The problem appeared due to a fact that stubs doesn't create a stack
frame, reusing the stack frame of the caller function. When building
stack traces, the current function is retrieved from PC, and its
callees are retrieved by traversing the stack backwards. Thus, for
stubs, the stub itself was discovered via PC, and then stub's caller's
caller was retrieved from stack.
To fix this problem, a pointer to JSFunction object is now captured
from the topmost stack frame, and is saved into stack trace log
record. Then a simple heuristics is applied whether a referred
function should be added to decoded stack, or not, to avoid reporting
the same function twice (from PC and from the pointer.)
BUG=553
TEST=added to mjsunit/tools/tickprocessor
Review URL: http://codereview.chromium.org/546089
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@3673 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
As this is only needed for internal profiling (not for DevTools),
the following approach had been chosen:
- during snapshot creation, positions of serialized objects inside
a snapshot are logged;
- then during V8 initialization, positions of deserealized objects
are logged;
- those positions are used for retrieving code objects names from
snapshot creation log, which needs to be supplied to tick processor
script.
Positions logging is controlled with the new flag: --log_snapshot_positions.
This flag is turned off by default, and this adds no startup penalty.
To plug this fix to Golem, the following actions are needed:
- logs created using 'mksnapshot' need to be stored along with VM images;
- tick processor script needs to be run with '--snapshot-log=...' cmdline
argument.
BUG=571
Review URL: http://codereview.chromium.org/551062
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@3635 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
This is needed to show calls to DOM in CPU profiles. I can think
of a better approach like adding specific functions into V8 API
for explicitly providing callback names and modifying bindings codegen
appropriately. My plan is as follows:
- submit this CL;
- implement anything I need to process log data and display DOM
calls in profiles;
- think again about adding specific functions and modifying bindings
codegen.
BUG=http://code.google.com/p/chromium/issues/detail?id=27613
Review URL: http://codereview.chromium.org/402100
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@3340 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
- don't engage the processing thread of CPU profiling until the first time profiling is resumed, this saves us a thread allocation for the majority of users;
- don't log shared libraries addresses: this is useless for JS-only profiling, and also consumes time on startup.
Review URL: http://codereview.chromium.org/340013
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@3154 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
Turned on with '--log-producers' flag, also needs '--noinline-new' (this is temporarily), '--log-code', '--log-gc'. Not all allocations are traced (I'm investigating.)
Stacks are stored using weak handles. Thus, when an object is collected, its allocation stack is deleted.
Review URL: http://codereview.chromium.org/267077
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@3069 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
When starting JS profiling under Chromium, a map from function addresses to function names is created. During it, for sourceful scripts, an attempt to access script source is made. This can cause a crash, if a source is an external string, which already has been disposed. We had a similar problem in the past with DebugGetLoadedScripts.
BUG=http://crbug.com/23768
TEST=test-log/Issue23768
Review URL: http://codereview.chromium.org/269003
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@3027 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
The profile is taken together with constructors profile. In theory, it
should represent a complete heap graph. However, this takes a lot of memory,
so it is reduced to a more compact, but still useful form. Namely:
- objects are aggregated by their constructors, except for Array and Object
instances, that are too hetereogeneous;
- for Arrays and Objects, initially every instance is concerned, but then
they are grouped together based on their retainer graph paths similarity (e.g.
if two objects has the same retainer, they are considered equal);
Review URL: http://codereview.chromium.org/200132
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@2903 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
I'm planning to use it in DevTools heap profiler. It is a common scenario in debugging memory leaks to enforce GC, then perform an operation, then enforce GC again to check for non-collected (that is, leaked) objects. Using the existing GC extension isn't possible because it doesn't exposed in the normal operation mode of Chromium.
Review URL: http://codereview.chromium.org/159787
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@2619 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
It is activated with '--log-gc' flag.
JS object size is calculated as its size + size of 'properties' and 'elements' arrays, if they are non-empty. This doesn't take maps, strings, heap numbers, and other shared objects into account.
As Soeren suggested, I've moved ZoneSplayTree from jsregexp to zone, and removed now empty jsregexp-inl header file.
Review URL: http://codereview.chromium.org/159504
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@2570 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
Also, add user time into heap sample begin events to make '--log-gc' flag alone sufficient for producing heap logs (previously, samples times were extracted from scavenge events which are only logged with '--log' flag).
Review URL: http://codereview.chromium.org/149611
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@2461 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
Previous implementation of sampler for OS X was copied from the Linux one. But BSD (OS X) and Linux has a very important difference in signal handling. LinuxThreads doesn't support the notion of process-directed signals. So, the SIGPROF signal was directed to the thread that installed the handler---the V8 thread. But on BSD, signal handling is implemented according to POSIX spec, where process-directed signal is to be handled by an arbitrary selected thread. By a coincidence, in V8's sample shell and in Chromium's test shell, V8's thread was picked almost every time, so sampling seemed working. But not in case of Chromium.
So, I've changed the implementation of profiler sampler to use the same scheme as on Windows---a dedicated thread with high priority is used to periodically pause and sample V8's thread.
Review URL: http://codereview.chromium.org/147150
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@2315 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
It seems that when calling a method that has two overloaded versions like this:
f(char* format, ...)
f(char* format, va_list args)
with a second pointer argument: f("format", pointer), the second version is picked up.
I've found a description of a similar issue here: http://bugs.gentoo.org/63112
So, to resolve this ambiguity, I've named such LogMessageBuilder's Append functions differently.
Review URL: http://codereview.chromium.org/125125
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@2172 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
Code addresses are now written as an offset from the previous address for ticks, code move and delete events. Employed backreference and RLE compression for code move and delete events. This gives additional 30% log size reduction for benchmarks run w/o snapshot.
Overall compression results (compared with the revision of V8 having no compression):
- V8: 70% size reduction for benchmarks run w/o snapshot (for reference, gzip gives 87%)
- Chromium: 65% size reduction for public html version of benchmarks (v4) (for reference, gzip gives 90%)
The one obvious opportunity for improving compression results in Chromium is to compress URLs of scripts.
Review URL: http://codereview.chromium.org/125114
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@2162 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
Two techniques are involved:
- compress repeated line ends (common stack beginnings) by using back references;
- do RLE compression of repeated tick events.
This gives only 5% size reduction on benchmarks run, but this is because tick events are only comprise 10% of file size. Under Chromium winnings are bigger because long repeated samples of idleness are now compressed into a single line.
Tickprocessor will be updated in the next patch.
Review URL: http://codereview.chromium.org/123012
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@2147 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
This is a trivial per-row compression:
- short aliases are introduced for events and code creation tags;
- in tick events, offsets are used instead of absolute addresses;
- removed 'code-allocation' event, as it seems not used.
The first two options are depend on the new flag: 'compress-log', which is off by default.
On benchmarks run w/o snapshot, this gives 45% log size reduction.
Review URL: http://codereview.chromium.org/119304
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@2122 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
My assumption that log initialization happens somewhere near the stack's bottom is true for V8's sample shell but isn't true for Chromium, causing many otherwise valid stack addresses to be thrown out. The solution proposed is to save stack pointer value for the outermost JS function in ThreadLocalTop similar to c_entry_fp.
Implemented only for IA-32. Currently I'm not dealing with profiling on ARM and x86-64 anyway.
Review URL: http://codereview.chromium.org/112082
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@2086 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
When profiler's memory buffer is filled up, profiling is stopped and it is ensured that the last record in the buffer is "profiler,\"pause\"" thus making the end of profiling session explicit. Otherwise DevTools Profiler would need to guess whether the current profiling session has been stopped.
Tested with Chromium.
Review URL: http://codereview.chromium.org/115859
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@2072 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
The goal of this change is to allow longer profiling sessions and preserve memory when profiler isn't started. The buffer starts with 64K and grows until it reaches the upper limit, which is currently set to 50MB --- according to my evaluations, this is enough for at least 20 minutes of GMail profiling. As we're planning to introduce compression for the profiler log, this time boundary will be significantly increased soon.
To make possible unit testing of the new component, I've factored out Logger's utility classes into a separate source file: log-utils.h/cc. Log and LogMessageBuilder are moved there from log.cc without any semantical changes.
Review URL: http://codereview.chromium.org/115814
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@2067 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
This issue was raised by Brett Wilson while reviewing my changelist for readability. Craig Silverstein (one of C++ SG maintainers) confirmed that we should declare one namespace per line. Our way of namespaces closing seems not violating style guides (there is no clear agreement on it), so I left it intact.
Review URL: http://codereview.chromium.org/115756
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@2038 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
This is intended to be used with Chromium. When in resource-saving mode, profiler doesn't consume any resources (sampler and logging is off) until resumed. Then again, when profiler is paused, sampling and logging are turned off.
Tested under Linux and Windows. Also have done preliminary testing with Chromium.
Review URL: http://codereview.chromium.org/113762
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@2036 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
The goal is to make possible having --prof flag always enabled in Chromium. Currently we can't do this because --prof causes compiler and gc to log code creations / moves / deletes which aren't needed until we start profiling. With LogCompiledFunctions it will be possible not to log anything until we start profiling. When started, the current map of compiled functions will be logged and compiler / gc logging will be enabled to update current state. When profling is stopped, logging will be turned off again.
Funny that testing code is actually much longer and complex than function code.
Review URL: http://codereview.chromium.org/112036
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@2009 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
This will enable reading profiler log in Chrome. The current implementation of memory buffer is trivial (fixed size buffer, no memory recycling) but enough to start end-to-end DevTools Profiler implementation. Later it will be enhanced.
Review URL: http://codereview.chromium.org/108011
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@1870 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
