to reflect the new license.
We understand that people may be surprised that we're moving the header
entirely to discuss the new license. We checked this carefully with the
Foundation's lawyer and we believe this is the correct approach.
Essentially, all code in the project is now made available by the LLVM
project under our new license, so you will see that the license headers
include that license only. Some of our contributors have contributed
code under our old license, and accordingly, we have retained a copy of
our old license notice in the top-level files in each project and
repository.
llvm-svn: 351636
Summary:
This change builds upon D54989, which removes memory allocation from the
critical path of the profiling implementation. This also changes the API
for the profile collection service, to take ownership of the memory and
associated data structures per-thread.
The consolidation of the memory allocation allows us to do two things:
- Limits the amount of memory used by the profiling implementation,
associating preallocated buffers instead of allocating memory
on-demand.
- Consolidate the memory initialisation and cleanup by relying on the
buffer queue's reference counting implementation.
We find a number of places which also display some problematic
behaviour, including:
- Off-by-factor bug in the allocator implementation.
- Unrolling semantics in cases of "memory exhausted" situations, when
managing the state of the function call trie.
We also add a few test cases which verify our understanding of the
behaviour of the system, with important edge-cases (especially for
memory-exhausted cases) in the segmented array and profile collector
unit tests.
Depends on D54989.
Reviewers: mboerger
Subscribers: dschuff, mgorny, dmgreen, jfb, llvm-commits
Differential Revision: https://reviews.llvm.org/D55249
llvm-svn: 348568
This reverts commit r348455, with some additional changes:
- Work-around deficiency of gcc-4.8 by duplicating the implementation of
`AppendEmplace` in `Append`, but instead of using brace-init for the
copy construction, use a placement new explicitly calling the copy
constructor.
llvm-svn: 348563
Continuation of D54989.
Additional changes:
- Use `.AppendEmplace(...)` instead of `.Append(Type{...})` to appease
GCC 4.8 with confusion on when an initializer_list is used as
opposed to a temporary aggregate initialized object.
llvm-svn: 348438
.. and also the follow-ups r348336 r348338.
It broke stand-alone compiler-rt builds with GCC 4.8:
In file included from /work/llvm/projects/compiler-rt/lib/xray/xray_function_call_trie.h:20:0,
from /work/llvm/projects/compiler-rt/lib/xray/xray_profile_collector.h:21,
from /work/llvm/projects/compiler-rt/lib/xray/xray_profile_collector.cc:15:
/work/llvm/projects/compiler-rt/lib/xray/xray_segmented_array.h: In instantiation of ‘T* __xray::Array<T>::AppendEmplace(Args&& ...) [with Args = {const __xray::FunctionCallTrie::mergeInto(__xray::FunctionCallTrie&) const::NodeAndTarget&}; T = __xray::FunctionCallTrie::mergeInto(__xray::FunctionCallTrie&) const::NodeAndTarget]’:
/work/llvm/projects/compiler-rt/lib/xray/xray_segmented_array.h:383:71: required from ‘T* __xray::Array<T>::Append(const T&) [with T = __xray::FunctionCallTrie::mergeInto(__xray::FunctionCallTrie&) const::NodeAndTarget]’
/work/llvm/projects/compiler-rt/lib/xray/xray_function_call_trie.h:517:54: required from here
/work/llvm/projects/compiler-rt/lib/xray/xray_segmented_array.h:378:5: error: could not convert ‘{std::forward<const __xray::FunctionCallTrie::mergeInto(__xray::FunctionCallTrie&) const::NodeAndTarget&>((* & args#0))}’ from ‘<brace-enclosed initializer list>’ to ‘__xray::FunctionCallTrie::mergeInto(__xray::FunctionCallTrie&) const::NodeAndTarget’
new (AlignedOffset) T{std::forward<Args>(args)...};
^
/work/llvm/projects/compiler-rt/lib/xray/xray_segmented_array.h: In instantiation of ‘T* __xray::Array<T>::AppendEmplace(Args&& ...) [with Args = {const __xray::profileCollectorService::{anonymous}::ThreadTrie&}; T = __xray::profileCollectorService::{anonymous}::ThreadTrie]’:
/work/llvm/projects/compiler-rt/lib/xray/xray_segmented_array.h:383:71: required from ‘T* __xray::Array<T>::Append(const T&) [with T = __xray::profileCollectorService::{anonymous}::ThreadTrie]’
/work/llvm/projects/compiler-rt/lib/xray/xray_profile_collector.cc:98:34: required from here
/work/llvm/projects/compiler-rt/lib/xray/xray_segmented_array.h:378:5: error: could not convert ‘{std::forward<const __xray::profileCollectorService::{anonymous}::ThreadTrie&>((* & args#0))}’ from
‘<brace-enclosed initializer list>’ to ‘__xray::profileCollectorService::{anonymous}::ThreadTrie’
/work/llvm/projects/compiler-rt/lib/xray/xray_segmented_array.h: In instantiation of ‘T* __xray::Array<T>::AppendEmplace(Args&& ...) [with Args = {const __xray::profileCollectorService::{anonymous}::ProfileBuffer&}; T = __xray::profileCollectorService::{anonymous}::ProfileBuffer]’:
/work/llvm/projects/compiler-rt/lib/xray/xray_segmented_array.h:383:71: required from ‘T* __xray::Array<T>::Append(const T&) [with T = __xray::profileCollectorService::{anonymous}::ProfileBuffer]
’
/work/llvm/projects/compiler-rt/lib/xray/xray_profile_collector.cc:244:44: required from here
/work/llvm/projects/compiler-rt/lib/xray/xray_segmented_array.h:378:5: error: could not convert ‘{std::forward<const __xray::profileCollectorService::{anonymous}::ProfileBuffer&>((* & args#0))}’ from ‘<brace-enclosed initializer list>’ to ‘__xray::profileCollectorService::{anonymous}::ProfileBuffer’
> Summary:
> This change makes the allocator and function call trie implementations
> move-aware and remove the FunctionCallTrie's reliance on a
> heap-allocated set of allocators.
>
> The change makes it possible to always have storage associated with
> Allocator instances, not necessarily having heap-allocated memory
> obtainable from these allocator instances. We also use thread-local
> uninitialised storage.
>
> We've also re-worked the segmented array implementation to have more
> precondition and post-condition checks when built in debug mode. This
> enables us to better implement some of the operations with surrounding
> documentation as well. The `trim` algorithm now has more documentation
> on the implementation, reducing the requirement to handle special
> conditions, and being more rigorous on the computations involved.
>
> In this change we also introduce an initialisation guard, through which
> we prevent an initialisation operation from racing with a cleanup
> operation.
>
> We also ensure that the ThreadTries array is not destroyed while copies
> into the elements are still being performed by other threads submitting
> profiles.
>
> Note that this change still has an issue with accessing thread-local
> storage from signal handlers that are instrumented with XRay. We also
> learn that with the testing of this patch, that there will be cases
> where calls to mmap(...) (through internal_mmap(...)) might be called in
> signal handlers, but are not async-signal-safe. Subsequent patches will
> address this, by re-using the `BufferQueue` type used in the FDR mode
> implementation for pre-allocated memory segments per active, tracing
> thread.
>
> We still want to land this change despite the known issues, with fixes
> forthcoming.
>
> Reviewers: mboerger, jfb
>
> Subscribers: jfb, llvm-commits
>
> Differential Revision: https://reviews.llvm.org/D54989
llvm-svn: 348346
Summary:
This change makes the allocator and function call trie implementations
move-aware and remove the FunctionCallTrie's reliance on a
heap-allocated set of allocators.
The change makes it possible to always have storage associated with
Allocator instances, not necessarily having heap-allocated memory
obtainable from these allocator instances. We also use thread-local
uninitialised storage.
We've also re-worked the segmented array implementation to have more
precondition and post-condition checks when built in debug mode. This
enables us to better implement some of the operations with surrounding
documentation as well. The `trim` algorithm now has more documentation
on the implementation, reducing the requirement to handle special
conditions, and being more rigorous on the computations involved.
In this change we also introduce an initialisation guard, through which
we prevent an initialisation operation from racing with a cleanup
operation.
We also ensure that the ThreadTries array is not destroyed while copies
into the elements are still being performed by other threads submitting
profiles.
Note that this change still has an issue with accessing thread-local
storage from signal handlers that are instrumented with XRay. We also
learn that with the testing of this patch, that there will be cases
where calls to mmap(...) (through internal_mmap(...)) might be called in
signal handlers, but are not async-signal-safe. Subsequent patches will
address this, by re-using the `BufferQueue` type used in the FDR mode
implementation for pre-allocated memory segments per active, tracing
thread.
We still want to land this change despite the known issues, with fixes
forthcoming.
Reviewers: mboerger, jfb
Subscribers: jfb, llvm-commits
Differential Revision: https://reviews.llvm.org/D54989
llvm-svn: 348335
Summary:
This change allows us to handle allocator exhaustion properly in the
segmented array implementation. Before this change, we relied on the
caller of the `trim` function to provide a valid number of elements to
trim. This change allows us to do the right thing in case the elements
to trim is greater than the size of the container.
Reviewers: mboerger, eizan
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D53484
llvm-svn: 344880
Summary:
In this change we apply `XRAY_NEVER_INSTRUMENT` to more functions in the
profiling implementation to ensure that these never get instrumented if
the compiler used to build the library is capable of doing XRay
instrumentation.
We also consolidate all the allocators into a single header
(xray_allocator.h) which sidestep the use of the internal allocator
implementation in sanitizer_common.
This addresses more cases mentioned in llvm.org/PR38577.
Reviewers: mboerger, eizan
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D51776
llvm-svn: 341647
Summary:
We avoid using dynamic memory allocated with the internal allocator in
the profile collection service used by profiling mode. We use aligned
storage for globals and in-struct storage of objects we dynamically
initialize.
We also remove the dependency on `Vector<...>` which also internally
uses the dynamic allocator in sanitizer_common (InternalAlloc) in favour
of the XRay allocator and segmented array implementation.
This change addresses llvm.org/PR38577.
Reviewers: eizan
Reviewed By: eizan
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D50782
llvm-svn: 339978
Summary:
This is a follow-on to D49217 which simplifies and optimises the
implementation of the segmented array. In this patch we co-locate the
book-keeping for segments in the `__xray::Array<T>` with the data it's
managing. We take the chance in this patch to actually rename `Chunk` to
`Segment` to better align with the high-level description of the
segmented array.
With measurements using benchmarks landed in D48879, we've identified
that calls to `pthread_getspecific` started dominating the cycles, which
led us to revert the change made in D49217 to use C++ thread_local
initialisation instead (it reduces the cost by a huge margin, since we
save one PLT-based call to pthread functions in the hot path). In
particular, this is in `__xray::getThreadLocalData()`.
We also took the opportunity to remove the least-common-multiple based
calculation and instead pack as much data into segments of the array.
This greatly simplifies the API of the container which hides as much of
the implementation details as possible. For instance, we calculate the
number of elements we need for the each segment internally in the Array
instead of making it part of the type.
With the changes here, we're able to get a measurable improvement on the
performance of profiling mode on top of what D48879 already provides.
Depends on D48879.
Reviewers: kpw, eizan
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D49363
llvm-svn: 337343
Summary:
This change simplifies the XRay Allocator implementation to self-manage
an mmap'ed memory segment instead of using the internal allocator
implementation in sanitizer_common.
We've found through benchmarks and profiling these benchmarks in D48879
that using the internal allocator in sanitizer_common introduces a
bottleneck on allocating memory through a central spinlock. This change
allows thread-local allocators to eliminate contention on the
centralized allocator.
To get the most benefit from this approach, we also use a managed
allocator for the chunk elements used by the segmented array
implementation. This gives us the chance to amortize the cost of
allocating memory when creating these internal segmented array data
structures.
We also took the opportunity to remove the preallocation argument from
the allocator API, simplifying the usage of the allocator throughout the
profiling implementation.
In this change we also tweak some of the flag values to reduce the
amount of maximum memory we use/need for each thread, when requesting
memory through mmap.
Depends on D48956.
Reviewers: kpw, eizan
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D49217
llvm-svn: 337342
Summary:
We found a bug while working on a benchmark for the profiling mode which
manifests as a segmentation fault in the profiling handler's
implementation. This change adds unit tests which replicate the
issues in isolation.
We've tracked this down as a bug in the implementation of the Freelist
in the `xray::Array` type. This happens when we trim the array by a
number of elements, where we've been incorrectly assigning pointers for
the links in the freelist of chunk nodes. We've taken the chance to add
more debug-only assertions to the code path and allow us to verify these
assumptions in debug builds.
In the process, we also took the opportunity to use iterators to
implement both `front()` and `back()` which exposes a bug in the
iterator decrement operation. In particular, when we decrement past a
chunk size boundary, we end up moving too far back and reaching the
`SentinelChunk` prematurely.
This change unblocks us to allow for contributing the non-crashing
version of the benchmarks in the test-suite as well.
Reviewers: kpw
Subscribers: mgorny, llvm-commits
Differential Revision: https://reviews.llvm.org/D48653
llvm-svn: 336644
Summary:
This change is part of the larger XRay Profiling Mode effort.
Here we implement an arena allocator, for fixed sized buffers used in a
segmented array implementation. This change adds the segmented array
data structure, which relies on the allocator to provide and maintain
the storage for the segmented array.
Key features of the `Allocator` type:
* It uses cache-aligned blocks, intended to host the actual data. These
blocks are cache-line-size multiples of contiguous bytes.
* The `Allocator` has a maximum memory budget, set at construction
time. This allows us to cap the amount of data each specific
`Allocator` instance is responsible for.
* Upon destruction, the `Allocator` will clean up the storage it's
used, handing it back to the internal allocator used in
sanitizer_common.
Key features of the `Array` type:
* Each segmented array is always backed by an `Allocator`, which is
either user-provided or uses a global allocator.
* When an `Array` grows, it grows by appending a segment that's
fixed-sized. The size of each segment is computed by the number of
elements of type `T` that can fit into cache line multiples.
* An `Array` does not return memory to the `Allocator`, but it can keep
track of the current number of "live" objects it stores.
* When an `Array` is destroyed, it will not return memory to the
`Allocator`. Users should clean up the `Allocator` independently of
the `Array`.
* The `Array` type keeps a freelist of the chunks it's used before, so
that trimming and growing will re-use previously allocated chunks.
These basic data structures are used by the XRay Profiling Mode
implementation to implement efficient and cache-aware storage for data
that's typically read-and-write heavy for tracking latency information.
We're relying on the cache line characteristics of the architecture to
provide us good data isolation and cache friendliness, when we're
performing operations like searching for elements and/or updating data
hosted in these cache lines.
Reviewers: echristo, pelikan, kpw
Subscribers: mgorny, llvm-commits
Differential Revision: https://reviews.llvm.org/D45756
llvm-svn: 331141
Chandler Carruth