This is a second attempt at submitting this changelist. The original one was

http://codereview.chromium.org/196053 It turns out, since none of our tests
abstract time correctly, unittests are very sensitive to subtle changes in
timing. This made some of the valgrind tests on Linux fail, and unfortunately,
neither my desktop nor the trybots could reproduce the problem reliably.

As far as I can tell, all the (design) bugs are in the unittests. The browser
is actually fine.

Tweaked the code a little more. Will resubmit and carefully monitor the
buildbots.


Original change description follows:

When converting between units of time or data types of different precision, 
we have to be careful to consistently round in the same direction. 

Timeout checks usually check if Now() is less or equal to a deadline in order 
to determine if a timeout has occurred. This correctly handles the case where 
actual sleep times are equal or longer than requested sleep times. 

But if we round down when setting the sleep delay, this can result in 
unnecessary and expensive looping. Make sure, we always round up when converting 
to a format with less precision. 

BUG=none 
TEST=none
Review URL: http://codereview.chromium.org/257044

git-svn-id: svn://svn.chromium.org/chrome/trunk/src@28801 0039d316-1c4b-4281-b951-d872f2087c98

1 files changed, 14 insertions, 2 deletions

diff --git a/webkit/glue/webkitclient_impl.cc b/webkit/glue/webkitclient_impl.cc
index 1639683..7b3dde924 100644
--- a/webkit/glue/webkitclient_impl.cc
+++ b/webkit/glue/webkitclient_impl.cc
@@ -2,6 +2,7 @@
 // source code is governed by a BSD-style license that can be found in the
 // LICENSE file.
 
+#include <math.h>
 #include "config.h"
 
 #include "FrameView.h"
@@ -267,12 +268,23 @@ void WebKitClientImpl::setSharedTimerFiredFunction(void (*func)()) {
 }
 
 void WebKitClientImpl::setSharedTimerFireTime(double fire_time) {
-  int interval = static_cast<int>((fire_time - currentTime()) * 1000);
+  // By converting between double and int64 representation, we run the risk
+  // of losing precision due to rounding errors. Performing computations in
+  // microseconds reduces this risk somewhat. But there still is the potential
+  // of us computing a fire time for the timer that is shorter than what we
+  // need.
+  // As the event loop will check event deadlines prior to actually firing
+  // them, there is a risk of needlessly rescheduling events and of
+  // needlessly looping if sleep times are too short even by small amounts.
+  // This results in measurable performance degradation unless we use ceil() to
+  // always round up the sleep times.
+  int64 interval = static_cast<int64>(
+      ceil((fire_time - currentTime()) * base::Time::kMicrosecondsPerSecond));
   if (interval < 0)
     interval = 0;
 
   shared_timer_.Stop();
-  shared_timer_.Start(base::TimeDelta::FromMilliseconds(interval), this,
+  shared_timer_.Start(base::TimeDelta::FromMicroseconds(interval), this,
                       &WebKitClientImpl::DoTimeout);
 }