Support InputMethodManager#updateCursorAnchorInfo for Android 5.0

content/renderer/render_widget.cc

Index: content/renderer/render_widget.cc
diff --git a/content/renderer/render_widget.cc b/content/renderer/render_widget.cc
index 592a57c0c1d3a076f4b8354c9affead6d7f5bdd3..197b5ab145230e84ab776c2a3e381a10fbe04643 100644
--- a/content/renderer/render_widget.cc
+++ b/content/renderer/render_widget.cc
@@ -87,6 +87,7 @@
 
 #if defined(OS_ANDROID)
 #include <android/keycodes.h>
+#include "base/android/build_info.h"
 #include "content/renderer/android/synchronous_compositor_factory.h"
 #include "content/renderer/android/synchronous_compositor_filter.h"
 #include "content/renderer/android/synchronous_compositor_output_surface.h"
@@ -1748,9 +1749,13 @@ ui::TextInputType RenderWidget::GetTextInputType() {
 
 void RenderWidget::UpdateCompositionInfo(bool should_update_range) {
 #if defined(OS_ANDROID)
-// TODO(yukawa): Start sending character bounds when the browser side
-// implementation becomes ready (crbug.com/424866).
-#else
+  const static bool is_cursor_anchor_info_supported =
+      base::android::BuildInfo::GetInstance()->sdk_int() >= 21 &&

[aelias_OOO_until_Jul13, 2016/02/10 08:23:42]

We shouldn't add an SDK version level check in the renderer.  I'd prefer we just
sent up the message unconditionally, even if K and below won't use it for
anything.

[kinaba, 2016/02/19 12:44:08]

I guess devices stuck at K or below tends to be older and slower one hence I
think
this is an optimization worth doing. (Originally the information was always
sent,
but later this #ifdef(OS_ANDROID) was added for avoiding the unnecessary flood
of IPCs.
I'd rather respect the history.) WDYT?

[jdduke (slow), 2016/02/19 16:28:07]

Seems reasonable. I suppose you could gait the flag itself on an SDK version
check in the browser, not sure if that's a cleaner approach though.

[aelias_OOO_until_Jul13, 2016/02/24 03:16:39]

Regardless of the history, I stand by what I said in my original comment. 
Forking this behavior based on Android version is unnecessary complexity.  <K
devices are also in the same ballpark of performance as L+ ones so there's no
reason to give them special optimizations, we don't do that anywhere else.  If
there's actually a performance problem here, then the solution would look
something like merging this into another of our IPCs like UpdateTextInputState,
which would benefit all Android versions, not to just exclude some subset of
devices while paying the cost others.

But personally I'm just not worried about the performance of this.  You don't
need to go out of your way to test performance of this patch as far as I'm
concerned, simply blinding landing and waiting to see if any perf sheriff
complains is good enough for me.  User typing rate is pretty slow (like 1
character every 100ms at most), and there are already heavy costs like raster
associated with it, so the cost of a single IPC per character (in the 100s of
microseconds cost) should really be lost in the noise here.

+      base::CommandLine::ForCurrentProcess()->HasSwitch(
+          switches::kEnableCursorAnchorInfo);
+  if (!is_cursor_anchor_info_supported)
+    return;
+#endif
   TRACE_EVENT0("renderer", "RenderWidget::UpdateCompositionInfo");
   gfx::Range range = gfx::Range();
   if (should_update_range) {
@@ -1767,7 +1772,6 @@ void RenderWidget::UpdateCompositionInfo(bool should_update_range) {
   composition_range_ = range;
   Send(new InputHostMsg_ImeCompositionRangeChanged(
       routing_id(), composition_range_, composition_character_bounds_));
-#endif
 }
 
 void RenderWidget::convertViewportToWindow(blink::WebRect* rect) {