "--enable-smooth-scrolling --disable-features=RafAlignedTouchInput,RafAlignedMouseInput"

Then opening wikipedia page and holding down key on keyboard to scroll.

Is there anything else that might be required to repro?

Interesting, this was on my Linux/NVidia machine.

Can you repro without those flags?

Actually, from looking at traces and recent commits, I have an idea of what might be causing this - do you have dcheck_always_on = true (or a debug build)?

The issue you're hitting seems to be from frequent calls to glGetIntegerv. Some code was added that might do this in https://codereview.chromium.org/2516063002, which is in the revision range, but would only be hit with DCHECKs enabled.

Ok, confirmed my guess from #13 - this reproduces with dcheck_always_on.

trchen@, it appears that the following logic introduced in https://codereview.chromium.org/2516063002 is significantly impacting performance with DCHECK enabled builds. The getIntegerv here causes a sync IPC to the GPU process - when this happens multiple times per frame the costs really add up. Is it possible to get most of this coverage with a unit test and avoid the DCHECK?

 3227 #if DCHECK_IS_ON()
 3228     // GetIntegerv doesn't modify the output array when GL context is lost or
 3229     // in test mock. Also our GL wrapper requires output to be initialized with
 3230     // 0 or -1. Assuming lost context if the output is still -1.
 3231     GLint actual_scissor[4] = {-1};
 3232     gl_->GetIntegerv(GL_SCISSOR_BOX, actual_scissor);
 3233     if (actual_scissor[0] == -1)
 3234       return;
 3235     DCHECK(scissor_rect.x() == actual_scissor[0] &&
 3236            scissor_rect.y() == actual_scissor[1] &&
 3237            scissor_rect.width() == actual_scissor[2] &&
 3238            scissor_rect.height() == actual_scissor[3]);
 3239 #endif

Thanks!

Yup, I have DCHECK always on - nice catch.

Tested on windows 10 touch laptop using chrome latest canary M57 #57.0.2937.0 and issue is not reproduced.

@MTV team---Could anyone from MTV team look into this , as inhouse team doesn't have Linux/NVidia touch machine .

Thanks!

If this is for DCHECK enabled build only, is it still Pri-1?

Also, ping trchen@ re #14.

We removed that code recently, and ya this is about dcheck builds only so I think we're good.