Stack iterator takes return address based on the frame pointer (ebp) and detects JS frames based on value at fp + StandardFrameConstants::kMarkerOffset. So in order the iterator to work correctly this values should be already setup for the current function. Stack frame is constructed at the very beginning of JS function code and destroyed before return. If sample is taken before before the frame construction is completed or after it was destroyed the stack iterator will wrongly think that FP points at the current functions frame base and will skip callers frame. To avoid this we mark code ranges where stack frame doesn't exist and completely ignore such samples.
This fixes cctest/test-cpu-profiler/CollectCpuProfile flakiness.
BUG=v8:2628
R=jkummerow@chromium.org
Review URL: https://codereview.chromium.org/14253015
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@14670 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
The new test checks full CPU profiling cycle: using public
V8 API it starts profiling, executes a script, stops profiling
and analyzes collected profile to check that its top-down
tree has expected strutcture. The script that is being profiled
is guaranteed to run > 200ms to make sure enough samples
are collected.
To avoid possible flakiness due to non-deterministic time required
to start new thread on varios OSs when Sampler and ProfilerEventsProcessor
threads are being started the main thread is blocked until the threads
are running.
Also I removed the heuristic in profile-generator.cc where we try
to figure out if the value on top of the sampled stack is return address
of some frameless stub invocation. The code periodically gives false positive
with the new test ending up in an extra node in the collected cpu profile.
After discussion with jkummerow@ we concluded that the logic is too fragile
and that we can address frameless stub invocations in a more reliable way
later should they have a noticeable effect on cpu profiling.
BUG=None
Review URL: https://codereview.chromium.org/13627002
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@14205 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
CPU profiler API is extended with methods that allow to retrieve individual samples from profile. Each sample is presented as a pointer to a node in the top-down profile tree. The samples will let us tie JS performance to time.
BUG=None
Review URL: https://codereview.chromium.org/12919002
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@13980 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
I tried to limit the use of v8::Isolate::GetCurrent() and v8::internal::Isolate::Current() as much as possible, but sometimes this would have involved restructuring tests quite a bit, which is better left for a separate CL.
BUG=v8:2487
Review URL: https://codereview.chromium.org/12716010
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@13953 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
The patch is based on the previous one that was rolled out: https://code.google.com/p/v8/source/detail?r=12985
On Linux sampling for CPU profiler is initiated on the profiler event processor thread, other platforms to follow.
CPU profiler continues to use SamplingCircularQueue, we will replave it with a single sample buffer when Mac and Win ports support profiling on the event processing thread.
When --prof option is specified profiling is initiated either on the profiler event processor thread if CPU profiler is on or on the SignalSender thread as it used to if no CPU profiles are being collected.
ProfilerEventsProcessor::ProcessEventsAndDoSample now waits in a tight loop, processing collected samples until sampling interval expires. To save CPU resources I'm planning to change that to use nanosleep as only one sample is expected in the queue at any point.
BUG=v8:2364
Review URL: https://codereview.chromium.org/12321046
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@13735 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
- perform CPU profiler sampling in the sampler thread as we used to;
- skip sampling in the sampling thread if processing thread is running;
- only install SIGPROF handler when CPU profiling is enabled.
BUG=v8:2364
Review URL: https://codereview.chromium.org/11231002
Patch from Sergey Rogulenko <rogulenko@google.com> and Andrey Kosyakov <caseq@chromium.org>.
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12985 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
The preprocessor defines ENABLE_LOGGING_AND_PROFILING and ENABLE_VMSTATE_TRACKING has been removed as these where required to be turned on for Crankshaft to work. To re-enable reducing the binary size by leaving out heap and CPU profiler a new set of defines needs to be created.
R=ager@chromium.org
BUG=v8:1271
TEST=all
Review URL: http://codereview.chromium.org//7350014
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@8622 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
Thread class was receiving an isolate parameter by default.
This approact violates the assumption that only VM threads
can have an associated isolate, and can lead to troubles,
because accessing the same isolate from different threads
leads to race conditions.
This was found by investigating mysterious failures of the
CPU profiler layout test on Linux Chromium. As almost all
threads were associated with some isolate, the sampler was
trying to sample them.
As a side effect, we have also fixed the DebuggerAgent test.
Thanks to Vitaly for help in fixing isolates handling!
R=vitalyr@chromium.org
BUG=none
TEST=none
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@8259 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
The main issue was due to multiple recompilations of functions. Now
code objects are grouped by function using SFI object address.
JSFunction objects are no longer tracked, instead we track SFI object
moves. To pick a correct code version, we now sample return addresses
instead of JSFunction addresses.
tools/{linux|mac|windows}-tickprocessor scripts differentiate
between code optimization states for the same function
(using * and ~ prefixes introduced earlier).
DevTools CPU profiler treats all variants of function code as
a single function.
ll_prof treats each optimized variant as a separate entry, because
it can disassemble each one of them.
tickprocessor.py not updated -- it is deprecated and will be removed.
BUG=v8/1087,b/3178160
TEST=all existing tests pass, including Chromium layout tests
Review URL: http://codereview.chromium.org/6551011
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@6902 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
As several pages can run in a single V8 instance, it is possible to
have functions from different security contexts intermixed in a single
CPU profile. To avoid exposing function names from one page to
another, filtering is introduced.
The basic idea is that instead of capturing return addresses from
stack, we're now capturing JSFunction addresses (as we anyway work
only with JS stack frames.) Each JSFunction can reach out for
context's security token. When providing a profile to a page, the
profile is filtered using the security token of caller page. Any
functions with different security tokens are filtered out (yes, we
only do fast path check for now) and their ticks are attributed to
their parents.
Review URL: http://codereview.chromium.org/2083005
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@4673 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
The simple formula "ms = ticks * sampler_interval" doesn't work,
because e.g. on Linux, the actual sampling rate can be 5 times
lower than the one set up in the code. To calculate actual sampling
rate, current time is periodically queried and processed along with
actual sampling ticks count.
Review URL: http://codereview.chromium.org/1539038
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@4427 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
