One of the things slowing us down is that ItaniumABILanguageRuntime class doesn't cache vtable to types in a map. This causes us, on every step, for every variable, to read the first pointer in a C++ type that could be dynamic and lookup the symbol, possibly in every symbol file (some symbols files on Darwin can end up having thousands of .o files when using DWARF in .o files, so thousands of .o files are searched each time).
This fix caches lldb_private::Address (the resolved vtable symbol address in section + offset format) to TypeAndOrName instances inside the one ItaniumABILanguageRuntime in a process. This allows caching of dynamic types and stops us from always doing deep searches in each file.
<rdar://problem/18890778>
llvm-svn: 270488
This is a pretty straightforward first pass over removing a number of uses of
Mutex in favor of std::mutex or std::recursive_mutex. The problem is that there
are interfaces which take Mutex::Locker & to lock internal locks. This patch
cleans up most of the easy cases. The only non-trivial change is in
CommandObjectTarget.cpp where a Mutex::Locker was split into two.
llvm-svn: 269877
The main issues were:
- Listeners recently were converted over to used by getting a shared pointer to a listener. And when they listened to broadcasters they would get a strong reference added to them meaning the listeners would never go away. This caused memory usage to increase and would cause performance issue if many steps were done.
- The lldb_private::Process private state thread had an issue where if a "stop" contol signal was attempted to be sent to that thread, it could end up not responding in 2 seconds and end up getting cancelled which might cause us to cancel a thread that had a mutex locked and it would deadlock the test.
This change makes broadcasters hold onto weak references to listeners. It also fixes some bad threading code that had races inside of it by making the m_events_mutex be non-recursive and getting rid of fragile use of a Predicate<bool> to say that new events are available, and replacing it with using the m_events_mutex with a new m_events_condition to control access to the events in a safer way.
The private state thread now uses a safer way to communicate that the control event has been received by the private state thread: it makes a EventDataReceipt instance that it attaches to the event that sends the control to the private state thread and used this to synchronize the fact that the private state thread has received the event instead of using a Predicate<bool> to convey the info. When the signal event is received, it will pull the event off of the queue in the private state thread and cause the EventData::DoOnRemoval() to be called, which will signal that the event has been received. This cleans up the signal delivery notification so it doesn't rely on a member variable of the process class to convey the info.
std::shared_ptr<EventDataReceipt> event_receipt_sp(new EventDataReceipt());
m_private_state_control_broadcaster.BroadcastEvent(signal, event_receipt_sp);
<rdar://problem/26256353> Listeners are being kept around longer than they should be due to recent changs
<rdar://problem/26256258> Private process state thread can be cancelled and cause deadlocks in test suite
llvm-svn: 269377
Summary:
This replaces the C-style "void *" baton of the child process monitoring functions with a more
C++-like API taking a std::function. The motivation for this was that it was very difficult to
handle the ownership of the object passed into the callback function -- each caller ended up
implementing his own way of doing it, some doing it better than others. With the new API, one can
just pass a smart pointer into the callback and all of the lifetime management will be handled
automatically.
This has enabled me to simplify the rather complicated handshake in Host::RunShellCommand. I have
left handling of MonitorDebugServerProcess (my original motivation for this change) to a separate
commit to reduce the scope of this change.
Reviewers: clayborg, zturner, emaste, krytarowski
Subscribers: lldb-commits
Differential Revision: http://reviews.llvm.org/D20106
llvm-svn: 269205
When stopping the private state thread, there was a race condition between the time the thread exits (resetting the HostThread object) and the time a Join was attempted, especially in the case of a timeout.
The previous workaround of copying the HostThread object is not enough, since on a Reset the internal thread stuff gets nulled out regardless of which HostThread object actually has Reset called on it, resulting in an attempt to dereference a null pointer on the subsequent call to Join from the copy as well.
Additionally, there was a race between the detach (called when stopping the process) and the stop itself, causing the stop to time out because it was waiting for the private state thread to see the stop state, but it had exited immediately after entering the detached state.
Patch by cameron314
Differential Revision: http://reviews.llvm.org/D19122
llvm-svn: 266733
This patch fixes a bunch of issues that show up on big-endian systems:
- The gnu_libstdcpp.py script doesn't follow the way libstdc++ encodes
bit vectors: it should identify the enclosing *word* and then access
the appropriate bit within that word. Instead, the script simply
operates on bytes. This gives the same result on little-endian
systems, but not on big-endian.
- lldb_private::formatters::WCharSummaryProvider always assumes wchar_t
is UTF16, even though it could also be UTF8 or UTF32. This is mostly
not an issue on little-endian systems, but immediately fails on BE.
Fixed by checking the size of wchar_t like WCharStringSummaryProvider
already does.
- ClangASTContext::GetChildCompilerTypeAtIndex uses uint32_t to access
the virtual base offset stored in the vtable, even though the size
of this field matches the target pointer size according to the C++
ABI. Again, this is mostly not visible on LE, but fails on BE.
- Process::ReadStringFromMemory uses strncmp to search for a terminator
consisting of multiple zero bytes. This doesn't work since strncmp
will stop already at the first zero byte. Use memcmp instead.
Differential Revision: http://reviews.llvm.org/D18983
llvm-svn: 266313
os to "ios" or "macosx" if it is unspecified. For environments
where there genuinely is no os, we don't want to errantly
convert that to ios/macosx, e.g. bare board debugging.
Change PlatformRemoteiOS, PlatformRemoteAppleWatch, and
PlatformRemoteAppleTV to not create themselves if we have
an unspecified OS. Same problem - these are not appropriate
platforms for bare board debugging environments.
Have Process::Attach's logging take place if either
process or target logging is enabled.
<rdar://problem/25592378>
llvm-svn: 265732
Summary:
There was a bug in linux core file handling, where if there was a running process with the same
process id as the id in the core file, the core file debugging would fail, as we would pull some
pieces of information (ProcessInfo structure) from the running process instead of the core file.
I fix this by routing the ProcessInfo requests through the Process class and overriding it in
ProcessElfCore to return correct data.
A (slightly convoluted) test is included.
Reviewers: clayborg, zturner
Subscribers: lldb-commits
Differential Revision: http://reviews.llvm.org/D18697
llvm-svn: 265391
We want to do a better job presenting errors that occur when evaluating
expressions. Key to this effort is getting away from a model where all
errors are spat out onto a stream where the client has to take or leave
all of them.
To this end, this patch adds a new class, DiagnosticManager, which
contains errors produced by the compiler or by LLDB as an expression
is created. The DiagnosticManager can dump itself to a log as well as
to a string. Clients will (in the future) be able to filter out the
errors they're interested in by ID or present subsets of these errors
to the user.
This patch is not intended to change the *users* of errors - only to
thread DiagnosticManagers to all the places where streams are used. I
also attempt to standardize our use of errors a bit, removing trailing
newlines and making clients omit 'error:', 'warning:' etc. and instead
pass the Severity flag.
The patch is testsuite-neutral, with modifications to one part of the
MI tests because it relied on "error: error:" being erroneously
printed. This patch fixes the MI variable handling and the testcase.
<rdar://problem/22864976>
llvm-svn: 263859
to each other. This should remove some infrequent teardown crashes when the
listener is not the debugger's listener.
Processes now need to take a ListenerSP, not a Listener&.
This required changing over the Process plugin class constructors to take a ListenerSP, instead
of a Listener&. Other than that there should be no functional change.
<rdar://problem/24580184> CrashTracer: [USER] Xcode at …ework: lldb_private::Listener::BroadcasterWillDestruct + 39
llvm-svn: 262863
(lldb) command_that_steps_process_thousands_of_times
As the "command_that_steps_process_thousands_of_times" could be a python command that resumed the process thousands of times and in doing so the IOHandlerProcessSTDIO would get pushed when the process resumed, and popped when it stoppped, causing the call to IOHandlerProcessSTDIO::Cancel(). Since the IOHandler thread is currently in IOHandlerEditline::Run() for the command interpreter handling the "command_that_steps_process_thousands_of_times" command, IOHandlerProcessSTDIO::Run() would never get called, even though the IOHandlerProcessSTDIO is on the top of the stack. This caused the command pipe to keep getting 1 bytes written each time the IOHandlerProcessSTDIO::Cancel() was called and eventually we will deadlock since the write buffer is full.
The fix here is to make sure we are in IOHandlerProcessSTDIO::Run() before we write anything to the command pipe, and just call SetIsDone(true) if we are not.
<rdar://problem/22361364>
llvm-svn: 262040
the xcode project file to catch switch statements that have a
case that falls through unintentionally.
Define LLVM_FALLTHROUGH to indicate instances where a case has code
and intends to fall through. This should be in llvm/Support/Compiler.h;
Peter Collingbourne originally checked in there (r237766), then
reverted (r237941) because he didn't have time to mark up all the
'case' statements that were intended to fall through. I put together
a patch to get this back in llvm http://reviews.llvm.org/D17063 but
it hasn't been approved in the past week. I added a new
lldb-private-defines.h to hold the definition for now.
Every place in lldb where there is a comment that the fall-through
is intentional, I added LLVM_FALLTHROUGH to silence the warning.
I haven't tried to identify whether the fallthrough is a bug or
not in the other places.
I haven't tried to add this to the cmake option build flags.
This warning will only work for clang.
This build cleanly (with some new warnings) on macosx with clang
under xcodebuild, but if this causes problems for people on other
configurations, I'll back it out.
llvm-svn: 260930
Seems that the patch was rebased on top of another change which obsoleted the
change but wasnt caught.
Thanks to nbjoerg for pointing this out!
llvm-svn: 258821
Address a couple of instances of -Wreturn-type warning from GCC. The switches
are covered, add an llvm_unreachable to the end of the functions to silence the
warning. NFC.
llvm-svn: 258546
Zachary introduced the 'default' case explicitly to placate a warning in
the Microsoft compiler but that broke clang with -Werror.
The new code should keep both compilers happy.
llvm-svn: 258212
with the one change that ThreadPlanStepOut::ThreadPlanStepOut
will now only advance the return address breakpoint to
the end of a source line, if we have source line debug information.
It will not advance to the end of a Symbol if we lack source line
information. This, or the recognition of the LEAVE instruction
in r257209, would have fixed the regression that Siva was seeing.
Both were good changes, so I've made both.
Original commit message:
Performance improvement: Change lldb so that it puts a breakpoint
on the first branch instruction after a function return (or the end
of a source line), instead of a breakpoint on the return address,
to skip an extra stop & start of the inferior process.
I changed Process::AdvanceAddressToNextBranchInstruction to not
take an optional InstructionList argument - no callers are providing
a cached InstructionList today, and if this function was going to
do that, the right thing to do would be to fill out / use a
DisassemblerSP which is a disassembler with the InstructionList for
this address range.
http://reviews.llvm.org/D15708
<rdar://problem/23309838>
llvm-svn: 257210
puts a breakpoint" it is causing a regression in the TestStepNoDebug
test case on ubuntu 14.04 with gcc 4.9.2. Thanks for the email
Siva. I'll recommit when I've figured out the regression.
llvm-svn: 257138
on the first branch instruction after a function return (or the end
of a source line), instead of a breakpoint on the return address,
to skip an extra stop & start of the inferior process.
I changed Process::AdvanceAddressToNextBranchInstruction to not
take an optional InstructionList argument - no callers are providing
a cached InstructionList today, and if this function was going to
do that, the right thing to do would be to fill out / use a
DisassemblerSP which is a disassembler with the InstructionList for
this address range.
http://reviews.llvm.org/D15708
<rdar://problem/23309838>
llvm-svn: 257117
Demangling complex Boost symbols can exhaust the default stack size. In practice, any thread that calls into LLDB functionality that touches symbols runs this risk. Guaranteeing a reasonable minimum for our own private state thread addressees some known scenarios debugging processes that make use of cpp-netlib.
llvm-svn: 255868
This is a resubmit of r254403, see that commit's message for context. This fixes an issue in the
original commit, where we would incorrectly interrupt the process if the interrupt request came
just as we were about to send the stopped event to the public.
llvm-svn: 254902
On android the symbols exposed by libdl (dlopen, dlclose, dlerror)
prefixed by "__dl_". This change moves the handling of process
load/unload to the platform object and override it for android to
handle the special prefix.
Differential revision: http://reviews.llvm.org/D11465
llvm-svn: 254504
Summary:
The following situation was occuring in TestAttachResume:
- we did a "continue" from a breakpoint (which involves a private start-stop to step over the
breakpoint)
- after receiving the stop-reply from the step-over, we issue a "detach" (which requires a
process interrupt)
- at this moment, the public state is "running", private state is "about-to-be-stopped" (the
stopped event was broadcast, but it was not received yet)
- StopForDestroyOrDetach (public thread) notes the public state is running, sends an interrupt
request to the private thread
- private thread gets the eBroadcastBitInterrupt (before the eStateStopped message), and asks the
process plugin to stop (via Halt())
- process plugin says it has nothing to do as the process is already stopped
- private thread shrugs and carries on. receives the stop event, restores the breakpoint and
resumes the process.
- after a while, the public thread times out and says it failed to stop the process
This patch does the following:
- splits Halt() into two functions, private and public, their usage depends on the context
- public Halt(): sends eBroadcastBitInterrupt to the private thread and waits for the Stop
event
- HaltPrivate(): asks the plugin to stop and makes a note that the halt was requested. When the
next stop event comes it sets the interrupt flag on it.
- removes HijackPrivateProcessEvents(), as the only user (old Halt()) has gone away
- removes the m_currently_handling_event hack, as the new Halt() does not need it
- adds a use_run_lock parameter to public Halt() and WaitForProcessToStop(). This was needed
because RunThreadPlan uses Halt() while holding the run lock and we don't want Halt() to take
it away from him.
Reviewers: clayborg, jingham
Subscribers: lldb-commits
Differential Revision: http://reviews.llvm.org/D14989
llvm-svn: 254403
in places where we check for Triple::IOS. They're mostly the same as far
as lldb is conerned.
.
Also add a base cass implementation for Process::IsAlive - Greg added this
last year but it didn't get upstreamed.
llvm-svn: 252227
Summary:
The code which was preventing the usage of the OS plugin while detach is in
progress also prevented us to update the thread list correctly. This resulted
in an empty thread list, which confused the detaching logic. Change the
condition do only do what it says (disable the usage of the OS plugin).
Reviewers: clayborg, jingham
Subscribers: lldb-commits
Differential Revision: http://reviews.llvm.org/D14201
llvm-svn: 251932
Summary:
ADB packets have a maximum size of 4k. This means the size of memory reads does not affect speed
too much (as long as it fits in one packet). Therefore, I am increasing the default memory read
size for android to 2k. This value is used only if the user has not modified the default
memory-cache-line-size setting.
Reviewers: clayborg, tberghammer
Subscribers: tberghammer, danalbert, srhines, lldb-commits
Differential Revision: http://reviews.llvm.org/D13812
llvm-svn: 250814
Most platforms have "/dev/null". Windows has "nul". Instead of
hardcoding the string /dev/null at various places, make a constant
that contains the correct value depending on the platform, and use
that everywhere instead.
llvm-svn: 250331
* ArchSpec::MergeFrom() would erroneously promote an unspecified
unknown to a specified unknown when both the ArchSpec and the merged
in ArchSpec were both unspecified unknowns. This no longer happens,
which fixes issues with global module cache lookup in some
situations.
* Added ArchSpec::DumpTriple(Stream&) that now properly prints
unspecified unknowns as '*' and specified unknows as 'unknown'.
This makes it trivial to tell the difference between the two.
Converted printing code over ot using DumpTriple() rather than
building from scratch.
* Fixed up a couple places that were not guaranteeing that an
unspecified unknown was recorded as such.
llvm-svn: 250253
set to true, but all plans run by RunThreadPlan need to have this set to false so they will
return control to RunThreadPlan without consulting plans higher on the stack.
Since this seems like a common error, I also modified RunThreadPlan to enforce this behavior.
<rdar://problem/22543166>
llvm-svn: 250084
Summary:
The following situation occured in TestAttachResume:
The inferior was stoped at a breakpoint and we did a continue, immediately followed by a detach.
Since there was a trap instruction under the IP, the continue did a step-over-breakpoint before
resuming the inferior for real. In some cases, the detach command was executed between these two
events (after the step-over stop, but before continue). Here, public state was running, but
private state was stopped. This caused a problem because HaltForDestroyOrDetach was checking the
public state to see whether it needs to stop the process (call Halt()), but Halt() was checking
the private state and concluded that there is nothing for it to do.
Solution: Instead of Halt() call SendAsyncInterrupt(), which will then cause Halt() to be
executed in the context of the private state thread. I also rename HaltForDestroyOrDetach to
reflect it does not call halt directly.
Reviewers: jingham, clayborg
Subscribers: lldb-commits
Differential Revision: http://reviews.llvm.org/D13056
llvm-svn: 248371
The Go runtime schedules user level threads (goroutines) across real threads.
This adds an OS plugin to create memory threads for goroutines.
It supports the 1.4 and 1.5 go runtime.
Differential Revision: http://reviews.llvm.org/D5871
llvm-svn: 247852
Greg Clayton