Enable 2 pending frames in the scheduler on ARM devices.

In the past eglSwapBuffers used to return early, before the swap actually happened and descheduling in the driver. This resulted in the scheduler beginning the next frame early and hitting the page flip deadline in time more often.

Freon returns swap ack when page flip actually happened, but with the side effect of scheduling shifted forward, resulting in more missed vblanks and reduced performance.

This change restores original performance on freon.

BUG=465599

Review URL: https://codereview.chromium.org/1059503003

Cr-Commit-Position: refs/heads/master@{#324098}

content/browser/compositor/gpu_surfaceless_browser_compositor_output_surface.cc

@@ -30,12 +30,14 @@ GpuSurfacelessBrowserCompositorOutputSurface::
       gpu_memory_buffer_manager_(gpu_memory_buffer_manager) {
   capabilities_.uses_default_gl_framebuffer = false;
   capabilities_.flipped_output_surface = true;
-  // TODO(alexst): Can't enable this on ARM since it hangs the device
-  // (crbug.com/470185). Alternatively remove this entirely once the compositor
-  // scheduling is optimized enough to work well with surfaceless.
-#if defined(ARCH_CPU_X86_FAMILY)
+  // Set |max_frames_pending| to 2 for surfaceless, which aligns scheduling
+  // more closely with the previous surfaced behavior.
+  // With a surface, swap buffer ack used to return early, before actually
+  // presenting the back buffer, enabling the browser compositor to run ahead.
+  // Surfaceless implementation acks at the time of actual buffer swap, which
+  // shifts the start of the new frame forward relative to the old
+  // implementation.
   capabilities_.max_frames_pending = 2;
-#endif
 
   gl_helper_.reset(new GLHelper(context_provider_->ContextGL(),
                                 context_provider_->ContextSupport()));