https://reviews.llvm.org/D70447 switched to `ManagedStatic` to work
around race conditions in MSVC runtimes and the MinGW runtime.
However, `ManagedStatic` is not suitable for other platforms.
However, this workaround is not suitable for other platforms (#66974).
lld::fatal calls exitLld(1), which calls `llvm_shutdown` to destroy
`ManagedStatic` objects. The worker threads will finish and invoke TLS
destructors (glibc `__call_tls_dtors`). This can lead to race conditions
if other threads attempt to access TLS objects that have already been
destroyed.
While lld's early exit mechanism needs more work, I believe Parallel.cpp
should avoid this pitfall as well.
Pull Request: https://github.com/llvm/llvm-project/pull/102989
PerThreadBumpPtrAllocator allows separating allocations by thread id.
That makes allocations race free. It is possible because
ThreadPoolExecutor class creates threads, keeps them until
the destructor of ThreadPoolExecutor is called, and assigns ids
to the threads. Thus PerThreadBumpPtrAllocator should be used with only
threads created by ThreadPoolExecutor. This allocator is useful when
thread safe BumpPtrAllocator is needed.
Reviewed By: MaskRay, dexonsmith, andrewng
Differential Revision: https://reviews.llvm.org/D142318
This patch changes check for nested TaskGroups so that it allows
parallel execution for TaskGroups. Following pattern would not work
parallelly with current check:
std::function<void()> Fn = [&]() {
parallel::TaskGroup tg;
tg.spawn([&]() { });
};
ThreadPool Pool;
Pool.async(Fn);
Pool.async(Fn);
Pool.wait();
One of the TaskGroup would work sequentially as current check
verifies overall number of TaskGroup. Two not nested
TaskGroups can work parallelly but current check prevents this.
Also this patch avoids parallel mode for TaskGroup
in parallel::strategy.ThreadsRequested == 1 case.
This patch is a followup of discussion from D142318
Differential Revision: https://reviews.llvm.org/D148984
That patch adds a check for threadIndex being used with only threads
created by ThreadPoolExecutor. This helps catch two types of errors:
1. If a thread is created not by ThreadPoolExecutor its index may clash
with the index of another thread. Using threadIndex, in that case, may
lead to a data race.
2. Index of the main thread(threadIndex == 0) currently clashes with
the index of thread0 in ThreadPoolExecutor threads. That may lead
to a data race if main thread and thread0 are executed concurrently.
This patch allows execution tasks on the main thread only in case
parallel::strategy.ThreadsRequested == 1. In all other cases,
assertions check that threadIndex != UINT_MAX(i.e. that task
is executed on a thread created by ThreadPoolExecutor).
Differential Revision: https://reviews.llvm.org/D148916
This patch allows to specify that some part of tasks should be
done in sequential order. It makes it possible to not use
condition operator for separating sequential tasks:
TaskGroup tg;
for () {
if(condition) ==> tg.spawn([](){fn();}, condition)
fn();
else
tg.spawn([](){fn();});
}
It also prevents execution on main thread. Which allows adding
checks for getThreadIndex() function discussed in D142318.
The patch also replaces std::stack with std::deque in the
ThreadPoolExecutor to have natural execution order in case
(parallel::strategy.ThreadsRequested == 1).
Differential Revision: https://reviews.llvm.org/D148728
The llvm::parallelFor() uses threads created by ThreadPoolExecutor as well as main thread.
The index for the main thread matches with the index for the first thread created by ThreadPoolExecutor.
It results in that getThreadIndex returns the same value for different threads.
To avoid thread index clashing - do not use main thread for llvm::parallelFor():
parallel::TaskGroup TG;
for (; Begin + TaskSize < End; Begin += TaskSize) {
TG.spawn([=, &Fn] {
for (size_t I = Begin, E = Begin + TaskSize; I != E; ++I)
Fn(I);
});
}
for (; Begin != End; ++Begin) <<<< executed by main thread.
Fn(Begin); <<<<
return; <<<<
Differential Revision: https://reviews.llvm.org/D142317
On Unix platforms, this wrapper function is inline, so it should
expand to the same direct access to the thread local variable. On
Windows, it's a non-inline function within Parallel.cpp, allowing
making the thread_local variable static.
Windows Native TLS doesn't support direct access to thread local
variables in a different DLL, and GCC/binutils on Windows occasionally
has problems with non-static thread local variables too.
This fixes mingw dylib builds with native TLS after
e6aebff67426fa0f9779a0c19d6188a043bf15e7.
At the same time, move the whole thread local variable within
#if LLVM_ENABLE_THREADS
to fix builds without threading support.
Differential Revision: https://reviews.llvm.org/D133759
* Change `Symbol::flags` to a `std::atomic<uint16_t>`
* Add `llvm::parallel::threadIndex` as a thread-local non-negative integer
* Add `relocsVec` to part.relaDyn and part.relrDyn so that relative relocations can be added without a mutex
* Arbitrarily change -z nocombreloc to move relative relocations to the end. Disable parallelism for deterministic output.
MIPS and PPC64 use global states for relocation scanning. Keep serial scanning.
Speed-up with mimalloc and --threads=8 on an Intel Skylake machine:
* clang (Release): 1.27x as fast
* clang (Debug): 1.06x as fast
* chrome (default): 1.05x as fast
* scylladb (default): 1.04x as fast
Speed-up with glibc malloc and --threads=16 on a ThunderX2 (AArch64):
* clang (Release): 1.31x as fast
* scylladb (default): 1.06x as fast
Reviewed By: andrewng
Differential Revision: https://reviews.llvm.org/D133003
We currently process one OutputSection at a time and for each OutputSection
write contained input sections in parallel. This strategy does not leverage
multi-threading well. Instead, parallelize writes of different OutputSections.
The default TaskSize for parallelFor often leads to inferior sharding. We
prepare the task in the caller instead.
* Move llvm::parallel::detail::TaskGroup to llvm::parallel::TaskGroup
* Add llvm::parallel::TaskGroup::execute.
* Change writeSections to declare TaskGroup and pass it to writeTo.
Speed-up with --threads=8:
* clang -DCMAKE_BUILD_TYPE=Release: 1.11x as fast
* clang -DCMAKE_BUILD_TYPE=Debug: 1.10x as fast
* chrome -DCMAKE_BUILD_TYPE=Release: 1.04x as fast
* scylladb build/release: 1.09x as fast
On M1, many benchmarks are a small fraction of a percentage faster. Mozilla showed the largest difference with the patch being about 1.03x as fast.
Differential Revision: https://reviews.llvm.org/D131247
Patch created by running:
rg -l parallelForEachN | xargs sed -i '' -c 's/parallelForEachN/parallelFor/'
No behavior change.
Differential Revision: https://reviews.llvm.org/D128140
* Merge parallel_for_each into parallelForEach (this removes 1 `Fn(...)` call)
* Change parallelForEach to use parallelForEachN
* Move parallelForEachN into Parallel.cpp
My x86-64 `lld` executable is 100KiB smaller.
No noticeable difference in performance.
Reviewed By: lattner
Differential Revision: https://reviews.llvm.org/D117510
--no-threads is a name copied from gold.
gold has --no-thread, --thread-count and several other --thread-count-*.
There are needs to customize the number of threads (running several lld
processes concurrently or customizing the number of LTO threads).
Having a single --threads=N is a straightforward replacement of gold's
--no-threads + --thread-count.
--no-threads is used rarely. So just delete --no-threads instead of
keeping it for compatibility for a while.
If --threads= is specified (ELF,wasm; COFF /threads: is similar),
--thinlto-jobs= defaults to --threads=,
otherwise all available hardware threads are used.
There is currently no way to override a --threads={1,2,...}. It is still
a debate whether we should use --threads=all.
Reviewed By: rnk, aganea
Differential Revision: https://reviews.llvm.org/D76885
The goal of this patch is to maximize CPU utilization on multi-socket or high core count systems, so that parallel computations such as LLD/ThinLTO can use all hardware threads in the system. Before this patch, on Windows, a maximum of 64 hardware threads could be used at most, in some cases dispatched only on one CPU socket.
== Background ==
Windows doesn't have a flat cpu_set_t like Linux. Instead, it projects hardware CPUs (or NUMA nodes) to applications through a concept of "processor groups". A "processor" is the smallest unit of execution on a CPU, that is, an hyper-thread if SMT is active; a core otherwise. There's a limit of 32-bit processors on older 32-bit versions of Windows, which later was raised to 64-processors with 64-bit versions of Windows. This limit comes from the affinity mask, which historically is represented by the sizeof(void*). Consequently, the concept of "processor groups" was introduced for dealing with systems with more than 64 hyper-threads.
By default, the Windows OS assigns only one "processor group" to each starting application, in a round-robin manner. If the application wants to use more processors, it needs to programmatically enable it, by assigning threads to other "processor groups". This also means that affinity cannot cross "processor group" boundaries; one can only specify a "preferred" group on start-up, but the application is free to allocate more groups if it wants to.
This creates a peculiar situation, where newer CPUs like the AMD EPYC 7702P (64-cores, 128-hyperthreads) are projected by the OS as two (2) "processor groups". This means that by default, an application can only use half of the cores. This situation could only get worse in the years to come, as dies with more cores will appear on the market.
== The problem ==
The heavyweight_hardware_concurrency() API was introduced so that only *one hardware thread per core* was used. Once that API returns, that original intention is lost, only the number of threads is retained. Consider a situation, on Windows, where the system has 2 CPU sockets, 18 cores each, each core having 2 hyper-threads, for a total of 72 hyper-threads. Both heavyweight_hardware_concurrency() and hardware_concurrency() currently return 36, because on Windows they are simply wrappers over std:🧵:hardware_concurrency() -- which can only return processors from the current "processor group".
== The changes in this patch ==
To solve this situation, we capture (and retain) the initial intention until the point of usage, through a new ThreadPoolStrategy class. The number of threads to use is deferred as late as possible, until the moment where the std::threads are created (ThreadPool in the case of ThinLTO).
When using hardware_concurrency(), setting ThreadCount to 0 now means to use all the possible hardware CPU (SMT) threads. Providing a ThreadCount above to the maximum number of threads will have no effect, the maximum will be used instead.
The heavyweight_hardware_concurrency() is similar to hardware_concurrency(), except that only one thread per hardware *core* will be used.
When LLVM_ENABLE_THREADS is OFF, the threading APIs will always return 1, to ensure any caller loops will be exercised at least once.
Differential Revision: https://reviews.llvm.org/D71775
Changed ThreadPoolExecutor to no longer use detached threads and instead
to join threads on destruction. This is to prevent intermittent crashing
on Windows when doing a normal full exit, e.g. via exit().
Changed ThreadPoolExecutor to be a ManagedStatic so that it can be
stopped on llvm_shutdown(). Without this, it would only be stopped in
the destructor when doing a full exit. This is required to avoid
intermittent crashing on Windows due to a race condition between the
ThreadPoolExecutor starting up threads and the process doing a fast
exit, e.g. via _exit().
The Windows crashes appear to only occur with the MSVC static runtimes
and are more frequent with the debug static runtime.
These changes also prevent intermittent deadlocks on exit with the MinGW
runtime.
Differential Revision: https://reviews.llvm.org/D70447
r179397 added Parallel.h and implemented it terms of concrt in 2013.
In 2015, a cross-platform implementation of the functions has appeared
and is in use everywhere but on Windows (r232419). r246219 hints that
<thread> had issues in MSVC2013, but r296906 suggests they've been fixed
now that we require 2015+.
So remove the concrt code. It's less code, and it sounds like concrt has
conceptual and performance issues, see PR41198.
I built blink_core.dll in a debug component build with full symbols and
in a release component build without any symbols. I couldn't measure a
performance difference for linking blink_core.dll before and after this
patch.
Differential Revision: https://reviews.llvm.org/D68820
llvm-svn: 374421
Summary:
Concurrent (e.g. nested) llvm::parallel::for_each() may lead to dead
locks. See PR35788 (fixed by rLLD322041) and PR41508 (fixed by D60757).
When parallel_for_each() is about to return, in ~Latch() called by
~TaskGroup(), a thread (in the default executor) may block in
Latch::sync() waiting for Count to become zero. If all threads in the
default executor are blocked, it is a dead lock.
To fix this, force serial execution if the current TaskGroup is not the
first one. For a nested llvm::parallel::for_each(), this parallelizes
the outermost loop and serializes inner loops.
Differential Revision: https://reviews.llvm.org/D61115
llvm-svn: 359182
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
We've been running doxygen with the autobrief option for a couple of
years now. This makes the \brief markers into our comments
redundant. Since they are a visual distraction and we don't want to
encourage more \brief markers in new code either, this patch removes
them all.
Patch produced by
for i in $(git grep -l '\\brief'); do perl -pi -e 's/\\brief //g' $i & done
Differential Revision: https://reviews.llvm.org/D46290
llvm-svn: 331272
But now include a check for CPU_COUNT so we still build on 10 year old
versions of glibc.
Original message:
Use sched_getaffinity instead of std:🧵:hardware_concurrency.
The issue with std:🧵:hardware_concurrency is that it forwards
to libc and some implementations (like glibc) don't take thread
affinity into consideration.
With this change a llvm program that can execute in only 2 cores will
use 2 threads, even if the machine has 32 cores.
This makes benchmarking a lot easier, but should also help if someone
doesn't want to use all cores for compilation for example.
llvm-svn: 314931
Summary:
This reverts D38481. The change breaks systems with older versions of glibc. It
injects a use of CPU_COUNT() from sched.h without checking to ensure that the
function exists first.
Reviewers:
Subscribers:
llvm-svn: 314922
The issue with std:🧵:hardware_concurrency is that it forwards
to libc and some implementations (like glibc) don't take thread
affinity into consideration.
With this change a llvm program that can execute in only 2 cores will
use 2 threads, even if the machine has 32 cores.
This makes benchmarking a lot easier, but should also help if someone
doesn't want to use all cores for compilation for example.
llvm-svn: 314809
This time it actually occurred to me to change the #defines
to actually test the pre-processed out codepath. Hopefully
this time it works.
llvm-svn: 302752
