| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This is a reland of accumulated patches including CLs
8414036, 8425010, 8425013, 8424013.
It fully enables tracking of objects, as seen in about:tracking
in teh release build.
The problem this had in its initaial landing centered on
Mac instability in base_unittest. We were encountering crash
on exit of the tests, with a stack dump involving pthreads.
r=rtenneti
BUG=101856
Review URL: http://codereview.chromium.org/8429009
git-svn-id: svn://svn.chromium.org/chrome/trunk/src@108026 0039d316-1c4b-4281-b951-d872f2087c98
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This is a re-land of:
http://codereview.chromium.org/8391019/
Committed: http://src.chromium.org/viewvc/chrome?view=rev&revision=107793
Original landing had trouble with message_loop_x.h, due to header
include ordering. I pulled out the structure that was really needed by
tracked_objects.h into a new file tracked_info.*. This allows tracked_objects
to inlude this tracked_info, but not have to include the message_loop.h
totality. I also removed a DCHECK that that was triggering on a test,
and added yet one more file (browser_main.cc) where I removed a #ifdef
for TRACKING_ALL_OBJECTS. The changes were minor, and I'm hoping to get
clear perf runs with tihs landing, so I'm going to TBR it and reland
early in the morning.
Comments from original landing:
Support is now controlled by the flag:
--enable-tracking
and the default is always on. To turn it off, use:
--enable-tracking=0
All profiler code is compiled now in release and official
builds (in addition to debug, where it was already active),
but most entry points can be disabled (turned into no-ops)
by a single const bool setting atop tracked_objects.cc (in
case folks want to revert the perf-impact of this change).
Transition to faster Now() service on Windows for the
profiler use only.
The TimeTicks::Now() function on Window uses locking
to get a 64 bit time value. This CL transitions
times used for profiling to more directly use a
32 bit Time interface, which is actually what drives the
64 bit TimeTicks. By using the smaller value, we avoid
the need for locks, or even atomic operations for the most
part in the tracking system. On linux, we just down-sample
the standard TimeTicks to 32 bits for consistency (clean
ability to snapshot asyncronously without atomics...
but I should verify that such is helpful to performance).
I've also put in yet more cleanup and refactoring.
tbr=rtenneti
bug=101856
Review URL: http://codereview.chromium.org/8414036
TBR=jar@chromium.org
Review URL: http://codereview.chromium.org/8430004
git-svn-id: svn://svn.chromium.org/chrome/trunk/src@107961 0039d316-1c4b-4281-b951-d872f2087c98
|
|
|
This is a re-land of:
http://codereview.chromium.org/8391019/
Committed: http://src.chromium.org/viewvc/chrome?view=rev&revision=107793
Original landing had trouble with message_loop_x.h, due to header
include ordering. I pulled out the structure that was really needed by
tracked_objects.h into a new file tracked_info.*. This allows tracked_objects
to inlude this tracked_info, but not have to include the message_loop.h
totality. I also removed a DCHECK that that was triggering on a test,
and added yet one more file (browser_main.cc) where I removed a #ifdef
for TRACKING_ALL_OBJECTS. The changes were minor, and I'm hoping to get
clear perf runs with tihs landing, so I'm going to TBR it and reland
early in the morning.
Comments from original landing:
Support is now controlled by the flag:
--enable-tracking
and the default is always on. To turn it off, use:
--enable-tracking=0
All profiler code is compiled now in release and official
builds (in addition to debug, where it was already active),
but most entry points can be disabled (turned into no-ops)
by a single const bool setting atop tracked_objects.cc (in
case folks want to revert the perf-impact of this change).
Transition to faster Now() service on Windows for the
profiler use only.
The TimeTicks::Now() function on Window uses locking
to get a 64 bit time value. This CL transitions
times used for profiling to more directly use a
32 bit Time interface, which is actually what drives the
64 bit TimeTicks. By using the smaller value, we avoid
the need for locks, or even atomic operations for the most
part in the tracking system. On linux, we just down-sample
the standard TimeTicks to 32 bits for consistency (clean
ability to snapshot asyncronously without atomics...
but I should verify that such is helpful to performance).
I've also put in yet more cleanup and refactoring.
tbr=rtenneti
bug=101856
Review URL: http://codereview.chromium.org/8414036
git-svn-id: svn://svn.chromium.org/chrome/trunk/src@107895 0039d316-1c4b-4281-b951-d872f2087c98
|
